diff --git a/HANDOFF.md b/HANDOFF.md index 513f41e..4eea921 100644 --- a/HANDOFF.md +++ b/HANDOFF.md @@ -35,7 +35,7 @@ On 12 October the advisor wrote: focus on the issues she and Prof. Thomas raised |---|---| | Research and evidence | **Complete.** Every number is rebuilt by `research/option_space_2026-10-01/rebuild_all.sh` (272 ledger checks pass). No further analysis is planned. | | Story | **Settled.** Golden pitch and five-part arc frozen (tags `golden-pitch-2026-10-08`, `narrative-arc-2026-10-09`). | -| Thesis text | **Started.** Chapter 1 drafted and revised over many rounds. Chapter 5 has an older draft to be rewritten. Chapters 2, 3, 4, 6, 7 not started. | +| Thesis text | **Started.** Chapter 1 drafted and revised over many rounds (introduction and 1.1 reviewed; 1.2 and 1.3 next). Chapters 2 and 3 have first full drafts (12 Oct, evening), awaiting the researcher's reading; a seam pass then aligned Chapters 1 to 3 with each other (vocabulary, the H2 and entry-test spine, order, numbers; see `EDIT_HISTORY.md`, last section). Chapter 5 has an older draft to be rewritten. Chapters 4, 6, 7 not started. | | Advisor | Progress deck and data package ready (`papers/IL_Advisor_Package_2026-10-12.zip`); the researcher uploads it. The advisor has **not yet seen** the change in what H2 means. | | Branch | All work is on `proposal/integrated-2026-09-16`, about 140 commits ahead of `main`. **`main` is stale** (it stops on 1 Oct, when PR #9 merged). | @@ -78,8 +78,10 @@ Frozen wording of the pitch: `research/option_space_2026-10-01/GOLDEN_PITCH.md` ## 6. Next steps, in order 1. Recheck section 1.2 (contributions) of chapter 1 against the ambition paragraph in 1.1. -2. Write chapter 3 (terms in one box, the growth formula, H1, H2, the entry test) and chapter 4 (data and tests). -3. Rewrite chapter 5 as ten sections, each answering one question, using `DRAFT_STORY_v2.md` for content and the writing rules for form. +2. Chapters 2 and 3 have first drafts (`ch2_literature.md`, `ch3_framework.md`) and Chapters 1 to 3 have had a seam pass; revise from the researcher's reading. **Chapter 4 has a first draft** (`ch4_data_methods.md`, 12 Oct night); it covers what it owed the reader: both groups for the H1 test (Indonesian platforms and the foreign benchmark); the entry-test rule; the data behind each test; how revenue movements are traced to fees and discounts (Objective 2); how the platform evidence is compared with national and payment figures (Objective 3). +2a. Chapters 4 and 5 were audited by an independent agent and cleaned on 13 Oct (see the last section of `EDIT_HISTORY.md`); it lists three decisions that touch Chapters 1 and 3 and the list Chapter 7 must carry. +2b. Chapter 4 open points: market values (4.7) are not promised in Chapters 1 or 3, so either add one clause to 3.4 and 1.3 or cut 4.7; the appendix it points to (platform list, seven traced windows, registered tests with dates) must be built; Table 4.1 labels Blibli's and Bukalapak's scope as the sources describe it, to be rechecked against the filings; household spending needs a reference. +3. Chapter 5 has a rewritten first draft (`ch5_results.md`, 13 Oct) with 11 sections; figures (exhibits) are not yet placed; Chapter 6 (rules and reach) is next. Original instruction for Chapter 5 kept for reference: rewrite as ten sections, each answering one question, using `DRAFT_STORY_v2.md` for content and the writing rules for form. 4. Write chapters 2, 6 and 7. Chapter 6 waits for the scheduled results below. 5. Insert citations; format to the YZU template (the researcher obtains it). @@ -109,6 +111,10 @@ Frozen wording of the pitch: `research/option_space_2026-10-01/GOLDEN_PITCH.md` - Wants honest assessments and pushback when a request would weaken the thesis; does not want reassurance. - Judges the writing by ear: expect to be corrected on flow and tone, and record each round in `EDIT_HISTORY.md`. - Cheap checks on existing data are always worth running; new analysis is not wanted now. +- **Readability comes before safety.** On 12 Oct (evening) the researcher said he does not want flags on the ambition paragraph ("technically no evidence, it is an educated guess, I am not going for safer options"), and that a safer claim that makes a sentence convoluted is the very thing Prof. Kong and Prof. Thomas complained about. So: the ambition and framing stay bold and are not flagged; support goes in notes or later chapters, never into hedged prose; flag a sentence only when it makes a different claim, once. This sits beside section 8: statements of what the data show keep their evidence strength; the framing is bold by design. +- **He reviews before and after.** Show the change as before and after in chat, apply only what he approves, commit once per round, log it in `EDIT_HISTORY.md`. To assess a passage: print the full text with numbered paragraphs, say what each paragraph does and whether it hands over to the next, then list only reading problems, ranked. +- **Vocabulary is fixed** (thesis README, Vocabulary rule): wedge, not gap; transaction value, with "the commerce a platform carries" as its plain companion; figure for a published number, measure for a yardstick; count of online sellers; entry test. New chapters use these words; run a sweep for the old ones before sending a chapter. +- **Style settled in this round** (full list in the thesis README, Writing rules): paragraph openers name their subject; a movement gets its mechanism in the paragraph that explains it; one currency (US$, rounded, World Bank rate for the year); no bare rupiah in the narrative. ## 10. Open items @@ -116,6 +122,8 @@ Frozen wording of the pitch: `research/option_space_2026-10-01/GOLDEN_PITCH.md` - Section 1.1 should be "brief, yet impactful" as standard prose; two attempts were rejected (`EDIT_HISTORY.md`, Part D). - Citations are not yet in the chapter text. - The YZU template is not yet in the repository. +- Chapter 1: the introduction and 1.1 were reviewed and revised on 12 Oct (evening); 1.2 and 1.3 were then rebuilt in the seam pass and have not yet been read by the researcher. Section 1.1 is 25 words longer than before (+6 percent) against the open wish for it to be brief; cut candidates are in `EDIT_HISTORY.md`. +- Chapter 3 now gives each link of the chain its own condition (inner: take rate steady; outer: national growth comes from sellers the platforms record; the count: survey keeps reaching the same sellers). The advisor has not yet seen how H2 is framed, so confirm this wording with the researcher before it goes to her. - Root-level files not part of the thesis are untracked and should be left alone (`side-academic-papers-*.zip`, `ssrn-7514022.pdf`, `Attraction-json_2026-10-09.zip`, a Chinese-named PDF). ## 11. Notes for a cloud or fresh session (added 12 Oct after a cloud session's first read) @@ -125,6 +133,7 @@ Frozen wording of the pitch: `research/option_space_2026-10-01/GOLDEN_PITCH.md` - **Numbers.** Take them only from `research/option_space_2026-10-01/FINDINGS.md`, the Part 1-7 files with their dated notes, and section 4 above. H1 now rests on 5 Indonesian and regional platforms against 27 abroad (p = 0.004); the proposal's 8-platform comparison is the earlier version. Anything from the 24 Sep proposal, the September manuscripts or `main`'s routing documents is out of date. - **Two notes exist only on `main`:** `docs/MEASUREMENT_PROXY_FRAMEWORK_2026-09-23.md` and `docs/PAPER_COMPARISON_BASELINE_2026-09-23.md`. Both predate the 27 Sep proposal: the first is the "adequate proxy" framing of the 24 Sep lineage (superseded by the stand-in conditions in Part 2), the second ranks the researcher's wider set of papers. Neither is needed for the thesis; leave them on `main`. - **Do not run or trust the old generators** (`scripts/build_thesis_manuscript.py`, proposal surgery scripts). The thesis is written by hand in `papers/thesis_2026-10/`. +- **Working with two sessions.** A cloud session works on its own branch (`claude/happy-pascal-uc2eya` this time), kept at the tip of the integrated branch. It sends changes back by pull request into the integrated branch when the researcher asks, never into `main`. Only one session edits the thesis files at a time; fetch before editing. - **First task:** section 6, step 1 (recheck section 1.2 of chapter 1), then chapters 3 and 4. Log each round of the researcher's feedback in `papers/thesis_2026-10/EDIT_HISTORY.md`. ## 12. Timeline diff --git a/papers/thesis_2026-10/EDIT_HISTORY.md b/papers/thesis_2026-10/EDIT_HISTORY.md index 406f1ee..7506a5e 100644 --- a/papers/thesis_2026-10/EDIT_HISTORY.md +++ b/papers/thesis_2026-10/EDIT_HISTORY.md @@ -41,6 +41,8 @@ One place for every concern raised about this thesis and what was done about it: | 12 Oct | A findings section in chapter 1 "gives spoilers" | Removed; short declaration of the answer kept | Addressed | | 12 Oct | The first paragraph of 1.1 must be the grand ambition; the question is the step toward it | Done | Addressed | | 12 Oct | Section 1.1 should be "brief, yet impactful" | Two attempts rejected (see Part D) | **OPEN** | +| 12 Oct, evening | On the ambition paragraph of 1.1: "technically no evidence, it is an educated guess, I do not want the flag"; "if getting a safer claim means convoluting, fuck safety"; "what we really need is readability while still maintaining quality". This is the line Prof. Kong and Prof. Thomas complained about | Ambition claims are not flagged. Rule recorded as "Readability before safety" in the thesis README and in `HANDOFF.md` section 9 | Addressed | +| 12 Oct, evening | Wants to see before and after, so he can check immediately | Preview-then-apply loop recorded in the thesis README and `HANDOFF.md` section 9 | Addressed | Standing constraints set by the researcher: no contact with BPS by email; BPS microdata and LSEG data are never shared or committed, only aggregates; payments and credentials are handled by the researcher. @@ -82,7 +84,431 @@ Standing constraints set by the researcher: no contact with BPS by email; BPS mi | 12 | "That aim cannot be reached directly…" then "For that, one question…" | Weird flow; then "For that" has no referent | "Measuring what the current figures miss begins with one question:" | | 13 | Section 1.1 cut from 406 to 257 words | Not what was meant: shorter paragraphs, not a shorter section | Reverted | | 14 | Section 1.1 as short paragraphs and bullet lists | Too non-standard for a thesis | Reverted to the five-paragraph prose version (commit `de9ca7d`) | +| 15 | Introduction paragraphs 2 to 5 still opened on back-references ("The difference…", "The same divergence…", "These figures…", "The difference matters…") although the openers rule was applied earlier | Preview shown as before and after; "all improvements, implement directly" | Openers now name their subject: "In 2023, traders…", "Tokopedia… shows the same split…", "Revenue and transaction value differ…", "Money and public decisions rest on figures like these…". One sentence added at the end of paragraph 2 ("The two official figures moved in opposite directions.") so "the same split" has a referent | +| 16 | Paragraph 4 explained why the levels differ, but paragraphs 2 and 3 had raised a movement (revenue up, sales down) | Approved | Mechanism added: a platform keeps only part of each sale and that part can change; at Tokopedia about 60 percent of the rise in revenue reflected smaller discounts and incentives (60.6%, Part 3) | +| 17 | Paragraph 7 used rupiah where paragraph 5 used US$ (Prof. Kong: one currency) | Approved | US$81 billion and US$31 billion; levy US$60 million and US$300 million. World Bank 2024 rate Rp15,855 per US$ (exact: US$81.3bn, US$30.7bn, US$64m, US$310m) | +| 18 | "In brief" said both figures "stopped moving with the activity", which can be read as H2 failing | Approved | "Each overstated growth" (Part 7's own phrase); H2 stated plainly ("National growth came mainly from more sellers, as H2 states, but about half of the 2023 rise in the seller count is not explained by first-time entry") | +| 19 | Claim flags on the ambition paragraph (trillions, "outside those figures altogether", "most is missing") | "I really do not want to hear or care about" those | Dropped. Ambition paragraph left exactly as it was | **OPEN: section 1.1 should be brief yet impactful.** The researcher's concern, in their words: the research question and objectives "should be brief, yet impactful", "not much reading". Two fixes failed (a shorter section; bullet points). The five-paragraph prose version stands as "fine enough". To revisit: how to make each paragraph of 1.1 land quickly while staying standard prose. Candidates not yet tried: tighter sentences inside the same five paragraphs; moving the two-halves explanation to chapter 3; splitting the long ambition paragraph in two. -**Also open in chapter 1:** the objectives keep the proposal's formal wording (the stiffest sentences in the section); citations not yet inserted; section 1.2 written before the ambition paragraph existed and not yet rechecked against it. +**Also open in chapter 1:** the objectives keep the proposal's formal wording (the stiffest sentences in the section); citations not yet inserted; section 1.2 written before the ambition paragraph existed and not yet rechecked against it (the researcher is going through the introduction and 1.1 first and has not reached 1.2). + +### Chapter 3 (12 Oct 2026, evening) + +| Round | What was tried | Researcher's response | Outcome | +|---|---|---|---| +| 1 | First full draft, on the proposal's and deck's order (deck slides 8 and 9: question and hypotheses, then "revenue = share kept x transaction value", then why D and not E). Opens on the chain of stand-ins; terms in one table; the identity D = (1 + g(V)) x (m1 - m0) / m0 with the Tokopedia example (take rate +68%, transaction value -9%, revenue +53%, D = 62 points); the wedge's two jobs (it follows the commerce; its size next to revenue sets how far a take-rate change swings growth), with a 2% and 17% illustration; H1 and H2; the entry test as a test of what "more businesses" means; further analyses; tests table | Awaiting | Drafted. Checked against the writing rules before sending: no "therefore", "indicator" or "rather than"; no bare rupiah; no mention of the proposal; openers name their subject; prose sentences average 15 words, longest 29 (a figure placeholder). Numbers recomputed: 1.682 x 0.911 - 1 = 53.2%; D = 62.1; 27.4 / 127.4 = 21.5%; 1/0.02 = 50 and 1/0.17 = 5.9 | + +Choices made that the researcher may want to change: +- H1 and H2 are stated in the plain wording already used in section 1.1, which keeps the meaning of the originals ("grow in proportion"; "comes mainly from more businesses selling online, and less from higher sales per business"). +- The 2% and 17% platforms are round numbers near the medians that chapter 5 reports (1.6% and 16.9%); they are labelled as an illustration, so the medians keep their home in chapter 5. +- The Figure 3.1 placeholder stands for the three nested circles of the oral deck (slide 8); the figure is not yet drawn. +- No citations yet (they are inserted after the chapters are written). + +### Chapter 2 (12 Oct 2026, evening) + +| Round | What was tried | Researcher's response | Outcome | +|---|---|---|---| +| 1 | The researcher asked whether Chapter 2 was being left alone. It was not: it had been put last in the plan because of citation risk. Written now, on the proposal's four sections (2.1 to 2.4) as the spine, lightened. Added the survey-count strand that the entry test needs (survey breaks, recall, small-firm exit, margins). Ends on the three open questions and hands over to Chapter 3 | Awaiting | Drafted. Checked before sending: no "therefore", "indicator", "Overall" or "rather than"; no regulation numbers or dates; institutions named with context (OECD, UNCTAD); sentences average 16 words without citations; 40 citations, each traced to `REFERENCES_WORKING.md` and used within its use limit | + +Choices made that the researcher may want to change: +- Dropped from the proposal's version: Medina and Schneider (2019) (reserve in the working list; the thesis makes no shadow-economy claim) and Momentum Works (a data source, cited in the data chapter). +- Kept despite being marked "reserve" in the working list: Barrios et al. (2022) and Denes et al. (2025), both Journal of Financial Economics, because Prof. Kong asked for finance citations. Cut them if the gig-economy angle feels off-topic. +- Added since the proposal: UNCTAD (2023, 2024), van den Brakel et al. (2008, 2020), Neter and Waksberg (1964), McKenzie and Paffhausen (2019), Bowen, Davis and Rajgopal (2002), Trueman, Wong and Zhang (2000), Brockmeyer and Hernandez (2016). +- Use limits respected: Hagiu and Wright is cited for platform versus reseller models, not for the IFRS principal and agent test; Givoly et al. for a general principle; Hummels and Klenow as trade terms; the OECD handbook and Ahmad and Schreyer for the separation of a platform's service from the sale, with no claim that GDP counts only platform revenue. +- Still open before submission (from the working list): check every citation against what the work supports, format to the YZU style, build the bibliography. + +Accuracy fix in Chapter 3 made in the same round: the terms table no longer says accounting rules always deduct incentives from revenue (IFRS 15 has an exception for payments for distinct goods or services). It now says "normally", and that Tokopedia reports the deduction as a separate line. + +### Seam pass across Chapters 1 to 3 (12 Oct 2026, evening) + +Why: the line-by-line read of Chapters 2 and 3 (sentence, paragraph, section, in order) and a check of how they line up against Chapter 1 found breaks that no single-line edit would fix, and showed that individual fixes would collide with each other. So the fixes were applied as clusters, each rewritten as whole paragraphs, and then re-read in context. Baseline: commit `b9dd114`. + +| Cluster | What was done | Where | +|---|---|---| +| Vocabulary | One word for one thing (see the Vocabulary rule in the thesis README): wedge, not gap; transaction value with "the commerce a platform carries" tied once; figure for a published number and measure for a yardstick; count of online sellers; entry test; chain only in Chapter 3; report for accounts and records for data | All three chapters; README | +| H2 and the entry test | The entry test is named in the first sentence of 1.1's design paragraph and in 1.3. Chapter 2's open questions now include H2's (do platform figures follow national growth). Chapter 3 opens from those questions and gives each link of the chain its condition: inner (take rate steady), outer (national growth comes from sellers the platforms record), and the count (survey keeps reaching the same sellers) | ch1 1.1 and 1.3; ch2 2.4; ch3 opening, 3.2, 3.3 | +| Order and structure | Chapter 1 paragraph 4 re-ordered (wedge named first, movement and Tokopedia last); Chapter 2 section 2.2 re-ordered (growth split before the survey weakness) with bridges into 2.2 and 2.3; Chapter 3 paragraph on the wedge split in two | ch1, ch2, ch3 | +| Numbers | One set across chapters: US$ at the World Bank 2024 rate; Tokopedia 53, 9, 60, 68 and 62 (0.91 x 0.68 = 0.62); growth defined as a fraction so the formula holds; the 127.4 step; 2 to 3 and 17 to 18 percent examples | ch1, ch3 | +| Repeats removed | UNCTAD "closest work" sentence four appearances down to two; the "cutting discounts" mechanism not repeated a third time in Chapter 3; Chapter 3's "two ways to grow" sentence no longer duplicates Chapter 2's | ch2, ch3 | +| Promises and homes | Objectives 2 and 3 now have a stated home (Chapter 4), named in 1.3 and 3.4; "payment indicators" became "payment figures"; "In brief" now answers the size half of the question | ch1, ch3 | +| Single lines | IFRS and GMV, GTV, TPV spelled out; a missing citation; "a study" for one study; the region named; the night-lights analogy explained; "from the inside" replaced by "whose underlying answers can be examined"; the wedge scope fixed ("at the marketplaces studied in Indonesia") | ch1, ch2, ch3 | + +Decisions taken on the researcher's "try it out", to be confirmed or reversed: +- The chain keeps the proposal's Figure 1 and each link has its own condition (the recommended option). The alternative, two stand-ins and no chain, would have changed how H2 reads. +- The size sentence in "In brief" is in. It adds length to 1.1 (see below). +- "Indicators" in Objective 3 became "figures" (it is a banned word, and Chapter 2 uses "figures"). + +Checks run after the rebuild: banned words none; no paragraph opens on a back-reference; the vocabulary sweep clean (the only "gap" left is the "research gap" heading); 41 citations in Chapter 2 each traced to `REFERENCES_WORKING.md`; every number recomputed; repeated phrases across chapters listed (only intended ones remain); sentence length (prose, citations excluded) averages 18.8, 15.2 and 15.3 words in Chapters 1, 2 and 3, with no sentence over 30 in Chapters 2 and 3. + +Size against the baseline: Chapter 1 +9 percent, Chapter 2 +8, Chapter 3 +12. **Section 1.1 grew by 28 words (+7 percent)**, against the researcher's open wish for it to be brief. Candidates to cut if wanted: the "A further analysis estimates…" sentence in the design paragraph (it is covered in 1.2 and 1.3), or the second sentence of "In brief". + +Chapter 4 now owes the reader (promised in 1.3, 3.2, 3.3 or 3.4): both groups for the H1 test (Indonesian platforms and the foreign benchmark); the entry-test rule; the data behind each test; how revenue movements are traced to fees and discounts (Objective 2); how the platform evidence is compared with national and payment figures (Objective 3). + +### Follow-up round after the second line-by-line read (12 Oct 2026, evening) + +The second sentence-by-sentence read (on the files after the seam pass) found 11 items, two of them introduced by the pass itself. All were fixed in one round and the touched paragraphs re-read with their neighbours. + +| Where | Problem found | Fix | +|---|---|---| +| ch3 opening, second paragraph | The sentence that brings in the count of online sellers started a new topic with no link to the outer-link condition before it (introduced by the seam pass) | "To check this, national growth is split into…" | +| ch3 opening, first paragraph | The seam pass dropped the line saying each figure stands in for the next (introduced by the seam pass) | "Each figure stands in for the next one out." restored | +| ch1 1.1 design paragraph | "is first-time entry" read awkwardly; the pointer to Chapter 3 sat before the further analysis | "comes from sellers entering for the first time"; the pointer dropped (1.3 already gives it), which also brings 1.1 back toward its old length | +| ch1 intro, paragraph on the wedge | "the driver's fare" arrived before ride-hailing was in view; the reason smaller discounts raise the share was missing; the paragraph was long | "the seller's receipts on a marketplace, the driver's fare on a ride-hailing app"; "which a platform pays for out of what it earns"; the paragraph split into the level (the wedge) and the movement (the share is not fixed) | +| ch1 1.2 | "it shows" followed a sentence whose subject was "they" | "the thesis shows / measures" | +| ch2 2.1, third paragraph | Weak join from revenue recognition to measures | "Revenue is not the only reported figure whose meaning depends on how it is computed." | +| ch2 2.2, first sentence | Promised "how large" but the section is about definitions and sources | "Measuring the wedge is a question of definitions and sources." | +| ch2 2.4 | The summary left out the survey-methods strand; "pieces" against "parts"; an orphan sentence on survey counts | Survey-methods sentence added to the summary and the orphan removed; "parts" | +| ch3 Tokopedia example | "0.91 x 0.68" did not show which factor was which | "(1 - 0.09) x 0.68 = 0.62" | + +After this round: Chapter 1 1,667 words, Chapter 2 1,280, Chapter 3 1,244 (baseline 1,520, 1,177, 1,095). **Section 1.1 is now 435 words, 25 more than the baseline (+6 percent)**, down from +28. No banned words; no sentence over 30 words in Chapters 2 and 3. + +### Chapter 4 first draft (12 Oct 2026, night) + +New file `ch4_data_methods.md`, about 3,100 words, eight sections: 4.1 platform data (two groups, Table 4.1 with the scope of every series), 4.2 the H1 test, 4.3 tracing revenue movements (Objective 2), 4.4 the BPS files and the entry test (Table 4.2), 4.5 channels, records and reach, 4.6 comparison with national and payment figures (Objective 3, Table 4.3), 4.7 market values, 4.8 traceability. + +Written from the repository records only: `h1_extended.py` and its tables (27 foreign platforms, 158 pairs; 5 Indonesian platforms, 15 pairs), `PREREGISTRATION.md` (entry-test rule, data checks, scenarios), `LICENSED_DATA.md` (rows, weights, checksums), `tokopedia_segment.py`, `wedge_split.py`, `m1_market_value.py`, Parts 3 to 6. No result is stated in the chapter; results stay in Chapter 5. + +Obligations checked against the promises in 1.3, 3.2, 3.3, 3.4 and 3.5: both H1 groups (4.1, 4.2); the entry-test rule with its 127.4 step and the cohort check (4.4); the data behind each test (Tables 4.1 to 4.3); tracing to fees and discounts (4.3); comparison with national and payment figures (4.6). Kong's asks: scope stated for every series, each source explained, the author's own measures marked, one currency. + +Checks: banned words none; no sentence over 33 words (two between 30 and 33 were split); every number traced to the table or script named above (pair counts, 21.5 and 23.9 percent thresholds, weights, 9,999 shuffles, 0.52 log change = 68 percent, Rp783 trillion about US$53 billion at the 2022 World Bank rate). + +Decisions for the researcher: (1) section 4.7 (market values) is outside what Chapters 1 and 3 promise; keep it with one added clause in 3.4, or cut it. (2) Length: Chapter 4 is longer than Chapter 3 by about 1,900 words; trimming candidates are 4.5 (World Bank paragraph) and the fourth H1 check in 4.2. (3) Spin-off tests, listing timing and the market-income pattern are left out, as the cut list puts them in reserve. + +### Chapter 5 first draft and a linear read of Chapters 4 and 5 (13 Oct 2026) + +`ch5_results.md` was replaced by a new draft (the older report-style draft stays in git history). It follows the story order of `DRAFT_STORY_v2.md` as eleven sections, with thesis prose instead of labelled beats: 5.1 wedge (Table 5.1), 5.2 H1 (Table 5.2), 5.3 tracing (Table 5.3), 5.4 market values, 5.5 the platforms in 2023 (Table 5.4), 5.6 what sellers reported, 5.7 count and entry test (Table 5.5), 5.8 who the extra sellers were (Table 5.6), 5.9 returners, 5.10 channels and records (Table 5.7), 5.11 why the national figures disagree. The 2026 rules and tax reach go to Chapter 6. Each result is marked registered or exploratory; every number comes from FINDINGS.md, Parts 3 to 6, the committed tables or `PREREGISTRATION.md`. + +The linear read (sentence, paragraph, section, in order; then against Chapters 1, 3 and 4) found these, all fixed: + +| Where | Problem | Fix | +|---|---|---| +| 5.1 | "the three largest platforms" is not shown; a summed wedge sat against Chapter 3's "defined for one platform"; "does not depend on the assumptions" overclaimed; the "three quarters" sentence was unclear | "three major platforms", "the sum of their invisible wedges", "changes little", sentence rewritten | +| 5.2 | the 5-v-27 test is not registered but read as if it were; the quarterly sentence contradicted itself on dates; "points of transaction value" confused | stated as not registered; dates fixed; "percentage points of the take rate" | +| 5.3 | "the other part" vague; Grab 2022-23 looked like an eighth window; no "so what" | volume part named; wording tied to the take-rate rise; closing sentence added | +| 5.4 | "second prediction" had no first | both predictions stated; handover to 5.5 added | +| 5.6 | marketplace-slice sentence belonged to 5.11; no handover | moved to 5.11; handover to 5.7 | +| 5.7 | bullet list in running text; "fourth check" appeared after three unlabelled ones; "cohorts" and "the count" confused; 5.8/5.9 order in the handover | prose; the four checks named in order; handover names both sections | +| 5.8 | inferred education codes mentioned (a limit) | removed; goes to Chapter 7 | +| 5.9 | the cited study is about exit, not restart | wording changed to exit | +| 5.10 | records paragraph contradicted itself on coverage; E7 prediction not stated | rewritten with the prediction and the 5.8-point fall | +| 5.11 | one 130-word paragraph; "4.4 times smaller" | split; "less than a quarter" | +| ch4 | no mention of the points comparison, the monthly-versus-annual reading or the overshoot rule, all used in Chapter 5; "earlier foreign platforms" and "first benchmark" both used; "the table says so" unclear | all added or unified | + +Decisions for the researcher: (1) 5.2 says plainly that the larger H1 test is not registered. (2) Chapter 7 must carry: inferred education codes in the 2023 file, the 19 sample rows in the top bracket, Bank Indonesia's method not public, the exploratory status of the 2023 accounting. (3) Figures: six exhibits exist (`exhibits/v2`) and none is placed. + +### Cleanup of Chapters 4 and 5 after an independent audit (13 Oct 2026) + +Why: the researcher asked for both chapters to reach Chapter 1's standard with no open problems. A fresh agent that had not seen my conclusions audited every number and claim in both chapters against the repository sources (53 findings; 13 categories of items verified clean, including all table cells in 5.2, 5.3, 5.4, 5.6 and 5.7). I checked its main claims against the sources myself (Shopee's transaction value is a Momentum Works estimate; GoTo's 2022 transaction value is an estimate; the 0.16 to 0.68 range is the final-consumer slice; Blibli's incentives are a proxy; the Tokopedia 60.6 percent and the per-unit split disagree) and all held. + +Substantive fixes: +| Where | Problem | Fix | +|---|---|---| +| 5.3 | Table 5.3 said Tokopedia was incentive-led while the text counted it as fee-led; the repo's two Tokopedia splits (60 percent from lower incentives by account lines; fees larger per unit of transaction value) disagree | One table of all seven windows on the per-unit basis; both Tokopedia views stated; "both agree on pricing, differ on which lever" | +| 5.1 | Table 5.1 labelled Shopee and Grab as the author's derived estimates; Shopee's value is Momentum Works; Grab's revenue is a direct country disclosure; 15.92 should be 15.93; "three quarters to four fifths" understated 84 percent | Labels, figure and range fixed; Sea's 11.5 percent rate added to the sensitivity | +| 5.11 | The 0.16 to 0.68 band was called the whole marketplace slice | Now the final-consumer slice, with low, middle and high (0.16, 0.68, 1.62); 0 to 13 percent growth range corrected to 0.2 to 13.2 | +| 5.7 | Table 5.5 mixed two exit scenarios; "published rise" was the weighted-file gap (819 vs 821 published); two different "caps"; US$ conversion broke the one-currency rule | Two-column table, labelled; 15.5 percent explained; US$53 to US$72 billion | +| 4.1, 5.2 | GoTo on-demand 2022 value is an estimate and the segment includes Vietnam to 2024; Grab is regional; Ozon's pair rule | Stated in Table 4.1 and under Tables 5.3, 5.4; the group is named "Indonesian and regional"; pair rule fixed | +| 5.2 | Table 5.2 mixed Mann-Whitney and shuffle p-values without saying so; "the reason is the slice" too strong | Method column; "most of the difference comes from the slice" | +| 5.10 | World Bank mean-sales failure (W3) omitted; second E7 version omitted; question-wording change and stable-coverage statements stated as fact; validation outcome never reported | All added or hedged | +| 4 and 5 | Method statements missing from Chapter 4 for Objective 1, the quarterly sample, the value side, the profile control, part-year selling, the returners scenario | Added to 4.1, 4.2 and 4.4 | +| labels | Several sections had no status label; "registered" list incomplete | Every section now marked registered, exploratory or descriptive | + +Also fixed: terms (incentives, platform-year, file versus survey year), the p-value and rank correlation defined at first use, rupiah removed from the narrative, sentences over 26 words split, "therefore" removed. + +Open for the researcher (these touch other chapters, so nothing was changed there): +1. **Chapter 1, paragraph 4:** "about 60 percent of the rise in revenue reflected smaller discounts and incentives" is the account-lines view. Per unit of transaction value, higher fees account for more. Proposed: "…about 60 percent of the extra net revenue came from smaller discounts and incentives, counting the lines of its accounts." +2. **Chapter 1, "In brief":** "the official count of online sellers also overstated growth" is stronger than Chapter 5, which cannot separate wider coverage from returning sellers. Proposed: "the growth in the count of online sellers is only partly explained by first-time entry." +3. **Chapter 3, 3.2:** H1 is "tested by asking whether growth divergence … is unusually large". The main test uses the log form. Proposed: add "measured in logs" and note the points form is a check. +4. **Chapter 7 must carry:** inferred education codes (2023 file); 19 sample rows in the top bracket; Bank Indonesia's method; the exploratory status of the 2023 accounting; the month question changing format; Grab's two vintages; the one-sided tests. +5. **Appendix still to build:** platform list; seven windows with the Blibli proxy; wedge by platform and year; registered tests with dates; World Bank description. +6. **Length:** Chapter 4 is now about 3,900 words and Chapter 5 about 4,200 including tables. Cut candidates are 4.5 (World Bank), 5.7's value paragraph, and the Table 5.2 narrower-test rows. + +### Chapters 4 and 5 rewritten in the voice of Chapters 1 and 2 (13 Oct 2026) + +The researcher's direction: the 27 September proposal is frozen, and Chapters 4 and 5 are where the post-proposal data and findings go, so they should read like the approved Chapters 1 and 2 (flowing paragraphs, each answering the last, plain words, numbers rounded in prose, exact values in tables). The audit-driven version had become a list of checks and caveats, which cut against "readability before safety". + +What changed: +- Chapter 4 follows the proposal's own spine (4.1 selecting the platform evidence, 4.2 comparing platform activity through time, 4.3 direct and constructed Indonesia measures, 4.4 evidence beyond platform accounts, 4.5 external evidence and comparisons) and adds 4.6 on labels, traceability and licensed data. It opens on the "two sides" idea from Chapter 1. +- Chapter 5 is nine sections, each opening on the finding and ending on the question the next one answers (wedge; H1; what moved the take rate; what markets valued; the platforms and the sellers in 2023; where the extra sellers came from; who they were and whether they returned; channels and records; why the national figures disagree). +- Proposal-era wording is gone from the prose ("the first version of the test" survives only as one status sentence; "first benchmark" became "chosen first"). +- The detail moved to `appendix_notes.md`: the 27-platform list, the seven windows in full, wedge by platform-year, every check on the test of H1 with both p-values, the quarterly sample, and all registered tests with outcomes. The audit's corrections stay in the main text where they change a claim (provenance labels, the two Tokopedia views, the GoTo estimate, the final-consumer slice). +- A mistake of mine found on the way: the World Bank 70 percent without written records belongs to unregistered firms that use social media with customers (29.9 percent keep records), not to all unregistered firms. Fixed in 5.8. +- Sentences split to about 15 to 18 words; Chapter 4 is about 3,000 words and Chapter 5 about 3,900 including tables (about 3,200 of prose). + +Still open from the audit: the two Chapter 1 sentences, the Chapter 3 wording of H1, "OECD" at first use in Chapter 2, market values in 1.3 and 3.4, and the list Chapter 7 must carry. + +### Check run after the Table 5.2 and heading changes (13 Oct 2026) + +Banned words, openers, sentence length, cross-references (sections, tables, appendices) and 22 number checks against the committed tables all pass, apart from two cells in Table 5.3 that I had rounded from already-rounded shares: Tokopedia's higher-fee share is 76 percent (not 77) and Grab 2022–24's commerce share is 42 percent (not 43). Fixed; the note under the table now says a row can add to 99 or 101. The earlier audit had matched these two cells at one decimal, so a check at whole-number rounding was the one that caught them. + +### Fixes after the assessment (13 Oct 2026) + +Done, at the researcher's request ("fix those", skipping what is not important): +- Chapter 1: "overstated growth" replaced (revenue grew faster than the commerce behind it; the seller-count finding stays in its own sentence); the Tokopedia 60 percent now reads "in Tokopedia's accounts"; citations added for the figures (AFP, BPS, Bank Indonesia, GoTo, UNCTAD, Ministry of Finance); 1.3 now mentions market values. +- Chapter 2: OECD spelled out at first use. Chapter 3: H1 sentence names the group (Indonesian and regional) and the log measure; 3.4 mentions market values. +- Chapter 4: the entry-test rule no longer repeats Chapter 3's arithmetic; the list of further checks is shorter. +- Chapter 5: about 3,700 words (from 4,000); paragraphs split, status tags kept only where a result is registered or exploratory ("unmarked results are descriptive"); the first line of 5.2 names the group; the thin-slice explanation now rests on arithmetic (log change is roughly the change in points over the take rate) and not on a non-significant test; 5.4 says what the data show and do not show about investors; the top-bracket detail, the provinces and the one-third/two-thirds figure moved to Appendix B. + +Not done (not judged important now): figures, bibliography, YZU template, Chapters 6 and 7. + +### Second independent audit of Chapters 4 and 5 (13 Oct 2026) + +A fresh reviewer re-checked every number and citation (21 checks in Chapter 4, 118 in Chapter 5, 187 appendix cells; all five figures reproduce). It found 41 problems. Fixed: +- **Wrong:** the foreign range of share kept (2.6 percent at Shopify to 71 percent at Zalando, not "25 percent at Mercado Libre"); Table 5.4 net entry 361, not 360; appendix units (Blibli in Rp billion; Shopee's assumed take rates 8.4, 10 and 10.8 percent); file paths in Appendix D; the claim about a palette script that is not in the repository. +- **Own-goods rule:** it applies to the 19 added platforms only. The eight chosen first are filtered on unchanged reporting scope. Ozon's one pair starts in a dropped year. New check `h1_thesis_drops.py`: without Ozon p = 0.003 (shuffle 0.022), without Zalando 0.005, without both 0.004. Disclosed in 4.1 and Appendix B; headline 5 against 27 unchanged. +- **Claims tied to evidence:** 5.2 now says the planned three-marketplace test gives p = 0.002 by pairs and 0.06 by platforms, and that the larger test is exploratory; the Bank Indonesia reconciliation holds in the low and middle scenarios and fails in the high one; the marketplace slices "agree on slow growth" became "the growth is small"; the second market-value run (34 platforms, p = 0.004, mixes segment and whole-company values) is reported next to the main run, and the market sample is said to differ from H1's 27. +- **Citations:** IFRS 15 now carries the agent and incentive claims ("normally deduct"); Hagiu and Wright only the marketplace–reseller line; Trueman et al. and Rajgopal et al. described as web-traffic studies; Rochet and Tirole no longer used to predict a thin take rate; BPS cited for the 819 thousand rise. +- **Smaller:** Table 5.3 caption, 61/39 percent for Tokopedia's account lines, Grab revenue 0.60, Momentum Works 3.5 labelled as calculated, cohort groups described as cumulative, E6 threshold stated, E7 second version inside the limit, rank correlation and standard error defined, "about 750 thousand" for the cohort check, Table 4.3 gained a row for adjacent figures, checksum claim narrowed, Chapter 3's "no national total" reconciled with the three-platform sum, figure axis and caption fixes, Figure 5.3 now reads its values from tables. + +Not changed, for the researcher: the second-run market result stays secondary; Tokopedia's "15 million merchants" is a company statement not checked in the repository (said so in 5.9); full company names for the tickers in A.1 are still to be added (ASAPQ not identified from the repository). + +### Chapters 1 to 4: openings rebuilt so each chapter stands alone (13 Oct 2026) + +The researcher's reading: Chapter 1 before 1.1 is long and repeats itself; paragraphs 12 and 13 should be plainer; Chapter 2 should begin with basics and justify its sections; each section should say what it shows and does not show; Chapter 3 should not start from the tail of Chapter 2; Chapter 4 should open with the reality of the data. +- **Chapter 1:** the text before 1.1 went from about 840 to about 620 words (wedge and Tokopedia merged; the stakes paragraph and the gap paragraph shortened; one sentence on statistical agencies dropped). Paragraph 12 now reads as a list of the tests in plain words; paragraph 13 leads with the finding. 1.3 is a roadmap that follows the argument (ideas, evidence, implications), not a bare list. +- **Chapter 2:** a three-paragraph opening says what is being studied and what knowledge is needed (why figures differ; how else the activity is measured; what the records are used for), which justifies 2.1 to 2.3. 2.1 now defines platform, platform economics, transaction value, revenue and revenue recognition in plain words before the literature. 2.2 and 2.3 open with basic definitions. Each section ends with what it shows and what it leaves out. 2.4 states the three open questions as a numbered list, each tied to the section that left it open. +- **Chapter 3:** opens fresh, with the three figures, the stand-in idea and the three questions restated in full and mapped to measures, hypotheses and sections. No sentence depends on having read Chapter 2. +- **Chapter 4:** opens with the state of the evidence (Shopee, Grab and Tokopedia disclosure; Blibli and Bukalapak incentives; BPS files, three rounds, no identifier, province only; the central bank's method not public), then why Indonesia still allows the tests, then the chapter map. "Sets the bar low on purpose" is stated. +- **Open for the researcher:** the YZU template has not been seen, so whether 1.3 and the heading pattern match it is unconfirmed. Section 2.4 and 3 now use a numbered list of three questions; say if you want prose instead. + +### Second pass on the openings, after the researcher asked for reasoning before editing (13 Oct 2026) + +The first pass ran mechanical checks only. A linear reading found: +- **Chapter 1 paragraph 3:** "The two differ because…" followed the 53 percent against minus 9 percent split and so seemed to explain the opposite movement. It explained only that the two are different amounts. Reworded; the movement is explained in the next paragraph. +- **Chapter 1:** paragraphs 7 and 8 and the 1.1 opening both said decisions rest on the figures. The government and rules material now sits once in paragraph 7 (money and rules) and once, concretely, in paragraph 8 (Indonesia, size and the levy). Paragraph 13 now says "in Indonesia" for the revenue finding, since it was not tested everywhere. +- **Chapter 2:** "explains why the difference can be large" was wrong (a fee share is large by definition); it now says revenue need not follow commerce, because pricing structure also matters. "Most platforms do not make or own what is sold" was false for resellers; fixed. The invisible wedge is no longer attributed to Chapter 1. "None shows" became "none of the studies reviewed here shows". The third kind of knowledge is now "how the records are used". +- **Chapter 3:** "the survey" was used before it was named (BPS). Question 2's elaboration did not match its heading; it now says why new sellers outside platform records matter. +- **Chapter 4:** "only Tokopedia gives both for Indonesia" was wrong for Blibli and Bukalapak, which are Indonesian companies; the opening now lists each series' scope. "Only one reports incentives" was wrong (Tokopedia, Grab and GoTo on-demand do; Blibli and Bukalapak do not). "Nobody follows the same seller" became "the files cannot". One sentence I wrote ("no database holds these figures") could not be checked and was removed. +- **Structure:** the proposal's Chapter 1 had only 1.1 and 1.2. Section 1.3 is my addition for the thesis, so its form is unconfirmed against the YZU template. The proposal's research gap named three things (how large, whether it persists, whether the choice changes the conclusion), not the three questions now in 2.4 and Chapter 3; Question 3 (rules) comes from the thesis extension. Flagged to the researcher. + +### Repetition cut in Chapters 1 to 4 (13 Oct 2026) + +Main points that appeared more than once, and where each now lives: +- **Chapter 1:** the "figures differ" point is in paragraph 2 (growth) and paragraph 7 (size) only. The investor and government stakes are in paragraph 5 once; the "rule reaches only platform sellers" point moved to 1.2, where it belongs to the tax contribution. Paragraph 6 lost its three-item summary of the literature (Chapter 2's job). 1.2 lost its closing sentence (the 1.1 ambition) and its repeated "why revenue moved" and "revenue depends on the measure" sentences. The text before 1.1 is now about 560 words (from 840). +- **Chapter 2:** 2.4 no longer re-summarises 2.1 to 2.3, because each section already ends with what it shows and leaves out; the list items are one sentence each. +- **Chapter 3:** list items shortened. +- **Chapter 4:** the scope of each series is stated once, in Table 4.1; the opening gives the headline and the paragraph after the table no longer repeats it. The file limits are stated in the opening only as "cannot follow the same seller"; 4.4 keeps the detail. +- **Kept on purpose:** the short definitions of transaction value and revenue in Chapters 2 and 3 (each chapter must stand alone, which conflicts with "define once"; the researcher to decide), and the three questions in both 2.4 and Chapter 3. + +### Independence read correctly (13 Oct 2026) + +The researcher's clarification: a chapter or paragraph standing on its own means it does its own job without leaning on a neighbour, not that it repeats what came before. I had read it as "restate the terms". Reversed: Chapter 2 no longer re-defines transaction value, revenue and the wedge (Chapter 1 introduces them and Table 3.1 defines them); Chapter 3's opening names the three figures in one line and leaves definitions to Table 3.1; the three questions in Chapter 3 now carry only their link and the section that answers them, while 2.4 carries the reasons. The earlier "kept on purpose" note on repeated definitions is withdrawn. + +### Checked against the agreed writing rules before editing (13 Oct 2026) + +The researcher asked for confirmation, not acceptance. Checked against `README.md` (rules agreed with the researcher) and the repository: +- "Stand alone" in the README means openers name their subject and do not lean on the previous paragraph ("these", "this literature", "both"); terms are defined once. It does not mean restating. This confirms the reading reached in the previous round. +- Violations found in my own drafting and fixed: bold-labelled numbered lists of the three questions in 2.4 and Chapter 3 (the README bars bullet lists and labelled beats in running text; they are now prose with "first, second, third"); three section endings in Chapter 2 opened with "This literature" or "The studies reviewed here"; 2.4's first sentence leaned on "each field reviewed here"; the "in brief" paragraph in 1.1 had grown to four sentences (the rule is a three-sentence declaration). +- Facts confirmed in the repository: Grab, GoTo on-demand and Tokopedia disclose incentives and Blibli does not (the panel and extraction notes); Sea's extraction carries no Indonesia split; no chapter mentions "the proposal". +- Open: the researcher once said Chapter 2's questions were "not in a list". The README rule and that comment pull apart; the questions are enumerated in prose for now. + +### The three questions as a list again (13 Oct 2026) + +The researcher preferred the list and thought the README rule had been read too widely. Checked: the rule's origin is the 1.1 bullet attempt ("too non-standard for a thesis") and the labelled beats of the story drafts, both argument text. A short numbered list of research questions is ordinary thesis practice and the later chapters refer back to the three questions, so the list is better there. Restored in 2.4 (question plus what the section leaves open) and Chapter 3 (question plus link and sections, so it does not repeat 2.4's reasons). README exception added. Prose stays for everything else. + +Note, same day: the researcher says the README rule was written because an early Chapter 2 or 1.2 draft was all bullets with no paragraphs, and that the restored list version is the one they wanted from the start. The repository records only the 1.1 bullet attempt (row 14), not the Chapter 2 or 1.2 one, so this origin is taken from the researcher. README wording changed to match. + +### Linear read of Chapters 1 to 4 openings, third pass (13 Oct 2026) + +Read sentence by sentence, then paragraph, section and chapter, against the README rules. Found and fixed: openers that leaned on the previous paragraph (Chapter 1 paragraph 5 "these figures" and paragraph 7 "both"; Chapter 2 "the chapter follows these three parts", "this test", "this idea", "Answering these questions", and the opening of 2.3 that leaned on 2.2); Chapter 2's closing paragraph repeated Chapter 1's data point and is now one handover sentence; "each figure" in Chapter 4 overstated (Shopee's value is an outside estimate), now "each company figure". Checked and left: the order of paragraphs in 2.2 (definitions differ, then numbers differ), and the Chapter 3 opening, which stands on its own. + +### Chapters 2 to 5 brought to one standard (13 Oct 2026) + +Checked and fixed against the README rules, in the order Chapter 2, 3, 4, 5: +- **Chapter 2:** the 2.1 paragraph on reported figures is split in two (comparable accounting and companies' own measures; investors looking past revenue), each with its own job. +- **Chapter 3:** the opening no longer explains the entry-test logic (3.2 and 3.3 do), and no longer defines BPS twice (Table 3.1 does). 3.2 states the take-rate condition once with the link to growth divergence. 3.3's "Chapter 4 gives the rule" and Chapter 4's "follows the rule in Chapter 3" no longer point at each other. 3.4 now states the market-value question, which Chapter 3 had left out. The 3.1 arithmetic on the wedge's second use stays; Chapter 5 now applies it rather than re-explaining it. +- **Chapter 4:** 4.1's first paragraph no longer re-defines revenue and transaction value; the own-goods rules are a paragraph of their own; 4.2's checks are a separate paragraph, and the unexplained "thin-slice explanation of Chapter 5" now reads "to see whether a thin take rate explains the result"; "it also reports" (ambiguous subject) fixed; 4.3 lost its repeat of the opening; the data-use sentence in 4.4 is said once; the World Bank sentence and the regression paragraph are separated; "fixed effect" and "clusters" are explained in plain words. +- **Chapter 5:** openers that leaned ("Little of this", "The entry test can", "The value side tells a similar story", "these sellers", "The World Bank surveys point the same way", "The 2024 figures show the same split") now name their subjects; the literature restatement in 5.1 is one clause with the citations; 5.2's long paragraph is split and its repeat of Chapter 3's wedge argument is cut; two paragraphs over 170 words are split. +- **Checked and left:** how Chapter 2 hands to Chapter 3 (three questions to the chain and its two links) and Chapter 3 to 4 (each test's data and rule) works. + +### Chapters 2 and 3 rebuilt on the researcher's reading (13 Oct 2026) + +The researcher's verdict on the previous version: Chapter 2 did not do its job. Chapter 1 already sets the question, so Chapter 2 should give the current academic consensus and several perspectives that try to answer it with what existed before this research, show where each answer fails, and let the gap follow from those failures. Chapter 3 should open with a declaration of its logic and anchor it (the figure first), and no paragraph may open on a term the reader has not met. +- **Chapter 2** is now organised by attempts to answer the thesis question: pricing theory and accounting (2.1), national statistics and survey methods (2.2), tax research on platform records (2.3). It opens by naming the four fields and testing each against one case (Tokopedia 2023). Each section states the field's consensus, develops it in full paragraphs, and ends with where its answer stops. 2.4 builds the three open questions from those stopping points. The two-group idea is introduced with examples (shoppers and merchants; riders and drivers) before it is used; the accounting passage says in plain words what a principal and an agent record and that discounts reduce revenue. Use limits from `REFERENCES_WORKING.md` respected (Hagiu and Wright as business model, not an accounting test; IFRS 15 "normally"; web-traffic studies named as such). A claim about UNCTAD comparing levels at one point in time was dropped because the audit says its full report is unchecked. +- **Chapter 3** opens with the declaration that a measure is a stand-in and the framework tests the links; the figure placeholder follows at once, then the conditions, then how the two links answer the research question, then the three questions with what measures each. 3.1 puts the formulas first and names each measure at its first use; m0 and m1 are defined before the formula that uses them; the Tokopedia example is followed by what each measure records before the wedge and growth divergence are contrasted. 3.2 opens by saying the two hypotheses test the two links. 3.3 states the entry test first as a block, then the arithmetic. 3.4 is four questions instead of a repeated list. 3.5 is unchanged (the researcher said it reads clearly). +- **Not yet done:** the Chapter 4 opening still reads as complaint ("thinner than the question deserves"); the researcher wants a disclaimer that is presented tactically. Chapters 4 and 5 are on hold until 2 and 3 are accepted. One known slip in 5.9 ("In 2024 it was less than a quarter…" with an unclear "it"). + +### Chapters 2 and 3, second attempt after the failure analysis (13 Oct 2026) + +Worked from the diagnosis, not from line fixes. Jobs fixed first. Chapter 2: give the best available answers to the thesis question, field by field, and show where each fails, so that the gap is a result. Chapter 3: declare the logic and its reason, anchor it with the figure, then measures, hypotheses and tests that answer Chapter 2's three questions. +- **Chapter 2** has four perspectives, one section each (2.1 pricing theory, 2.2 accounting, 2.3 national statistics, 2.4 platform records and tax), then 2.5 the gap. Each opens with the field's consensus, develops it in paragraphs of its own (the two groups introduced with examples, the mechanism of price structure, the principal and agent distinction with a worked ride of 100 and a driver's 80, the voucher rule explained), then states what the field says about the thesis question and where it breaks down. The three open questions are built from those breaking points. +- **Chapter 3** opens with the stand-in idea and why it matters (investors, governments and analysts read the figures as signs, the link is rarely tested). The figure paragraph names H1 and H2 on the links. 3.2 opens by naming the two checks and the hypothesis that states each. 3.3 opens with why the count has to be checked. +- **Cold-reader check run:** first use of each term compared with where it is introduced (script run in session, not kept). Fixed: the figure labels H1 and H2 before the hypotheses existed; transaction value in Chapter 2 now glossed at first use. +- **Known weak points:** 3.4 may still repeat Chapter 1 and Chapter 4 and could be cut; Chapter 2 has grown to about 2,300 words; the Tokopedia readings in the Chapter 2 opening are the author's application of each field and not quotations. + +### Independence audit of Chapters 1 to 5 (13 Oct 2026) + +Rule, as the researcher states it: the idea runs linearly, but every chapter, section and paragraph opening must stand on its own, naming its subject, never opening on the previous text. Links between chapters live at the end of the earlier chapter, inside a paragraph that does other work. + +Audit method: read the opening sentence of every paragraph in Chapters 1 to 5 as a cold reader; flag back-references to chapters, sections and earlier text, leaning starters (this, these, both, each, the first, the same, because, also, either, again), terms used before they are introduced, and answers to a heading phrased as a question. Fixed: +- **Chapter 2:** the chapter no longer opens with "Chapter 1 asked"; it states the question as its own subject. Section openers that leaned ("The result is that…", "The same standard…", "These rules…", "Applied to the thesis question…") now name their subject. The closing one-line pointer to Chapter 3 is a full paragraph. +- **Chapter 3:** "the research question" was used without being stated in the chapter; it is now stated before use. "Each link", "The threshold" and "Four further questions" now name their subjects. +- **Chapter 4:** the opening is rewritten as a statement of what the evidence takes and why Indonesia allows the tests (no complaint tone; no "thinner than the question deserves"); "the pairs", "own goods", "the second step", "the extra sellers", "the first objective" and "the third objective" are all introduced where first used. +- **Chapter 5:** headings phrased as questions were answered with "they" or "it" (5.2, 5.3, 5.7); fixed. 5.9 no longer opens "Chapter 1 began"; the unclear "it" in the 2024 sentence is fixed; "again" and "either" openers fixed. +- **Chapter 1:** "the same split", "the figures", "the question" at the starts of three paragraphs now name their subjects. Left for the researcher: the opening of the research-question paragraph in 1.1 ("Measuring what the current figures miss…"), which deliberately bridges from the ambition paragraph the researcher approved. + +### Chapters 2 and 3 rebuilt on the proposal's spine (13 Oct 2026) + +The researcher asked whether the proposal had been checked. It had not, properly, and the thesis README's own method (start from the proposal's paragraph order, lighten, insert new material where the line reaches it) had been broken by three rebuilds with invented frames. The proposal's Chapters 2 and 3 already had the structure the researcher kept asking for: sections that follow the platform's anatomy (it carries a transaction, records part as revenue, keeps a record of the whole), section endings on what the literature leaves unmeasured, a gap that maps onto the research question (how large, whether it persists, whether the choice of figure changes the conclusion), a Chapter 3 that opens by declaring the framework's logic before the figure, numbered formulas with "where" definitions, measures grounded in the literature, and hypothesis development that reasons from the literature to each hypothesis. Its fault was readability (Prof. Kong) and confusing terms (Prof. Thomas). +- **Chapter 2** keeps the proposal's order and section titles and expands each step: what platform economics studies and its central finding, the price-structure result with an example, the platform–reseller distinction, revenue recognition with a worked ride example and the voucher rule with its exception, reporting choices, comparability and operating measures, investor evidence; official statistics, sources, night lights, margins, survey breaks, recall error and exit; digital records, third-party compliance evidence with what each study found, and the platform rules. Each section closes on what remains unmeasured. 2.4 keeps the proposal's three-part gap; the reach of rules is stated as a further question that extends the thesis. +- **Chapter 3** keeps the proposal's declaration, figure note, pairing rule, equations (1) to (4) with "where" definitions, the E example, D as the central test, and hypothesis development from the literature. Added where the line reaches them: the terms table (Prof. Thomas), the take-rate form of D and the Tokopedia illustration, the thin-slice paragraph (Prof. Thomas's "just subtraction"), Table 3.2 placed before the hypotheses so they are anchored, and the entry test after H2's stated limit. The further analyses close the chapter in one paragraph. +- Register: no rhetorical questions, vignettes, invented frames or announcements; studies are reported by what they find; openers name their subject; "indicator", "therefore", "rather than" and "Overall" avoided. + +### Chapter 2 applied in the three-part form (13 Oct 2026) + +After two previews the researcher confirmed the bones: an introduction that follows the platform's steps; each section in three parts (the situation, what the research explains, where it stops and why); and a gap anchored to those stopping points, with Table 2.1 and one question per section. The "where it stops" paragraphs explain what each body of work was built to answer and what would be needed, so they read as reasons, not as criticism. Cut for focus: Barrios et al., Denes et al., Kleven, Kreiner and Saez 2016, Slemrod 2019, Brockmeyer and Hernandez, La Porta and Shleifer, Ulyssea, De Franco et al. + +Queued as expansions, agreed not to be structural faults: +- A short finance section on how investors and analysts read platform figures (valuation multiples, disclosure of operating measures, value of non-financial metrics; reserve sources Liu, Nissim and Thomas; Damodaran; SEC 2020), which would also supply a source for "investors and analysts often read revenue as a sign of commerce". +- A targeted search for empirical studies of platform take rates over time, to confirm or narrow gap question 1. +- Checking each study finding against the paper (Kleven et al. 2011, Pomeranz 2015, Naritomi 2019, Berg et al. 2020, Bowen et al. 2002) and the reference list as a whole. +- Open choice: keep the reach of platform-based rules as gap question 3, or present it as an extension. + +### Two corrections from the literature search (13 Oct 2026) + +From `research/option_space_2026-10-01/LIT_SEARCH_2026-10-13.md`: (1) gap question 1 overclaimed; commission studies exist (Li and Wang on delivery fee caps; a creator-platform study; an Oxford study tracking an Uber take rate over 2016-2024), but they study commissions as prices and their effects on participants. The 2.1 closing, Table 2.1 and gap item 1 now claim only what holds: no study reviewed sets reported revenue against reported transaction value to test whether revenue stays a fair guide, and none covers Indonesian platforms. (2) Pomeranz (2015) was misdescribed; now: the VAT paper trail deters evasion and enforcement spreads up the supply chain. Expansions (commission studies in 2.1, a finance section, the Bowen detail) remain queued. + +### Chapter 3 rebuilt as the bridge from the gap to the tests (13 Oct 2026) + +The researcher kept Chapter 2 as applied (three-part sections, gap anchored to the shortfalls; the queued Chapter 2 expansion was not applied) and asked for a complete Chapter 3: since this is the thesis and not the proposal, it may carry more than the proposal did. Chapter 3 now states how the literature and the gap become the approach. +- **Opening:** the declaration stays, with its basis named (test a figure against the activity it stands for; Henderson et al. 2012). The paragraph restating the research question became a numbered map of the three gap questions of 2.4 to the tools that answer each. Gap question 3 (reach of rules) stays a gap question, measured by the channel split. +- **Concerns raised in Chapter 2 now used where they apply:** Givoly et al. moved from the H1 reasoning to the comparison rules (a measure tells less when its computation changes); the own-goods rule (IFRS 15; Hagiu and Wright) stated in 3.1; van den Brakel et al. cited for the survey break in 3.5; Neter and Waksberg used for the direction of recall error (it favours passing the entry test, so a shortfall is cautious); McKenzie and Paffhausen give the reason the age of sellers bears on returners; exits only raise the share needed, so 21.5 percent is the smallest share that can produce the rise. +- **Added:** equations (5) and (6), splitting revenue growth into volume, fee rate and incentive rate (Objective 2); equation (7), the H2 margin identity with its decision rule; a new 3.6 with Table 3.3 (tracing, channels and reach, national figures, market values) and equation (8) with the two registered market-value predictions, anchored to the investor line in 2.1. Table 3.1 gains the fee rate and the incentive rate. +- **Form:** first draft too dense (paragraphs up to 181 words), a second too spread (64 paragraphs, many one-sentence pieces); the applied version joins the second without changing a sentence (checked by script): 51 paragraphs, longest 92 words, "where" definitions joined to what each measure means. Sections renumbered 3.1 to 3.6; no other chapter cites a Chapter 3 section number. +- **Open:** the researcher finds the chapter trustworthy but heavy to read, probably because of its eight equations and many symbols; accepted for now, cold read due. If still heavy, consider moving equation (8) to Chapter 4. Chapter 4 now repeats the revenue split (4.2) and the market regression (4.5) and should be cut to data and procedure; Chapter 1.3 may add the further analyses to its line on Chapter 3. One sentence is the author's reasoning, not a cited finding: "Growth from the take rate has a limit, because a platform cannot raise its share of each sale without end." + +### Cross-chapter read of Chapters 1 to 3 (13 Oct 2026) + +Read as one text after Chapter 3 was accepted. The line holds (Chapter 1 names what an answer needs, each Chapter 2 section stops there, Chapter 3 supplies the tools; the research question and the three gap questions match). Overlaps left by chapter-by-chapter work, fixed with the researcher's approval: +- Chapter 2's closing paragraph repeated Chapter 3's map; cut, and the handover moved into the sentence that introduces the three questions. Chapter 2 now ends on question 3. +- Chapter 1's three objectives and the three gap questions were never linked; Chapter 3 now says which objective serves which question, and that the third question is answered by a further analysis. +- Chapter 1.2 promises to set each recent rule against the figure it acts on; Chapter 3.3 now says where a rule acts in equation (6) (commission caps on the fee rate, discount limits on the incentive rate), pointing to Chapter 6. +- Chapter 3.2's "two established results" paragraph repeated 3.1 and 3.4; cut. +- Chapter 1.3 now names the four further analyses; Chapter 2's intro says "online sales across the country" where "the sales" was ambiguous. +- Left for the reference pass: "investors and analysts often read revenue as a sign of commerce" (2.1) has no source, and Chapter 3.6 relies on it. +- Practice from now on: a cross-chapter read after each chapter is accepted. + +### Chapter 4 restructured after Chapter 3 (13 Oct 2026) + +- **Opening:** the researcher wanted the disclaimer first, short and neutral, not an apology. After two rejected drafts ("No ready-made dataset…" opened on an absence; "The evidence this thesis needs is hard to obtain" stated a feeling and read as an excuse), the chapter now opens: "The figures this thesis compares come from sources that were never designed to be read together. Platforms publish in their own formats, and the answers of individual sellers are released only under a data-use agreement. The methods in this chapter are built around these conditions." The data-use agreement is said once. "The aim is modest…" dropped so the bar is not lowered twice. Rule learned: a disclaimer states the condition as a fact and lets the method answer it. +- **Repetition of Chapter 3 removed:** the definition rule, the own-goods theory, the log change, the revenue split, fees minus incentives, the H2 identity, the comparison rules and the market-value question now point to Sections 3.1, 3.3, 3.4 and 3.6 and equations (5) to (8). Kept in Chapter 4 because Chapter 3 lacks them: the business-mix condition on the definition rule, the overlap rule for comparing sources. +- **Order follows Chapter 5:** platform side (data, wedge, H1, tracing, market values), then the survey (files and H2, entry test and extra sellers, channels, records and reach), then national figures and labels. Sections 4.1 to 4.10; only internal reference changed (4.2 to 4.3 in 4.10). +- **Checked by script:** every data number kept; dropped only citations now carried by Chapter 3 (IFRS 15, Hummels and Klenow, Trueman et al., van den Brakel). Paragraphs 98 words or fewer. +- **Survey-file naming made consistent** (a file is named by its round; the 2024 file describes 2023): the cohort check now says "the 2024 file"; the Bank Indonesia test says "the 2023 file, which describes 2022"; Chapter 3's entry-test box now says "the survey file that describes 2023". Chapter 5 still says "the 2023 survey" in places; fix in its pass. + +### Chapter 5: plan and a sample, not yet applied (13 Oct 2026) + +- **Plan agreed:** the proposal is a floor for Chapters 4 and 5, not a spine (the thesis evidence is much larger). Chapter 5 follows Figure 3.1's two links: inner link 5.1 to 5.4 (wedge, H1, what moved the take rate, markets), outer link 5.5 to 5.10 (2023 platforms and sellers, H2 and the entry test, what the rise contains, who the extra sellers were, channels and records, why the national figures disagree), ending back on Chapter 1's 2023 scene. Each section opens with its answer and ends on what stays open (no rhetorical questions). Table 5.1 lists every main result with its label and checks. Rounded numbers in prose, exact values in tables. Two halves named in the opening, no part headings; short two-sentence close; GDP-scale sentence (2.92 percent, ledger C39) left out until the researcher decides. +- **Sample written and polished in four passes:** opening, Table 5.1, 5.1 and 5.2, in `drafts/ch5_sample_2026-10-13.md`. Awaiting the researcher's reading for tone against Chapters 1 to 4. +- **Correction found:** the current 5.2 reports the narrower test planned in advance only as the platform shuffle (p = 0.06, "comes close without passing"). The Mann–Whitney test on one median per platform passes (p = 0.024; ledger C47, `build_numbers.py`). The sample reports both; Appendix B needs the 0.024 row. +- **New count:** `research/option_space_2026-10-01/h1_direction_count.py`: take rate rose in 13 of 15 Indonesian and regional pairs, opposite directions in 2; abroad it rose in 103 of 158 pairs. Size, not direction, separates the groups. +- **Still to fix in Chapter 5:** "about 750 thousand" new sellers in 2024 should read "at least 754 thousand" (FINDINGS); survey years named by activity year; section references 4.6 to 4.10 and 4.4 to 4.8; four rhetorical questions; repeats of Chapters 3 and 4; dense paragraphs in 5.4 to 5.9. + +### Chapter 5: full draft in the agreed shape, not yet applied (14 Oct 2026) + +`drafts/ch5_full_2026-10-14.md` continues the 13 Oct sample through 5.3 to 5.10 (ten sections, Tables 5.1 to 5.5, Figures 5.1 to 5.5), polished in four passes (sentence length, paragraph size and number count, openers, a reread for flow). Awaiting the researcher's reading; `ch5_results.md` is unchanged. +- **New evidence on the full thesis groups:** `research/option_space_2026-10-01/take_rate_share_full.py` (`tables/take_rate_share_full.json`). The take rate accounts for 0.66 of a revenue move at the median Indonesian and regional pair (15 pairs) and 0.24 abroad (158 pairs, 27 platforms; 0.22 with one median per platform); about 95 percent of a wedge move comes from volume in both groups. (A first run reported 139 foreign pairs because the script took the earlier year's take rate after filtering; fixed 14 Oct before use.) This replaces the proposal-era 3-against-8 figures (66.5 and 26.7 percent) in 5.3; they agree. +- **Checked:** Chapter 1's 2024 levels (BPS about US$81 billion, Bank Indonesia about US$31 billion) against `tables/yardsticks_bi_bps.csv`, now given their results home in 5.10; payment growth from ledger C50 to C54; "at least 754 thousand" new sellers in 2024. +- **Removed from Chapter 5 because Chapters 3 and 4 carry them:** the literature sentence in 5.1, the thin-share arithmetic (now a pointer to 3.2), the incentive-deduction rule, the telescoping explanation (pointer to 3.5), the cohort definition, the list of market-value platforms (Chapter 4.5), "associations, not stock returns" (Chapter 4.5). +- **To do when applying:** Appendix B row for the narrower test's Mann–Whitney p = 0.024. + +### Chapters 1 to 5 aligned, and Chapter 5 applied (14 Oct 2026) + +- **Chapter 5** replaced by the reviewed draft (`drafts/ch5_full_2026-10-14.md`): answer-first sections on Figure 3.1's two links, Table 5.1 of results and evidence, rounded numbers in prose. Appendix references renumbered (Tables 5.2 to 5.4); Appendix B gains the narrower test's Mann–Whitney row (p = 0.0242, ledger C47). +- **Alignment of Chapters 3 to 5:** Chapter 3's further analyses reordered (tracing, market values, channels, national figures) to match Chapters 4 and 5, wording unchanged; Chapter 4.9 now states the 2023 comparison of six platform measures with household spending that 5.5 reports; 5.4 names β_V and β_m of equation (8); "thin share" became "thin take rate" (one term); Chapter 3's entry-test box reads "BPS's survey of online sellers for 2023" (the file name arrives in Chapter 4). +- **Readability, judged by reading in context, not by counts** (the researcher's correction): kept the splits that help the line (Chapter 1's "two things… One is… The other is…", the 2024 BPS and Bank Indonesia figures, "In brief" as two points, the contribution sentence; Chapter 2's named openers and the 2.2 stopping point split into its two halves; Chapter 5 rejoins that tie a claim to its reason). Reverted the ones that broke it: the levy contrast (one sentence), the colon in "why this matters", "…are choices, so…", the 2.1 "where it stops" paragraph (kept whole as the researcher's device), "The median change was zero" (own sentence), a colon that promised a list. +- **Bug found and fixed:** the new take-rate share script first used 139 foreign pairs; corrected to all 158 (0.66 here against 0.24 abroad). + +### Open for a later pass on Chapter 5 (14 Oct 2026, the researcher's decision to revisit after reading Chapters 3 to 5 together) + +1. The middle of Chapter 5 (5.6 to 5.9, the checks) may read close to Chapter 4's procedural voice; the verdict-first rhythm rests mainly on each section's first sentence. +2. The return to Chapter 1's opening scenes is one sentence in 5.10. A fuller bookend is possible, but must only set facts side by side (the 2023 break sat in chat and social selling; the same year the government banned sales on social media) without claiming one caused the other. +3. Economic scale: most results are percentages; US$ amounts appear mainly in 5.1, 5.7 and 5.10. The GDP-scale sentence (2.92 percent, ledger C39, magnitude only) is still the researcher's decision. + +### Chapters 6 and 7: first drafts, not yet applied (14 Oct 2026) + +`drafts/ch6_draft_2026-10-14.md` and `drafts/ch7_draft_2026-10-14.md`, written from Parts 6 and 7 (with their dated notes), `PREREGISTRATION.md` A and C, `tables/part6_checks.json`, `tables/tax_base_rulers.csv`, `tables/src/policy_timeline_web.csv` and `tables/src/platform_seller_rules_intl.csv`, in the voice of Chapters 1 to 5 and polished in three reading passes. Awaiting the researcher's reading; the chapter files are not created yet. +- **Chapter 6 shape:** opening with the tax as headline; 6.1 Table 6.1 (seven rules; regulation numbers only in the table) and equation (6) (a fee cap pushes the take rate down, an incentive limit pushes it up); 6.2 cap and discount rule; 6.3 marketplace tax reach (at most about 1 in 20 sellers and 17 to 24 percent of value; record of nearly one seller in five); 6.4 statistics rules (platform data to BPS, the 2026 census frame); 6.5 registered predictions (Table 6.2) with a marked results slot; 6.6 beyond Indonesia (OECD, DAC7, Philippines, Viet Nam; US delivery caps). Order changed from the plan: inner link to outer, as in Chapter 5, so the tax is 6.3; the statistics rules got their own section. +- **Chapter 7 shape:** opening answer (a figure is a fair guide while a condition holds); 7.1 the answer at both links; 7.2 what it adds, in Chapter 1.2's order plus research; 7.3 predictions that failed (E5, Q1, M1, E7, W3, RBI), matching Appendix C (S1, S2 and the discount-leverage rule are not in Appendix C and are left out); 7.4 limits, carrying the whole audit list (education codes, 19 top-bracket rows, Bank Indonesia's method, the exploratory 2023 accounting, the month question, Grab's filing vintages, one-sided tests) and the caveats kept out of the story; 7.5 next steps from the proposal's Section 8; a closing return to the 2023 figures. +- **To confirm before keeping:** the source of "two-wheel trips under 6 percent of Grab's rides business" (graded C in `cap_prediction.py`; marked in the draft). New citations for the reference check: GoTo Gojek Tokopedia (2026, 2Q26 call), Pasardana (2025), Katadata (2025), CNN Indonesia (2026), Republika (2024), Bureau of Internal Revenue (2023), Government of Viet Nam (2025), Li and Wang (2025, found 13 Oct, not yet read in full). +- **Later:** fill 6.5 after 27 Oct, early Nov and mid Nov; Appendix C rows for each outcome; 7.5's first step updated once scored. + +### Pending: fixes to the Chapter 6 and 7 drafts, previewed 15 Oct 2026, not yet approved + +Found by checking Chapters 1 to 7 against each other and against the proposal. Apply only after the researcher approves; the exact before and after texts were shown in chat and are summarised here. +1. **6.3 and 7.4, BPS's "marketplace" includes food and ride apps** (in 2023, 47 percent of BPS's marketplace sellers used a shopping marketplace and 43 percent only Gojek or Grab; FINDINGS.md). Add after the value-reach sentence: "Both figures use BPS's marketplace category, which also counts food and ride apps, so both are upper limits." Replace the record sentence ("nearly one seller in five") with a range of about 8 to 18 percent of online sellers (0.47 × 17.8 to 17.8), depending on which apps the tax office designates. In 7.4, the tax-reach limit also names the food and ride apps. +2. **7.2, ownership of the measures:** follow Chapter 3 (only the ecosystem ratio is the author's own): "It tracks the invisible wedge and growth divergence over time, and its ecosystem ratio is the author's own measure." +3. **4.8, add at the end:** Chapter 6 sets each rule against the part of the figures it acts on; each legal text is read where published and labelled by status (primary text read, announced, news only). +4. **6.3, after "Exempt sellers file a statement as well.":** "Tax research finds that records held by a third party improve compliance where they exist (Section 2.3)." +5. **7.4, first paragraph, add the proposal's dropped boundaries:** "The splits of Table 5.4 are accounting, not causes. Marketplace sellers keep financial statements more often, but the thesis does not show that marketplaces lead them to." +6. **7.5, new paragraph before "The last step":** "The platform sample can grow as well. Lazada and TikTok Shop can join the test of H1 once they report Indonesian revenue and transaction value." +- Kept, after reconsideration: the "two thirds against a quarter" sentence in 7.2 (the one finding at economic scale for investors). +- Still open: source of "two-wheel trips under 6 percent of Grab's rides business" (6.2, marked in the draft). + +**Assessment shown to the researcher (15 Oct, the drafter's own judgement):** proposal about 6/10, thesis as drafted about 8/10, ceiling 8.5 to 9 after the reference check, Figure 3.1, a cold read and the 6.5 scoring. Largest integrity risk: the unchecked reference list. Largest content risk for a finance degree: the finance part is thin (M2 an association, M1 failed, stock returns deferred); framing the finance contribution deliberately in 1.2 and 7.2 is the researcher's decision. + +**Branch note:** this work is on `claude/happy-pascal-uc2eya`, 49 commits ahead of `proposal/integrated-2026-09-16`, which AGENTS.md and this README still name as the working branch. + +### Fixes applied from the 15 Oct materials (15 Oct 2026, the researcher: "you can try and all, and we'll consolidate") + +Source of every change: `research/option_space_2026-10-01/materials_2026-10-15/FIX_LIST.md` and the files it points to. +- **Chapters 1 to 5 (29 edits).** + - Citations split: BPS 2025a/2025b; Bank Indonesia 2023 with KabarBursa 2024 for the 2023 fall; GoTo 2024a (press release) and 2024b (annual report). Tokopedia's 53 percent is now "net revenue". New citations added for the social-media ban and TikTok Shop's closure. + - Trueman et al. is "beyond earnings"; Hummels and Klenow is levels; Kleven is "employers and banks". Hagiu and Wright moved; Table 4.3 now cites the tax office. + - 2.1's unsourced "investors read revenue as a sign of commerce" replaced with a sourced sentence (Bowen, Davis and Rajgopal 2002; Davis 2002). + - Two factual errors fixed: 5.1's ecosystem-ratio reading and 5.4's source of market values. + - Cold-read sentences fixed: 1 opening (share price), 4.7 (23.9 percent), 5.2 (thin take rates belong to the marketplaces), 5.3 (Tokopedia line count), 5.7 (the untraced 53 to 59 percent), 5.9 (rounded multiples). + - Figure 3.1 placed. 1.2 now states the registered market-value test. 4.8 now says where the rule evidence comes from. +- **Chapter 6 draft rewritten on the checked facts.** + - The tax started on 1 Oct 2026 through Shopee, Tokopedia, Lazada and Blibli; food and ride apps are not designated. + - The seller record covers at most about 8 percent of online sellers. + - Grab is "6 percent of mobility GMV", about 2 percent of on-demand. + - Presidential Regulation 27/2026 released on 8 Oct. + - The discount rule binds all e-commerce operators. + - GoTo's take rate rose to about 22 percent from about 18. + - Fee events now cite named sources (Katadata's promotional page dropped). + - Li and Wang wording; the Philippines and OECD wording. + - "Since 2023" for the platform-data rule. + - Table 6.2: the tax-start row withdrawn, P5's three outcomes stated, the P1/P5 reading reworded. +- **Chapter 7 draft.** + - 7.2 states ownership of the measures as Chapter 3 does, and states the market-value result. + - The tax record wording; 7.3 reports the withdrawn prediction. + - 7.4 adds the dropped proposal boundaries and names the survey files by round. + - 7.4 states the cap texts and food and ride apps as limits. + - 7.5 updated for the tax start and Lazada and TikTok Shop. +- **Integrity records.** + - `PREREGISTRATION.md` addendum of 15 Oct withdraws the tax-start prediction: its outcome was public on 1 Oct, before the 2 Oct registration. + - Dated note on frozen Part 6. + - Appendix C rows split. + - Seven verified events added to `tables/src/policy_timeline_web.csv`; existing rows kept. +- **New file:** `references.md`, 62 entries merged from the reference check, with 14 details still marked [UNVERIFIED]. A script checked that every in-text key has an entry. +- **Checks:** `rebuild_all.sh` ran clean (209 checks passed in the log, none failing). Its float-noise rewrites of tables and exhibits were reverted. +- **Not changed, on purpose:** the 1.1 ambition paragraph (bold by design under the writing rules), the citation style (APA until assembly) and `cap_prediction.py` (a registered file; its "< 6%" note is superseded by the Part 6 dated note). +- **Still open:** the foreign platforms' incentive accounting check (agent stopped at the usage limit; to resume), the 6.5 scoring from 27 Oct, the Chinese abstract and the YZU assembly. +- **15 Oct, later: foreign platforms' incentive accounting checked** (`materials_2026-10-15/INCENTIVE_ACCOUNTING.md`, `incentive_accounting_check.py`; exploratory, designed after the H1 result). Eight of the 27 book shopper discounts mainly as marketing; their take rates are not steadier (median 0.060 against 0.075 for the 11 that deduct; two-sided p = 0.90). Without the eight, H1's Mann–Whitney p = 0.0035 (main test 0.0038, reproduced). Two rows added to Appendix B; three sentences added to 7.4. Yatra changed its policy in FY2019 without restating; its median barely moves. Partly read: Rakuten, Zalando, Shopify, Alibaba. +- **16 Oct: the researcher's handoff of an outside review (ChatGPT), applied.** + - Scope kept. Integrity: X3 withdrawn as a test (published e-filing shares settled it at registration); E5 disclosure that BPS's 2024 Appendix 12 table was first used on 8 Oct, after the 7 Oct run; the discount-leverage rule (L) reported (P1 holds on 2 of 2, P3 fails at 0.117 against 0.10); the F3 recode disclosed (13.0 to 39.9 percent). All in the `PREREGISTRATION.md` addendum of 16 Oct, Appendix C and 7.3. + - Coverage-adjusted H2 added as exploratory (`research/option_space_2026-10-01/h2_coverage_adjusted.py`, `tables/h2_coverage_adjusted.json`). Correction to the review: removing only the extra count gives 46.6 and 33.6 percent, but removing those sellers' count and sales together gives about 53 to 55 and 41 to 43 percent. The thesis reports the consistent version: H2 sits on its threshold. Text in Table 5.1, 5.7, 7.1 and the abstract. + - H1 bounded to 2022–24 in 5.2 and 7.3; the listing and young-platform check brought into 5.2 (median 0.07 against 0.08, p = 0.24). + - Cap: the mechanical size (at most about 0.25 points on Grab's take rate; about 1 to 1.4 on GoTo's) stated in 6.2; causal wording removed. + - 5.7 to 5.9 prose tightened; companies introduced in 4.0; appendix tickers replaced by names; audit factual fixes (Tokopedia merchants, GoTo "controlled" Tokopedia, Katadata source replaced). + - New `abstract.md`; references 62 to 65; reading-copy pipeline rebuilt (`reading/`), every page of the 55-page PDF checked by eye. + - Still open: finance paragraphs from `materials_2026-10-15/FINANCE_FRAMING.md`, abbreviations list, Chinese abstract (researcher), [UNVERIFIED] references, 6.5 scoring from 27 Oct, YZU assembly, about thirteen sentences over 30 words. +- **16 Oct, later: rest of the review handoff that can be done now.** The researcher confirmed the review's ten points are the whole feedback list. + - Finance framing (item 5) from `materials_2026-10-15/FINANCE_FRAMING.md`. 1.2 now says why the wedge matters for value (market values track revenue; Bowen, Davis and Rajgopal 2002). It adds that when the take rate moves, market value per dollar of revenue and per dollar of commerce change by different amounts. 7.2 repeats this with the 0.6 headline. 5.4 adds Rajgopal et al. (2003) and Davis (2002) on revenue booked at the full sale amount. 3.6 adds Rajgopal et al. + - Long sentences (item 7): the 17 sentences over 30 words split, except the 1.1 ambition sentences and the research question, which are kept on purpose. + - New `symbols.md`, "Symbols and Abbreviations", for the YZU list of symbols (item 10). + - 3.6: one closing sentence on Chapter 4 and one on Chapter 6 dropped (each said only what comes next); Figure 5.5 moved below the paragraph that introduces it. + - Reading copy: Inter bundled as a web font (no Type 3 fonts left in the PDF), figures kept with their captions, short tables kept whole (`reading/post.py`). 56 pages, every page checked by eye. +- **16 Oct, evening: reading copy restyled after the researcher found the look too busy.** One serif face throughout (Source Serif 4), black on white, no colour rules or shaded boxes; chapter titles centred; three-line tables with captions above; hypotheses as indented italic; contents with page numbers and dotted leaders (`reading/toc_pages.py`, two-pass build). 64 pages, every page checked by eye. No text changed. +- **16 Oct, night: number re-audit of everything changed since 15 Oct** (`materials_2026-10-15/AUDIT_NUMBERS_16OCT.md`): 461 items, 456 match, 2 rounding-ok, 3 mismatch, none without a source; 16 of the 18 problems of 15 Oct fixed. Applied: + - Abstract: "three times" no longer tied to 2022–24 (the ratio uses all pairs); "about half" now credited to the exploratory count, while the registered test finds too few new sellers. + - Ch7: "Indonesian and regional" for the two-thirds figure; "above all from 2022 to 2024"; small growth "in two of three scenarios"; the H2 coverage reading labelled exploratory. + - Ch5: the shuffle uses one median per platform; the listing check names its 15 foreign platforms; the 0.7 million marketplace sellers marked as the author's calculation. + - Ch6: GoTo "at most about 1.4 points" with GoTo's own 0.9 estimate; Grab's 13 percent dated to the second quarter of 2026; "who apply for the discount" dropped (not in the cited source). + - Table A.4 labels name the source of each column (Shopee: outside estimate and scenario). + - Left for the researcher: 1.1 "Decisions worth trillions of dollars" (N2), the researcher's own wording. +- **16 Oct, night: references verified** (`materials_2026-10-15/REFERENCE_VERIFY_16OCT.md`). 15 flagged entries checked at the primary source: 6 verified, 8 corrected (dates, numbers, full titles; tax office n.d. is SP-1/2025 of 20 Jan 2025; SNA 2025 issued by the European Commission and four others), 2 left with a named secondary source (Philippine RR 16-2023 via GMA News; Momentum Works via Katadata, Setyowati 2024). Added Setyowati (2024, 2025) and World Bank (2026) for the exchange rate; OECD 2021 link fixed. In text: "Agence France-Presse 2023", "European Commission et al. 2025", "Directorate General of Taxes 2025", the World Bank rate cited in 4.1, Momentum Works cited through Katadata, and Table 6.1's legal texts given author-year sources. No [UNVERIFIED] note remains. Still open: the 13 proposal references the thesis dropped (listed in the report), the AFP/BPS short forms, and the citation style (APA until YZU decides). +- **Reading copy, page by page at full size (60 pages).** Fixed after reading: four tickers in A.1 named (Nebius Group, formerly Yandex; Trip.com; Tuniu; Uxin); Figure 5.3 moved below the paragraph that follows it (page 31 was half empty); Ch7 trimmed by two lines so its last page no longer holds three lines (licensed-data sentence now points to 4.10; the Lazada sentence joined to the last step). diff --git a/papers/thesis_2026-10/README.md b/papers/thesis_2026-10/README.md index b831e69..998c407 100644 --- a/papers/thesis_2026-10/README.md +++ b/papers/thesis_2026-10/README.md @@ -16,10 +16,22 @@ The thesis written in full. Same title, research question, hypotheses and measur - **Write linearly.** Each paragraph answers what the previous one left open. Test for every paragraph: does it make the next one necessary, and does it add to the stakes? A paragraph that only states what exists is "almost duh" and must be rewritten or cut. Deliver whole chapters; do not spend a turn per paragraph. - **Zoom in and out** to build stakes: a concrete Indonesian scene, then the general point, then global stakes with amounts and the cost of being wrong, then Indonesia with its own amounts. - **Openers stand alone.** No paragraph or section opens on a word that leans on the previous one ("both", "these readers", "these figures", "described above", "for that"). Name the subject. -- **Register.** Statements of fact in a neutral academic tone; no rhetorical lines, no labelled beats, no bullet lists in the running text of a section. Rounded figures in the prose (41%, 5%, 53%, 9%); exact values in tables and notes. No regulation numbers or exact dates in the narrative. Sentences of about 15 to 18 words with connecting words; do not chop to 11 words. +- **Register.** Statements of fact in a neutral academic tone; no rhetorical lines, no labelled beats, no section written as bullets in place of paragraphs (an early Chapter 2 or 1.2 draft was all bullets, which is why this rule exists, per the researcher, 13 Oct 2026). Prose carries the argument. A short numbered list is fine where items are parallel and later chapters refer back to them, such as the three research questions in 2.4 and at the start of Chapter 3. Rounded figures in the prose (41%, 5%, 53%, 9%); exact values in tables and notes. No regulation numbers or exact dates in the narrative. Sentences of about 15 to 18 words with connecting words; do not chop to 11 words. - **Chapter 1.** No "main findings" section (it spoils the tension); a three-sentence declaration of the answer is enough. The first paragraph of 1.1 is the grand ambition, stated as boldly as the subject allows; the research question follows as the step needed to reach it. Contributions are tied to the three groups who rely on the figures. - **Claims.** Each claim at its evidence strength. Never write that the extra 2023 sellers "were there all along" or "were only counted that year", that the gap is "too large to be chance", or that returning sellers are ruled out. Say "about half of the increase is not explained by first-time entry" and "the additional count was concentrated among older, long-established small businesses". The growth identity is D = (1 + gV)(m1 - m0)/m0. Do not mention "the proposal" in thesis prose. - **Terms.** Define each once, in plain words (transaction value, revenue, take rate, invisible wedge, ecosystem ratio, growth divergence). Avoid "indicator", "Overall,", "therefore", "rather than". Do not use the sentence "of every 100 rupiah, about 7 appear as revenue". +- **Readability before safety** (12 Oct, evening). The test for a sentence is how easy it is to read, then whether it is right. A safer wording that makes a sentence heavier is rejected: that is what Prof. Kong and Prof. Thomas complained about. The ambition and the framing are bold by design and are not flagged for lacking proof. Back them with a source in a note or a later chapter when that is cheap, never by hedging the prose. Statements of what the data show keep their evidence strength (the Claims rule above). Flag a sentence only when it makes a different claim; say so once and offer the closest supportable version. +- **A movement gets its mechanism.** When a paragraph shows a movement (revenue up while sales fell), the paragraph that explains it says how it happened (the platform kept more of each sale; at Tokopedia, smaller discounts and incentives), not only that the measures differ. +- **One currency.** Running text uses US$, rounded, converted at the World Bank official rate for the year (`research/option_space_2026-10-01/tables/src/worldbank_fx_2022_2024.json`; 2024: Rp15,855 per US$). No bare rupiah in the narrative. +- **Show before you change.** Preview the change as before and after in chat; apply only what the researcher approves; then one commit, and a row in `EDIT_HISTORY.md`. When asked to assess a passage, print the full text first with numbered paragraphs, then say what each paragraph does and whether it hands over to the next, then list only reading problems (flow, openers, number density, currency), ranked. +- **Vocabulary** (settled in the 12 Oct seam pass; one word for one thing, in every chapter): + - *transaction value* is the term for what is bought through a platform; "the commerce a platform carries" is its plain companion (tied once, in Table 3.1). "Sales" means a single sale, or BPS's national online sales, never a platform's transaction value. + - *invisible wedge* or *the wedge* is transaction value minus revenue. Never "gap" for that quantity ("research gap" is a different thing). + - *figure* is a published or reported number; *measure* is a yardstick, or one of the three the thesis builds (wedge, ecosystem ratio, growth divergence). + - *count of online sellers*, then *the count*; *entry test*; *first-time entry*; *H1*, *H2*. + - *chain* is used only in Chapter 3, for revenue, platform commerce and national commerce. + - *report* for what an accounting statement shows; *records* for a platform's data on sales. + - *take rate* for the share of each sale a platform keeps; *US$* for money in running text. - **The committee's points** and where each is answered: see Part A of `EDIT_HISTORY.md`. `EDIT_HISTORY.md` records each round of edits, the researcher's response and any open concern; read it before revising a chapter. @@ -27,5 +39,13 @@ The thesis written in full. Same title, research question, hypotheses and measur | File | Chapter | Status | |---|---|---| | `ch1_introduction.md` | 1 Introduction | drafted 12 Oct on the proposal's and deck's flow (opening, 1.1 question and objectives, 1.2 contributions, 1.3 structure); open concern on 1.1 in `EDIT_HISTORY.md` | -| `ch5_results.md` | 5 Results | early draft of 5.1-5.5 in the older report style; to be rewritten to the v3 structure (ten question sections, `research/option_space_2026-10-01/DRAFT_STORY_v2.md`) | -| ch2-ch4, ch6, ch7 | | not started | +| `ch5_results.md` | 5 Results | rewritten 14 Oct on Figure 3.1's two links: Table 5.1 (main results and evidence), inner link 5.1 to 5.4 (wedge, H1, what moved the take rate, markets), outer link 5.5 to 5.10 (2023 platforms and sellers, H2 and the entry test, what the rise contains, who the extra sellers were, channels and records, why the national figures disagree). Awaiting the researcher's reading with Chapters 3 and 4 | +| `ch3_framework.md` | 3 Framework and Hypotheses | rewritten and accepted for now 13 Oct: opening declaration and map of the three gap questions, 3.1 terms and rules for comparison, 3.2 three measures (equations 1 to 4), 3.3 tracing a movement (equations 5 and 6), 3.4 hypotheses (equation 7), 3.5 the entry test, 3.6 further analyses (Table 3.3, market-value equation 8). Figure 3.1 still to be drawn; a cold read is still due (see `EDIT_HISTORY.md`) | +| `ch2_literature.md` | 2 Literature Review | first full draft 12 Oct (evening), awaiting the researcher's reading: on the proposal's four sections (2.1 pricing and revenue recognition, 2.2 measuring the online economy, 2.3 digital records and third-party information, 2.4 the gap), with the survey-count strand added (breaks, recall, exit, margins). Citations are author-year, every one traced to `REFERENCES_WORKING.md`; the bibliography itself is not yet built | +| `ch4_data_methods.md` | 4 Data and Methods | restructured 13 Oct to follow Chapter 5's order and to point to Chapter 3 instead of repeating it: short opening disclaimer, 4.1 platform data, 4.2 the wedge in Indonesia, 4.3 testing H1, 4.4 tracing, 4.5 market values, 4.6 survey files and H2, 4.7 entry test and extra sellers, 4.8 channels, records and reach, 4.9 comparing national figures, 4.10 labels and licensed data | +| `appendix_notes.md` | Appendices A to C | working draft 13 Oct: platform lists, the seven windows, wedge by platform-year, every check on the test of H1, registered tests with outcomes; the detail that Chapters 4 and 5 point to | +| ch6, ch7 | 6 The Rules and How Far They Reach; 7 Conclusion | drafts in `drafts/ch6_draft_2026-10-14.md` and `drafts/ch7_draft_2026-10-14.md`, revised 15 Oct on checked sources (tax started 1 Oct 2026; tax-start prediction withdrawn); 6.5 has a results slot for the 27 Oct to mid-Nov scoring | +| `references.md` | References | merged and checked 15 Oct; verified at primary sources 16 Oct, 67 entries, none marked unverified; APA until the YZU style is settled | +| `symbols.md` | Symbols and Abbreviations | added 16 Oct for the YZU list of symbols | +| `abstract.md` | Abstract | English draft 16 Oct (about 280 words, keywords); the Chinese abstract, keywords and title page are left for the researcher | +| `reading/` | Reading copy | `build_reading.sh` builds `thesis_reading_copy_.pdf` (A4, Source Serif 4, checked page by page) and an editing DOCX from the chapters above; latest 16 Oct, 60 pages, plain thesis style | diff --git a/papers/thesis_2026-10/abstract.md b/papers/thesis_2026-10/abstract.md new file mode 100644 index 0000000..c242127 --- /dev/null +++ b/papers/thesis_2026-10/abstract.md @@ -0,0 +1,15 @@ +# Abstract + +**The Invisible Ledger: Quantifying the Invisible Wedge in Indonesia's Platform Economy** + +Figures about the digital economy stand for activity that no one observes directly. A platform's revenue stands for the commerce it carries, and a survey's count of sellers stands for the businesses selling online. This thesis asks how much activity remains invisible when platforms are measured through their revenue, and when that choice changes the conclusion about growth. It studies Indonesia, using platform reports, BPS's survey of online sellers and national figures. + +Revenue shows only a small part of the commerce. In 2023, three major platforms in Indonesia kept about US$3 billion of the US$43 billion that passed through them. Revenue also stopped tracking the commerce when platforms changed the share of each sale they kept. Indonesian and regional platforms changed that share about three times as much as 27 platforms abroad. Most of their large changes came between 2022 and 2024. So the hypothesis that revenue grows in proportion to the commerce is rejected. + +The national count of online sellers rose by 27 percent in 2023. A test registered before the survey files were opened finds too few new sellers to produce that rise. A later, exploratory count shows that first-time entry explains only about half of it. More businesses account for most of the growth in online sales on the published figures. If the rest of the rise is read as wider survey coverage, that conclusion sits on its threshold. + +The two results describe one problem. A figure stops tracking the activity it stands for when the link between them changes, through a platform's take rate or a survey's reach. The thesis also measures how far Indonesia's recent platform rules can reach: the new marketplace tax can collect from at most about 1 in 20 online sellers. + +**Keywords:** digital platforms; take rate; e-commerce measurement; transaction value; online sellers; Indonesia + +*[Chinese abstract (中文摘要), Chinese keywords and the Chinese title page are required by YZU and are left for the researcher. The proposal's Chinese title was 隱形帳簿：量化印尼平台經濟中的隱形缺口.]* diff --git a/papers/thesis_2026-10/appendix_notes.md b/papers/thesis_2026-10/appendix_notes.md new file mode 100644 index 0000000..26b0b00 --- /dev/null +++ b/papers/thesis_2026-10/appendix_notes.md @@ -0,0 +1,180 @@ +# Appendices + +These appendices hold the detail that the main text points to. Every number comes from a table in the project's repository, named in the notes in small type. Nothing here changes a result in Chapter 5. + +## Appendix A. Platforms, windows and the wedge by year + +### A.1 The foreign group (27 platforms) + +Pairs are consecutive years that pass the rules of Chapter 4, section 4.1: for the platforms added, the later year is not one in which own goods exceed half of revenue; for the eight chosen first, the reporting scope did not change. Ozon's one pair starts in a year that fails the own-goods rule (Appendix B). The median is the platform's median size of yearly change in log take rate. Source: `tables/spinoff_panel.csv`, `tables/transitions_master.csv`. + +| Platform | Chosen first or added | Pairs | Median size of yearly change in log take rate | +|---|---|---|---| +| ACV Auctions | Added | 5 | 0.15 | +| Waitr | Added | 6 | 0.10 | +| Alibaba | Added | 7 | 0.21 | +| Despegar | Added | 9 | 0.07 | +| 1stDibs | Added | 6 | 0.03 | +| Expedia | Added | 11 | 0.03 | +| Fiverr | Added | 7 | 0.05 | +| Grubhub | Added | 6 | 0.09 | +| Hepsiburada | Added | 5 | 0.03 | +| Kaspi.kz | Added | 4 | 0.08 | +| Mogu | Added | 9 | 0.21 | +| Ozon | Added | 1 | 0.18 | +| PDD Holdings | Added | 4 | 0.13 | +| RB Global | Added | 11 | 0.06 | +| The RealReal | Added | 8 | 0.04 | +| Vivid Seats | Added | 6 | 0.07 | +| Upwork | Added | 8 | 0.04 | +| Vacasa | Added | 5 | 0.03 | +| Yatra | Added | 11 | 0.03 | +| Etsy | Chosen first | 3 | 0.07 | +| Jumia | Chosen first | 3 | 0.16 | +| Mercado Libre | Chosen first | 2 | 0.06 | +| Rakuten | Chosen first | 1 | 0.02 | +| Sea / Shopee | Chosen first | 7 | 0.25 | +| Shopify | Chosen first | 5 | 0.05 | +| Zalando | Chosen first | 4 | 0.02 | +| eBay | Chosen first | 4 | 0.03 | + +Left out after examination: Dingdong, JD.com, Jumei, Liquidity Services, Newegg and Vipshop, because their own goods are most of revenue in every year; and four platforms that disclose no transaction value (Nebius Group, formerly Yandex; Trip.com; Tuniu; and Uxin). + +### A.2 The Indonesian and regional group + +| Platform | Pairs | Median size of yearly change in log take rate | Share of pairs with a change of 0.10 or less | +|---|---|---|---| +| Blibli third-party retail | 4 | 0.25 | 0.00 | +| Bukalapak Group | 3 | 0.14 | 0.33 | +| GoTo on-demand | 3 | 0.09 | 0.67 | +| Grab on-demand | 4 | 0.20 | 0.50 | +| Tokopedia | 1 | 0.52 | 0.00 | + +### A.3 The seven windows traced (Table 5.4 in full) + +Shares are of the log growth of revenue. Source: `tables/growth_decomposition.csv`, built from `tables/l_panel.csv` by `growth_decomposition.py`. Tokopedia, GoTo: audited statements. Grab: annual filings, summed over mobility and deliveries. Blibli: incentives are a proxy (gross profit before discounts less net revenue) and stay out of the main sample. GoTo's 2022 transaction value is an estimate (2022 first and second quarters backed out of 2023 year-on-year comparisons). + +| Platform and window | Growth of revenue (%) | Growth of transaction value (%) | Take rate, start (%) | Take rate, end (%) | More commerce | Higher fees | Lower incentives | +|---|---|---|---|---|---|---|---| +| Grab on-demand, 2021–24 | 320.5 | 62.3 | 5.34 | 13.83 | 33.7 | 33.8 | 32.5 | +| Grab on-demand, 2022–24 | 85.8 | 30.1 | 9.69 | 13.83 | 42.5 | 10.6 | 46.9 | +| GoTo on-demand, 2023–25 | 114.2 | 22.4 | 10.62 | 18.58 | 26.6 | 6.1 | 67.3 | +| GoTo on-demand, 2022–25 | 110.7 | 9.6 | 9.66 | 18.58 | 12.3 | 36.1 | 51.7 | +| Tokopedia, 2022–23 | 53.2 | -8.9 | 1.48 | 2.48 | -21.9 | 76.5 | 45.4 | +| Blibli third-party retail, 2022–25 | 679.4 | 43.6 | 0.54 | 2.92 | 17.6 | 62.6 | 19.8 | +| Blibli third-party retail, 2022–23 | 465.3 | 34.7 | 0.54 | 2.25 | 17.2 | 20.1 | 62.7 | + +Tokopedia, 2022–23, line by line (Rp million, `tables/tokopedia_segment.csv`): gross revenue 8,143,239 to 8,988,909; incentives 4,112,320 to 2,813,719; net revenue 4,030,919 to 6,175,190. Net revenue rose by 2,144,271, of which lower incentives account for 60.6 percent and higher gross revenue for 39.4 percent (about 60 and 40 percent in the text). + +### A.4 Wedge and ecosystem ratio by platform-year, Indonesian series + +Original currency and units, as in the filings: Rp million for Tokopedia and Bukalapak, Rp billion for Blibli, and US$ billion for Grab and Shopee, whose Indonesian amounts are derived or estimated. Shopee's revenue assumes a take rate of 8.4 percent in 2022, 10 percent in 2023 and 10.8 percent in 2024. Table 5.2 converts Tokopedia's 2023 rupiah amounts to US dollars. Take rate and ecosystem ratio carry no currency. Each label names the source of the transaction value and of the revenue in that row. Source: `tables/take_rate_levels.csv`. + +| Series | Year | Transaction value | Revenue | Wedge | Take rate (%) | Ecosystem ratio | Label | +|---|---|---|---|---|---|---|---| +| Grab | 2021 | 1.88 | 0.08 | 1.80 | 4.20 | 22.8 | Revenue: country disclosure; value: derived | +| Grab | 2022 | 3.83 | 0.28 | 3.55 | 7.19 | 12.9 | Revenue: country disclosure; value: derived | +| Grab | 2023 | 5.38 | 0.60 | 4.78 | 11.24 | 7.9 | Revenue: country disclosure; value: derived | +| Shopee | 2022 | 18.68 | 1.57 | 17.11 | 8.40 | 10.9 | Value: outside estimate; revenue: scenario | +| Shopee | 2023 | 21.52 | 2.15 | 19.37 | 10.00 | 9.0 | Value: outside estimate; revenue: scenario | +| Shopee | 2024 | 25.99 | 2.81 | 23.18 | 10.80 | 8.3 | Value: outside estimate; revenue: scenario | +| Tokopedia | 2022 | 273,146,166 | 4,030,919 | 269,115,247 | 1.48 | 66.8 | Direct segment disclosure | +| Tokopedia | 2023 | 248,836,156 | 6,175,190 | 242,660,966 | 2.48 | 39.3 | Direct segment disclosure | +| Blibli third-party retail | 2021 | 15,746 | 122 | 15,624 | 0.77 | 128.1 | Segment disclosure, includes travel | +| Blibli third-party retail | 2022 | 37,050 | 199 | 36,851 | 0.54 | 185.2 | Segment disclosure, includes travel | +| Blibli third-party retail | 2023 | 49,912 | 1,125 | 48,787 | 2.25 | 43.4 | Segment disclosure, includes travel | +| Blibli third-party retail | 2024 | 54,210 | 1,384 | 52,826 | 2.55 | 38.2 | Segment disclosure, includes travel | +| Blibli third-party retail | 2025 | 53,187 | 1,551 | 51,636 | 2.92 | 33.3 | Segment disclosure, includes travel | +| Bukalapak Group | 2020 | 85,081,843 | 1,351,664 | 83,730,179 | 1.59 | 61.9 | Company-wide disclosure | +| Bukalapak Group | 2021 | 122,618,794 | 1,869,122 | 120,749,672 | 1.52 | 64.6 | Company-wide disclosure | +| Bukalapak Group | 2022 | 153,738,958 | 3,618,366 | 150,120,592 | 2.35 | 41.5 | Company-wide disclosure | +| Bukalapak Group | 2023 | 164,380,065 | 4,438,269 | 159,941,796 | 2.70 | 36.0 | Company-wide disclosure | + +## Appendix B. The checks on the test of H1 + +### B.1 The checks + +The main test is in Table 5.3. This appendix lists each check with both p-values where both exist. Sources: `h1_extended.py` (`tables/h1_extended_results.json`), `tables/test_battery.csv`, `tables/h1_quarterly_results.json`, `tables/h1_points_vs_log.json`, `tables/h1_thesis_drops.json`. The two rows on how platforms book shopper discounts come from `research/option_space_2026-10-01/materials_2026-10-15/incentive_accounting_check.py` (classification and filing quotes in `INCENTIVE_ACCOUNTING.md`); their shuffle uses a different seed. + +| Check | Indonesian and regional (median) | Foreign (median) | Mann–Whitney p | Shuffle p | +|---|---|---|---|---| +| Main test, 5 against 27 | 0.201 | 0.062 | 0.0038 | 0.0221 | +| 5 against the 8 chosen first | 0.201 | 0.055 | 0.0225 | 0.0497 | +| 5 against the 19 added | 0.201 | 0.070 | 0.0046 | 0.0175 | +| Without the 8 foreign platforms that book shopper discounts as marketing costs (exploratory, 15 Oct 2026) | 0.201 | 0.062 | 0.0035 | 0.0066 | +| Only the 11 foreign platforms that deduct shopper discounts from revenue (exploratory, 15 Oct 2026) | 0.201 | 0.075 | 0.0043 | 0.0052 | +| Growth divergence in points, 5 against 26 | 27.0 | 6.5 | 0.0055 | | +| Without Ozon, 26 foreign platforms | 0.201 | 0.060 | 0.0030 | 0.0215 | +| Without Zalando, 26 foreign platforms | 0.201 | 0.064 | 0.0045 | 0.0238 | +| Without Ozon and Zalando, 25 foreign platforms | 0.201 | 0.062 | 0.0036 | 0.0230 | +| Without Blibli third-party retail | 0.169 | | 0.0137 | 0.0364 | +| Without Bukalapak Group | 0.225 | | 0.0049 | 0.0070 | +| Without GoTo on-demand | 0.225 | | 0.0030 | 0.0071 | +| Without Grab on-demand | 0.194 | | 0.0094 | 0.0166 | +| Without Tokopedia | 0.169 | | 0.0164 | 0.0409 | +| Three marketplaces against the 8 chosen first, comparing pairs of years, absolute growth divergence in points | 38.917 | 7.105 | 0.0022 | 0.0606 | +| Three marketplaces against the 8 chosen first, one median per platform, absolute growth divergence in points (claims ledger C47, `build_numbers.py`) | | | 0.0242 | | +| Three marketplaces against the 8 chosen first, comparing pairs of years, absolute growth divergence in points (without Sea) | 38.917 | 4.192 | <0.0001 | 0.0083 | +| Three marketplaces against the 8 chosen first, comparing pairs of years, absolute change in log take rate | 0.252 | 0.055 | 0.0037 | 0.0303 | +| Three marketplaces against the 8 chosen first, comparing pairs of years, absolute change in log take rate (without Sea) | 0.252 | 0.033 | 0.0001 | 0.0167 | +| Three marketplaces, dropping all three 2022–23 comparisons (5 pairs), absolute growth divergence in points | 14.411 | 7.105 | 0.0248 | 0.2222 | +| Three marketplaces, dropping the largest comparison (7 pairs), absolute growth divergence in points | 15.738 | 7.105 | 0.0061 | 0.0606 | +| Three marketplaces against all 40 foreign pairs, absolute growth divergence in points | 38.917 | 7.373 | 0.0014 | 0.0121 | +| Five Indonesian and regional series including constructed Grab and Shopee (12 pairs) against the 8 chosen first, absolute growth divergence in points | 42.017 | 7.105 | 0.0006 | 0.0225 | +| Tokopedia and GoTo on-demand as one company, 4 against 27 | | | 0.003 | 0.007 | +| Quarterly pairs, 3 against 8 (same quarter a year earlier) | 0.071 | 0.049 | 0.19 | 0.33 | + +Growth in revenue and growth in transaction value had a correlation of 0.10 in the Indonesian main sample and 0.98 in the foreign sample excluding Sea (`tables/claims_ledger.csv`, C48 and C49). + +The quarterly sample is Grab on-demand, GoTo on-demand and Blibli against Sea, Airbnb, Booking, DoorDash, eBay, Etsy, Lyft and Uber. Airbnb and Booking use trailing four-quarter sums because they book revenue at check-in. DoorDash pairs ending in 2025Q4 or later are dropped because of an acquisition. + +In the pairs of the Indonesian and regional group, the take rate stayed within about 10 percent (0.10 in log terms) in one third of pairs, against 67 to 69 percent in the foreign groups. The version of the test with growth divergence in points has 26 foreign platforms because Ozon has no pair in which both years pass the own-goods rule. The main test keeps Ozon's pair, whose later year passes, and the first rows above show that dropping it, or Zalando, leaves the result in place. + +The comparison in points of the take rate (exploratory, `h1_points_vs_log.py`): median yearly change 1.01 points in the Indonesian and regional group and 0.98 points in the foreign group (26 platforms), Mann–Whitney p = 0.65; median share kept 1.59 percent against 16.93 percent. + + +### B.2 Where the 2023 break sits (exploratory) + +At the top of the size distribution, sellers with revenue above Rp50 billion a year (about US$3.4 million at the 2022 rate) were 0.5 percent of BPS's marketplace sellers in 2022 and held between 18 and 39 percent of the marketplace slice's value, resting on 19 sample rows (`tables/part5_checks.json`). + +Across the 34 provinces, count growth had a rank correlation of 0.10 (p = 0.59) with the share of new sellers and 0.35 (p = 0.04) with growth in sample size. In 23 provinces count growth exceeded what entry can explain, and these provinces carry 98.5 percent of the increase. The standard error of the national count implied by BPS's published relative standard errors is about 1.4 percent. The 306 thousand newly counted sellers are about 4.6 times the standard error of the change, and 2.3 times if the true error were twice as large (`tables/bps_frame_check.json`, `tables/rq_closing_checks.json`). + +## Appendix C. Registered tests and their outcomes + +A registered test was written down, dated and stored in the repository before the data it concerns were opened. The dates are those of the commits in `research/option_space_2026-10-01/PREREGISTRATION.md`. Exact wording and thresholds are in that file. + +| Test | What it predicted | Registered | Outcome | +|---|---|---|---| +| E1 | Sellers without financial statements hold a smaller share of online value than of the count | 2 Oct 2026 | Holds, as a range (25.7 to 40.4 percent of value against 79.3 percent of the count) | +| E2 | The marketplace share of online value in 2022 is above the published 2023 figure | 2 Oct 2026 | Holds, as a range (22.6 to 25.6 percent) | +| E2b, E2c | Marketplace use did not rise between 2020 and 2024 | 2 Oct and 3 Oct 2026 | Hold (21.6, 19.8, 17.8 and 17.2 percent) | +| E3 | Shopee reaches at least as many sellers as Tokopedia | 2 Oct 2026 | Holds (7.5 against 2.2 percent of all sellers) | +| E4 | Entrants use chat or social media at least as often as earlier sellers | 2 Oct 2026 | Holds, with no power (nearly everyone uses them) | +| E5 | At least 21.5 percent of 2023 sellers began in 2023 (the entry test) | 2 Oct 2026 | Falsified (13.45 percent). Disclosure: BPS's 2024 publication, held at registration, has a table of 2024 sellers by start year; if the seller population were stable from 2023 to 2024, it would cap the 2023 share near 18 percent. The table was first used on 8 Oct, after the test ran on 7 Oct | +| E6 | The median change in sales of sellers already online is at most 10.36 percent | 2 Oct 2026 | Holds (median 0 percent) | +| E7 | The share of value held by sellers without statements fell by at most 5 points from 2020 to 2022 | 3 Oct 2026 | Falsified, narrowly (5.8 points; 3.3 on the second version) | +| E8 | Sellers with revenue of Rp300 million a year or more hold over half of marketplace value | 3 Oct 2026 | Holds (92.2 percent, middle scenario, an upper bound) | +| RBI | The marketplace part sold to final consumers is below Bank Indonesia's 2022 figure | 3 Oct 2026 | Holds in the low (0.16) and middle (0.68) scenarios; fails in the high one (1.62) | +| W1, W2 | Unregistered firms using social media keep written records and take digital payments more often | 3 Oct 2026 | Hold (29.9 against 12.3 percent; 11.5 against 2.7 percent) | +| W3 | Their mean annual sales are below Rp100 million (about US$6,600) | 3 Oct 2026 | Falsified (mean Rp111 million; median Rp60 million) | +| F1, F2 | Formal firms with a website take more e-payments and file taxes online more often | 3 Oct 2026 | Hold | +| F3 | More than one third of formal firms have neither a website nor online tax filing | 3 Oct 2026 | Holds (39.9 percent), after a disclosed coding fix: the first run read "yes, partially" as no for online filing and gave 13.0 percent; the questionnaire counts it as yes | +| X3 | Indonesia's share of firms with neither is the highest of the countries compared | 3 Oct 2026 | Withdrawn as a test (addendum of 16 Oct 2026): published e-filing shares seen at registration already capped every other country below Indonesia. Reported as descriptive (Malaysia 9.1, Thailand 8.7) | +| X2 | Firms with a website file taxes online more often, in at least 3 of 4 countries | 3 Oct 2026 | Holds in 4 of 4 | +| Q1 | The test of H1 with quarterly pairs rejects H1 | 7 Oct 2026, before it was run, but after the quarterly series had been built and the annual result was known, so it is a robustness check | Falsified (p = 0.19) | +| M1 | Markets value growth from more transactions more than growth from a higher take rate | 8 Oct 2026, before the market values were pulled | Falsified in the main sample (p = 0.20) and with extreme values capped (p = 0.16); holds in the second run of 34 platforms (p = 0.004), which mixes segment figures with whole-company values | +| M2 | Market value moves with transaction value | 8 Oct 2026 | Holds (p = 0.002) | +| L (P1 to P3) | Large take-rate moves driven by discounts occur only where incentives are at least as large as revenue (L of 1 or more) | 2 Oct 2026 | Thin support out of sample: P1 holds on 2 of 2 cases; P3 fails at the main threshold (median move 0.117 against 0.10); rank correlation 0.20 (p = 0.43) | +| P1 to P5, R1 to R3 | Predictions on the third-quarter 2026 results of GoTo, Grab and Sea | 2 and 3 Oct 2026 | Not yet scored; first results are due on 27 Oct 2026 | +| Marketplace tax start | The tax start slips again | 2 Oct 2026 | Withdrawn: collection began on 1 Oct 2026, before the prediction was registered (addendum of 15 Oct 2026) | +| BPS marketplace share | The next published marketplace share is at most 15.79 percent | 2 Oct 2026 | Not yet scored | + +## Appendix D. The values behind Figures 5.3 to 5.5 + +**Growth in 2023, percent (Figure 5.3).** Household spending +9.4; Tokopedia transaction value −8.9; GoTo on-demand transaction value −10.5 (rests on an estimated 2022 value); Bank Indonesia e-commerce −4.7; Momentum Works marketplace estimate +3.5 (calculated from its published levels; its 2024 release states 3.7); Bukalapak transaction value +6.9; Blibli third-party retail transaction value +34.7; Tokopedia revenue +53.2; BPS e-commerce value +40.6; BPS online sellers +27.4. Sources: `tables/consistency_grid_growth.csv`, `tables/yardsticks_bi_bps.csv`, `tables/wedge_over_time.csv`, `tables/h1_groundwork.json`, company filings. + +**Who the extra sellers were, percent (Figure 5.4).** Share of the 2022 count and share of the extra count: business opened before 2010, 18.3 and 40.5; owner aged 50 or over, 28.9 and 41.2; outside Java, 23.6 and 36.3; food and lodging or making goods, 38.5 and 53.0. Source: `tables/bps_newly_counted_profile.json`. + +**Marketplace use and financial statements, percent of sellers (Figure 5.5).** Marketplace use: 21.6 (2020), 19.8 (2022), 17.8 (2023), 17.2 (2024, published). Complete financial statements: 23.5 (2020), 20.7 (2022), 15.2 (2023). Source: `tables/bps_descriptives.json`. + +The figures are drawn from the tables named above (`papers/thesis_2026-10/figures/make_figures.py`). In each figure, a legend or label names every colour. diff --git a/papers/thesis_2026-10/ch1_introduction.md b/papers/thesis_2026-10/ch1_introduction.md index 1965b96..7fe9fbb 100644 --- a/papers/thesis_2026-10/ch1_introduction.md +++ b/papers/thesis_2026-10/ch1_introduction.md @@ -1,43 +1,43 @@ # Chapter 1. Introduction -Digital commerce has changed how goods and services are bought and sold in Indonesia. Sellers in small towns reach buyers across the country, businesses operate from mobile phones, and payments are made by QR code. The figures used to measure this activity, however, differ on both its size and its growth. +Digital commerce has changed how goods and services are bought and sold in Indonesia. Sellers in small towns reach buyers across the country, businesses run from mobile phones, and customers pay by QR code. The figures that measure this activity disagree on both its size and its growth. -The difference was visible in 2023. Traders at Tanah Abang, a major wholesale market in Jakarta, reported that their customers had moved to sellers on TikTok Shop. The government then banned transactions on social media platforms, and TikTok Shop closed in Indonesia shortly afterwards. The national statistics office, BPS, later reported that online sales rose by about 41 percent that year. The central bank's measure of online shopping fell by about 5 percent over the same period. +In 2023, traders at Tanah Abang, a major wholesale market in Jakarta, said their customers had moved to sellers on TikTok Shop (Agence France-Presse 2023). The government then banned transactions on social media platforms, and TikTok Shop closed in Indonesia soon after (Ministry of Trade of the Republic of Indonesia 2023; Mandagie 2023). That year the national statistics office, BPS, reported that online sales rose by about 41 percent (BPS 2025a). The central bank's measure of online shopping fell by about 5 percent (Bank Indonesia 2023; KabarBursa 2024). The two official figures moved in opposite directions. -The same divergence appears within a single company. Between 2022 and 2023, Tokopedia's revenue rose by 53 percent while the value of transactions on the platform fell by 9 percent. Both figures come from the company's own reports. Measured by revenue, the platform grew; measured by transaction value, it contracted. +Tokopedia, an Indonesian online marketplace, shows revenue and commerce moving apart inside one company. Between 2022 and 2023, its net revenue rose by 53 percent while the value of transactions on the platform, its transaction value, fell by 9 percent (GoTo Gojek Tokopedia 2024a, 2024b). The two are different amounts. A platform keeps only part of each sale, as commissions and fees. The rest is the seller's receipts, the driver's fare or another amount that passes through. This thesis calls the difference the invisible wedge. The term means what a revenue-only measure leaves out. It does not mean activity that goes unrecorded. -These figures differ because they measure different things. A platform that handles 100 units of sales may keep only a small part as its own revenue, in the form of commissions and fees. The rest belongs to the sale itself: the seller's receipts, the driver's fare and other amounts that pass through. An analyst reading the platform's accounts sees the small part, while one measuring the commerce it carries sees the whole. This thesis calls the difference the invisible wedge. The term refers to what a revenue-only measure leaves out, not to activity that goes unrecorded. +The share a platform keeps is not fixed. When it rises, revenue can grow even while transaction value falls. In Tokopedia's accounts, about 60 percent of the extra revenue came from smaller discounts and incentives, which a platform pays for out of what it earns (GoTo Gojek Tokopedia 2024b). -The difference matters because money and public decisions depend on these figures, and not only in Indonesia. Businesses in 43 economies sold almost US$27 trillion online in 2022, by the United Nations' estimate, and platforms carry a growing part of that trade. The companies that run them are among the most valuable in the world, and they are valued largely on the growth they report. If part of that growth comes from a platform keeping more of each sale, and not from more commerce, then part of what investors pay for is a pricing decision. Governments face the same uncertainty. They use these figures to judge how large their digital economies are, and a growing number now ask platforms to report what their sellers earn, or to collect tax from them. A rule built on one figure reaches only what that figure covers. Each of these readers is acting on a number without knowing how far it can be relied on. +Platform and online sales figures carry weight well beyond Indonesia. Businesses in 43 economies sold almost US$27 trillion online in 2022, by the United Nations' estimate (UNCTAD 2024), and platforms carry a growing part of that trade. Investors value the companies that run them largely on reported growth. If part of that growth is a platform keeping more of each sale, then part of the share price rests on a pricing decision. Governments use the same figures to size their digital economies and to tax online sellers. -Existing research gives investors and governments limited help in checking the figures they use. It explains why the figures differ: platforms set prices on both sides of a market, accounting rules decide how much of each sale counts as revenue, and national accounts separate a platform's own service from the commerce it carries. Statistical agencies have also begun to set business surveys beside platform transaction values. What this work does not show is when a figure stops moving with the activity it is used to represent, or by how much. Answering that requires two things that are rarely available together: a platform's revenue and transaction value reported over several years, and a national count of sellers that can be examined from the inside. +Research explains why platform revenue and national sales figures can differ, but not when a figure stops moving with the activity it represents, or by how much. Answering that takes two things that are rarely found together. One is a platform's revenue and transaction value over several years. The other is a national count of online sellers whose underlying answers can be examined. -Indonesia is one of the few markets where both kinds of evidence are available. Several of its platforms report their transaction value and their revenue each year, and BPS asks online sellers when they began selling and how their sales changed. The stakes are also concrete there. For 2024, the state's two main estimates of online sales are about Rp1,290 trillion and about Rp490 trillion, so the size of the market is uncertain by a factor of more than two. The government is nevertheless writing rules for that market, on platform commissions, on discounts and on tax. Its new levy on marketplace sellers would raise about Rp1 trillion a year on one measure of the market and about Rp5 trillion on another, before exemptions. Indonesia is also a market where platform figures and survey figures are likely to part, because most of its online sellers work outside the marketplaces, through messaging and social media applications. The events of 2023 suggest that they did part. +In Indonesia, several platforms report revenue and transaction value each year, and BPS asks online sellers when they began and how their sales changed. The stakes are concrete. For 2024, BPS gives about US$81 billion for online sales and the central bank about US$31 billion, more than two to one (BPS 2025b; Bank Indonesia 2025). The government has set new rules on platform commissions, discounts and tax. Its tax on marketplace sellers can raise about US$60 million a year on one measure of the market and about US$300 million on another, before exemptions (Ministry of Finance of the Republic of Indonesia 2025; author's calculation). Most online sellers work through messaging and social media, outside the marketplaces, so platform and survey figures are likely to part. The events of 2023 suggest that they did. ## 1.1 Research question and objectives This thesis sets out to change how the digital economy is measured. Most online commerce is missing from the platform figures on which companies are valued and sellers are taxed, and millions of the people who sell online are outside those figures altogether. Decisions worth trillions of dollars rest on numbers that show a fraction of the activity they are taken to describe. The ambition of this thesis is to bring the rest into view: to show how much of the online economy the current figures miss, where it is, and who is in it. -Measuring what the current figures miss begins with one question: how much economic activity remains invisible when digital platforms are measured through their reported revenue, and when does that choice change the conclusion about growth? The question has two halves. The first concerns size, which is how much of the activity lies outside the figure. The second concerns growth, which is whether the figure still moves with the activity once prices or coverage change. A large gap would matter less if it stayed in proportion; the second half asks whether it did. +Measuring what the current figures miss begins with one question: how much economic activity remains invisible when digital platforms are measured through their reported revenue, and when does that choice change the conclusion about growth? The question has two halves. The first concerns size, which is how much of the activity lies outside the figure. The second concerns growth, which is whether the figure still moves with the activity once prices or coverage change. A large wedge would matter less if it stayed in proportion; the second half asks whether it did. -The research question is addressed through three objectives. The first is to measure the invisible wedge across the platform histories available from 2020 to 2025. The second is to reconcile selected large movements with disclosed changes in monetization, customer incentives, revenue recognition and reporting scope. The third is to compare the platform evidence with national e-commerce and payment indicators, to test whether they give the same account of economic scale and growth. +The research question is addressed through three objectives. The first is to measure the invisible wedge across the platform histories available from 2020 to 2025. The second is to reconcile selected large movements with disclosed changes in monetization, customer incentives, revenue recognition and reporting scope. The third is to compare the platform evidence with national e-commerce and payment figures, to test whether they give the same account of economic scale and growth. -Two hypotheses test whether revenue and the national figures move with the activity behind them. H1 states that a platform's revenue grows in proportion to its transaction value. H2 states that national e-commerce growth comes mainly from more businesses selling online, and less from higher sales per business. Chapter 3 sets out both in full. A further analysis estimates how many sellers a rule based on platform records can reach. +The thesis tests two hypotheses. H1 says that a platform's revenue grows in proportion to its transaction value. H2 says that national e-commerce growth comes mainly from more businesses selling online, and less from higher sales per business. An entry test then asks whether a rise in the count of online sellers comes from sellers entering for the first time. A further analysis estimates how many sellers a rule built on platform records can reach. -In brief, the thesis finds that platform revenue and the official count of sellers both stopped moving with the activity behind them in the period studied. Platform revenue rose mainly because platforms kept a larger share of each sale, and about half of the 2023 increase in the official count of sellers was not first-time entry. Chapters 5 and 6 establish how large these effects were and what follows from them. +In brief, revenue shows only a small part of the commerce that platforms carry. In BPS's survey, the marketplaces carry only about a fifth of online sales value. In Indonesia, platform revenue also grew faster than the commerce behind it, mainly because platforms kept a larger share of each sale and not because buyers bought more. National growth came mainly from more sellers, as H2 states, but about half of the 2023 rise in the count of sellers is not explained by first-time entry. ## 1.2 Contributions The thesis makes three contributions, one for each group that relies on these figures: investors, statistics offices, and the authorities that tax and regulate online selling. -For those who value platforms, it shows why the size of the invisible wedge matters for growth and not only for scale. The larger the wedge is relative to revenue, the more a given change in fees or discounts moves revenue growth. This gives investors a direct check on reported growth: whether it came from more commerce or from the platform keeping a larger share of each sale. +For those who value platforms, the thesis shows how small a part of the commerce a platform's revenue records. This matters because market values of internet firms track revenue (Bowen, Davis and Rajgopal 2002). It matters for growth as well as scale: the larger the wedge is next to revenue, the more a small change in fees or discounts moves revenue growth. When the take rate moves, market value per dollar of revenue and market value per dollar of commerce change by different amounts. The thesis also traces why revenue moved, which gives investors a check on whether reported growth came from more commerce or from a larger share of each sale. A prediction registered in advance then tests how market value moves with each of the two. -For those who measure the economy, it provides a test of what an increase in a national count of online sellers represents. Each seller in the survey reports the year they began selling online, so the survey itself shows how much of an increase is first-time entry. The test was specified before the survey files were opened. It bears directly on a market whose official size estimates differ by a factor of more than two. +For those who measure the economy, the thesis shows where the commerce outside the platforms sits and tests what a rise in the count of online sellers represents. Most online sellers are small businesses that keep no financial statements. Each reports the year it began selling online, so the survey shows how much of a rise is first-time entry. The test was specified before the survey files were opened. -For those who design tax and regulation, it measures how far rules based on platform records reach. Where most sellers work outside the platforms, such rules cover a minority of them, and the revenue they can raise depends on which measure of the market is correct. Because several countries have adopted similar rules, the result is relevant beyond Indonesia. +For those who design tax and regulation, the thesis measures how far rules based on platform records reach. A rule built on platform records reaches only the sellers who use platforms, so where most sellers work outside them, such rules cover a minority. The thesis also sets each recent rule, from the cap on ride-hailing commissions to the limit on discounts, against the figure it acts on. Several other countries have adopted similar rules, so the same question applies there. -Earlier work has already compared platform transaction values with national e-commerce statistics. The contribution of this thesis is the connected evidence for one country, drawn from the platforms' accounts, the sellers' survey responses and the rules built on both. +Earlier work has already compared platform transaction values with national e-commerce statistics. The contribution of this thesis is connected evidence for one country. It draws on the platforms' accounts, the sellers' own survey responses and the rules that act on those figures. Across all three, one point recurs. A figure can stop tracking the activity it stands for when the link between them changes, through a platform's take rate or a survey's reach. ## 1.3 Structure of the thesis -Chapter 2 reviews the literature on platform revenue, the measurement of online selling and the use of platform records for tax. Chapter 3 defines the measures and develops the hypotheses. Chapter 4 describes the data and the tests. Chapter 5 presents the results, beginning with the platforms and moving to the national figures. Chapter 6 examines how far Indonesia's recent rules reach and what the findings imply for other countries. Chapter 7 concludes and states the limitations. +The thesis moves from ideas to evidence and then to what follows from it. Chapter 2 reviews why platform revenue differs from the commerce behind it, how online selling is measured, and what platform records can show. It ends with the questions the literature leaves open. Chapter 3 turns those questions into measures, two hypotheses, the entry test and four further analyses. Chapter 4 describes the data, which come from platform reports, BPS survey files and national and market figures, and the tests that use them. Chapter 5 gives the results, first for the platforms and then for the national figures. Chapter 6 asks how far Indonesia's recent rules reach and what the findings mean for other countries. Chapter 7 concludes and states the limits of the evidence. diff --git a/papers/thesis_2026-10/ch2_literature.md b/papers/thesis_2026-10/ch2_literature.md new file mode 100644 index 0000000..bcd3fa2 --- /dev/null +++ b/papers/thesis_2026-10/ch2_literature.md @@ -0,0 +1,49 @@ +# Chapter 2. Literature Review + +Digital platforms sit between economic activity and the figures used to measure it. A platform carries a sale between a buyer and a seller. It records part of the payment as its own revenue, and it keeps a record of the whole sale. Different bodies of research study these steps. Platform economics and accounting explain why revenue is only part of each payment (Section 2.1). Official statistics explain how online sales across the country are measured (Section 2.2). Research on third-party information explains what a platform's records can show to tax authorities (Section 2.3). Each body of research explains why the figures differ, but none measures how far they differ in practice. Section 2.4 sets out this research gap. + +## 2.1 Platform economics and revenue recognition + +A platform reports two figures that describe its business: the value of the commerce it carries, and the revenue it keeps from them. Investors price that revenue: market values of internet firms track it, and share prices respond to revenue news beyond earnings news (Bowen, Davis and Rajgopal 2002; Davis 2002). Revenue is a fair guide to the commerce only if it keeps a steady share of that commerce. + +Platform economics explains why revenue is only a small share of the commerce. A platform serves two groups that need each other, such as shoppers and merchants, or riders and drivers. It sets its charges for both groups together and earns fees for arranging trades, not the value of each sale (Rochet and Tirole 2003; Armstrong 2006; Evans and Schmalensee 2016). The level of its charges and the split between the two groups are choices, so the platform can take more or less revenue from the same commerce (Rochet and Tirole 2003; Caillaud and Jullien 2003). + +Accounting rules add a second reason. Under the international standard, International Financial Reporting Standard 15 (IFRS 15), a company that only arranges a sale records its fee and not the full payment. If a rider pays 100 for a trip and the driver keeps 80, such a company records 20. Discounts and vouchers that it pays for are normally deducted from its revenue (International Accounting Standards Board 2014). So revenue can rise when a platform raises its fees or cuts its discounts, with no change in what customers buy. + +Platform economics and accounting explain why the share is small and why it can move, but they leave its actual size and movement to be measured. The pricing models were built to explain how a platform sets its charges, so they predict the direction of a change, not its size in a given market. The accounting studies asked how internet firms around 2000 chose to report revenue, not how far that revenue drifted from the commerce behind it (Bowen, Davis and Rajgopal 2002). Studies of how investors valued those firms, and how analysts forecast their revenue, relied on web traffic as their measure of activity (Trueman, Wong and Zhang 2000, 2001; Rajgopal, Venkatachalam and Kotha 2003). Transaction value also needs care, because each company defines it for itself. A figure of this kind loses information when its definition changes (Givoly, Li, Lourie and Nekrasov 2019). Testing whether revenue stays a fair guide needs the reported revenue and transaction value of the same platforms, over several years and on one definition. The studies reviewed here do not make that comparison, and none covers Indonesian platforms. + +## 2.2 Measuring the digital economy + +A country's online commerce is larger than what platforms carry, because many sellers trade outside platforms. Several producers measure it, and each counts something different. Industry reports add up the value of goods sold through platforms (Google, Temasek and Bain 2024). Statistics offices, such as BPS in Indonesia, survey businesses about their online sales (BPS 2025a, 2025b). Central banks count electronic payments (Bank Indonesia 2025). The resulting values change with the definition used (UNCTAD 2023, 2024). + +Official statistics address the question in two ways. National accounts separate the sale between a buyer and a seller from the platform's fee, and they count only the fee as the platform's output (Ahmad and Schreyer 2016; International Monetary Fund 2018; Organisation for Economic Co-operation and Development [OECD] 2023; European Commission et al. 2025). Research on measurement also offers a method: test a figure against the activity it stands for, as studies of night-time lights do for national income (Henderson, Storeygard and Weil 2012). Growth in online sales can be split in the same spirit, into more sales by existing sellers and more sellers entering. Trade economists split a country's exports in this way, into more of each good and more kinds of goods (Hummels and Klenow 2005). + +Official statistics and measurement research say what to count and how to check one figure against another. Neither follows platform figures and national figures through time in one country. National accounting guidance defines the platform's output, and it does not track how platform commerce and national online sales grow. The test of one figure against another has been applied to national income and night-time lights. Among the studies reviewed here, it has not been applied to platform commerce and national online sales. + +The growth split has its own difficulty, because it rests on a count of sellers. Such counts can mislead for known reasons. A change in survey design can create a break that looks like real change (van den Brakel, Smith and Compton 2008; van den Brakel, Zhang and Tam 2020). People tend to report past events as more recent than they were (Neter and Waksberg 1964). Small firms that stop are more often young firms with young owners (McKenzie and Paffhausen 2019). Published totals cannot show whether a rise in the count reflects new sellers or a wider survey, because they do not record when each seller began. That question needs the answers of the sellers themselves. + +## 2.3 Digital records and third-party information + +A platform records every sale it carries: who sold what, to whom, for how much and when. Governments have begun to use these records for tax. The OECD, a group of mostly high-income countries, set out model rules under which platforms report what their sellers earn (OECD 2020). The European Union applies these rules through its DAC7 directive (European Union 2021). Indonesia goes further and requires marketplaces to collect tax from their sellers (Ministry of Finance of the Republic of Indonesia 2025). + +Tax research explains why such records matter. Reports from third parties improve compliance. In Denmark, evasion was close to zero on income that employers and banks reported, and substantial on income that taxpayers reported themselves (Kleven et al. 2011). In Chile, the paper trail of the value-added tax deters evasion, and enforcement spreads up the supply chain (Pomeranz 2015). In Brazil, rewards for asking for receipts raised the sales that firms reported (Naritomi 2019). Digital records can also hold information that conventional records lack (Berg, Burg, Gombovic and Puri 2020). + +Tax research shows that platform records are valuable where they exist. Its evidence comes from settings where a third party already sees the transaction, such as an employer paying wages or a firm buying from another firm. Online selling differs, because many sellers trade through messaging apps and social media, where no marketplace records the sale (BPS 2025a). How far a rule built on platform records can reach depends on how much selling happens on platforms. The studies reviewed here do not estimate that share for Indonesia. + +## 2.4 Research gap + +The shortfalls identified in Sections 2.1 to 2.3 define the research gap. Table 2.1 sets them side by side. + +**Table 2.1. What the literature establishes and where it stops** + +| Section | What it establishes | Main sources | Where it stops | +|---|---|---|---| +| 2.1 Platform economics and accounting | Revenue is a fee that is a small share of the commerce, and the share can move with pricing, discounts and accounting | Rochet and Tirole 2003; Armstrong 2006; IASB 2014 | Whether reported revenue stays a fair guide to reported transaction value has not been tested, and not for Indonesian platforms | +| 2.2 Official statistics and surveys | Platform commerce and platform revenue differ; sources differ; a count of sellers can break | Ahmad and Schreyer 2016; Henderson et al. 2012; van den Brakel et al. 2008 | Platform and national figures have not been followed through time in one country, and published totals cannot show what a rise in the count of sellers contains | +| 2.3 Third-party information | Platform records show sales, and third-party reports improve compliance | Kleven et al. 2011; OECD 2020 | How much of online selling platform records cover has not been estimated for Indonesia | + +Each shortfall becomes one question for the thesis, and Chapter 3 sets out the tools that answer each. + +1. **How large is the difference between transaction value and revenue, and does it last?** This follows from Section 2.1. It explains why the share is small and can move, but it does not test whether revenue stays a fair guide to the commerce. +2. **Does the choice of figure change the conclusion about growth?** This follows from Section 2.2. It shows that figures differ, but it does not test whether they grow for the same reasons, or what a rise in the count of sellers contains. +3. **How far can rules built on platform records reach?** This follows from Section 2.3. It shows that platform records are valuable, but not how much of online selling they cover. diff --git a/papers/thesis_2026-10/ch3_framework.md b/papers/thesis_2026-10/ch3_framework.md new file mode 100644 index 0000000..2e1f775 --- /dev/null +++ b/papers/thesis_2026-10/ch3_framework.md @@ -0,0 +1,201 @@ +# Chapter 3. Framework and Hypotheses + +Platform activity can be observed through several figures, and these figures do not measure the same economic object. Transaction value measures the commerce a platform carries. Revenue measures the part of that commerce the platform keeps. National statistics measure online sales across all sellers, whether or not they use a platform. + +The framework tests each figure against the activity it is used to represent, as studies of night-time lights do for national income (Henderson, Storeygard and Weil 2012). It starts from the difference between the value a platform processes and the revenue it recognizes. It then asks whether that relationship stays stable over time, and whether figures from other sources give the same account of online activity. Figure 3.1 shows the structure. + +Figure 3.1 places the three quantities inside one another. Revenue sits inside platform commerce, because a platform keeps only part of each sale. Platform commerce sits inside all online commerce, because many sellers trade outside platforms. The first hypothesis, H1, tests the inner link between revenue and platform commerce. The second, H2, tests the outer link between platform commerce and national online commerce. + +![Figure 3.1](figures/fig3_1_two_links.png) + +**Figure 3.1. Three nested quantities and the two links.** Revenue sits inside platform commerce, which sits inside all online commerce. H1 tests the inner link and H2 the outer link. + +*Note:* The figure is illustrative, and the sizes are not to scale. The platform circle differs from the marketplace channel in BPS's survey of online sellers. + +The research question asks how much activity remains invisible when platforms are measured through revenue, and when that choice changes the conclusion about growth. The literature leaves three parts of it open (Section 2.4). The framework answers each with its own tools: + +1. **How large the difference between transaction value and revenue is, and whether it lasts:** three measures, a split of revenue growth, and H1 (Sections 3.2 to 3.4). +2. **Whether the choice of figure changes the conclusion about growth:** H2, the entry test and a comparison of national figures (Sections 3.4 to 3.6). +3. **How far rules built on platform records reach:** how online selling divides between marketplaces and other channels (Section 3.6). + +The three objectives set in Chapter 1 sit inside this map. Measuring the wedge and tracing its large moves serve the first question, and comparing platform figures with national and payment figures serves the second. The third question is answered by a further analysis. + +## 3.1 Terms and rules for comparison + +Table 3.1 defines the terms used throughout the thesis, each once, with the symbol used in the formulas. + +**Table 3.1. Terms and symbols** + +| Term | Symbol | Meaning | +|---|---|---| +| Transaction value | V | Everything bought through a platform in a period, also called the commerce the platform carries. Companies call it gross merchandise value (GMV), gross transaction value (GTV) or total payment volume (TPV); the thesis treats these as one measure and notes where a company's scope differs. | +| Revenue | R | What the platform recognizes as its own income in its accounts. | +| Take rate | m | Revenue divided by transaction value: the share of each sale that the platform keeps. | +| Customer incentives | | Discounts, vouchers and other promotions that a platform pays for. Accounting rules normally deduct them from revenue, and Tokopedia reports the deduction as a separate line. | +| Fee rate | f | The platform's revenue before customer incentives, divided by transaction value. | +| Incentive rate | s | Customer incentives divided by transaction value. | +| Invisible wedge | W | Transaction value minus revenue: what a revenue figure leaves out. It is defined for one platform. The thesis sums it over three platforms in Indonesia for 2023 and computes no national total. | +| Ecosystem ratio | E | The wedge divided by revenue. It equals one divided by the take rate, minus one. This is the author's own measure, built for this thesis. | +| Growth | g | The change from one year to the next as a share of the first year: 0.09 means 9 percent. | +| Growth divergence | D | Growth of revenue minus growth of transaction value in the same year: 0.62 means 62 percentage points. | +| Online seller | | A business that the national statistics office, BPS, counts as selling online; the text also calls it a seller or a business. The count of online sellers is BPS's estimate of how many there are in a year. | +| First-time entry | | A seller that began selling online in the year being measured. | + +Three rules decide which figures can be set against each other. First, transaction value and revenue are paired only when they refer to the same period and the same business. A take rate built from figures of different scope would measure the mismatch, not the platform. + +Second, each platform is compared only with itself from one year to the next. Companies define transaction value in their own ways, and a disclosed measure tells less when the way it is computed changes (Givoly, Li, Lourie and Nekrasov 2019). Within one company the definition cancels out of the yearly change, as long as the company keeps it. + +Third, companies that mainly sell goods themselves are kept out. Such a company reports the whole sale as revenue, while a platform that only arranges the sale reports its fee (International Accounting Standards Board 2014). The two business models differ in kind (Hagiu and Wright 2015), and their take rates cannot be compared. + +Every figure also keeps a label for where it comes from. Figures that companies report are kept apart from figures that the thesis constructs, and both are kept apart from national statistics and payment figures. Chapter 4 applies these rules platform by platform. + +## 3.2 Three measures + +The invisible wedge measures, in currency, the part of transaction value that lies outside a platform's revenue: + +``` +W_it = V_it − R_it (1) +``` + +where V_it is the transaction value of platform i in year t and R_it is the revenue it recognizes for the same year and business. The wedge is not an accounting error, because both amounts may appear in the same company report. It is a measure of transaction value, not of income or value added, so it does not measure missing output. + +To compare platforms of different size, the thesis also expresses the wedge relative to revenue, as the ecosystem ratio: + +``` +E_it = W_it / R_it = 1 / m_it − 1 (2) +``` + +where m_it = R_it / V_it is the take rate, the share of each sale that the platform keeps. The ecosystem ratio is the author's own measure, and it restates the take rate. An ecosystem ratio of 9 means that 9 units of transaction value lie outside revenue for every unit recorded as revenue. Equivalently, the platform keeps 10 percent of each sale. + +A single ratio does not show whether the relationship holds over time, so the thesis defines growth divergence: + +``` +D_it = g(R_it) − g(V_it) (3) +``` + +where g(X) is the growth of X from one year to the next, as a share of the first year. A positive D means that revenue grew faster than transaction value, so the take rate rose. A negative D means the reverse. If revenue were a stable proxy for transaction value, D would stay close to zero. Revenue equals the take rate times transaction value, so D can also be written with the take rate alone: + +``` +D_it = (1 + g(V_it)) × (m_it − m_i,t−1) / m_i,t−1 (4) +``` + +where m_i,t−1 is the take rate in the earlier year and m_it the take rate in the later year. If the take rate is steady, D is zero and revenue grows exactly as fast as transaction value. + +Tokopedia between 2022 and 2023 illustrates the three measures. Its take rate rose from about 1.5 to about 2.5 percent, so it kept about two thirds more of each sale. Its transaction value fell by about 9 percent and its revenue rose by about 53 percent, so D was 0.91 × 0.68 = 0.62, or 62 percentage points. Its wedge stayed at about 98 percent of transaction value and fell with it, by about 10 percent. + +The three measures do different jobs. The wedge gives the size of what revenue leaves out, and at a marketplace that keeps a few percent of each sale it is almost all of transaction value. The ecosystem ratio expresses that size relative to revenue. Growth divergence carries the central test: whether revenue and transaction value keep the same relationship from year to year. + +The ecosystem ratio also governs how far a change in the take rate moves revenue growth. Equation (4) shows that D is the change in the take rate divided by its starting level. + +A platform that keeps 2 percent of each sale has an ecosystem ratio of 49. If its transaction value is flat, a rise in its take rate to 3 percent raises its revenue by 50 percent. A platform that keeps 17 percent has an ecosystem ratio near 5, and a rise to 18 percent raises its revenue by about 6 percent. So a large ecosystem ratio means that ordinary changes in fees or discounts can produce large swings in reported revenue growth. + +## 3.3 Tracing a movement + +A large divergence calls for an explanation, and the same definitions supply it. Measured in logarithms, the growth of revenue splits exactly into the growth of transaction value and the change in the take rate: + +``` +Δln R_it = Δln V_it + Δln m_it (5) +``` + +where Δln X is the change in the natural logarithm of X from one year to the next. The first term is the volume effect, more or less commerce. The second is the take-rate effect, a larger or smaller share of each sale. A log change treats a rise and a fall by the same factor as equal in size, so the two effects can be compared directly. The take rate itself has two parts: + +``` +m_it = f_it − s_it (6) +``` + +where f_it is the fee rate and s_it the incentive rate, both measured against transaction value. A take rate can rise because a platform charges more or because it spends less on discounts and vouchers. Accounting rules normally deduct such incentives from revenue (International Accounting Standards Board 2014). So a cut in incentives raises revenue even when no fee changes. + +Equations (5) and (6) together divide a movement in revenue into more commerce, higher fees and lower incentives. A change in the mix of a company's businesses is a further source, read from its reports where they describe one. The split shows how a movement happened. It does not show why the company chose it, and company explanations are read as reports. + +Equation (6) also shows where a rule acts. A cap on commissions acts on the fee rate, and a limit on discounts acts on the incentive rate. Chapter 6 sets Indonesia's recent rules against these parts of the take rate. + +## 3.4 Hypotheses + +Two hypotheses follow from the research question. H1 asks whether revenue follows the commerce carried within a platform. H2 asks where national online growth comes from, which decides whether the commerce on platforms can stand for national online commerce. Together they test the two links in Figure 3.1. Table 3.2 pairs each hypothesis, and the entry test of Section 3.5, with its main test and with the result that would count against it. + +**Table 3.2. Hypotheses, tests and what would count against them** + +| Test | What it asks | Main test | What would count against it | +|---|---|---|---| +| H1 | Do revenue and transaction value grow together within a platform? | Size of yearly changes in the take rate, Indonesian and regional platforms against listed platforms abroad | Changes in Indonesia far larger than abroad | +| H2 | Does national e-commerce value grow because of more businesses or because of higher sales per business? | BPS's national e-commerce value split into the number of businesses and sales per business | Higher sales per business accounting for most of the growth | +| Entry test | Does a rise in the count of online sellers come from first-time entry? | Share of 2023 sellers in BPS's survey files who began selling online in 2023 | Fewer new sellers than the rise requires | + +The share of transaction value that a platform keeps is not fixed. A platform sets its charges on both sides of the market together, and both the level and the split are choices (Rochet and Tirole 2003). Platforms often hold down the price to one side to draw it in. They can raise that price once both sides have joined (Evans and Schmalensee 2016). + +Accounting rules move the share in two ways. The discounts a platform pays for come out of its revenue. A platform that sells goods itself, a different business model (Hagiu and Wright 2015), records more of each sale than one that only arranges it (International Accounting Standards Board 2014). Both can change as a platform changes its business model or scope. + +Revenue follows transaction value only while the take rate holds steady, that is, while D stays near zero. In that case revenue would be a workable proxy for the commerce a platform carries, and the ecosystem ratio would not change. The first hypothesis states this case. + +> **H1.** Transaction value and revenue recognized by the same platform grow in proportion over time. + +Growth rates never match exactly, so H1 is tested by comparison. The test compares the size of yearly changes in the take rate, measured in logs as in equation (5), between two groups. One group is the Indonesian and regional platforms. The other is listed platforms abroad, where revenue and transaction value broadly move together. Changes far larger in Indonesia would count against H1. Chapter 4 describes the two groups and the test. + +National e-commerce value can grow in two ways. Existing businesses can sell more, which is the intensive margin. More businesses can sell online, which is the extensive margin. Trade economists split a country's exports in the same way, into more of each good and more kinds of goods (Hummels and Klenow 2005). + +The distinction matters for measurement. Growth driven by new sellers can take place in channels that platform accounts do not cover, such as social media and instant messaging, which BPS reports are widely used (BPS 2025a). Platform figures can then miss that growth even when every platform reports accurately. + +BPS publishes both the number of businesses selling online, N, and their total online sales value, Y. Since Y equals N times the sales per business, its growth in logarithms splits exactly in two: + +``` +Δln Y_t = Δln N_t + Δln (Y_t / N_t) (7) +``` + +where the first term is the extensive margin and the second the intensive margin. H2 holds if the first term is more than half of the total. The second hypothesis states the case in which more businesses drive growth. + +> **H2.** The growth of national e-commerce value comes mainly from more businesses selling online, and less from higher sales per business. + +The H2 split is arithmetic. It does not assign new businesses to channels, and it does not show whether the businesses in the count are new. + +## 3.5 The entry test + +H2 rests on BPS's count of businesses selling online, and a rise in that count can mean two different things. Either more businesses started selling, or the survey reached businesses that were already selling. A change in how a survey finds its respondents can create a break in a series that looks like real change (van den Brakel, Smith and Compton 2008). + +The published figures cannot tell the two apart. The survey can, because it asks every seller in which year it began selling online. The thesis checks what the count contains with a test written down before the survey files were opened. + +> **Entry test.** In BPS's survey of online sellers for 2023, at least 21.5 percent of online sellers must report that they began selling online in 2023. + +The 21.5 percent threshold comes from arithmetic. The published count of online sellers rose by 27.4 percent in 2023. Suppose every extra seller was new and no seller stopped. Then for every 100 sellers a year earlier there were 127.4, and 27.4 of those were new. New sellers would make up 21.5 percent of all sellers, which is 27.4 divided by 127.4. + +Sellers who stopped would have to be replaced as well, which only raises the share needed. The threshold is the smallest share of new sellers that can produce the rise. The test does not test H2 itself. It tests what "more businesses" means. + +Sellers' memory works in favour of passing the test. People tend to report past events as more recent than they were (Neter and Waksberg 1964). A seller who began in 2022 but reports 2023 raises the share of new sellers. Recall error of this kind makes the threshold easier to reach, so a share below it is a cautious finding. + +If the share of new sellers falls short, part of the rise is something other than first-time entry. Two readings are open. The survey may have reached more of the sellers who were already there. Or sellers may have come back after a pause, since a returning seller keeps an old start year. Small firms that stop are more often young firms with young owners (McKenzie and Paffhausen 2019). So the age of the sellers behind the rise bears on the second reading. Chapter 5 weighs the two readings with that evidence. + +## 3.6 Further analyses + +Four further analyses complete the framework without forming hypotheses of their own. Each answers a part of the research question that H1, H2 and the entry test leave open. Table 3.3 lists them. + +**Table 3.3. Further analyses** + +| Analysis | What it asks | Main evidence | +|---|---|---| +| Tracing large movements | What moved revenue when it parted from transaction value? | Equations (5) and (6), applied to company reports | +| Market values | Do markets value growth from more commerce differently from growth in the take rate? | Year-end market values of listed platforms | +| Channels and reach | How much online selling passes through marketplaces, and how far can rules built on platform records reach? | BPS's channel shares and survey files | +| Comparing national figures | Do national figures tell the same story of growth? | BPS, Bank Indonesia, industry estimates and household spending | + +Tracing large movements applies the split of Section 3.3. Wherever H1 finds a large change in the take rate, the thesis divides the change in revenue into its parts. It also reads the company's reports for changes in the mix of businesses and in reporting scope. + +The market-value analysis asks whether investors, who price revenue (Section 2.1), treat its two sources alike. Equation (5) shows that revenue can grow in two ways, through more commerce or through a larger take rate. Early studies of internet companies found that markets priced measures of web activity beyond reported earnings (Trueman, Wong and Zhang 2000; Rajgopal, Venkatachalam and Kotha 2003). + +The two sources of growth need not be worth the same. Growth from the take rate has a limit, because a platform cannot raise its share of each sale without end. A higher price to one side of a platform can also reduce how much that side uses it (Rochet and Tirole 2003). + +The thesis relates the yearly change in a platform's market value, MV, to the two sources of revenue growth: + +``` +Δln MV_it = β_V Δln V_it + β_m Δln m_it + year effect + error (8) +``` + +where β_V shows how market value moves with transaction value, and β_m how it moves with the take rate. Two predictions were written down before the market values were collected. The first is that market value moves with transaction value, so β_V is positive. The second is that markets value growth from more commerce more than growth from a higher take rate, so β_V is larger than β_m. + +The channel analysis answers the third question of Section 2.4. A rule built on platform records reaches only the sellers whose sales pass through a platform. Research on tax shows that reports from a third party improve compliance where they exist (Kleven et al. 2011). + +The reach of a rule built on platform records depends on how online selling divides between marketplaces and channels that no marketplace records, such as messaging apps and social media. The thesis measures that division by sellers and by sales value. It also counts the marketplace sellers who may be large enough to fall under the marketplace tax. + +Record keeping is the second part of this analysis. The thesis compares how often marketplace sellers and other sellers keep financial statements. The comparison bears on the information that platform records can add. + +The comparison of national figures extends the outer link in Figure 3.1. Platform reports, BPS's survey, Bank Indonesia's statistics and industry estimates count different things. Their values change with the definition used (UNCTAD 2024). + +If two national figures move in opposite directions in one year, the conclusion about growth depends on the figure chosen. The thesis compares growth rates, each in its own unit, and levels only after reconciling what each source covers. Household spending is a further point of comparison. diff --git a/papers/thesis_2026-10/ch4_data_methods.md b/papers/thesis_2026-10/ch4_data_methods.md new file mode 100644 index 0000000..fb231ca --- /dev/null +++ b/papers/thesis_2026-10/ch4_data_methods.md @@ -0,0 +1,139 @@ +# Chapter 4. Data and Methods + +The figures this thesis compares come from sources that were never designed to be read together. Platforms publish in their own formats, and the answers of individual sellers are released only under a data-use agreement. The methods in this chapter are built around these conditions. + +On the platform side, the thesis reads revenue and transaction value from annual reports, audited statements and earnings releases. The extraction tables keep the source file and the quoted passage beside every value. The Indonesian platforms are three online marketplaces, Tokopedia, Bukalapak and Blibli, and GoTo, the group that runs the Gojek app for rides and deliveries. GoTo also controlled Tokopedia until 2024 and still holds a stake in it. Two regional companies, both based in Singapore, complete the group: Grab, which runs rides and deliveries across Southeast Asia, and Sea, whose marketplace is Shopee. Where a platform does not report Indonesia alone, as with Shopee and Grab, the thesis states what it derives or borrows and labels it. + +The national side rests on BPS's survey of online sellers. BPS publishes a few totals, and the thesis obtained the answers behind them for three survey rounds, which show when each seller began selling online and how much it sold. The central bank's own figure has no published method, so it serves only as a point of comparison. + +Setting the two sides against each other for one country is what makes the tests possible. Indonesia allows it. Five platform series report revenue and transaction value in the same year, and three of them report what they pay out in incentives. The survey asks every online seller when it began. The tests are built to ask only what these sources can answer, and each limit is stated where it applies. + +The chapter describes the evidence on each side and the tests that use it. It starts with the platforms: the data, the wedge in Indonesia, the test of H1, the tracing of large moves and market values. It then turns to the survey: H2, the entry test, and the channels and records of sellers. It ends with the comparison of national figures and with how every number is labelled and traced. + +## 4.1 The platform data + +The basic observation is one platform in one year, which this thesis calls a platform-year. Revenue comes from the company's financial statements and transaction value from its annual report or earnings release. Both refer to the same year and, as closely as the filings allow, the same business. + +Companies define transaction value in different ways, and some include shipping charges, digital goods or travel. The thesis keeps each company's own definition and records it. Under the second rule of Section 3.1, the definition cancels out of a take rate's yearly changes. This holds as long as the company keeps both the definition and the mix of its businesses. + +Two groups of platforms are compared. The Indonesian and regional group has five platforms, listed in Table 4.1. Three are Indonesian marketplaces, and two are the on-demand businesses of Grab and GoTo, which cover rides and deliveries. The foreign group has 27 listed platforms in marketplaces, travel and ticketing, freelance work, resale and food delivery. Eight were chosen first, and 19 were added to widen the comparison. Ten more were examined and left out, six because they sell mostly their own goods and four because they do not disclose transaction value. + +**Table 4.1. The Indonesian and regional group of platforms** + +| Platform series | What it covers | Years | Yearly pairs | How it is labelled | +|---|---|---|---|---| +| Tokopedia | The e-commerce segment of GoTo, an Indonesian marketplace | 2022–2023 | 1 | Direct segment disclosure | +| Blibli | Blibli's third-party retail segment, which includes online travel | 2021–2025 | 4 | Direct segment disclosure | +| Bukalapak | The whole Bukalapak group, mainly in Indonesia, including Mitra, its online-to-offline business | 2020–2023 | 3 | Direct company disclosure | +| Grab on-demand | Grab's rides and deliveries across Southeast Asia, not Indonesia alone | 2020–2024 | 4 | Direct segment disclosure | +| GoTo on-demand | GoTo's rides and deliveries segment, in Indonesia and Vietnam to 2024 | 2022–2025 | 3 | Direct segment disclosure; the 2022 transaction value is an estimate | + +Table 4.1 states the scope of each series because the series differ, and only Tokopedia is an Indonesian segment alone. The thesis never adds the series together, and each platform enters the tests on its own. Sea, whose Shopee business has Indonesia as a main market, sits in the foreign group. + +Platform-years are compared in pairs, and three rules keep the pairs comparable. Two consecutive platform-years form a pair, and nothing else does. A missing year stays missing and is never filled in. Annual and quarterly pairs are never pooled. A take rate is a ratio of two amounts in one currency, so the tests need no exchange rate. Amounts in running text are in US dollars at the World Bank's average rate for the year (World Bank 2026). + +The third rule of Section 3.1, which keeps out companies that mainly sell their own goods, is applied in two ways. For the 19 platforms added, the thesis drops any year in which own goods exceed half of revenue. For the eight chosen first, it keeps a pair only when the company's reporting scope did not change between the two years. + +Two platforms sit close to that line. Zalando is one of the eight and sells part of what it carries, so Appendix B repeats the test without it. One added pair, Ozon from 2021 to 2022, ends in a kept year but starts in a dropped one. Appendix B repeats the test without Ozon as well. The result is 15 pairs for the Indonesian and regional group and 158 for the foreign group. + +## 4.2 Measuring the wedge in Indonesia + +The first objective, the size of the invisible wedge, needs amounts for Indonesia alone. The thesis sums the transaction value and revenue of three major platforms in Indonesia in 2023: Tokopedia, Shopee and Grab. The wedge and the ecosystem ratio are then computed from the sums. + +Each platform reaches the sum by its own route. Tokopedia reports both amounts for its segment. Shopee's transaction value is an outside estimate by Momentum Works, a research firm that tracks online commerce in Southeast Asia. Its revenue assumes a take rate of 10 percent, which the thesis varies from 9 to 11 percent. Grab reports its Indonesian revenue, and the thesis derives its Indonesian transaction value from Grab's company-wide ratio of the two amounts. The thesis drops each platform in turn, and it repeats the exercise for every platform-year between 2020 and 2025 that the filings allow. + +## 4.3 Testing H1 + +The test of H1 asks how far each platform's take rate moves from one year to the next. For each pair of years it takes the change in the logarithm of the take rate, the take-rate term of equation (5). This change is growth divergence measured in logs, and it does not depend on how fast transaction value grew. Tokopedia's log change in 2023 is 0.52, which is a rise by a factor of 1.68, or 68 percent. + +A move in either direction counts against H1, so the test uses the size of each yearly change in the take rate. Each platform gets one number, the median size of its yearly changes, because the pairs of one platform are not independent of one another. + +The test compares five numbers for the Indonesian and regional group with 27 for the foreign group. It asks whether the first group ranks higher than chance would give, using a one-sided Mann–Whitney test (Mann and Whitney 1947). It repeats the question by shuffling the group labels 9,999 times. The p-value is the chance of a result at least this lopsided if the groups did not differ. + +Appendix B lists the checks on this test. They drop each platform in turn and treat Tokopedia and GoTo as one company. They repeat the test with growth divergence in points as Chapter 3 defines it. They compare the change in the take rate in percentage points, to see whether a thin take rate explains the result. They narrow the test to the three marketplaces, correlate revenue growth with transaction value growth, and repeat the test with quarterly pairs. + +## 4.4 Tracing the moves + +The second objective, to reconcile large movements in revenue with disclosed changes in fees, discounts and scope, is answered in two steps. The first divides each movement into volume and take rate with equation (5). The take-rate share is the size of the take-rate change divided by the sum of the sizes of both changes. The same split is made for the wedge. + +The second step splits the take-rate change with equation (6). Some companies disclose both their gross fees and their incentives. For them, the thesis splits the growth of revenue over a window into three parts: more commerce, a change in the fee rate and a change in the incentive rate. A change in the mix of businesses is recorded from the company's reports. + +The sources differ by company. Tokopedia's audited segment note gives gross revenue, incentives and net revenue, and the thesis checks them against the audited consolidated statements. For Tokopedia the thesis also reports the change line by line in rupiah, the usual way to read accounts, and says where the two views differ. + +Grab is read from its annual filings and GoTo from its audited statements. Blibli does not report incentives, so its split rests on a proxy built from its releases. Seven windows are traced, and Appendix A lists them. The thesis also records where accounting treatment or reporting scope matters, such as the deduction of incentives from revenue and the travel business inside Blibli's retail segment. Company explanations are read as reports and are not tested. + +## 4.5 Market values + +The market-value analysis of Section 3.6 uses year-end market values from the financial data service LSEG, under the university licence. The main sample has 28 listed platforms and 138 platform-years, made up of Bukalapak and 27 platforms abroad, and it uses whole-company figures only. It is not the list of 27 used for H1. It adds Airbnb, Booking, DoorDash, Lyft and Uber, and it lacks Despegar, Mercado Libre, Ozon, Rakuten and Sea. + +Equation (8) is estimated with a fixed effect for each year, which removes what all platforms share in that year. Standard errors are clustered by company, which allows for the repeated observations of one company. Each prediction names a direction, so both tests are one-sided. The test was written down before the market values were pulled. The result is an association, and market value also changes when a company issues shares, so the thesis does not call it a stock return. A second run adds five segment series and Mercado Libre as a robustness check. Grab, GoTo and Sea enter only there, because their reported transaction value covers only part of the business that the market value prices. + +## 4.6 The survey files and H2 + +The national evidence comes from the survey of online sellers that BPS runs. BPS publishes the number of businesses selling online, their total online sales and the shares by channel (BPS 2025a, 2025b). The survey rounds are named for the survey year, but their questions refer to the year before. The thesis names a file by its round and a year by the year of activity, so the file for the 2024 round describes 2023. Each round lists businesses in sampled areas and updates its sampling frame, so a change in who is found can change the count. + +H2 uses the published figures and the split of equation (7). BPS prints two counts of businesses for 2023, so the thesis computes the 2023 to 2024 split with both and reports both. The 2022 to 2023 split uses the count that the survey weights match. + +The published figures cannot tell a new seller from one counted for the first time, but the survey files can. Table 4.2 lists them. Each row is a business that sold online in the activity year. Its survey weight shows how many businesses it represents, so the weights add up to the national count. + +Two checks come before any test. The weighted total of each file must fall within 1 percent of the BPS count. The 2024 file must also reproduce two published shares of sellers who keep complete financial statements, 28.63 and 12.25 percent. All three files pass. The files carry no business identifier across rounds and no area variable below the province. So the rounds compare groups of sellers and not the same sellers over time. + +**Table 4.2. The three BPS survey files** + +| Survey round | Activity year | Rows | Weighted total | BPS count | +|---|---|---|---|---| +| E-Commerce 2021 | 2020 | 11,848 | 2,361,423 | 2,361,423 | +| E-Commerce 2023 | 2022 | 31,724 | 2,995,986 | 2,995,879* | +| E-Commerce 2024 | 2023 | 39,736 | 3,814,799 | 3,816,750 | + +*Implied by the published count for 2023 and its published growth of 27.4 percent. + +## 4.7 The entry test and the extra sellers + +The entry test asks whether at least 21.5 percent of the 2023 sellers in the 2024 file began selling online in 2023, a threshold that Chapter 3 derives. BPS also prints a second count for 2023, but the survey weights match the first, so the threshold stays at 21.5 percent. The test was written down and dated in the project's code archive, called the repository, before the files were opened. + +A failed entry test raises a second question: how much of the rise does first-time entry not explain? The sellers behind that part are called the extra sellers. A cohort is the group of sellers that began selling online by a given year. If two surveys count the same population, a cohort can only shrink between them, because sellers stop and no new seller can join it. + +The cohort check compares the 2024 file's count of sellers who began before 2023 with the whole 2022 count. It does so twice. The first time it assumes that no seller stopped. The second time it assumes that sellers stopped at about 5 percent a year, the rate at which one cohort shrank between the 2020 and 2022 surveys. + +Further checks ask where the extra sellers sit. They look across provinces and channels, using rank correlations, and at sellers' own reports of their sales. They use the sampling error that BPS publishes. They profile who the extra sellers are, with each group's count in 2023 compared with what its own yearly change in 2020 to 2022 would give. They look at part-year selling and test a scenario of returners. + +The cohort check and the further checks were designed after the entry test failed, so Chapter 5 labels them exploratory. The exception is the median change in sellers' sales, which was registered. The files cannot rule out returners, because a seller who stopped and returned keeps an old start year. + +## 4.8 Channels, records and reach + +Three analyses beyond the entry test also use the survey files. One asks which channels sellers use and how sales divide between them. The published tables split the 2023 to 2024 rise in value between marketplaces and other channels, and BPS counts food and ride apps in its marketplace category. + +Value by channel rests on revenue brackets, which only the 2021 and 2023 files give. The thesis assigns each bracket a low, a middle and a high amount. The amounts were fixed before the file was opened. A validation rule applies: the middle scenario's 2022 total must lie between half and double BPS's published total of about US$53 billion. + +The 2023 questionnaire asks for monthly revenue but the bracket limits are annual size classes, so both readings are computed and the one that passes is kept. If the lower brackets alone already exceed the total, value shares are reported as ranges. + +Whether a seller keeps complete financial statements is the subject of a second analysis. It bears on what platform records can add, because a seller with no statements leaves fewer formal records. World Bank Enterprise Surveys add a check on records. They cover unregistered firms in six Indonesian cities in 2023, and formal firms with five or more employees in seven countries. + +A third analysis builds on the first two and is developed in Chapter 6. It counts the marketplace sellers whose revenue is above a bracket line and expresses the count as a share of all online sellers. At the 2022 rate, BPS's brackets divide at about US$20,000 a year. The marketplace tax exempts individual sellers up to about US$34,000. So the count above BPS's line is an upper limit on the sellers the tax can reach. Chapter 6 also sets each recent rule against the part of the figures it acts on. The thesis reads each legal text where it is published. It labels each rule by the status of its text: primary text read, announced, or reported in the news only. + +## 4.9 Comparing national figures + +The third objective, to compare the platform evidence with national and payment figures, rests on the sources that Table 4.3 lists, with what each one counts. Beyond the rules of Section 3.1, the comparison is closest where the coverage of two sources overlaps. Payment figures serve as context, because they show how people pay and not what they buy online. + +The 2023 file, which describes 2022, lets the thesis separate the part of BPS's marketplace sales that went to final consumers. It compares that part with Bank Indonesia's figure for the same year, in a test written down before the file was opened. The thesis sets the 2023 growth of six measures of buying through platforms beside household spending, each in its own unit. It does the same with quarterly platform growth. + +**Table 4.3. Sources compared in Objective 3** + +| Source | What it covers | How it counts | +|---|---|---| +| Platform reports | Everything bought through one platform | The platform's own records | +| BPS online-seller survey | Every business that sold online, including chat and social media | Listing in sampled areas, scaled up by survey weights | +| Bank Indonesia e-commerce statistics | Transactions of large e-commerce firms (Bank Indonesia 2025) | Data sent to the central bank; the method is not public | +| Industry estimates | Shopping marketplaces in Indonesia or Southeast Asia, from the research firm Momentum Works (2024), as quoted by Setyowati (2024), and the e-Conomy report (Google, Temasek and Bain 2024) | Outside estimates | +| Household spending | All household purchases | BPS national accounts | +| Payment figures | Payments made through systems such as QRIS, the national QR-code system | Payment records, in person and online | +| Adjacent figures | Parcels shipped in Southeast Asia, the value-added tax collected on foreign digital services (Directorate General of Taxes 2025), and the number of merchants a company reports | Used as context only, since none measures Indonesian e-commerce | + +## 4.10 Labels, traceability and licensed data + +Every number in the thesis follows one path: source document, extract, analytic row, result. Each extract keeps the company's currency and units, and each value carries a label: direct company disclosure, direct segment disclosure, external estimate, derived value, conversion or scenario. The labels are never mixed in one sample. A restatement or a conflict between two reports is reported and never overwritten. A single script rebuilds every table that depends only on committed files, and 272 automatic checks confirm that each recorded result still matches its table. + +The thesis also separates tests by when they were written. A registered test was written down, dated and stored in the repository before the data were opened. This applies to the entry test, the scenario rules for value, the comparison with Bank Indonesia and the market-value test, among others. Appendix C lists them with their dates. An exploratory test was designed after some results were seen. The test with the larger groups in Section 4.3 belongs here, because it was set up after a first test on the three marketplaces. A descriptive result reports numbers without a test. + +Three sources are licensed and are not in the repository: the BPS survey files, the LSEG market values and the World Bank survey files. The repository records the size and checksum of each BPS and LSEG file and of the World Bank cross-country query, so anyone who obtains the same files can check them. It holds the code that reads them and the national totals that the code produces, and no business-level record. BPS is cited as the source of every survey figure, in line with the data-use agreement. Chapter 7 states what these data cannot show. diff --git a/papers/thesis_2026-10/ch5_results.md b/papers/thesis_2026-10/ch5_results.md index 7233383..0f2040d 100644 --- a/papers/thesis_2026-10/ch5_results.md +++ b/papers/thesis_2026-10/ch5_results.md @@ -1,82 +1,237 @@ # Chapter 5. Results -This chapter follows the numbers from the platforms outward. It starts with what a platform keeps from each sale, then asks whether its revenue followed its sales (H1). It then turns to 2023, the year the headline figures described a boom, and to what sellers themselves reported. The national count of sellers (H2) follows in section 5.6. +The results follow the two links of Figure 3.1. On the inner link, revenue showed a small and unsteady part of the commerce that platforms carried. Platforms changed how much of each sale they kept, and revenue moved with those choices (Sections 5.1 to 5.4). On the outer link, the national count of online sellers rose by 27 percent in 2023. About half of that rise is not explained by sellers entering for the first time, and whether H2 holds depends on how that half is read (Sections 5.5 to 5.10). + +Every result in this chapter carries one of the three labels of Section 4.10: registered, exploratory or descriptive. A result given without a label is descriptive. Table 5.1 gives each main result with its label, the checks behind it and the section that reports it. + +**Table 5.1. Main results and the evidence behind them** + +| Result | Label | Main checks | Section | +|---|---|---|---| +| *Inner link: revenue and platform commerce* | | | | +| Revenue was about 7 percent of what three major platforms carried in 2023; the wedge was about US$40 billion | Descriptive | Shopee's take rate varied; each platform dropped in turn | 5.1 | +| Take rates moved about three times as much in Indonesia and the region as abroad, so H1 is rejected | Exploratory; the narrower test, set before the wider one, points the same way | Each platform dropped in turn; Ozon and Zalando dropped; change measured in points; quarterly pairs do not reject | 5.2 | +| The thin take rate of Indonesian and regional platforms explains much of that difference | Exploratory | Changes in percentage points compared across groups | 5.2 | +| Pricing, not more commerce, moved revenue | Descriptive | All 15 pairs against 158 abroad; seven traced windows; two views of Tokopedia | 5.3 | +| Market value moves with transaction value | Registered; holds | Extreme values capped (winsorised); a second run | 5.4 | +| Markets value growth from a higher take rate less | Registered; does not hold in the main sample | Holds only in a second run that mixes segment figures | 5.4 | +| *Outer link: platform commerce and all online commerce* | | | | +| Buying through platforms, and sellers' own sales, did not boom in 2023 | Descriptive; registered for sellers' sales, holds | Six platform measures against household spending | 5.5 | +| More businesses account for most national growth, so H2 holds | Descriptive | Both BPS counts for 2023; two periods | 5.6 | +| Too few 2023 sellers were new to produce the rise in the count (13.45 against 21.5 percent) | Registered; the entry test fails | Recall error would favour passing | 5.6 | +| About half of the rise is not explained by first-time entry | Exploratory | With and without exits; cohorts; provinces; sampling error | 5.7 | +| Read as wider coverage, the extra count leaves H2 near its threshold (about 42 to 54 percent) | Exploratory | Count and sales removed together; two exit rates | 5.7 | +| The extra count leaned toward long-running businesses; returning sellers cannot be ruled out | Exploratory | Part-year selling | 5.8 | +| Most online selling happens outside marketplaces, and records are thinner there | Registered and descriptive | Four survey years; World Bank surveys | 5.9 | +| National figures agree only where they cover the same sellers | Registered, with a mixed outcome, and descriptive | Three value scenarios | 5.10 | ## 5.1 How much revenue leaves out -A platform keeps only a small part of what passes through it. In 2023, Tokopedia, Shopee and Grab together carried US$43.23 billion of transaction value. Their combined revenue was US$3.16 billion. The invisible wedge, transaction value minus revenue, was therefore US$40.07 billion. Revenue showed about 7 rupiah of every 100 that buyers spent. +Revenue showed only a small part of what Indonesia's major platforms carried. In 2023, Tokopedia, Shopee and Grab together carried about US$43 billion of transaction value and kept about US$3 billion as revenue. Revenue was about 7 percent of that commerce, and the invisible wedge was about US$40 billion. The ecosystem ratio was about 13, so almost 13 dollars of commerce passed outside revenue for every dollar the platforms kept. Table 5.2 shows each platform. -The size of the wedge does not depend on any one assumption. Shopee's Indonesian revenue rests on an assumed take rate of 9 to 11 percent. Across that range, the wedge moves only between US$39.86 billion and US$40.29 billion. Dropping any one platform still leaves a wedge of at least US$20.70 billion. +**Table 5.2. The invisible wedge of three platforms, 2023 (US$ billion)** -The wedge is also stable as a share of what the platforms carry. For the three platforms that report both numbers themselves, it was 97 to 99 percent of transaction value in every year. It therefore moves almost one for one with transaction value. Tokopedia in 2023 shows this clearly: transaction value fell 9 percent and the wedge fell 10 percent. +| Platform | Transaction value | Revenue | Wedge | Ecosystem ratio | How it is labelled | +|---|---|---|---|---|---| +| Tokopedia | 16.33 | 0.41 | 15.93 | 39.3 | Both direct segment disclosure | +| Shopee, Indonesia | 21.52 | 2.15 | 19.37 | 9.0 | Transaction value is an outside estimate (Momentum Works); revenue is derived at an assumed take rate of 10 percent | +| Grab, Indonesia | 5.38 | 0.60 | 4.78 | 7.9 | Revenue is a direct country disclosure; transaction value is derived from Grab's company-wide ratio | +| Total | 43.23 | 3.16 | 40.07 | 12.7 | Sum of three platforms, not a national total | -Revenue is the small part left over, and it can move on its own. In the same year, Tokopedia's revenue rose 53 percent. +*Note:* Rounded amounts may not add. With Shopee's take rate at 9 or 11 percent, the total wedge is US$40.3 or US$39.9 billion; at Sea's company-wide rate of 11.5 percent, it is US$39.8 billion. Source: company filings; Momentum Works (2024), as quoted by Setyowati (2024); author's calculations. -The platforms themselves carry only part of online selling. In BPS's own survey of online businesses, sales through marketplaces and ride-hailing apps were about 16 to 26 percent of online sales value. The rest moved through chat apps, social media and sellers' own websites. A measure built on the platforms therefore sees a thin slice of a minority of online selling. +The ratios differ because the amounts were obtained in different ways. Tokopedia reports both amounts, and its ratio is about 39. Shopee's ratio of 9 follows from the assumed take rate, and Grab's ratio of about 8 is its company-wide ratio. Neither is separate evidence for Indonesia. Blibli and Bukalapak also report both amounts, with ratios of about 43 and 36, close to Tokopedia's. They are left out of the sum because Blibli's figure includes online travel and Bukalapak reports for its whole group. -So the first half of the research question has a plain answer. Revenue shows about 7 percent of what passes through the platforms, and the platforms carry about a fifth of online sales. The next question is whether that thin slice still tracks growth. +The size of the wedge does not depend on the assumed rates. A take rate for Shopee between 9 and 11 percent keeps the total near US$40 billion. Dropping any one platform still leaves more than US$20 billion. For platforms that report both amounts, the wedge was 97 to 99 percent of transaction value in every year. It moved almost one for one with the commerce, and Appendix A gives the wedge by platform and year. -## 5.2 Did revenue grow with transaction value? (H1) +Platform commerce is itself only part of online commerce. In BPS's survey, marketplaces carry about a fifth of online sales value (Section 5.9), so platform revenue shows a smaller part still. A small share can still be a fair guide to the commerce, as long as it keeps its size from year to year. -H1 states that transaction value and revenue grow in proportion over time. If it holds, revenue is a fair guide to the commerce a platform carries. The test compares how much the take rate moves in Indonesia with how much it moves at platforms abroad. The take rate is revenue divided by transaction value. +## 5.2 Testing H1 -H1 is rejected. Five Indonesian and regional platforms were compared with 27 listed platforms abroad, one result per firm. In a typical year, the Indonesian take rate moved by about 20 percent of its own level. Abroad, it moved by about 6 percent. The difference is significant at the firm level (Mann-Whitney, p = 0.004). +Revenue did not keep pace with transaction value at the Indonesian and regional platforms, and the thesis rejects H1. At the typical platform in this group, the take rate moved by about 20 percent a year, up or down. At the typical foreign platform it moved by about 6 percent, so the moves were about three times as large. Table 5.3 gives the test and its checks. -No single platform drives the result. Dropping each Indonesian platform in turn leaves p at or below 0.016. The proposal's own measure, growth divergence, gives the same answer: 27.0 percentage points in Indonesia against 6.5 abroad (p = 0.005). Treating GoTo's two segments as one company leaves p = 0.003. +**Table 5.3. The test of H1** -The size of the gap shows how far revenue and sales parted. For the three platforms in the main sample, the typical absolute growth divergence was 38.9 percentage points. Abroad, it was 7.1 points. Revenue stayed close to transaction value in only 12.5 percent of Indonesian comparisons, against 69 percent abroad. +| Test | Method | Indonesian and regional group | Foreign group | p-value | +|---|---|---|---|---| +| Yearly change in log take rate, 5 platforms against 27 | Mann–Whitney | 0.20 | 0.06 | 0.004 | +| The same | Shuffle | 0.20 | 0.06 | 0.022 | +| Growth divergence in points, 5 platforms against 26 | Mann–Whitney | 27.0 | 6.5 | 0.005 | +| Drop each platform in turn, or treat Tokopedia and GoTo as one company | Mann–Whitney | 0.17 to 0.23 | 0.06 | 0.003 to 0.016 | +| Quarterly pairs, 3 platforms against 8 | Mann–Whitney | 0.071 | 0.049 | 0.19 | -The main sample on its own, with eight comparisons, gives p = 0.06 on a firm-level test. That test uses growth divergence against the proposal's eight foreign platforms; it is a different test, not a stricter one. One registered check did not hold. With quarterly data, the difference shrinks and is not significant (p = 0.19). The quarterly series mostly cover 2023 to 2026, after the largest price changes had happened. +*Note:* Group values are medians of each platform's median yearly change. Appendix B lists every check, including the tests without Ozon and Zalando. -So revenue did not grow in proportion to transaction value in Indonesia. Most of the separation happened in 2022 to 2024. The next section asks why. +The main test compares the five platforms with 27 abroad, one median per platform. Both methods of Section 4.3 reject equal moves, with p = 0.004 and p = 0.022. No single platform carries the result. Dropping any one, or treating Tokopedia and GoTo as one company, keeps the Mann–Whitney p-value at 0.016 or below. The main test was set up after a first, narrower test, so it is exploratory. -## 5.3 Why revenue moved: similar price changes, a much thinner slice +A narrower test, set before the wider one, compares the three Indonesian marketplaces with the eight foreign platforms chosen first. It points the same way. With one median per platform, the Mann–Whitney test rejects equal moves (p = 0.024). A shuffle of platform labels, using one median per platform, comes close without passing (p = 0.06), since three platforms give a test little power. In that sample, revenue growth and transaction value growth had a correlation of 0.10 in Indonesia and 0.98 abroad, excluding Sea. -The obvious explanation is that Indonesian platforms changed their prices far more than platforms abroad. The data do not show that. Measured in rupiah per 100 of transaction value, the typical yearly change in what a platform keeps was about one rupiah. That held in Indonesia and abroad alike. +The direction of the moves is less telling than their size. Revenue grew faster than transaction value in 13 of the 15 Indonesian and regional pairs. In two of them, revenue and transaction value moved in opposite directions. Abroad, take rates also rose in about two thirds of pairs, but by much less. What sets the Indonesian and regional platforms apart is the size of the moves, which is what the test measures. -The difference lies in the slice itself. The Indonesian marketplaces keep under 2 rupiah of every 100 sold. A typical platform abroad keeps about 17. A change of one rupiah is small against a slice of 17 and large against a slice of 2. The same price move therefore shifts Indonesian revenue growth about three times as much (Figure 5.1). +Quarterly pairs do not reject H1 (p = 0.19), contrary to a prediction registered before the test was run. The quarterly Indonesian pairs mostly cover 2023 to 2026, after most of the repricing, while the large moves came between 2022 and 2024. The rejection of H1 describes that period. -This gives the wedge its job in the thesis. Growth divergence follows directly from the change in the take rate: +The large moves were not simply a mark of young platforms, an exploratory check. The Indonesian and regional platforms made their largest moves one or two years after listing on the stock market. Foreign platforms, by contrast, did not move their take rates more in their first three years after listing than later (median 0.07 against 0.08, p = 0.24). This comparison uses the 15 foreign platforms with years in both periods. Nor did foreign platforms in lower-income markets move more in 2022 to 2024: one of seven did, against all four Indonesian and regional platforms with enough years. The repricing was an Indonesian and regional episode of those years. -> D = (1 + gV) × (m1 − m0) / m0 +The thin take rate of Indonesian and regional platforms explains much of the threefold difference, an exploratory finding. Measured in percentage points, both groups changed their take rates by about 1 point a year. The difference between them is nowhere near large enough to count (p = 0.65). -Here m0 and m1 are the take rates in the two years and gV is growth in transaction value. The term 1 / m0 equals 1 + E0, one plus the ecosystem ratio. The larger the wedge is next to revenue, the more an ordinary price change moves revenue growth. A thin slice helps explain why Indonesian revenue swung so much. This comparison was made after the H1 results and is exploratory. +The two groups differ in the take rate they start from. The typical Indonesian and regional platform keeps about 1.6 percent of each sale, against about 17 percent abroad. The thinnest take rates belong to the three marketplaces, while Grab and GoTo keep shares closer to those abroad. As Section 3.2 showed, the same 1-point rise then has very different effects. It lifts revenue by more than half at the thinner take rate and by about 6 percent at the thicker one. Figure 5.1 shows the three measures side by side. The test shows that take rates moved, but not what the platforms changed to move them. -The companies' own accounts show what the price changes were. Between 2022 and 2024, the platforms paid for fewer discounts and charged higher fees. Two forces move revenue: how much people buy, and how much of each sale the platform keeps. In a typical Indonesian year, the second was about two thirds of the movement. Abroad, it was about a quarter. +![Figure 5.1](figures/fig5_1_thin_slice.png) -The cases differ in their mix: +**Figure 5.1. The thin take rate.** Each dot is one platform's median over its years. Left: the share of each sale that the platform keeps, on a log scale. Middle: the yearly change in that share, in percentage points. Right: the same change relative to the share itself, as a log change in percent. Black bars mark group medians. The foreign group has 26 platforms here and 27 in the main test, because one platform has no pair of usable years for the change in points. Source: company filings; author's calculations. Exploratory. -- **Tokopedia, 2022 to 2023.** Transaction value fell 8.9 percent and net revenue rose 53.2 percent. Customer incentives fell from Rp4.11 trillion to Rp2.81 trillion. Lower incentives account for 60.6 percent of the rise in net revenue (Figure 5.2). -- **Grab, 2022 to 2023.** Its take rate rose 3.9 points; repricing accounts for 3.5 of them. -- **GoTo's ride and delivery business, 2023 to 2025.** Revenue rose 114 percent; fewer discounts account for 67 percent of that rise. -- **Blibli, 2022 to 2023.** Revenue rose 465 percent; fewer discounts account for 63 percent, and part of the rest is a shift toward travel. +## 5.3 What moved the take rate -So in 2022 to 2024, Indonesian platform revenue mostly reported pricing decisions. It did not mainly report more buying. +Pricing, not more commerce, moved revenue at the Indonesian and regional platforms. Pricing here means a change in the take rate, through fees or incentives. With equation (5), the take rate accounts for about two thirds of a revenue move at the typical pair, against about a quarter abroad. The wedge, by contrast, follows the commerce: in both groups, about 95 percent of a move in the wedge comes from transaction value. -## 5.4 What market value followed +Where companies disclose both their fees and their incentives, equation (6) splits the growth of revenue further. Table 5.4 gives the seven traced windows. Fees and incentives are measured per unit of transaction value, and a negative share means that a part pulled revenue down. The windows of one platform overlap, so they are not seven independent cases. -Revenue that grows through pricing could still matter to investors. I tested this before collecting the market data, with two predictions registered in advance. Across 28 listed platforms and 138 firm-years, changes in market value moved with growth in transaction value (b = 0.60, p = 0.002). This is an association, not a measure of how investors decide. +**Table 5.4. What moved revenue in seven Indonesian and regional windows (share of the growth of revenue, percent)** -The second prediction did not hold in the main sample. Revenue growth that came from a higher take rate had a smaller estimated effect (b = 0.33). The difference from volume growth is not significant (p = 0.20). It becomes significant only when segment series are added, which mixes segment figures with group market values. +| Platform and window | Growth of revenue | More commerce | Higher fees | Lower incentives | +|---|---|---|---|---| +| Tokopedia, 2022–23 | +53% | −22 | 76 | 45 | +| Blibli third-party retail, 2022–23 | +465% | 17 | 20 | 63 | +| Blibli third-party retail, 2022–25 | +679% | 18 | 63 | 20 | +| Grab on-demand, 2021–24 | +321% | 34 | 34 | 33 | +| Grab on-demand, 2022–24 | +86% | 42 | 11 | 47 | +| GoTo on-demand, 2022–25 | +111% | 12 | 36 | 52 | +| GoTo on-demand, 2023–25 | +114% | 27 | 6 | 67 | -So market value moved with buying. Whether investors priced revenue from higher fees differently is not settled. +*Note:* Shares are rounded, so a row can add to 99 or 101. Source: company filings; author's calculations (Appendix A). -## 5.5 2023: the headline figures and the sellers +Higher fees and lower incentives together outweighed more commerce in every window. Lower incentives were the larger of the two in four windows and higher fees in two. In the seventh, Grab from 2021 to 2024, the two were nearly equal. Three series need a note. GoTo's 2022 transaction value is an estimate. Blibli does not report incentives, so its split rests on a proxy, and part of its 2022–23 rise comes from travel and digital products. Bukalapak's reports do not allow the parts to be separated. -The platform results meet the national figures in 2023. That year, the headline figures described a boom. Tokopedia's revenue rose 53.2 percent. BPS later reported that the value of online sales grew 40.6 percent. It also reported 27.4 percent more online sellers, about 819,000 more (Figure 5.3). +Tokopedia can be read in two ways, and both point to pricing. Its incentives fell by about a third in 2023. Counted line by line in its accounts, lower incentives gave about 60 percent of the extra revenue and higher gross fees about 40 percent. That count does not adjust for the fall in transaction value, which lowered both lines. Measured per unit of transaction value, as in Table 5.4, higher fees account for more of the rise in the take rate. The two views differ only on which lever led (Figure 5.2). -Measures of buying through the platforms told a different story. Household spending grew 9.4 percent in nominal terms. Five of six Indonesia-only measures of buying through platforms grew more slowly than that. Three of them fell: +![Figure 5.2](figures/fig5_2_tokopedia_two_views.png) -- Tokopedia's transaction value fell 8.9 percent. -- GoTo's ride and delivery transaction value fell 10.5 percent. -- Bank Indonesia's e-commerce figure fell 4.7 percent. +**Figure 5.2. Tokopedia, 2022 to 2023: two ways to split the rise in revenue.** Left: the change in net revenue split by the lines of the accounts. Right: the growth of revenue split per unit of transaction value, where a negative share pulled revenue down. Source: GoTo annual report 2023, segment note; author's calculations. -The exception was Blibli, whose transaction value grew 34.7 percent. +Tokopedia is not the only case of revenue rising while the commerce shrank. From 2024 to 2025, Blibli's transaction value fell by about 2 percent while its revenue rose by about 12 percent. In the windows traced here, revenue growth was not a fair guide to the commerce the platforms carried. Whether markets treated it as one is a separate question. -Sellers described a hard year. BPS asks each business how its online sales changed. Among businesses already selling online before 2023, 24 percent said up, 44 percent said about the same, and 32 percent said down. Their average reported change was a fall of 4.3 percent. Asked about their main problem, 41 percent named a lack of customers, up from 35 percent the year before. +## 5.4 What markets valued -Where two measures look at the same place, they agree. BPS's own estimate of sales through marketplaces grew little in 2023: 0.2 percent in the middle scenario, at most 13 percent. That is in line with the platforms. The national boom sits outside the marketplaces. If online selling grew strongly that year, it grew among sellers who use chat apps and social media. The official count of those sellers is the subject of H2. +Across 28 listed platforms and 138 platform-years, market value moved with transaction value, a registered result. A 1 percent rise in transaction value goes with a rise of about 0.6 percent in market value (β_V in equation (8), p = 0.002). Section 4.5 describes this sample, which differs from the 27 foreign platforms of H1. ---- +The second registered prediction does not hold in the main sample. It expected markets to value growth from a higher take rate less than growth from more commerce. A 1 percent rise in the take rate goes with a rise of about 0.3 percent (β_m). That estimate is too uncertain to tell from zero. The gap between the two estimates is not large enough to count (p = 0.20), or p = 0.16 with extreme values capped. -*Figures for this part: 5.1 `exhibits/v2/1_thin_slice.png`; 5.2 `exhibits/v2/2_tokopedia_2023.png`; 5.3 `exhibits/6_verdict_2023.png` (paths under `research/option_space_2026-10-01/`). Sources: `FINDINGS.md` sections 1-3 and the 2023 verdict; Part 3 with its dated notes; `h1_extended.py`, `h1_points_vs_log.py`, `wedge_split.py`, `growth_decomposition.py`, `m1_market_value.py`, `gel_checks.py`; GoTo annual report 2023, Note 29.* +A second run adds five segment series, including Grab, GoTo and Sea, and Mercado Libre. There the gap counts (p = 0.004), with about 0.6 for transaction value against about 0.1 for the take rate. That run sets segment figures against whole-company market values, as the registration warned, so the thesis does not rely on it. The answer depends on which platforms are in the sample. + +Market value follows the commerce a platform carries. Early studies found a similar link, with markets pricing measures of web activity beyond earnings (Trueman, Wong and Zhang 2000; Rajgopal, Venkatachalam and Kotha 2003). Here the measure of activity is transaction value, which the platforms report themselves. How a firm records revenue matters too. After the 2000 crash, revenue explained less of the prices of internet firms that booked the full sale amount as revenue (Davis 2002). The data do not show that markets valued growth from pricing any less. Revenue and transaction value come from the platforms' own reports, and market value from share prices. The national figures give a separate view of the same commerce. + +## 5.5 The platforms and the sellers in 2023 + +Buying through platforms did not boom in 2023. Household spending rose by about 9 percent in current prices that year. Of six measures of buying through platforms, five grew more slowly, and three fell. Blibli is the one exception. Figure 5.3 sets the measures side by side, each in its own unit. + +The figures that described a boom were of other kinds. Tokopedia's revenue rose by 53 percent and BPS's online sales value by 41 percent, while the buying that platforms carried grew slowly or fell. Figure 5.3 shows the two kinds in different colours. + +Sellers' own accounts of their sales show no boom either, a registered result. BPS asks each business how its online sales changed. For businesses already selling, the registered prediction set an upper limit on the median change. The limit was about 10 percent, the published growth in sales per business. The median change was zero. + +![Figure 5.3](figures/fig5_3_growth_2023.png) + +**Figure 5.3. Growth in 2023 (percent).** Blue bars show measures of buying; orange bars show the figures that described a boom. Each measure is in its own unit. Sources: company reports, Bank Indonesia, Momentum Works and BPS; household spending from national accounts. GoTo's change rests on an estimated 2022 value, and Momentum Works' 3.5 is the author's calculation from its published levels (its 2024 release states 3.7). Appendix D gives the values. + +Fewer sellers reported a rise than a year earlier. About a quarter reported a rise and about a third a fall. A year before, more than a third had reported a rise and a quarter a fall. The mean change was about −4 percent. Lack of demand became the main obstacle, named by 41 percent of sellers against 35 percent in 2022. The published count of online sellers nonetheless rose by 27 percent in 2023. A rise of that size, in a year without a boom in sales, needs explaining. + +## 5.6 H2 and the entry test + +H2 holds on BPS's published figures, but the registered entry test shows that first-time entry cannot explain the whole rise in the count. From 2022 to 2023, online sales value grew by 41 percent and the number of businesses by 27 percent. More businesses account for 71 percent of that growth, and higher sales per business for the rest. + +H2 holds again from 2023 to 2024. Value grew by 17 percent, and more businesses account for 90 percent of the growth, or 71 percent on BPS's other count for 2023. The split of equation (7) is arithmetic, and it cannot tell a new seller from one counted for the first time. + +The registered entry test separates new sellers from sellers counted for the first time, and it fails. In the survey for 2023 (the 2024 file), about 13 percent of sellers began selling online that year. The test required 21.5 percent. Even with no sellers stopping, that many new sellers would raise the count by at most about 16 percent, against a published rise of 27 percent. + +Recall error does not explain the shortfall. As Section 3.5 noted, sellers who report a start date later than the true one make new sellers look more numerous. Such errors push the share toward passing, so the shortfall is, if anything, understated. Part of the 2023 rise in the count is something other than first-time entry, and its size is the next result. + +## 5.7 What the rise in the count contains + +About half of the 2023 rise in the count is not explained by first-time entry, an exploratory result. Table 5.5 sets out the accounting. In the survey for 2023, sellers who began before 2023 outnumber the whole count for 2022 by about 306 thousand. They make up 37 percent of the rise if no seller stopped, and 56 percent if sellers stopped at about 5 percent a year. + +**Table 5.5. The 2023 rise in the count of online sellers (thousands)** + +| Part of the rise | If no seller stopped | If sellers stopped at 5 percent a year | +|---|---|---| +| Rise between the weighted totals of the two files | 819 (100%) | 819 (100%) | +| Net entry: sellers who began in 2023, less those who stopped | 513 (63%) | 361 (44%) | +| Earlier sellers newly counted above the whole 2022 count | 306 (37%) | 458 (56%) | + +*Note:* BPS's published counts give a rise of 821 thousand (BPS 2025a). Source: BPS survey files, weighted; author's calculations. + +The cohort counts show the break directly. A cohort should only shrink between surveys, and the cohorts did so in every interval except 2023. Sellers who began by 2020 fell by about 10 percent between the surveys for 2020 and 2022. In the survey for 2023 they rose by 6 percent. + +The count behaved normally again in BPS's published table for 2024. The cohort that began before 2020 shrank by about 10 percent. At least 754 thousand new sellers appeared, more than the rise of 584 thousand. + +The break was uneven across places and channels. Across the 34 provinces, growth in the count had almost no link to the share of new sellers, a rank correlation of 0.10. It had only a weak link to growth in the sample, at 0.35. A rank correlation compares the order of provinces on two measures and runs from −1 to 1. The break sat in chat and social selling. Among sellers who began by 2022, sellers using only chat or social media rose by 13 percent, and marketplace users by 0.3 percent. + +Sampling noise is very unlikely to explain the break. BPS's published error rates put the usual size of the sampling noise in the national count at about 1.4 percent. The 306 thousand extra sellers are about 4.6 times that noise in the change between the two years. The calculation applies the 2024 edition's error rates to both years and treats the two surveys as independent. + +Most of the 2023 rise in sales value cannot be traced to what sellers reported. BPS's online sales value rose by 41 percent in rupiah, from about US$53 billion to about US$72 billion at each year's exchange rate. The thesis splits that rise by giving each kind of seller its typical size in the 2022 file. New sellers, at the size of those that began in 2022, account for 25 to 30 percent of the rise. The extra earlier sellers, at the size of chat and social sellers, account for 15 to 17 percent. The typical seller already online reported no growth, with a mean change of about −4 percent when each seller counts once. So 53 to 59 percent of the rise has no source in the sellers' answers. A few large sellers growing fast could fill that part, and the files cannot show whether they did. + +The extra count also bears on H2, in an exploratory follow-up. H2's split treats every seller in the count as having entered. Suppose instead that the sellers not explained by entry were reached by wider coverage, and that their sales came with them. Take both out of the 2023 figures. More businesses then account for about 54 percent of the growth if no seller stopped, and about 42 percent if sellers stopped at 5 percent a year. H2 requires more than 50 percent. So H2 holds on the published series, but it sits on its threshold once the extra count is read as coverage. Taking out the sellers but keeping their sales gives 34 to 47 percent, though that version treats the count and the value unevenly. + +The checks point to a survey that reached more chat and social sellers in 2023, the kind of break that Section 3.5 described. The files cannot separate that reading from sellers returning after a pause. Who the extra sellers were bears on the choice between the two. + +## 5.8 Who the extra sellers were + +The extra count leaned toward long-running businesses, older owners and places outside Java, an exploratory result. Each group's extra count compares two numbers (Section 4.7). The first is the group's 2023 count. The second is the 2022 count grown at the group's own yearly rate from 2020 to 2022. The extra count is the first less the second. The extra sellers total about 450 thousand, close to the 458 thousand in Table 5.5. Figure 5.4 compares each group's share of the extra count with its share of the 2022 count. + +![Figure 5.4](figures/fig5_4_extra_sellers.png) + +**Figure 5.4. Who the extra sellers were.** The extra sellers are the 2023 count above what each group's normal yearly change in 2020 to 2022 would give, about 450 thousand in total. Source: BPS survey files, weighted; author's calculations. Exploratory. + +The extra count did not lean toward young businesses. A lean toward men in the raw counts disappears once each group's normal change is taken out. This profile bears on whether the extra sellers were returning after a pause. + +Part-year selling weighs against returners. A returning seller would usually sell for only part of the year, unless it restarted in January, yet part-year selling fell in 2023. Among sellers who began by 2021, those selling for part of the year fell from about 470 thousand to about 320 thousand. The same fall appears for the groups that began by 2020 and by 2022. + +The age profile weighs against returners as well. Small firms that stop are more often young firms with young owners (McKenzie and Paffhausen 2019). Returners would then make the extra count lean young. It leaned old. + +Returning sellers still cannot be ruled out. Returners equal to 10 to 15 percent of the 2022 count would remove the unexplained part, which is a scenario and not an estimate. The files carry no linked histories, and BPS changed how it asks about months between the rounds. Exits, or a move to full-year selling, could also offset part-year returners. Where online sellers sell, and what records they keep, complete the survey evidence. + +## 5.9 Channels and records + +Most online selling happens outside marketplaces, and records are thinner there. The path of marketplace use, the value held by sellers without financial statements and the World Bank results for Indonesia were registered. The comparison within the group that began before 2020 is exploratory. + +Nearly all online sellers, about 96 percent, use chat or social media. Marketplace use fell every time it was measured, from about 22 percent of sellers in 2020 to about 17 percent in 2024. BPS worded the question slightly differently for 2023 and 2024. + +Marketplaces carry only about a fifth of online sales value. The share was 23 to 26 percent in 2022 on the survey files. In BPS's published figures it was 18 percent in 2023 and 16 percent in 2024. + +Almost all of the recent growth in value came from outside the marketplaces. From 2023 to 2024, BPS's marketplace value grew by about 1.5 percent and other channels by about 21 percent. BPS's marketplace category is itself broad: 43 percent of its sellers use only Gojek or Grab, and 47 percent use a shopping marketplace. + +Records are thinner outside the marketplaces. In 2023, about 29 percent of marketplace sellers kept complete financial statements, against about 12 percent of other sellers. Figure 5.5 shows the share of all sellers who keep them falling. The fall from 2020 to 2022 came while the survey appears to have counted the same kinds of seller. The fall in 2023 may partly reflect the chat-only sellers added to the count. + +![Figure 5.5](figures/fig5_5_marketplace_and_records.png) + +**Figure 5.5. Marketplace use and financial statements, 2020 to 2024.** Marketplace use for 2020, 2022 and 2023 comes from the survey files and for 2024 from BPS's published figure. Financial statements come from the survey files. BPS worded the marketplace question slightly differently for 2023 and 2024, and the survey has no point for 2021. Source: BPS; author's calculations. + +The fall reaches long-running sellers too, an exploratory finding. Among sellers who began before 2020, the share keeping complete statements fell from about 23 percent in 2022 to about 17 percent in 2023. + +Sellers without statements hold far less of the value than of the count. On the bracket scenarios of Section 4.8, they held 26 to 40 percent of online value in 2022, against 79 percent of the count. The validation rule asks the total built from the brackets to lie within half and double BPS's total. The annual reading passes, just inside the limit at about twice BPS's total, while the monthly reading gives about 23 times. The lower brackets alone exceed BPS's total, so the shares are ranges. + +One registered prediction on value failed narrowly. It expected the value share of sellers without statements to fall by no more than 5 points between 2020 and 2022. The share fell by 5.8 points on the main version and by 3.3 points on a second version. + +World Bank surveys of Indonesian firms show the same thin records. In six Indonesian cities, 27 percent of unregistered firms use social media with their customers. Among those, 70 percent keep no written records, and their median sales are about US$4,000 a year. A registered prediction that their mean sales would stay below about US$6,600 failed, at about US$7,300. + +Among formal firms, about 40 percent in Indonesia have neither a website nor online tax filing. That is the highest share of seven countries, against about 9 percent in Malaysia and in Thailand, a descriptive comparison. + +The split by channel bears on the third question of Section 2.4, because a rule built on platform records reaches only the marketplace part. Chapter 6 measures that reach. The same split also marks where the national figures part. + +## 5.10 Why the national figures disagree + +The national figures agree that growth was small only where they cover the same sellers, and they do not agree on scale. In 2023, Bank Indonesia's e-commerce figure fell by about 5 percent while BPS's rose by about 41 percent. That is the disagreement with which Chapter 1 began, and the third objective asks what lies behind it. + +Where Bank Indonesia's figure and BPS's cover the same ground, the growth is small. From 2022 to 2023, BPS's own marketplace slice grew by about 0.2 percent in the middle scenario. Across the scenarios it grew by 0.2 to 13 percent. From 2023 to 2024 it grew by about 1.5 percent. Most of the disagreement sits outside the marketplaces, among the sellers that only BPS reaches. + +A registered test on size gives a mixed answer. It predicted that the part of BPS's marketplace sales that went to final consumers in 2022 would fall below Bank Indonesia's figure. The prediction holds in the low and middle scenarios, at 0.16 and 0.68 times Bank Indonesia's figure. It fails in the high one, at 1.62 times. + +In 2024 the figures divide by coverage. Those that cover mainly platforms grew by 5 to 7 percent, and BPS's grew by about 17 percent. The adjacent figures of Table 4.3 point the same way as BPS but do not measure Indonesian e-commerce. Parcels shipped in Southeast Asia grew by about 25 percent, and the tax on foreign digital services by about 25 percent. + +Payment figures grew faster than any of these from 2023 to 2024. E-money shopping grew by about 30 percent, mobile-banking purchases by 83 percent and QRIS payments by 187 percent. They are context only, because they include in-person payments and the switch away from cash. + +Levels cannot be reconciled. In 2024, BPS put online sales at about US$81 billion and Bank Indonesia at about US$31 billion. In 2023, BPS's marketplace slice was about US$13 billion, against about US$38 billion for Tokopedia and the Shopee estimate together. In 2024 it was less than a quarter of Momentum Works' marketplace estimate of about US$57 billion (Setyowati 2025). + +Platform figures and the seller survey count different things. Platform transaction value includes shipping, digital goods, travel and offline sales that a seller survey does not count in the same way. The counts of sellers differ too. Tokopedia alone reported more than 14 million merchants in 2023 (GoTo Gojek Tokopedia 2024b). The survey files put marketplace sellers at about 0.7 million that year, on the author's calculation. + +Platform commerce does not always keep pace with household spending. In each of the four quarters to mid-2026, GoTo's on-demand transaction value grew more slowly than household spending. In the last, it grew by 2 percent against 8 percent, while its revenue grew by about 20 percent. Grab's and Shopee's figures grew faster, but they cover the region. + +The results answer the research question at two levels. On the inner link, revenue showed a small part of the commerce and stopped being a fair guide when the take rate moved. On the outer link, the 2023 rise in the count of sellers was only partly first-time entry. The figures part most where only the survey looks. Chapter 6 turns to the rules that Indonesia is writing. diff --git a/papers/thesis_2026-10/drafts/ch5_full_2026-10-14.md b/papers/thesis_2026-10/drafts/ch5_full_2026-10-14.md new file mode 100644 index 0000000..0104b69 --- /dev/null +++ b/papers/thesis_2026-10/drafts/ch5_full_2026-10-14.md @@ -0,0 +1,232 @@ +# Chapter 5. Results + +The results follow the two links of Figure 3.1. On the inner link, revenue showed a small and unsteady part of the commerce that platforms carried. Platforms changed how much of each sale they kept, and revenue moved with those choices (Sections 5.1 to 5.4). On the outer link, the national count of online sellers rose by 27 percent in 2023. About half of that rise did not come from sellers entering for the first time (Sections 5.5 to 5.10). + +Every result in this chapter carries one of the three labels of Section 4.10: registered, exploratory or descriptive. A result given without a label is descriptive. Table 5.1 gives each main result with its label, the checks behind it and the section that reports it. + +**Table 5.1. Main results and the evidence behind them** + +| Result | Label | Main checks | Section | +|---|---|---|---| +| *Inner link: revenue and platform commerce* | | | | +| Revenue was about 7 percent of what three major platforms carried in 2023; the wedge was about US$40 billion | Descriptive | Shopee's take rate varied; each platform dropped in turn | 5.1 | +| Take rates moved about three times as much in Indonesia and the region as abroad, so H1 is rejected | Exploratory; the narrower test planned in advance points the same way | Each platform dropped in turn; Ozon and Zalando dropped; change measured in points; quarterly pairs do not reject | 5.2 | +| The thin take rate of Indonesian and regional platforms explains much of that difference | Exploratory | Changes in percentage points compared across groups | 5.2 | +| Pricing, not more commerce, moved revenue | Descriptive | All 15 pairs against 158 abroad; seven traced windows; two views of Tokopedia | 5.3 | +| Market value moves with transaction value | Registered; holds | Extreme values trimmed; a second run | 5.4 | +| Markets value growth from a higher take rate less | Registered; does not hold in the main sample | Holds only in a second run that mixes segment figures | 5.4 | +| *Outer link: platform commerce and all online commerce* | | | | +| Buying through platforms, and sellers' own sales, did not boom in 2023 | Descriptive; registered for sellers' sales, holds | Six platform measures against household spending | 5.5 | +| More businesses account for most national growth, so H2 holds | Descriptive | Both BPS counts for 2023; two periods | 5.6 | +| Too few 2023 sellers were new to produce the rise in the count (13.45 against 21.5 percent) | Registered; the entry test fails | Recall error would favour passing | 5.6 | +| About half of the rise is not explained by first-time entry | Exploratory | With and without exits; cohorts; provinces; sampling error | 5.7 | +| The extra count leaned toward long-running businesses; returning sellers cannot be ruled out | Exploratory | Part-year selling | 5.8 | +| Most online selling happens outside marketplaces, and records are thinner there | Registered and descriptive | Four survey years; World Bank surveys | 5.9 | +| National figures agree only where they cover the same sellers | Registered, with a mixed outcome, and descriptive | Three value scenarios | 5.10 | + +## 5.1 How much revenue leaves out + +Revenue showed only a small part of what Indonesia's major platforms carried. In 2023, Tokopedia, Shopee and Grab together carried about US$43 billion of transaction value and kept about US$3 billion as revenue. Revenue was about 7 percent of that commerce, and the invisible wedge was about US$40 billion. The ecosystem ratio was about 13, so almost 13 dollars of commerce passed through the platforms for every dollar of revenue. Table 5.2 shows each platform. + +**Table 5.2. The invisible wedge of three platforms, 2023 (US$ billion)** + +| Platform | Transaction value | Revenue | Wedge | Ecosystem ratio | How it is labelled | +|---|---|---|---|---|---| +| Tokopedia | 16.33 | 0.41 | 15.93 | 39.3 | Both direct segment disclosure | +| Shopee, Indonesia | 21.52 | 2.15 | 19.37 | 9.0 | Transaction value is an outside estimate (Momentum Works); revenue is derived at an assumed take rate of 10 percent | +| Grab, Indonesia | 5.38 | 0.60 | 4.78 | 7.9 | Revenue is a direct country disclosure; transaction value is derived from Grab's company-wide ratio | +| Total | 43.23 | 3.16 | 40.07 | 12.7 | Sum of three platforms, not a national total | + +*Note:* Rounded amounts may not add. With Shopee's take rate at 9 or 11 percent, the total wedge is US$39.9 or US$40.3 billion; at Sea's company-wide rate of 11.5 percent, it is US$39.8 billion. Source: company filings; Momentum Works (2024); author's calculations. + +The ratios differ because the amounts were obtained in different ways. Tokopedia reports both amounts, and its ratio is about 39. Shopee's ratio of 9 follows from the assumed take rate, and Grab's ratio of about 8 is its company-wide ratio. Neither is separate evidence for Indonesia. Blibli and Bukalapak also report both amounts, with ratios of about 43 and 36, close to Tokopedia's. They are left out of the sum because Blibli's figure includes online travel and Bukalapak reports for its whole group. + +The size of the wedge does not depend on the assumed rates. A take rate for Shopee between 9 and 11 percent keeps the total near US$40 billion. Dropping any one platform still leaves more than US$20 billion. For platforms that report both amounts, the wedge was 97 to 99 percent of transaction value in every year. It moved almost one for one with the commerce, and Appendix A gives the wedge by platform and year. + +Platform commerce is itself only part of online commerce. In BPS's survey, marketplaces carry about a fifth of online sales value (Section 5.9), so platform revenue shows a smaller part still. A small share can still be a fair guide to the commerce, as long as it keeps its size from year to year. + +## 5.2 Testing H1 + +Revenue did not keep pace with transaction value at the Indonesian and regional platforms, and the thesis rejects H1. At the typical platform in this group, the take rate moved by about 20 percent a year, up or down. At the typical foreign platform it moved by about 6 percent, so the moves were about three times as large. Table 5.3 gives the test and its checks. + +**Table 5.3. The test of H1** + +| Test | Method | Indonesian and regional group | Foreign group | p-value | +|---|---|---|---|---| +| Yearly change in log take rate, 5 platforms against 27 | Mann–Whitney | 0.20 | 0.06 | 0.004 | +| The same | Shuffle | 0.20 | 0.06 | 0.022 | +| Growth divergence in points, 5 platforms against 26 | Mann–Whitney | 27.0 | 6.5 | 0.005 | +| Drop each platform in turn, or treat Tokopedia and GoTo as one company | Mann–Whitney | 0.17 to 0.23 | 0.06 | 0.003 to 0.016 | +| Quarterly pairs, 3 platforms against 8 | Mann–Whitney | 0.071 | 0.049 | 0.19 | + +*Note:* Group values are medians of each platform's median yearly change. Appendix B lists every check, including the tests without Ozon and Zalando. + +The main test compares the five platforms with 27 abroad, one median per platform. Both methods of Section 4.3 reject equal moves, with p = 0.004 and p = 0.022. No single platform carries the result. Dropping any one, or treating Tokopedia and GoTo as one company, keeps the Mann–Whitney p-value at 0.016 or below. The main test was set up after a first, narrower test, so it is exploratory. + +The test planned in advance was narrower: the three Indonesian marketplaces against the eight foreign platforms chosen first. It points the same way. With one median per platform, the Mann–Whitney test rejects equal moves (p = 0.024). The shuffle of platform labels comes close without passing (p = 0.06), since three platforms give a test little power. In that sample, revenue growth and transaction value growth had a correlation of 0.10 in Indonesia and 0.98 abroad, excluding Sea. + +The direction of the moves is less telling than their size. Revenue grew faster than transaction value in 13 of the 15 Indonesian and regional pairs. In two of them, revenue and transaction value moved in opposite directions. Abroad, take rates also rose in about two thirds of pairs, but by much less. What sets the Indonesian and regional platforms apart is the size of the moves, which is what the test measures. + +Quarterly pairs do not reject H1 (p = 0.19), contrary to a prediction registered before the test was run. The quarterly Indonesian pairs mostly cover 2023 to 2026, after most of the repricing, while the large moves came between 2021 and 2024. The rejection of H1 describes that period. + +The thin take rate of Indonesian and regional platforms explains much of the threefold difference, an exploratory finding. Measured in percentage points, both groups changed their take rates by about 1 point a year. The difference between them is nowhere near large enough to count (p = 0.65). + +The two groups differ in the take rate they start from. The typical Indonesian and regional platform keeps about 1.6 percent of each sale, against about 17 percent abroad. As Section 3.2 showed, the same 1-point rise then has very different effects. It lifts revenue by more than half at the thinner take rate and by about 6 percent at the thicker one. Figure 5.1 shows the three measures side by side. The test shows that take rates moved, but not what the platforms changed to move them. + +![Figure 5.1](figures/fig5_1_thin_slice.png) + +**Figure 5.1. The thin take rate.** Each dot is one platform's median over its years. Left: the share of each sale that the platform keeps, on a log scale. Middle: the yearly change in that share, in percentage points. Right: the same change relative to the share itself, as a log change in percent. Black bars mark group medians. The foreign group has 26 platforms here and 27 in the main test, because one platform has no pair of usable years for the change in points. Source: company filings; author's calculations. Exploratory. + +## 5.3 What moved the take rate + +Pricing, not more commerce, moved revenue at the Indonesian and regional platforms. With equation (5), the take rate accounts for about two thirds of a revenue move at the typical pair, against about a quarter abroad. The wedge, by contrast, follows the commerce: in both groups, about 95 percent of a move in the wedge comes from transaction value. + +Where companies disclose both their fees and their incentives, equation (6) splits the growth of revenue further. Table 5.4 gives the seven traced windows. Fees and incentives are measured per unit of transaction value, and a negative share means that a part pulled revenue down. The windows of one platform overlap, so they are not seven independent cases. + +**Table 5.4. What moved revenue in seven Indonesian and regional windows (share of the growth of revenue, percent)** + +| Platform and window | Growth of revenue | More commerce | Higher fees | Lower incentives | +|---|---|---|---|---| +| Tokopedia, 2022–23 | +53% | −22 | 76 | 45 | +| Blibli third-party retail, 2022–23 | +465% | 17 | 20 | 63 | +| Blibli third-party retail, 2022–25 | +679% | 18 | 63 | 20 | +| Grab on-demand, 2021–24 | +321% | 34 | 34 | 33 | +| Grab on-demand, 2022–24 | +86% | 42 | 11 | 47 | +| GoTo on-demand, 2022–25 | +111% | 12 | 36 | 52 | +| GoTo on-demand, 2023–25 | +114% | 27 | 6 | 67 | + +*Note:* Shares are rounded, so a row can add to 99 or 101. Source: company filings; author's calculations (Appendix A). + +Higher fees and lower incentives together outweighed more commerce in every window. Lower incentives were the larger of the two in four windows and higher fees in two. In the seventh, Grab from 2021 to 2024, the two were nearly equal. Three series need a note. GoTo's 2022 transaction value is an estimate. Blibli does not report incentives, so its split rests on a proxy, and part of its 2022–23 rise comes from travel and digital products. Bukalapak's reports do not allow the parts to be separated. + +Tokopedia can be read in two ways, and both point to pricing. Its incentives fell by about a third in 2023. Counted line by line in its accounts, lower incentives gave about 60 percent of the extra revenue and higher gross fees about 40 percent. That count leaves the fall in volume inside both lines. Measured per unit of transaction value, as in Table 5.4, higher fees account for more of the rise in the take rate. The two views differ only on which lever led (Figure 5.2). + +![Figure 5.2](figures/fig5_2_tokopedia_two_views.png) + +**Figure 5.2. Tokopedia, 2022 to 2023: two ways to split the rise in revenue.** Left: the change in net revenue split by the lines of the accounts. Right: the growth of revenue split per unit of transaction value, where a negative share pulled revenue down. Source: GoTo annual report 2023, segment note; author's calculations. + +Tokopedia is not the only case of revenue rising while the commerce shrank. From 2024 to 2025, Blibli's transaction value fell by about 2 percent while its revenue rose by about 12 percent. In the windows traced here, revenue growth was not a fair guide to the commerce the platforms carried. Whether markets treated it as one is a separate question. + +## 5.4 What markets valued + +Across 28 listed platforms and 138 platform-years, market value moved with transaction value, a registered result. A 1 percent rise in transaction value goes with a rise of about 0.6 percent in market value (β_V in equation (8), p = 0.002). Section 4.5 describes this sample, which differs from the 27 foreign platforms of H1. + +The second registered prediction does not hold in the main sample. It expected markets to value growth from a higher take rate less than growth from more commerce. A 1 percent rise in the take rate goes with a rise of about 0.3 percent (β_m). That estimate is too uncertain to tell from zero. The gap between the two estimates is not large enough to count (p = 0.20), or p = 0.16 with extreme values trimmed. + +A second run adds three segment series and three more whole companies. There the gap counts (p = 0.004), with about 0.6 for transaction value against about 0.1 for the take rate. That run sets segment figures against whole-company market values, as the registration warned, so the thesis does not rely on it. The answer depends on which platforms are in the sample. + +Market value follows the commerce a platform carries. Early studies found a similar link, with markets pricing measures of web activity beyond revenue (Trueman, Wong and Zhang 2000). The data do not show that markets valued growth from pricing any less. Revenue, transaction value and market value all come from the platforms' own reports. The national figures give a separate view of the same commerce. + +## 5.5 The platforms and the sellers in 2023 + +Buying through platforms did not boom in 2023. Household spending rose by about 9 percent in current prices that year. Of six measures of buying through platforms, five grew more slowly, and three fell. Blibli is the one exception. Figure 5.3 sets the measures side by side, each in its own unit. + +The figures that described a boom were of other kinds. Tokopedia's revenue rose by 53 percent and BPS's online sales value by 41 percent, while the buying that platforms carried grew slowly or fell. Figure 5.3 shows the two kinds in different colours. + +![Figure 5.3](figures/fig5_3_growth_2023.png) + +**Figure 5.3. Growth in 2023 (percent).** Blue bars show measures of buying; orange bars show the figures that described a boom. Each measure is in its own unit. Sources: company reports, Bank Indonesia, Momentum Works and BPS; household spending from national accounts. GoTo's change rests on an estimated 2022 value, and Momentum Works' 3.5 is the author's calculation from its published levels (its 2024 release states 3.7). Appendix D gives the values. + +Sellers' own accounts of their sales show no boom either, a registered result. BPS asks each business how its online sales changed. For businesses already selling, the registered prediction set an upper limit on the median change. The limit was about 10 percent, the published growth in sales per business. The median change was zero. + +Fewer sellers reported a rise than a year earlier. About a quarter reported a rise and about a third a fall. A year before, more than a third had reported a rise and a quarter a fall. The mean change was about −4 percent. Lack of demand became the main obstacle, named by 41 percent of sellers against 35 percent in 2022. The published count of online sellers nonetheless rose by 27 percent in 2023. A rise of that size, in a year without a boom in sales, needs explaining. + +## 5.6 H2 and the entry test + +H2 holds on BPS's published figures, but the registered entry test shows that first-time entry cannot explain the whole rise in the count. From 2022 to 2023, online sales value grew by 41 percent and the number of businesses by 27 percent. More businesses account for 71 percent of that growth, and higher sales per business for the rest. + +H2 holds again from 2023 to 2024. Value grew by 17 percent, and more businesses account for 90 percent of the growth, or 71 percent on BPS's other count for 2023. The split of equation (7) is arithmetic, and it cannot tell a new seller from one counted for the first time. + +The registered entry test separates new sellers from sellers counted for the first time, and it fails. In the survey for 2023 (the 2024 file), about 13 percent of sellers began selling online that year. The test required 21.5 percent. Even with no sellers stopping, that many new sellers would raise the count by at most about 16 percent, against a published rise of 27 percent. + +Recall error does not explain the shortfall. As Section 3.5 noted, sellers who report a start date later than the true one make new sellers look more numerous. Such errors push the share toward passing, so the shortfall is, if anything, understated. Part of the 2023 rise in the count is something other than first-time entry, and its size is the next result. + +## 5.7 What the rise in the count contains + +About half of the 2023 rise in the count is not explained by first-time entry, an exploratory result. Table 5.5 sets out the accounting. In the survey for 2023, sellers who began before 2023 outnumber the whole count for 2022 by about 306 thousand. They make up 37 percent of the rise if no seller stopped, and 56 percent if sellers stopped at about 5 percent a year. + +**Table 5.5. The 2023 rise in the count of online sellers (thousands)** + +| Part of the rise | If no seller stopped | If sellers stopped at 5 percent a year | +|---|---|---| +| Rise between the weighted totals of the two files | 819 (100%) | 819 (100%) | +| Net entry: sellers who began in 2023, less those who stopped | 513 (63%) | 361 (44%) | +| Earlier sellers newly counted above the whole 2022 count | 306 (37%) | 458 (56%) | + +*Note:* BPS's published counts give a rise of 821 thousand (BPS 2025). Source: BPS survey files, weighted; author's calculations. + +The cohort counts show the break directly. A cohort should only shrink between surveys, and the cohorts did so in every interval except 2023. Sellers who began by 2020 fell by about 10 percent between the surveys for 2020 and 2022. In the survey for 2023 they rose by 6 percent. + +The count behaved normally again in BPS's published table for 2024. The cohort that began before 2020 shrank by about 10 percent. At least 754 thousand new sellers appeared, more than the rise of 584 thousand. + +The break was uneven across places and channels. Across the 34 provinces, growth in the count had almost no link to the share of new sellers, a rank correlation of 0.10. It had only a weak link to growth in the sample, at 0.35. A rank correlation compares the order of provinces on two measures and runs from −1 to 1. The break sat in chat and social selling. Among sellers who began by 2022, sellers using only chat or social media rose by 13 percent, and marketplace users by 0.3 percent. + +Sampling noise does not explain the break. BPS's published error rates put the usual size of the sampling noise in the national count at about 1.4 percent. The 306 thousand extra sellers are about 4.6 times that noise in the change between the two years. The calculation applies the 2024 edition's error rates to both years and treats the two surveys as independent. + +Most of the 2023 rise in value has no source in the answers of existing sellers. BPS's online sales value rose from about US$53 billion to about US$72 billion. Entrants account for 25 to 30 percent of that rise, and sellers not explained by entry for 15 to 17 percent. Existing sellers reported a mean change of about −4 percent, so the remaining 53 to 59 percent cannot be traced to them. + +The checks point to a survey that reached more chat and social sellers in 2023, the kind of break that Section 3.5 described. The files cannot separate that reading from sellers returning after a pause. Who the extra sellers were bears on the choice between the two. + +## 5.8 Who the extra sellers were + +The extra count leaned toward long-running businesses, older owners and places outside Java, an exploratory result. Each group's extra count is its 2023 count above what its own yearly change from 2020 to 2022 would give (Section 4.7). The extra sellers total about 450 thousand, close to the 458 thousand in Table 5.5. Figure 5.4 compares each group's share of the extra count with its share of the 2022 count. + +![Figure 5.4](figures/fig5_4_extra_sellers.png) + +**Figure 5.4. Who the extra sellers were.** The extra sellers are the 2023 count above what each group's normal yearly change in 2020 to 2022 would give, about 450 thousand in total. Source: BPS survey files, weighted; author's calculations. Exploratory. + +The extra count did not lean toward young businesses. A lean toward men in the raw counts disappears once each group's normal change is taken out. This profile bears on whether the extra sellers were returning after a pause. + +Part-year selling weighs against returners. A returning seller would usually sell for only part of the year, unless it restarted in January, yet part-year selling fell in 2023. Among sellers who began by 2021, those selling for part of the year fell from about 470 thousand to about 320 thousand. The same fall appears for the groups that began by 2020 and by 2022. + +The age profile weighs against returners as well. Small firms that stop are more often young firms with young owners (McKenzie and Paffhausen 2019). Returners would then make the extra count lean young. It leaned old. + +Returning sellers still cannot be ruled out. Returners equal to 10 to 15 percent of the 2022 count would remove the unexplained part, which is a scenario and not an estimate. The files carry no linked histories, and BPS changed how it asks about months between the rounds. Exits, or a move to full-year selling, could also offset part-year returners. Where online sellers sell, and what records they keep, complete the survey evidence. + +## 5.9 Channels and records + +Most online selling happens outside marketplaces, and records are thinner there. The path of marketplace use, the value held by sellers without financial statements and the World Bank results were registered. The comparison within the group that began before 2020 is exploratory. + +Nearly all online sellers, about 96 percent, use chat or social media. Marketplace use fell every time it was measured, from about 22 percent of sellers in 2020 to about 17 percent in 2024. BPS worded the question slightly differently for 2023 and 2024. + +Marketplaces carry only about a fifth of online sales value. The share was 23 to 26 percent in 2022 on the survey files. In BPS's published figures it was 18 percent in 2023 and 16 percent in 2024. + +Almost all of the recent growth in value came from outside the marketplaces. From 2023 to 2024, BPS's marketplace value grew by about 1.5 percent and other channels by about 21 percent. BPS's marketplace category is itself broad: 43 percent of its sellers use only Gojek or Grab, and 47 percent use a shopping marketplace. + +![Figure 5.5](figures/fig5_5_marketplace_and_records.png) + +**Figure 5.5. Marketplace use and financial statements, 2020 to 2024.** Marketplace use for 2020, 2022 and 2023 comes from the survey files and for 2024 from BPS's published figure. Financial statements come from the survey files. BPS worded the marketplace question slightly differently for 2023 and 2024, and the survey has no point for 2021. Source: BPS; author's calculations. + +Records are thinner outside the marketplaces. In 2023, about 29 percent of marketplace sellers kept complete financial statements, against about 12 percent of other sellers. Figure 5.5 shows the share of all sellers who keep them falling. The fall from 2020 to 2022 came while the survey appears to have counted the same kinds of seller. The fall in 2023 may partly reflect the chat-only sellers added to the count. + +The fall reaches long-running sellers too, an exploratory finding. Among sellers who began before 2020, the share keeping complete statements fell from about 23 percent in 2022 to about 17 percent in 2023. + +Sellers without statements hold far less of the value than of the count. On the bracket scenarios of Section 4.8, they held 26 to 40 percent of online value in 2022, against 79 percent of the count. The annual reading of the brackets passes the validation rule, at 1.95 times BPS's total, while the monthly reading fails at 23 times. The lower brackets alone exceed BPS's total, so the shares are ranges. + +One registered prediction on value failed narrowly. It expected the value share of sellers without statements to fall by no more than 5 points between 2020 and 2022. The share fell by 5.8 points on the main version and by 3.3 points on a second version. + +World Bank surveys of Indonesian firms show the same thin records. In six Indonesian cities, 27 percent of unregistered firms use social media with their customers. Among those, 70 percent keep no written records, and their median sales are about US$4,000 a year. A registered prediction that their mean sales would stay below about US$6,600 failed, at about US$7,300. + +Among formal firms, about 40 percent in Indonesia have neither a website nor online tax filing. That is the highest share of seven countries, against about 9 percent in Malaysia and in Thailand. + +The split by channel bears on the third question of Section 2.4, because a rule built on platform records reaches only the marketplace part. Chapter 6 measures that reach. The same split also marks where the national figures part. + +## 5.10 Why the national figures disagree + +The national figures agree on growth only where they cover the same sellers, and they do not agree on scale. In 2023, Bank Indonesia's e-commerce figure fell by about 5 percent while BPS's rose by about 41 percent. That is the disagreement with which Chapter 1 began, and the third objective asks what lies behind it. + +Where Bank Indonesia's figure and BPS's cover the same ground, the growth is small. From 2022 to 2023, BPS's own marketplace slice grew by about 0.2 percent in the middle scenario. Across the scenarios it grew by 0.2 to 13 percent. From 2023 to 2024 it grew by about 1.5 percent. Most of the disagreement sits outside the marketplaces, among the sellers that only BPS reaches. + +A registered test on size gives a mixed answer. It predicted that the part of BPS's marketplace sales that went to final consumers in 2022 would fall below Bank Indonesia's figure. The prediction holds in the low and middle scenarios, at 0.16 and 0.68 times Bank Indonesia's figure. It fails in the high one, at 1.62 times. + +In 2024 the figures divide by coverage. Those that cover mainly platforms grew by 5 to 7 percent, and BPS's grew by about 17 percent. The adjacent figures of Table 4.3 point the same way as BPS but do not measure Indonesian e-commerce. Parcels shipped in Southeast Asia grew by about 25 percent, and the tax on foreign digital services by about 25 percent. + +Payment figures grew faster than any of these from 2023 to 2024. E-money shopping grew by about 30 percent, mobile-banking purchases by 83 percent and QRIS payments by 187 percent. They are context only, because they include in-person payments and the switch away from cash. + +Levels cannot be reconciled. In 2024, BPS put online sales at about US$81 billion and Bank Indonesia at about US$31 billion. In 2023, BPS's marketplace slice was about US$13 billion, against about US$38 billion for Tokopedia and the Shopee estimate together. In 2024 it was less than a quarter of Momentum Works' marketplace estimate of about US$57 billion. + +Platform figures and the seller survey count different things. Platform transaction value includes shipping, digital goods, travel and offline sales that a seller survey does not count in the same way. The counts of sellers differ too. Tokopedia alone said that it had about 15 million merchants in 2022, while BPS counts about 0.6 million marketplace sellers. + +Platform sales do not always keep pace with household spending. In each of the four quarters to mid-2026, GoTo's on-demand transaction value grew more slowly than household spending. In the last, it grew by 2 percent against 8 percent, while its revenue grew by 21 percent. Grab's and Shopee's figures grew faster, but they cover the region. + +The results answer the research question at two levels. On the inner link, revenue showed a small part of the commerce and stopped being a fair guide when the take rate moved. On the outer link, the 2023 rise in the count of sellers was only partly first-time entry. The figures part most where only the survey looks. Chapter 6 turns to the rules that Indonesia is writing. diff --git a/papers/thesis_2026-10/drafts/ch5_sample_2026-10-13.md b/papers/thesis_2026-10/drafts/ch5_sample_2026-10-13.md new file mode 100644 index 0000000..58fe0dd --- /dev/null +++ b/papers/thesis_2026-10/drafts/ch5_sample_2026-10-13.md @@ -0,0 +1,78 @@ +# Chapter 5. Results + +The results follow the two links of Figure 3.1. On the inner link, revenue showed a small and unsteady part of the commerce that platforms carried. Platforms changed how much of each sale they kept, and revenue moved with those choices (Sections 5.1 to 5.4). On the outer link, the national count of online sellers rose by 27 percent in 2023. About half of that rise did not come from sellers entering for the first time (Sections 5.5 to 5.10). + +Every result in this chapter carries one of the three labels of Section 4.10: registered, exploratory or descriptive. Table 5.1 gives each main result with its label, the checks behind it and the section that reports it. + +**Table 5.1. Main results and the evidence behind them** + +| Result | Label | Main checks | Section | +|---|---|---|---| +| *Inner link: revenue and platform commerce* | | | | +| Revenue was about 7 percent of what three major platforms carried in 2023; the wedge was about US$40 billion | Descriptive | Shopee's take rate varied; each platform dropped in turn | 5.1 | +| Take rates moved about three times as much in Indonesia and the region as abroad, so H1 is rejected | Exploratory; the narrower test planned in advance points the same way | Each platform dropped in turn; Ozon and Zalando dropped; change measured in points; quarterly pairs do not reject | 5.2 | +| The thin share kept by Indonesian and regional platforms explains much of that difference | Exploratory | Changes in percentage points compared across groups | 5.2 | +| Pricing, not more commerce, moved revenue in the traced windows | Descriptive | Two views of Tokopedia; seven windows | 5.3 | +| Market value moves with transaction value | Registered; holds | Extreme values trimmed; a second run | 5.4 | +| Markets value growth from a higher take rate less | Registered; does not hold in the main sample | Holds only in a second run that mixes segment figures | 5.4 | +| *Outer link: platform commerce and all online commerce* | | | | +| Buying through platforms, and sellers' own sales, did not boom in 2023 | Descriptive; registered for sellers' sales, holds | Six platform measures against household spending | 5.5 | +| More businesses account for most national growth, so H2 holds | Descriptive | Both BPS counts for 2023; two periods | 5.6 | +| Too few 2023 sellers were new to produce the rise in the count | Registered; the entry test fails | Recall error would favour passing | 5.6 | +| About half of the rise is not explained by first-time entry | Exploratory | With and without exits; cohorts; provinces; sampling error | 5.7 | +| The extra count leaned toward long-running businesses; returning sellers cannot be ruled out | Exploratory | Part-year selling | 5.8 | +| Most online selling happens outside marketplaces, and records are thinner there | Registered and descriptive | Four survey years; World Bank surveys | 5.9 | +| National figures agree only where they cover the same sellers | Registered, with a mixed outcome, and descriptive | Three value scenarios | 5.10 | + +## 5.1 How much revenue leaves out + +Revenue showed only a small part of what Indonesia's major platforms carried. In 2023, Tokopedia, Shopee and Grab together carried about US$43 billion of transaction value and kept about US$3 billion as revenue. Revenue was about 7 percent of that commerce, and the invisible wedge was about US$40 billion. The ecosystem ratio was about 13, so almost 13 dollars of commerce passed through the platforms for every dollar of revenue. Table 5.2 shows each platform. + +**Table 5.2. The invisible wedge of three platforms, 2023 (US$ billion)** + +| Platform | Transaction value | Revenue | Wedge | Ecosystem ratio | How it is labelled | +|---|---|---|---|---|---| +| Tokopedia | 16.33 | 0.41 | 15.93 | 39.3 | Both direct segment disclosure | +| Shopee, Indonesia | 21.52 | 2.15 | 19.37 | 9.0 | Transaction value is an outside estimate (Momentum Works); revenue is derived at an assumed take rate of 10 percent | +| Grab, Indonesia | 5.38 | 0.60 | 4.78 | 7.9 | Revenue is a direct country disclosure; transaction value is derived from Grab's company-wide ratio | +| Total | 43.23 | 3.16 | 40.07 | 12.7 | Sum of three platforms, not a national total | + +*Note:* Rounded amounts may not add. With Shopee's take rate at 9 or 11 percent, the total wedge is US$39.9 or US$40.3 billion; at Sea's company-wide rate of 11.5 percent, it is US$39.8 billion. Source: company filings; Momentum Works (2024); author's calculations. + +The ratios differ because the amounts were obtained in different ways. Tokopedia reports both amounts, and its ratio is about 39. Shopee's ratio of 9 follows from the assumed take rate, and Grab's ratio of about 8 is its company-wide ratio. Neither is separate evidence for Indonesia. Blibli and Bukalapak also report both amounts, with ratios of about 43 and 36, close to Tokopedia's. They are left out of the sum because Blibli's figure includes online travel and Bukalapak reports for its whole group. + +The size of the wedge does not depend on the assumed rates. A take rate for Shopee between 9 and 11 percent keeps the total near US$40 billion. Dropping any one platform still leaves more than US$20 billion. For platforms that report both amounts, the wedge was 97 to 99 percent of transaction value in every year. It moved almost one for one with the commerce. Appendix A gives the wedge by platform and year. + +Platform commerce is itself only part of online commerce. In BPS's survey, marketplaces carry about a fifth of online sales value (Section 5.9), so platform revenue shows a smaller part still. A small share can still be a fair guide to the commerce, as long as it keeps its size from year to year. + +## 5.2 Testing H1 + +Revenue did not keep pace with transaction value at the Indonesian and regional platforms, and the thesis rejects H1. At the typical platform in this group, the take rate moved by about 20 percent a year, up or down. At the typical foreign platform it moved by about 6 percent, so the moves were about three times as large. Table 5.3 gives the test and its checks. + +**Table 5.3. The test of H1** + +| Test | Method | Indonesian and regional group | Foreign group | p-value | +|---|---|---|---|---| +| Yearly change in log take rate, 5 platforms against 27 | Mann–Whitney | 0.20 | 0.06 | 0.004 | +| The same | Shuffle | 0.20 | 0.06 | 0.022 | +| Growth divergence in points, 5 platforms against 26 | Mann–Whitney | 27.0 | 6.5 | 0.005 | +| Drop each platform in turn, or treat Tokopedia and GoTo as one company | Mann–Whitney | 0.17 to 0.23 | 0.06 | 0.003 to 0.016 | +| Quarterly pairs, 3 platforms against 8 | Mann–Whitney | 0.071 | 0.049 | 0.19 | + +*Note:* Group values are medians of each platform's median yearly change. Appendix B lists every check, including the tests without Ozon and Zalando. + +The main test compares the five platforms with 27 abroad, one median per platform. Both methods of Section 4.3 reject equal moves, with p = 0.004 and p = 0.022. No single platform carries the result. Dropping any one, or treating Tokopedia and GoTo as one company, keeps the Mann–Whitney p-value at 0.016 or below. The main test was set up after a first, narrower test, so it is exploratory. + +The test planned in advance was narrower: the three Indonesian marketplaces against the eight foreign platforms chosen first. It points the same way. With one median per platform, the Mann–Whitney test rejects equal moves (p = 0.024). The shuffle of platform labels comes close without passing (p = 0.06). Three platforms give a test little power. In that sample, revenue growth and transaction value growth had a correlation of 0.10 in Indonesia and 0.98 abroad, excluding Sea. + +The direction of the moves is less telling than their size. Revenue grew faster than transaction value in 13 of the 15 Indonesian and regional pairs. In two of them, revenue and transaction value moved in opposite directions. Abroad, take rates also rose in about two thirds of pairs, but by much less. What sets the Indonesian and regional platforms apart is the size of the moves, which is what the test measures. + +Quarterly pairs do not reject H1 (p = 0.19), contrary to a prediction registered before the test was run. The quarterly Indonesian pairs mostly cover 2023 to 2026, after most of the repricing, while the large moves came between 2021 and 2024. The rejection of H1 describes that period. + +The thin share that Indonesian and regional platforms keep explains much of the threefold difference, an exploratory finding. Measured in percentage points, both groups changed their take rates by about 1 point a year. The difference between them is nowhere near large enough to count (p = 0.65). + +The two groups differ in the share they start from. The typical Indonesian and regional platform keeps about 1.6 percent of each sale, against about 17 percent abroad. As Section 3.2 showed, the same 1-point rise then has very different effects. It lifts revenue by more than half at the thinner share and by about 6 percent at the thicker one. Figure 5.1 shows the three measures side by side. The test shows that take rates moved, but not what the platforms changed to move them. + +![Figure 5.1](figures/fig5_1_thin_slice.png) + +**Figure 5.1. The thin share.** Each dot is one platform's median over its years. Left: the share of each sale that the platform keeps, on a log scale. Middle: the yearly change in that share, in percentage points. Right: the same change relative to the share itself, as a log change in percent. Black bars mark group medians. The foreign group has 26 platforms here and 27 in the main test, because one platform has no pair of usable years for the change in points. Source: company filings; author's calculations. Exploratory. diff --git a/papers/thesis_2026-10/drafts/ch6_draft_2026-10-14.md b/papers/thesis_2026-10/drafts/ch6_draft_2026-10-14.md new file mode 100644 index 0000000..aa37390 --- /dev/null +++ b/papers/thesis_2026-10/drafts/ch6_draft_2026-10-14.md @@ -0,0 +1,113 @@ +# Chapter 6. The Rules and How Far They Reach + +Indonesia's government is writing rules that act on the figures this thesis measures. In 2026 it capped a ride-hailing commission and limited discounts in online trade. Since October 2026, four marketplaces have collected a tax from their sellers. Platforms must also send data to BPS, and BPS has run a new economic census. Each rule acts on one part of what Chapter 5 measured, and each reaches only that part. + +The marketplace tax is the rule aimed most directly at online sellers. It can collect from at most about 1 in 20 online sellers and about a fifth of online sales value. It also gives the state a record of every seller on the four marketplaces, including those who pay nothing. This chapter measures what each rule can reach, using the data of Chapter 5. What each rule changes can be seen only after it takes effect, and Section 6.5 sets out the predictions that will be scored on those results. + +## 6.1 The rules and what each acts on + +The recent rules act on three parts of the figures: the platforms' take rate, the sales of sellers inside marketplaces, and BPS's coverage. Table 6.1 lists each rule with the part it acts on and the section of Chapter 5 that measures that part. + +**Table 6.1. Indonesia's recent rules and what each acts on** + +| Rule | What it does | What it acts on | Measured in | Status of the text | +|---|---|---|---|---| +| Commission cap (Transport Minister Decree KP-PHB 532/2026; Presidential Regulation 27/2026) | Caps the commission on two-wheel passenger trips at 8 percent from 1 July 2026 | The fee rate, for one service | 5.3 | Decree in force from July 2026; presidential regulation released 8 October 2026; neither text public when this thesis was written | +| Discount rule (Trade Minister Regulation 19/2026, Art. 18) | Requires every online trade operator, including ride-hailing apps that sell goods, to prevent repeated price subsidies that hurt local products and endless below-cost discounts | The incentive rate | 5.3 | Primary text read | +| Marketplace tax (Finance Minister Regulation 37/2025) | Designated marketplaces collect 0.5 percent of domestic sellers' gross sales; individuals with sales up to Rp500 million a year are exempt; every seller gives a tax number or national ID | Sellers' sales inside marketplaces | 5.9 | Primary text read; collected by Shopee, Tokopedia, Lazada and Blibli since 1 October 2026 | +| Small-business final tax (Government Regulation 20/2026) | Keeps a 0.5 percent final tax on turnover up to Rp4.8 billion a year; excludes content creators | Small sellers in every channel | Not measured | Primary text read | +| Platform data to BPS (BPS Regulation 4/2023) | E-commerce operators must send BPS their revenue, transactions and numbers of sellers and buyers, with sanctions | BPS's coverage | 5.10 | Primary text read | +| Economic census 2026 | Door-to-door count of all businesses, including online selling, affiliates and influencers | BPS's sampling frame | 5.7 | News reports; results not yet released | +| Ban on payments on social media (Trade Minister Regulation 31/2023, carried into Regulation 19/2026) | Bars social media platforms from handling payments for sales | Where online sales are paid | 5.9 | Primary text read | + +*Note:* At the 2024 rate, Rp500 million is about US$32,000 and Rp4.8 billion about US$300,000. Section 4.8 converts the exemption at the 2022 rate, about US$34,000, to match the survey year. Sources: the legal texts (BPS-Statistics Indonesia 2023; Government of the Republic of Indonesia 2026; Ministry of Finance of the Republic of Indonesia 2025; Ministry of Trade of the Republic of Indonesia 2023, 2026). The texts behind the commission cap were not public; its terms come from GoTo Gojek Tokopedia (2026) and CNN Indonesia (2026b). + +Equation (6) shows where the first two rules act. The take rate is the fee rate less the incentive rate. A cap on a fee pushes the take rate down. A limit on incentives pushes it up. The two rules act on opposite sides of the same take rate. + +The marketplace tax acts further out, on the sales of sellers inside marketplaces. It leaves the take rate alone, but it depends on the marketplaces' records of each sale. The platform-data rule and the census act on the national figure: the first on what BPS receives, the second on where BPS looks. + +Two rules serve as context. The first is the 2023 ban on payments on social media, the step with which Chapter 1 began. Selling through chat and social media remained the norm, and in 2023 about 96 percent of online sellers used them. The second is the small-business tax, which applies in every channel. The survey files cannot identify the content creators it excludes, so the thesis does not measure its reach. + +The sections that follow take the other rules in the order of Figure 3.1, from the take rate outward. + +## 6.2 What the commission cap and the discount rule can move + +The commission cap reaches a small part of the take rates this thesis measures. It followed protests by ride-hailing drivers, who since 2025 had asked that platforms keep at most 10 percent of each fare, down from 20 (Ronal 2025). From July 2026, a platform may keep at most 8 percent of the fare on a two-wheel ride. The cap does not cover car rides, food delivery or marketplaces. The presidential regulation released in October 2026 asks another ministry to set delivery fees by the end of the year. + +The trips under the cap are a small share of each company's business. At GoTo, the two-wheel ride service brings about 7 percent of group net revenue (GoTo Gojek Tokopedia 2026). At Grab, two-wheel rides in Indonesia are about 6 percent of the transaction value of its rides business across the region (Grab Holdings Limited 2026). That is about 2 percent of Grab's on-demand transaction value. Both companies report rides and deliveries as one on-demand take rate, so the cap acts on a slice of the figure Chapter 5 measured. + +The arithmetic of the cap shows how small that slice is. At Grab, the cut from 20 to 8 percent applies to trips worth about 2 percent of on-demand transaction value. If nothing else changes, it lowers the on-demand take rate by at most about a quarter of a point. Grab's on-demand take rate was about 13 percent in the second quarter of 2026. At GoTo, where two-wheel rides weigh more, the same arithmetic gives at most about 1.4 points on a take rate of about 22 percent. GoTo's own cost estimate implies about 0.9 points. In both cases the wedge grows by the same amount. The cap moves money from the platform's revenue to the drivers' fares, inside the same transaction value. + +The players around the cap have different stakes. Drivers asked for it, and the President announced the limit at a May Day rally in 2026, saying the cut "must be below 10 percent" (CNN Indonesia 2026b). The platforms lose part of the commission unless they recover it elsewhere, for example by cutting incentives. The cap brings the state no revenue; its tax on online sellers works through the marketplaces instead (Section 6.3). + +A cap on one fee does not fix the take rate, because the platform still sets its incentives. At GoTo, on-demand revenue rose by about 20 percent in the second quarter of 2026 while transaction value grew by 2 percent. The company attributed the rise to its mix of services and to "rationalization of incentives as we prepared for the implementation of the 8% commission rate" (GoTo Gojek Tokopedia 2026). Its take rate rose to about 22 percent in that quarter, the last before the cap, from about 18 percent a year earlier. + +The discount rule acts on the lever that moved take rates most often. Lower incentives were the larger part of the rise in the take rate in four of the seven windows of Table 5.4. The rule binds every online trade operator, including ride-hailing apps that sell food or goods. Whether it reaches ride fares, which the transport ministry sets, is unclear. + +Where incentives can be followed over time, much of that lever had been pulled before the rule came. Tokopedia, the only marketplace that reported its incentives, cut them by about a third in 2023, its last full year of separate reporting. At GoTo's on-demand business, incentives fell from about 11 percent of transaction value in 2023 to about 4 percent in 2025. Grab still spent about 10 percent of on-demand transaction value on incentives in the four quarters to early 2026, across the region. The thesis did not register a test of the discount rule, because its effect on the take rate cannot be told apart from changes in fees. + +Marketplace fees moved the other way. Shopee, Tokopedia and TikTok Shop added a fixed fee to every order in mid-2025 (Islamiati 2025). In 2026, Shopee and TikTok Shop, which took over Tokopedia in 2024, raised their seller fees again (CNN Indonesia 2026a; Setyowati 2026). The government declined to cap these fees. It later required marketplaces to halve the service fee for small sellers of local products (Ayudiana 2026). The rules on the take rate cap one fee on one ride-hailing service and leave the largest marketplace fees free. What the companies' take rates did after the cap is the subject of the predictions in Section 6.5. + +## 6.3 How far the marketplace tax reaches + +The marketplace tax can collect from few online sellers, but it gives the state a record of every seller on the four marketplaces that collect it. These marketplaces collect 0.5 percent of each domestic seller's gross sales and pass it to the tax office. Individual sellers with sales up to about US$34,000 a year are exempt (Section 4.8). Collection started in October 2026 through Shopee, Tokopedia, Lazada and Blibli (Directorate General of Taxes 2026c). It had begun once in August 2026, then stopped, and the tax collected was returned (Directorate General of Taxes 2026a, 2026b). + +At most about 1 in 20 online sellers falls under the tax. In 2022, marketplace sellers with sales above BPS's bracket line of about US$20,000 a year were about 5 percent of all online sellers. The exemption sits higher, at about US$34,000, so the true share is smaller. + +The value the tax reaches is larger, but it is still a minority of online sales. Marketplaces carried 23 to 26 percent of online sales value in 2022, and sellers above the bracket line held 77 to 94 percent of that. So the tax reaches at most about 18 to 24 percent of online sales value, about a fifth. Both figures use BPS's marketplace category, which also counts food and ride apps. Those apps are not among the four collectors, so both figures are upper limits for a second reason. + +The record reaches further than the tax. Every domestic seller on a designated marketplace must give the marketplace a tax number or national ID and an address before being paid. Exempt sellers file a statement as well. Tax research finds that records held by a third party improve compliance where they exist (Section 2.3). + +The record still covers a minority of sellers. Marketplace sellers were about 18 percent of online sellers in 2023. Only 47 percent of them used a shopping marketplace, and most of the rest sold only through Gojek or Grab. So the record covers at most about 8 percent of online sellers, including the many who pay nothing. + +The tax leaves most sellers where they were. Sellers off the marketplaces sell through chat, social media or their own websites, and they keep complete financial statements less often (Section 5.9). Few of them plan to move. In the survey for 2023, about 13 percent of sellers off the marketplaces wanted to join one. + +The revenue at stake is small. At 0.5 percent, the tax would raise about US$60 million a year on BPS's marketplace sales for 2024, before exemptions. On e-Conomy's estimate of e-commerce transaction value, about five times larger, it would raise about US$300 million. Even the larger amount is about 0.2 percent of the government's tax target for 2026. + +The tax adds a record more than revenue. Whether sellers move away from the four marketplaces now that collection has started will be known only from later data. + +## 6.4 What the statistics rules can change + +The rules on statistics act on the outer link, where the national figures parted in 2023. Since 2023, e-commerce operators must send BPS their revenue, transactions and numbers of sellers and buyers. Operators that do not comply face sanctions. By late 2024, 61 operators had sent data and 75 registered operators had not (Risalah 2024). BPS's publication for 2024 describes its figures as survey-based and does not mention the platform data. + +Full compliance would still leave most online selling to the survey. Platform data describe the marketplace slice, which holds about a fifth of online sales value and under a fifth of sellers (Section 5.9). The chat and social sellers, where the national figures part (Section 5.10), would still be counted only by BPS's survey. + +The economic census of 2026 acts on where the survey looks. BPS's e-commerce survey draws its areas from the economic census of 2016. The 2026 census visited homes and businesses from June to August 2026. It asked about online selling and counted affiliates and influencers. Its results had not been released when this thesis was written. + +A new frame can move the count of sellers without any change in selling. About half of the 2023 rise was not explained by first-time entry (Section 5.7). When BPS adopts a frame from the 2026 census, counts on either side of the change will need the same care. The entry test of Section 3.5 can be repeated on the first survey that uses the new frame. + +## 6.5 What the registered predictions test + +Nine predictions on the rules and on what they act on were registered before any result for the third quarter of 2026 existed. Eight concern the third-quarter results of GoTo, Grab and Sea. Six of them bear on the commission cap, one on marketplace fees and one on buying through platforms. The ninth concerns BPS's next figure for the marketplace share. Table 6.2 lists them with the date each is scored. + +**Table 6.2. Registered predictions on the 2026 rules** + +| Prediction | What it expects | Bears on | Scored | Outcome | +|---|---|---|---|---| +| P1 | GoTo's on-demand take rate falls by 0.3 to 2.5 points from the second quarter of 2026 | Commission cap | Late October 2026 | [to be filled] | +| P3 | GoTo's on-demand transaction value grows by 0 to 12 percent on a year earlier | Commission cap | Late October 2026 | [to be filled] | +| P5 | GoTo's on-demand transaction value grows by at least 5 percent; the prediction fails only below 2 percent | Commission cap, with spending on growth | Late October 2026 | [to be filled] | +| R1 | GoTo's on-demand revenue growth less transaction value growth lies between −0.5 and +12.5 points; it follows from P1 and P3 | Commission cap | Late October 2026 | [to be filled] | +| R2 | GoTo's on-demand transaction value grows more slowly than household spending | Buying through platforms | Early November 2026 | [to be filled] | +| P2 | Grab's on-demand take rate stays within −0.6 to +0.4 points of the second quarter | Commission cap | Mid-November 2026 | [to be filled] | +| P4 | Grab's change in the rides take rate, less its change in the deliveries take rate, is at most +0.3 points | Commission cap | Mid-November 2026 | [to be filled] | +| R3 | Shopee's revenue growth less its transaction value growth is at least +10 points | Marketplace fees | Mid-November 2026 | [to be filled] | +| Marketplace share | BPS's next marketplace share of online sales value is at most 15.79 percent | Reach of marketplace rules | Next BPS release | [to be filled] | + +*Note:* Ranges as registered; registration dates are in Appendix C. A tenth prediction, that the start of the marketplace tax would slip again, is withdrawn: collection had begun on 1 October 2026, a day before the prediction was registered. Sources for scoring: company results releases; BPS national accounts and e-commerce publication. + +The predictions describe company numbers, not the effect of the cap. Both companies report rides and deliveries together, so any effect of the cap is diluted in their figures. P1 and P5 are read together. If both hold, GoTo met the cap partly by spending more to grow. If P1 holds and P5 fails, GoTo let the lower commission show in its take rate without spending more to grow. P4 has limited power, because its range cannot separate the cap from ordinary quarter-to-quarter noise. R2 and R3 concern single companies, and Shopee's figures cover Asia and Brazil. + +> **[Results slot]** To be written after the GoTo results (late October), the BPS national accounts (early November) and the Grab and Sea results (mid-November). For each outcome: the result against its range in Table 6.2 and Appendix C, then one paragraph on what the outcomes show about the cap and one on marketplace fees. + +The outcomes will be reported whatever their direction, as the other registered tests are. + +## 6.6 Beyond Indonesia + +Other countries are building rules on platform records too, and the same question of reach applies to them. The OECD's model rules ask platforms to report what their sellers earn, and a later module added the sale of goods (OECD 2020, 2021). The European Union applies such rules through its DAC7 directive, with reports on activity from 2023 (European Union 2021). The Philippines has marketplaces and payment apps withhold tax from sellers' payouts, with small sellers exempt (Bureau of Internal Revenue 2023). Viet Nam has had platforms with a payment function withhold tax for household sellers since 2025 (Government of Viet Nam 2025). + +Indonesia's design goes further than reporting, because its marketplaces collect the tax as well as identify the seller. Every one of these rules reaches only the sales that pass through a platform. Where most selling runs through platforms, a rule built on their records reaches most sellers. Where it runs through chat and social media, as in Indonesia, it reaches a minority. + +Commission caps have been tried elsewhere as well. Fourteen US cities and states capped the commissions that delivery platforms charged independent restaurants. The platforms raised delivery fees for consumers and showed chain restaurants first more often. Independent restaurants lost orders and revenue, while chains gained (Li and Wang 2025). In the terms of equation (6), a cap on one fee left the platform free to change another. Indonesia's cap also covers one fee on one service. + +The thesis does not measure the split between channels outside Indonesia. Any statistics office that asks sellers which channels they use can make the same count, and with it measure how far its own rules reach. diff --git a/papers/thesis_2026-10/drafts/ch7_draft_2026-10-14.md b/papers/thesis_2026-10/drafts/ch7_draft_2026-10-14.md new file mode 100644 index 0000000..bb7e51e --- /dev/null +++ b/papers/thesis_2026-10/drafts/ch7_draft_2026-10-14.md @@ -0,0 +1,77 @@ +# Chapter 7. Conclusion + +Each figure of Indonesia's online economy is a fair guide to another only while a condition holds. Revenue follows transaction value while the take rate stays steady. A survey's count of sellers follows entry while the survey's reach stays steady. Between 2022 and 2024 the platforms' take rates moved, and revenue parted from the commerce it carried. In 2023 about half of the rise in BPS's count of online sellers was not explained by first-time entry. + +This chapter answers the research question at both links of Figure 3.1. It then says what the answer adds for those who use the figures, reports the predictions that failed, states the limits of the evidence and sets out the next steps. + +## 7.1 The answer to the research question + +The research question asked how much economic activity remains invisible when digital platforms are measured through their reported revenue, and when that choice changes the conclusion about growth. The answer to the first half is that most of the activity remains outside revenue. In 2023, three major platforms in Indonesia kept about US$3 billion of the US$43 billion that passed through them. The invisible wedge was about US$40 billion, and it moved almost one for one with the commerce. + +Platform commerce is itself only part of online commerce. In BPS's survey, marketplaces carry about a fifth of online sales value. Any measure built on platforms, whether revenue or transaction value, leaves out most online sales, which move through chat, social media and sellers' own websites. + +The answer to the second half differs by link. On the inner link, the choice of revenue changed the conclusion, above all from 2022 to 2024. The Indonesian and regional platforms changed their take rates about three times as much as platforms abroad, measured against the take rate itself, so H1 is rejected. They keep a much thinner share of each sale, so price changes of a similar size in points moved their revenue far more. In the windows traced, pricing moved revenue more than commerce did. + +On the outer link, H2 holds on BPS's published figures: more businesses account for most of the growth in online sales value. The registered entry test failed, and about half of the 2023 rise in the count is not explained by first-time entry. The extra count leaned toward long-running businesses. Read as wider coverage, an exploratory check, the extra count leaves H2 on its threshold. With those sellers and their sales taken out, more businesses account for about 42 to 54 percent of the growth, against the 50 percent H2 requires. In the same year, five of six measures of buying through platforms grew more slowly than household spending, and the median seller already online reported no change in sales. Reading BPS's 2023 figures as a boom in online selling overstates what sellers and platforms reported. + +Where the national figures cover the same sellers, they agree on small growth in two of three scenarios. They part outside the marketplaces, among the chat and social sellers that only BPS reaches. The figure a reader picks decides whether 2023 looks like a stall or a boom. Market value, for its part, moved with transaction value, and the data do not show that markets valued growth from pricing any less. + +## 7.2 What the thesis adds + +For investors, the thesis adds a check that can be run on any platform that reports both amounts. Revenue growth splits into more commerce and a larger take rate with two public figures (equation (5)). At the Indonesian and regional platforms, the take rate accounted for about two thirds of a revenue move at the typical pair. Abroad it accounted for about a quarter. The thinner the take rate, the more a small change in fees or incentives moves revenue growth. The same split shows when market value per dollar of revenue and per dollar of commerce will disagree. Across 28 listed platforms, market value moved with transaction value, as a registered prediction stated. A 1 percent rise in transaction value went with about 0.6 percent more market value. Earlier studies used web traffic as the measure of activity; here it is the transaction value that platforms report themselves. + +For statistics offices, the thesis adds a test of what a count of sellers represents. The entry test needs no linked records, only a question on when each business began selling online. Applied to BPS's survey, it showed that the 2023 rise in the count was only partly first-time entry. The same test can mark a break in any survey that asks the question. + +For tax authorities and regulators, the thesis measures how far rules built on platform records reach. In Indonesia, the marketplace tax can collect from at most about 1 in 20 online sellers and about a fifth of online sales value. It also gives the state a record of every seller on the four marketplaces that collect it, at most about 8 percent of online sellers. Each of the other recent rules acts on one part of the take rate or of the national figure, and Chapter 6 maps each to the part it reaches. + +For research, the thesis adds connected evidence for one country: the platforms' accounts, the sellers' own answers and the rules that act on both. It tracks the invisible wedge and growth divergence over time, and its ecosystem ratio is the author's own measure. All three use only figures that platforms already publish, so they can be computed for any platform that reports transaction value. + +## 7.3 Predictions that failed + +Several registered predictions failed, and the failures are part of the answer. The entry test is the clearest case. It required at least 21.5 percent of the sellers counted for 2023 to have begun that year, and the survey gave about 13 percent. That failure is the main result of the outer link. + +The test of H1 with quarterly pairs did not reject H1 (p = 0.19). The quarterly series cover mostly 2023 to 2026, after most of the repricing. So the rejection of H1 describes a period, 2022 to 2024, and not a lasting trait of Indonesian platforms. + +The second market-value prediction failed in the main sample (p = 0.20). It expected markets to value growth from pricing less than growth from more commerce. It held only in a second run that mixes segment figures with whole-company market values. Whether investors price the two sources of growth differently remains open. + +Three predictions on sellers' records and sales failed narrowly or in part. The value share of sellers without financial statements fell by 5.8 points between 2020 and 2022, beyond the registered limit of 5. On a second version it fell by only 3.3 points. The mean sales of unregistered firms that use social media were about US$7,300 a year, above the registered limit of about US$6,600, while their median was about US$4,000. The comparison with Bank Indonesia's figure held in two of three value scenarios and failed in the third. + +A registered rule on discounts gave only thin support. It expected large take-rate moves only where a platform's incentives were at least as large as its revenue. Out of sample, the rule pointed the right way, but it rested on two cases and failed at its main threshold. + +Appendix C lists every registered test with its outcome. The predictions on the 2026 rules are scored in Section 6.5. One of them, that the start of the marketplace tax would slip again, was withdrawn, because the tax had started a day before the prediction was registered. A comparison of seven countries on firms' digital records was withdrawn as a test for a similar reason. Published figures seen at registration already settled it, so it is reported as descriptive. + +## 7.4 Limits of the evidence + +The invisible wedge is the part of each sale that a platform passes on to sellers, drivers and others. It is not missing output, unrecorded income or unpaid tax. The thesis does not find BPS's figures wrong in general, nor that total online buying in Indonesia fell. It finds that one year's count of sellers moved for reasons other than first-time entry, and that the figures part where they cover different sellers. The splits of Table 5.4 are accounting, not causes. Marketplace sellers keep financial statements more often, but the thesis does not show that marketplaces lead them to. + +The platform evidence rests on a small sample. Three Indonesian marketplaces report both revenue and transaction value. The main test of H1 adds the on-demand businesses of Grab and GoTo, which mix rides and deliveries, and Grab's figures cover the whole region. That test was set up after a narrower first test, so it is exploratory. The tests of H1 are one-sided, because each asks whether the Indonesian and regional group moves more. A two-sided test would about double each p-value, and the main Mann–Whitney result would still be below 1 percent. + +The platform figures carry their own limits. Shopee's and Grab's Indonesian amounts rest on an outside estimate, an assumed take rate or a company-wide ratio, and they stay out of the main tests. Grab revised some figures between filings. The thesis takes each year from the latest filing that reports it, so one series can mix filings from before and after a revision. Blibli's incentives are a proxy, and Bukalapak's reports do not allow the parts of a revenue move to be separated. Company explanations of why revenue moved are reported as statements and are not tested. Foreign platforms also differ in how they book shopper discounts. Eight of the 27 record them mainly as marketing costs, not as lower revenue, which could steady their take rates. Their take rates were not steadier, and without them the test of H1 still gives p = 0.004, an exploratory check (Appendix B). + +The survey files describe sellers in one year, not businesses followed over time. No business identifier links the files, so the thesis cannot follow a seller from one round to the next, and returning sellers cannot be ruled out. Revenue appears only in brackets, and only in the 2021 and 2023 files, which describe 2020 and 2022, so value shares are ranges. The 2024 file, which describes 2023, has no revenue amounts. At the top, sellers with sales above about US$3.4 million a year rest on 19 sample rows in the survey for 2022. Yet they hold 18 to 39 percent of the marketplace value. + +Two changes in the questions also limit the survey checks. BPS changed how it asks about the months of selling between the rounds, which affects the check on part-year sellers. The education codes of the survey for 2023 were matched to the earlier groups by their shares, so education there is an inferred grouping. + +The accounting of the 2023 rise is exploratory, because it was designed after the files were opened. Its range depends on how many sellers stopped, and it assumes that no seller returned after a pause. Returners equal to 10 to 15 percent of the 2022 count would remove the unexplained part. The check on sampling noise applies one year's published error rates to both years and treats the two surveys as independent. The files carry no sampling-area or design variables, and the thesis did not contact BPS. So a change in BPS's frame cannot be told apart from other changes in coverage, and the break covers a single year. + +The national figures are read as published. Bank Indonesia does not publish its method. Momentum Works and e-Conomy publish estimates, not counts. Payment figures include payments made in person. The World Bank's survey of unregistered firms covers six cities, not the whole country. + +The market-value result is an association across 28 listed platforms. Market value also changes when a company issues shares, so the result is not a stock return. The thesis does not claim that transaction value causes market value to move. + +The rules are mapped, not evaluated. The legal texts behind the commission cap were not public when the thesis was written, so its terms come from official statements, news reports and GoTo's earnings call. The reach of the marketplace tax is an upper limit, because BPS's bracket line sits below the exemption and BPS's marketplace category also counts food and ride apps. The thesis measures what each rule can reach and makes no claim about its effects. + +The licensed files stay outside the repository, which holds the code and national totals (Section 4.10). The evidence concerns Indonesia, with regional series labelled as such. + +## 7.5 Next steps + +The first step is to score the registered predictions on the 2026 results, as Section 6.5 sets out. The outcomes will show what the companies' take rates did after the commission cap. Later data will show whether sellers moved away from the four marketplaces that collect the tax. + +Reconciling the levels of BPS's and the platforms' figures is the second step. It needs common definitions of what counts as a sale, including shipping, digital goods, travel and offline sales. The platform data that BPS now collects could supply the platform side, if BPS publishes them. + +The third step concerns the 2026 economic census. When BPS adopts a frame from it, the entry test can be repeated on the first survey that uses the new frame. A business identifier that links survey rounds would settle the question of returning sellers. + +The fourth step is a stock-return extension. A test of share prices around results announcements would ask whether investors react differently to growth from commerce and growth from pricing. It needs announcement timing and raw prices that can be reproduced. + +The platform sample can grow as well: Lazada and TikTok Shop can join the test of H1 once they report Indonesian figures. The last step takes the channel count abroad. Any country whose statistics office asks sellers which channels they use can measure how much online selling its platform rules reach. + +The thesis began with figures that moved apart in 2023. BPS's online sales rose by about 41 percent while Bank Indonesia's measure fell by about 5 percent, and Tokopedia's revenue rose while its transaction value fell. The figures disagreed because they measure different things. Revenue is what a platform keeps. Bank Indonesia's figure comes from large e-commerce firms, and BPS's survey counts the sellers it reaches. Each stops standing for another once a take rate or a survey's reach changes. A reader who checks which condition holds can use any of them. diff --git a/papers/thesis_2026-10/figures/fig3_1_two_links.png b/papers/thesis_2026-10/figures/fig3_1_two_links.png new file mode 100644 index 0000000..be07606 Binary files /dev/null and b/papers/thesis_2026-10/figures/fig3_1_two_links.png differ diff --git a/papers/thesis_2026-10/figures/fig5_1_thin_slice.png b/papers/thesis_2026-10/figures/fig5_1_thin_slice.png new file mode 100644 index 0000000..67100b1 Binary files /dev/null and b/papers/thesis_2026-10/figures/fig5_1_thin_slice.png differ diff --git a/papers/thesis_2026-10/figures/fig5_2_tokopedia_two_views.png b/papers/thesis_2026-10/figures/fig5_2_tokopedia_two_views.png new file mode 100644 index 0000000..ce6531e Binary files /dev/null and b/papers/thesis_2026-10/figures/fig5_2_tokopedia_two_views.png differ diff --git a/papers/thesis_2026-10/figures/fig5_3_growth_2023.png b/papers/thesis_2026-10/figures/fig5_3_growth_2023.png new file mode 100644 index 0000000..352d1ba Binary files /dev/null and b/papers/thesis_2026-10/figures/fig5_3_growth_2023.png differ diff --git a/papers/thesis_2026-10/figures/fig5_4_extra_sellers.png b/papers/thesis_2026-10/figures/fig5_4_extra_sellers.png new file mode 100644 index 0000000..ab6a254 Binary files /dev/null and b/papers/thesis_2026-10/figures/fig5_4_extra_sellers.png differ diff --git a/papers/thesis_2026-10/figures/fig5_5_marketplace_and_records.png b/papers/thesis_2026-10/figures/fig5_5_marketplace_and_records.png new file mode 100644 index 0000000..a8254d1 Binary files /dev/null and b/papers/thesis_2026-10/figures/fig5_5_marketplace_and_records.png differ diff --git a/papers/thesis_2026-10/figures/make_fig3_1.py b/papers/thesis_2026-10/figures/make_fig3_1.py new file mode 100644 index 0000000..1eb2116 --- /dev/null +++ b/papers/thesis_2026-10/figures/make_fig3_1.py @@ -0,0 +1,37 @@ +"""Figure 3.1 of the thesis (draft, 15 Oct 2026): the three nested quantities and the two links tested by H1 and H2. +Illustrative only, not to scale. Same palette and ink as make_figures.py. Run from the repository root: +python3 papers/thesis_2026-10/figures/make_fig3_1.py""" +import pathlib +import matplotlib +matplotlib.use("Agg") +import matplotlib.pyplot as plt +from matplotlib.patches import Circle, FancyArrowPatch + +OUT = pathlib.Path(__file__).resolve().parent +BLUE, ORANGE, INK, INK2 = "#2a78d6", "#eb6834", "#0b0b0b", "#52514e" +plt.rcParams.update({"font.size": 9, "font.family": "DejaVu Sans", "savefig.dpi": 300}) + +fig, ax = plt.subplots(figsize=(7.2, 3.9)) +ax.set_xlim(0, 13.2); ax.set_ylim(0, 7); ax.set_aspect("equal"); ax.axis("off") +# circles share a bottom point so each sits inside the next +base = 0.4 +for r, face, edge, lw in [(3.2, "#eef4fc", BLUE, 1.6), (2.0, "#dbe8fa", BLUE, 1.6), (0.9, "#fbe3d9", ORANGE, 1.8)]: + ax.add_patch(Circle((3.6, base + r), r, facecolor=face, edgecolor=edge, linewidth=lw)) +ax.text(3.6, base + 2 * 3.2 - 0.45, "All online commerce", ha="center", va="top", fontsize=10, fontweight="bold", color=INK) +ax.text(3.6, base + 2 * 3.2 - 0.85, "every channel, including chat\nand social media", ha="center", va="top", fontsize=8, color=INK2) +ax.text(3.6, base + 2 * 2.0 - 0.35, "Platform commerce", ha="center", va="top", fontsize=10, fontweight="bold", color=INK) +ax.text(3.6, base + 2 * 2.0 - 0.72, "transaction value, V", ha="center", va="top", fontsize=8, color=INK2) +ax.text(3.6, base + 0.9 + 0.12, "Revenue", ha="center", va="center", fontsize=9.5, fontweight="bold", color=INK) +ax.text(3.6, base + 0.9 - 0.3, "R", ha="center", va="center", fontsize=8, color=INK2) +ax.text(3.6, base + 2.45, "invisible wedge W = V − R", ha="center", va="center", fontsize=8, color=INK2, style="italic") + +def link(y_from, y_to, label, sub, y_text): + ax.add_patch(FancyArrowPatch((7.3, y_from), (7.3, y_to), arrowstyle="<->", mutation_scale=12, color=INK, lw=1.2)) + ax.plot([3.8, 7.3], [y_from, y_from], color=INK2, lw=0.6, ls=":") + ax.plot([3.8, 7.3], [y_to, y_to], color=INK2, lw=0.6, ls=":") + ax.text(7.55, y_text + 0.2, label, ha="left", va="center", fontsize=10, fontweight="bold", color=INK) + ax.text(7.55, y_text - 0.3, sub, ha="left", va="center", fontsize=8, color=INK2) + +link(base + 1.8, base + 4.0, "H1, the inner link", "Does revenue grow in proportion\nto transaction value?", base + 2.75) +link(base + 4.0, base + 6.4, "H2, the outer link", "Does national online growth come\nmainly from more businesses?", base + 5.2) +fig.savefig(OUT / "fig3_1_two_links.png", bbox_inches="tight", facecolor="white"); plt.close(fig) diff --git a/papers/thesis_2026-10/figures/make_figures.py b/papers/thesis_2026-10/figures/make_figures.py new file mode 100644 index 0000000..f98b8c8 --- /dev/null +++ b/papers/thesis_2026-10/figures/make_figures.py @@ -0,0 +1,125 @@ +"""Figures for Chapter 5 of the thesis (13 Oct 2026). Built only from committed tables; no licensed data are read. +Palette: reference categorical slots 1 (blue #2a78d6) and 2 (orange #eb6834), validated on a white surface (CVD dE 24.7, normal 33.6, contrast >= 3:1), +plus neutral grey. Colour is not fixed to one entity across figures; inside each figure a legend or a label names every mark, so colour is never the only +cue. Text is neutral ink. Run from the repository root: python3 papers/thesis_2026-10/figures/make_figures.py""" +import json, pathlib +import matplotlib +matplotlib.use("Agg") +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd + +ROOT = pathlib.Path(__file__).resolve().parents[3] +T = ROOT / "research/option_space_2026-10-01/tables" +OUT = pathlib.Path(__file__).resolve().parent +BLUE, ORANGE, GREY, INK, INK2, GRID = "#2a78d6", "#eb6834", "#a3a29d", "#0b0b0b", "#52514e", "#e4e3df" +plt.rcParams.update({"font.size": 9, "font.family": "DejaVu Sans", "axes.edgecolor": GRID, "axes.labelcolor": INK2, "xtick.color": INK2, + "ytick.color": INK, "axes.spines.top": False, "axes.spines.right": False, "axes.spines.left": False, + "figure.facecolor": "white", "axes.facecolor": "white", "axes.grid": True, "grid.color": GRID, "grid.linewidth": 0.6, + "axes.axisbelow": True, "savefig.dpi": 300}) + +def save(fig, name): + fig.savefig(OUT / name, bbox_inches="tight", facecolor="white"); plt.close(fig) + +# ---- Figure 5.1: the thin slice ------------------------------------------------------------------------------------------------------- +f = pd.read_csv(T / "h1_points_vs_log_firms.csv") +f["change_pct"] = 100 * f.change_log +rng = np.random.default_rng(7) +fig, axes = plt.subplots(1, 3, figsize=(7.2, 2.7), sharey=True) +panels = [("share_kept_pct", "What the platform keeps\n(percent of each sale)", True), ("change_points", "Yearly change in what it keeps\n(percentage points)", False), + ("change_pct", "Yearly change relative to\nwhat it keeps (percent)", False)] +rows = [("Indonesian and regional (5)", f[f.group == "Indonesia"], ORANGE, 1), ("Foreign (26)", f[f.group == "Foreign"], BLUE, 0)] +for ax, (col, lab, logx) in zip(axes, panels): + for name, d, c, y in rows: + jit = y + rng.uniform(-0.16, 0.16, len(d)) + ax.scatter(d[col], jit, s=22, color=c, edgecolor="white", linewidth=0.8, zorder=3) + m = d[col].median() + ax.plot([m, m], [y - 0.32, y + 0.32], color=INK, lw=2, zorder=4) + ax.text(m, y + 0.38, f"median {m:.1f}", ha="center", va="bottom", fontsize=8, color=INK) + if logx: ax.set_xscale("log"); ax.set_xticks([1, 3, 10, 30, 100]); ax.set_xticklabels(["1", "3", "10", "30", "100"]) + ax.set_xlabel(lab); ax.set_ylim(-0.6, 1.8); ax.set_yticks([0, 1]); ax.tick_params(axis="y", length=0); ax.grid(axis="y", visible=False) +axes[0].set_yticklabels(["Foreign (26)", "Indonesian and\nregional (5)"]) +fig.tight_layout(); save(fig, "fig5_1_thin_slice.png") + +# ---- Figure 5.2: Tokopedia, two ways to split the rise --------------------------------------------------------------------------------- +s = pd.read_csv(T / "tokopedia_segment.csv", index_col=0) +dg = s.loc[2023] - s.loc[2022] +inc_share, gross_share = -dg.incentives / dg.net * 100, dg.gross / dg.net * 100 +g = pd.read_csv(T / "growth_decomposition.csv"); g = g[(g.firm == "Tokopedia")].iloc[0] +vol, fee, dis = 100 * g.share_from_volume, 100 * g.share_from_fee, 100 * g.share_from_discount_cut +fig, (a, b) = plt.subplots(1, 2, figsize=(7.2, 2.5), gridspec_kw={"width_ratios": [1, 1.25]}) +a.barh([1, 0], [inc_share, gross_share], color=[ORANGE, BLUE], height=0.5, zorder=3) +for y, v in zip([1, 0], [inc_share, gross_share]): a.text(v + 1.5, y, f"{v:.0f}", va="center", color=INK) +a.set_yticks([1, 0]); a.set_yticklabels(["Lower incentives", "Higher gross fees"]); a.set_xlim(0, 80); a.set_xlabel("Share of the extra net revenue (percent)") +a.set_title("By lines of the accounts", fontsize=9, color=INK, loc="left", fontweight="bold") +cols = [GREY, BLUE, ORANGE]; vals = [vol, fee, dis]; labs = ["More commerce", "Higher fees", "Lower incentives"] +b.barh([2, 1, 0], vals, color=cols, height=0.5, zorder=3) +for y, v in zip([2, 1, 0], vals): b.text(v + (2 if v > 0 else -2), y, f"{v:.0f}".replace("-", "−"), va="center", ha="left" if v > 0 else "right", color=INK) +b.axvline(0, color=INK2, lw=0.8, zorder=4); b.set_yticks([2, 1, 0]); b.set_yticklabels(labs); b.set_xlim(-40, 95) +b.set_xlabel("Share of the growth of revenue (percent)"); b.set_title("Per unit of transaction value", fontsize=9, color=INK, loc="left", fontweight="bold") +for ax in (a, b): ax.grid(axis="y", visible=False); ax.tick_params(axis="y", length=0) +fig.tight_layout(); save(fig, "fig5_2_tokopedia_two_views.png") + +# ---- Figure 5.3: 2023 growth, each measure in its own unit ---------------------------------------------------------------------------- +# Values come from committed tables: consistency_grid_growth.csv (BPS, Bank Indonesia, GoTo, Momentum Works, household spending), wedge_over_time.csv +# (Tokopedia, Bukalapak, Blibli) and bps_descriptives.json (published seller count). Momentum Works' 3.5 is the grid's calculation from published levels. +grid = pd.read_csv(T / "consistency_grid_growth.csv", index_col=0)["2023"] +wot = pd.read_csv(T / "wedge_over_time.csv"); wot = wot[wot.year == 2023].set_index("firm") +sellers = json.load(open(T / "bps_descriptives.json"))["published_count_growth_pct"] +items = [("Household spending", grid["Official household spending (nominal)"], GREY, "Reference"), + ("Tokopedia transaction value", wot.loc["Tokopedia e-commerce segment", "g_tv_pct"], BLUE, "Buying through platforms"), + ("GoTo on-demand transaction value", grid["GoTo on-demand GTV"], BLUE, "Buying through platforms"), + ("Bank Indonesia e-commerce", grid["Bank Indonesia e-commerce"], BLUE, "Buying through platforms"), + ("Momentum Works marketplace estimate", grid["Momentum Works Indonesia GMV"], BLUE, "Buying through platforms"), + ("Bukalapak transaction value", wot.loc["Bukalapak Group", "g_tv_pct"], BLUE, "Buying through platforms"), + ("Blibli third-party retail transaction value", wot.loc["Blibli 3P Retail", "g_tv_pct"], BLUE, "Buying through platforms"), + ("Tokopedia revenue", wot.loc["Tokopedia e-commerce segment", "g_revenue_pct"], ORANGE, "Figures that described a boom"), + ("BPS e-commerce value", grid["BPS e-commerce"], ORANGE, "Figures that described a boom"), + ("BPS online sellers (count)", sellers, ORANGE, "Figures that described a boom")] +order = [i for i in items if i[3] == "Figures that described a boom"] + [items[0]] + sorted([i for i in items if i[3] == "Buying through platforms"], key=lambda t: -t[1]) +order = sorted(order[:3], key=lambda t: -t[1]) + order[3:] +fig, ax = plt.subplots(figsize=(7.2, 3.9)) +ys = np.arange(len(order))[::-1] +ax.barh(ys, [o[1] for o in order], color=[o[2] for o in order], height=0.62, zorder=3) +for y, o in zip(ys, order): ax.text(o[1] + (1.0 if o[1] >= 0 else -1.0), y, f"{o[1]:+.1f}".replace("-", "−"), va="center", ha="left" if o[1] >= 0 else "right", color=INK, fontsize=8.5) +ax.axvline(0, color=INK2, lw=0.9, zorder=4); ax.set_yticks(ys); ax.set_yticklabels([o[0] for o in order]); ax.tick_params(axis="y", length=0) +ax.set_xlim(-24, 62); ax.set_xlabel("Growth in 2023 (percent); each measure in its own unit"); ax.grid(axis="y", visible=False) +from matplotlib.patches import Patch +ax.legend(handles=[Patch(color=ORANGE, label="Figures that described a boom"), Patch(color=GREY, label="Reference"), Patch(color=BLUE, label="Buying through platforms")], + loc="lower right", frameon=False, fontsize=8.5) +fig.tight_layout(); save(fig, "fig5_3_growth_2023.png") + +# ---- Figure 5.4: who the extra sellers were ------------------------------------------------------------------------------------------ +p = json.load(open(T / "bps_newly_counted_profile.json"))["profile"] +sec = p["sector"] +groups = [("Business opened before 2010", p["business opened"]["before 2010"]["share_of_2022_count_pct"], p["business opened"]["before 2010"]["share_of_beyond_normal_pct"]), + ("Owner aged 50 or over", p["owner age"]["50 and over"]["share_of_2022_count_pct"], p["owner age"]["50 and over"]["share_of_beyond_normal_pct"]), + ("Outside Java", p["island"]["outside Java"]["share_of_2022_count_pct"], p["island"]["outside Java"]["share_of_beyond_normal_pct"]), + ("Food and lodging, or making goods", sec["food and lodging"]["share_of_2022_count_pct"] + sec["making goods"]["share_of_2022_count_pct"], + sec["food and lodging"]["share_of_beyond_normal_pct"] + sec["making goods"]["share_of_beyond_normal_pct"])] +fig, ax = plt.subplots(figsize=(7.2, 2.9)); y0 = np.arange(len(groups))[::-1]; h = 0.34 +ax.barh(y0 + h / 2, [g_[2] for g_ in groups], height=h, color=ORANGE, zorder=3, label="Share of the extra sellers counted in 2023") +ax.barh(y0 - h / 2, [g_[1] for g_ in groups], height=h, color=BLUE, zorder=3, label="Share of all sellers in the 2022 count") +for y, g_ in zip(y0, groups): + ax.text(g_[2] + 1, y + h / 2, f"{g_[2]:.1f}", va="center", color=INK, fontsize=8.5); ax.text(g_[1] + 1, y - h / 2, f"{g_[1]:.1f}", va="center", color=INK, fontsize=8.5) +ax.set_yticks(y0); ax.set_yticklabels([g_[0] for g_ in groups]); ax.tick_params(axis="y", length=0); ax.grid(axis="y", visible=False) +ax.set_xlim(0, 70); ax.set_xlabel("Percent"); ax.legend(loc="lower left", bbox_to_anchor=(-0.02, 1.0), ncol=1, frameon=False, fontsize=8.5) +fig.tight_layout(); save(fig, "fig5_4_extra_sellers.png") + +# ---- Figure 5.5: marketplace use and financial statements -------------------------------------------------------------------------------- +d = json.load(open(T / "bps_descriptives.json")) +yrs_m = [2020, 2022, 2023, 2024]; mk = [d["marketplace_share"]["2020"], d["marketplace_share"]["2022"], d["marketplace_share"]["2023"], d["marketplace_share"]["2024_published"]] +yrs_b = [2020, 2022, 2023]; bk = [d["books_share"]["2020"], d["books_share"]["2022"], d["books_share"]["2023"]] +fig, ax = plt.subplots(figsize=(6.2, 2.9)) +ax.plot(yrs_m, mk, color=BLUE, lw=2, marker="o", ms=7, mec="white", mew=1.5, zorder=3, label="Sellers who use a marketplace") +ax.plot(yrs_b, bk, color=ORANGE, lw=2, marker="o", ms=7, mec="white", mew=1.5, zorder=3, label="Sellers who keep complete financial statements") +for x, v in zip(yrs_m, mk): + up = x == 2023 + ax.text(x, v + (0.8 if up else -0.8), f"{v:.1f}", ha="center", va="bottom" if up else "top", color=INK, fontsize=8.5) +for x, v in zip(yrs_b, bk): + down = x == 2023 + ax.text(x, v + (-0.8 if down else 0.8), f"{v:.1f}", ha="center", va="top" if down else "bottom", color=INK, fontsize=8.5) +ax.set_xticks(yrs_m); ax.set_ylim(10, 28); ax.set_ylabel("Percent of online sellers"); ax.legend(loc="upper right", frameon=False, fontsize=8.5) +ax.grid(axis="x", visible=False) +fig.tight_layout(); save(fig, "fig5_5_marketplace_and_records.png") +print("ok") diff --git a/papers/thesis_2026-10/reading/.gitignore b/papers/thesis_2026-10/reading/.gitignore new file mode 100644 index 0000000..f373c13 --- /dev/null +++ b/papers/thesis_2026-10/reading/.gitignore @@ -0,0 +1 @@ +reading.tmp.html diff --git a/papers/thesis_2026-10/reading/build_reading.sh b/papers/thesis_2026-10/reading/build_reading.sh new file mode 100644 index 0000000..4609bb0 --- /dev/null +++ b/papers/thesis_2026-10/reading/build_reading.sh @@ -0,0 +1,23 @@ +#!/usr/bin/env bash +# Rebuilds the reading copy from the chapter files: a typeset PDF (HTML + print CSS, rendered by Chromium) and a Word file +# for editing. Run from anywhere: bash papers/thesis_2026-10/reading/build_reading.sh +# Needs pandoc, python3, node with Playwright (preinstalled Chromium). +set -euo pipefail +cd "$(dirname "$0")" +D=$(date +%Y-%m-%d) +FILES="abstract.md symbols.md ch1_introduction.md ch2_literature.md ch3_framework.md ch4_data_methods.md ch5_results.md drafts/ch6_draft_2026-10-14.md drafts/ch7_draft_2026-10-14.md appendix_notes.md references.md" +{ + printf -- '---\ntitle: "The Invisible Ledger: Quantifying the Invisible Wedge in Indonesia'"'"'s Platform Economy"\nsubtitle: "Master'"'"'s thesis, reading copy. Chapters 6 and 7 are drafts, and Section 6.5 waits for results due from 27 October 2026."\nauthor:\n - "Christopher Ongko (王新福)"\n - "Advisor: Prof. De-Rong Kong (孔德蓉)"\n - "College of Management, Yuan Ze University"\ndate: "Draft of %s"\nlang: en\n---\n\n' "$D" + python3 prep.py $(for f in $FILES; do printf '../%s ' "$f"; done) +} > reading.tmp.md +pandoc reading.tmp.md --from markdown --to html5 --standalone --toc --toc-depth=2 --section-divs --css style.css --embed-resources \ + --resource-path=.. --metadata title-prefix="" -o reading.tmp.html +python3 post.py reading.tmp.html +node print_pdf.js reading.tmp.html first_pass.tmp.pdf +python3 toc_pages.py first_pass.tmp.pdf reading.tmp.html +rm -f first_pass.tmp.pdf +node print_pdf.js reading.tmp.html thesis_reading_copy_$D.pdf +{ printf -- '---\ntitle: "The Invisible Ledger (editing copy)"\nsubtitle: "Draft of %s"\n---\n\n' "$D"; for f in $FILES; do sed -E 's/^!\[[^]]*\]\(/![](/' "../$f"; printf '\n\n'; done; } > edit.tmp.md +pandoc edit.tmp.md --resource-path=.. -o thesis_editing_copy_$D.docx --toc --toc-depth=2 +rm -f reading.tmp.md edit.tmp.md +echo "built thesis_reading_copy_$D.pdf, thesis_editing_copy_$D.docx (HTML kept as reading.tmp.html for checking)" diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-400-italic.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-400-italic.woff2 new file mode 100644 index 0000000..0419209 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-400-italic.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-400-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-400-normal.woff2 new file mode 100644 index 0000000..f3704cf Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-400-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-600-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-600-normal.woff2 new file mode 100644 index 0000000..03abaed Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-600-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-700-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-700-normal.woff2 new file mode 100644 index 0000000..723f20a Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-greek-700-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-400-italic.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-400-italic.woff2 new file mode 100644 index 0000000..4d96a37 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-400-italic.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-400-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-400-normal.woff2 new file mode 100644 index 0000000..bab3088 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-400-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-600-italic.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-600-italic.woff2 new file mode 100644 index 0000000..8b9f95c Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-600-italic.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-600-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-600-normal.woff2 new file mode 100644 index 0000000..360cbb2 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-600-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-700-italic.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-700-italic.woff2 new file mode 100644 index 0000000..f65ba72 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-700-italic.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-700-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-700-normal.woff2 new file mode 100644 index 0000000..70dd991 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-700-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-400-italic.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-400-italic.woff2 new file mode 100644 index 0000000..2d111b7 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-400-italic.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-400-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-400-normal.woff2 new file mode 100644 index 0000000..fda5610 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-400-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-600-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-600-normal.woff2 new file mode 100644 index 0000000..ab91e95 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-600-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-700-normal.woff2 b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-700-normal.woff2 new file mode 100644 index 0000000..b6d22d9 Binary files /dev/null and b/papers/thesis_2026-10/reading/fonts/source-serif-4-latin-ext-700-normal.woff2 differ diff --git a/papers/thesis_2026-10/reading/post.py b/papers/thesis_2026-10/reading/post.py new file mode 100644 index 0000000..33baddb --- /dev/null +++ b/papers/thesis_2026-10/reading/post.py @@ -0,0 +1,12 @@ +"""Post-processes the pandoc HTML before printing: keeps each figure on one page with its caption and note, +and keeps short tables (10 body rows or fewer, under about 1,500 characters) from splitting across pages. Edits the file in place.""" +import re, sys +p = sys.argv[1] +h = open(p, encoding="utf-8").read() +h = re.sub(r'(

]*/>

\s*

.*?

(\s*

.*?

)?)', + r'
\1
', h, flags=re.S) +def keep(m): + t = m.group(0) + return t.replace("]+>", "", t)) < 1500 else t +h = re.sub(r"", keep, h, flags=re.S) +open(p, "w", encoding="utf-8").write(h) diff --git a/papers/thesis_2026-10/reading/prep.py b/papers/thesis_2026-10/reading/prep.py new file mode 100644 index 0000000..e640cc5 --- /dev/null +++ b/papers/thesis_2026-10/reading/prep.py @@ -0,0 +1,21 @@ +"""Turns the chapter Markdown into print-ready Markdown: equations as centred blocks with real subscripts, caption and note +paragraphs tagged for styling, figure alt text dropped so captions print once. Reads files named on the command line, writes stdout.""" +import re, sys, html +def sub(expr): + expr = html.escape(expr) + return re.sub(r"_((?:i,t−1)|[A-Za-z0-9]+)", r"\1", expr) +def eq(m): + body, no = m.group(1).rstrip(), m.group(2) + return f'
{sub(body)}{no}
' +out = [] +for f in sys.argv[1:]: + t = open(f, encoding="utf-8").read() + t = re.sub(r"```\n(.+?)\s+(\(\d+\))\s*\n```", eq, t) + t = re.sub(r"^!\[[^]]*\]\(", "![](", t, flags=re.M) + t = t.replace("β_V", "βV").replace("β_m", "βm") + t = re.sub(r"(?\2", t) + t = re.sub(r"^\*\*(Table [0-9A-Z.]+\.)\s*(.*?)\*\*\s*$", r'

\1 \2

', t, flags=re.M) + t = re.sub(r"^\*\*(Figure [0-9.]+\.[^*]*)\*\*(.*)$", lambda m: f'

{m.group(1)}{m.group(2)}

', t, flags=re.M) + t = re.sub(r"^\*Note:\*(.*)$", lambda m: f'

Note:{m.group(1)}

', t, flags=re.M) + out.append(t.rstrip() + "\n\n") +sys.stdout.write("".join(out)) diff --git a/papers/thesis_2026-10/reading/print_pdf.js b/papers/thesis_2026-10/reading/print_pdf.js new file mode 100644 index 0000000..aef89b1 --- /dev/null +++ b/papers/thesis_2026-10/reading/print_pdf.js @@ -0,0 +1,13 @@ +// Renders the reading-copy HTML to PDF with page numbers. Usage: node print_pdf.js in.html out.pdf +const { chromium } = require('/opt/node-tools/node_modules/playwright'); +(async () => { + const [inp, out] = process.argv.slice(2); + const browser = await chromium.launch(); + const page = await browser.newPage(); + await page.goto('file://' + require('path').resolve(inp), { waitUntil: 'networkidle' }); + await page.pdf({ path: out, format: 'A4', printBackground: true, preferCSSPageSize: true, displayHeaderFooter: true, + headerTemplate: '
', + footerTemplate: '
', + margin: { top: '22mm', bottom: '20mm', left: '21mm', right: '21mm' } }); + await browser.close(); +})(); diff --git a/papers/thesis_2026-10/reading/reference_compact.docx b/papers/thesis_2026-10/reading/reference_compact.docx new file mode 100644 index 0000000..54f78a6 Binary files /dev/null and b/papers/thesis_2026-10/reading/reference_compact.docx differ diff --git a/papers/thesis_2026-10/reading/style.css b/papers/thesis_2026-10/reading/style.css new file mode 100644 index 0000000..15f5622 --- /dev/null +++ b/papers/thesis_2026-10/reading/style.css @@ -0,0 +1,98 @@ +/* Print styling for the reading copy: one serif face (Source Serif 4, bundled in fonts/, SIL Open Font License), black on white, + three-line tables, captions above tables and below figures as YZU asks. A4. */ +@font-face { font-family: "Thesis Serif"; font-style: italic; font-weight: 400; src: url(fonts/source-serif-4-greek-400-italic.woff2) format("woff2"); unicode-range: U+0370-0377,U+037A-037F,U+0384-038A,U+038C,U+038E-03A1,U+03A3-03FF; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 400; src: url(fonts/source-serif-4-greek-400-normal.woff2) format("woff2"); unicode-range: U+0370-0377,U+037A-037F,U+0384-038A,U+038C,U+038E-03A1,U+03A3-03FF; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 600; src: url(fonts/source-serif-4-greek-600-normal.woff2) format("woff2"); unicode-range: U+0370-0377,U+037A-037F,U+0384-038A,U+038C,U+038E-03A1,U+03A3-03FF; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 700; src: url(fonts/source-serif-4-greek-700-normal.woff2) format("woff2"); unicode-range: U+0370-0377,U+037A-037F,U+0384-038A,U+038C,U+038E-03A1,U+03A3-03FF; } +@font-face { font-family: "Thesis Serif"; font-style: italic; font-weight: 400; src: url(fonts/source-serif-4-latin-400-italic.woff2) format("woff2"); unicode-range: U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+2074,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 400; src: url(fonts/source-serif-4-latin-400-normal.woff2) format("woff2"); unicode-range: U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+2074,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD; } +@font-face { font-family: "Thesis Serif"; font-style: italic; font-weight: 600; src: url(fonts/source-serif-4-latin-600-italic.woff2) format("woff2"); unicode-range: U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+2074,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 600; src: url(fonts/source-serif-4-latin-600-normal.woff2) format("woff2"); unicode-range: U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+2074,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD; } +@font-face { font-family: "Thesis Serif"; font-style: italic; font-weight: 700; src: url(fonts/source-serif-4-latin-700-italic.woff2) format("woff2"); unicode-range: U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+2074,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 700; src: url(fonts/source-serif-4-latin-700-normal.woff2) format("woff2"); unicode-range: U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+2074,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD; } +@font-face { font-family: "Thesis Serif"; font-style: italic; font-weight: 400; src: url(fonts/source-serif-4-latin-ext-400-italic.woff2) format("woff2"); unicode-range: U+0100-02AF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 400; src: url(fonts/source-serif-4-latin-ext-400-normal.woff2) format("woff2"); unicode-range: U+0100-02AF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 600; src: url(fonts/source-serif-4-latin-ext-600-normal.woff2) format("woff2"); unicode-range: U+0100-02AF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF; } +@font-face { font-family: "Thesis Serif"; font-style: normal; font-weight: 700; src: url(fonts/source-serif-4-latin-ext-700-normal.woff2) format("woff2"); unicode-range: U+0100-02AF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF; } + +@page { size: A4; margin: 24mm 23mm 21mm 23mm; } +html { font-size: 10.5pt; } +body { font-family: "Thesis Serif", "DejaVu Serif", serif; color: #111; background: #fff; line-height: 1.44; + margin: 0; padding: 0; max-width: none; hyphens: manual; text-rendering: optimizeLegibility; } +p { margin: 0 0 0.55em; orphans: 3; widows: 3; } +a { color: inherit; text-decoration: none; } +strong { font-weight: 600; } + +/* title page */ +header#title-block-header { break-after: page; text-align: center; padding-top: 60mm; } +header#title-block-header .title { font-size: 18pt; font-weight: 700; line-height: 1.3; margin: 0 0 8mm; } +header#title-block-header .subtitle { font-size: 10.5pt; font-style: italic; margin: 0 0 30mm; } +header#title-block-header .author { font-size: 11pt; margin: 0 0 1.5mm; } +header#title-block-header .date { font-size: 10pt; margin-top: 20mm; } + +/* contents, with page numbers added in a second pass (toc_pages.py) */ +nav#TOC { break-after: page; font-size: 10.5pt; } +nav#TOC::before { content: "Contents"; display: block; text-align: center; font-weight: 700; font-size: 16pt; margin: 0 0 10mm; } +nav#TOC ul { list-style: none; margin: 0; padding: 0; } +nav#TOC > ul > li { margin: 3mm 0 0; } +nav#TOC > ul > li > a { font-weight: 600; } +nav#TOC ul ul { padding-left: 8mm; } +nav#TOC ul ul li { margin: 0.4mm 0; } +nav#TOC a { display: flex; align-items: baseline; } +nav#TOC a .lead { flex: 1; border-bottom: 1px dotted #999; margin: 0 2mm; transform: translateY(-0.25em); } +nav#TOC a .pg { font-weight: 400; min-width: 6mm; text-align: right; } + +/* headings */ +h1, h2, h3 { font-family: inherit; color: #111; break-after: avoid; page-break-after: avoid; } +h1 { font-size: 16pt; font-weight: 700; text-align: center; line-height: 1.3; margin: 0 0 9mm; padding-top: 6mm; break-before: page; } +h2 { font-size: 12pt; font-weight: 700; margin: 6mm 0 2mm; } +h3 { font-size: 10.5pt; font-weight: 700; font-style: italic; margin: 4.5mm 0 1.5mm; } + +/* captions and notes */ +p.tcap { font-size: 10pt; margin: 5mm 0 2mm; break-after: avoid; line-height: 1.4; } +p.tcap .lbl, p.fcap strong { font-weight: 700; } +p.fcap { font-size: 9.5pt; margin: 2mm 0 5mm; line-height: 1.4; } +p.note { font-size: 8.8pt; color: #333; line-height: 1.4; margin: 1.5mm 0 5mm; } + +/* tables: three rules, no shading */ +table { width: 100%; border-collapse: collapse; table-layout: auto; font-size: 9pt; line-height: 1.35; margin: 0 0 1mm; + border-top: 1.2px solid #111; border-bottom: 1.2px solid #111; } +col { width: auto !important; } +thead { display: table-header-group; } +th { text-align: left; font-weight: 600; padding: 1.4mm 2mm 1.2mm; border-bottom: 0.6px solid #111; vertical-align: bottom; } +td { padding: 0.9mm 2mm; vertical-align: top; font-variant-numeric: tabular-nums; } +tbody tr:first-child td { padding-top: 1.6mm; } +tbody tr:last-child td { padding-bottom: 1.6mm; } +td:first-child, th:first-child { min-width: 26mm; padding-left: 0; } +td:last-child, th:last-child { padding-right: 0; } +tr { break-inside: avoid; page-break-inside: avoid; } +td em { font-style: italic; } +table.keep { break-inside: avoid; page-break-inside: avoid; } + +/* figures */ +p > img { display: block; margin: 4mm auto 0; max-width: 100%; max-height: 86mm; height: auto; } +div.figblock { break-inside: avoid; page-break-inside: avoid; } + +/* equations */ +div.eq { display: flex; justify-content: center; align-items: baseline; position: relative; margin: 3mm 0 3.5mm; break-inside: avoid; font-size: 10.5pt; } +div.eq .eq-no { position: absolute; right: 0; } +div.eq sub, p sub, td sub { font-size: 72%; } + +/* hypotheses, the entry test and the results slot: indented, no box */ +blockquote { margin: 3mm 10mm 4mm; padding: 0; font-style: italic; } +blockquote strong { font-style: normal; } +blockquote p { margin: 0; } + +/* lists */ +ol, ul { margin: 0 0 0.7em 1.1em; padding-left: 0.6em; } +li { margin: 0.2em 0; } + +/* file names in the appendices */ +code { font-family: "DejaVu Sans Mono", monospace; font-size: 7.8pt; color: #444; } + +/* references: hanging indent */ +section#references p { padding-left: 7mm; text-indent: -7mm; font-size: 9.5pt; line-height: 1.4; margin-bottom: 1.6mm; } + +/* space after a table, closer when a note follows */ +table { margin-bottom: 4.5mm; } +table + p.note { margin-top: -3mm; } diff --git a/papers/thesis_2026-10/reading/thesis_editing_copy_2026-10-16.docx b/papers/thesis_2026-10/reading/thesis_editing_copy_2026-10-16.docx new file mode 100644 index 0000000..7be36a8 Binary files /dev/null and b/papers/thesis_2026-10/reading/thesis_editing_copy_2026-10-16.docx differ diff --git a/papers/thesis_2026-10/reading/thesis_reading_copy_2026-10-16.pdf b/papers/thesis_2026-10/reading/thesis_reading_copy_2026-10-16.pdf new file mode 100644 index 0000000..ed66ad0 Binary files /dev/null and b/papers/thesis_2026-10/reading/thesis_reading_copy_2026-10-16.pdf differ diff --git a/papers/thesis_2026-10/reading/toc_pages.py b/papers/thesis_2026-10/reading/toc_pages.py new file mode 100644 index 0000000..6a9e527 --- /dev/null +++ b/papers/thesis_2026-10/reading/toc_pages.py @@ -0,0 +1,29 @@ +"""Second pass of the reading-copy build: finds the page of each heading in a first-pass PDF and writes the page numbers, +with dotted leaders, into the contents list of the HTML. Usage: python3 toc_pages.py first_pass.pdf reading.tmp.html""" +import html, re, subprocess, sys +pdf, page_html = sys.argv[1], sys.argv[2] +n = int(re.search(r"Pages:\s+(\d+)", subprocess.run(["pdfinfo", pdf], capture_output=True, text=True).stdout).group(1)) +pages = [subprocess.run(["pdftotext", "-f", str(i), "-l", str(i), pdf, "-"], capture_output=True, text=True).stdout for i in range(1, n + 1)] +norm = lambda s: re.sub(r"\s+", " ", s.replace("’", "'")).strip() +lines = [[norm(l) for l in p.splitlines() if l.strip()] for p in pages] +h = open(page_html, encoding="utf-8").read() +# contents pages: early pages listing at least four numbered sections +toc_end = max(i for i in range(min(6, n)) if sum(bool(re.match(r"^\d\.\d+ ", l)) for l in lines[i]) >= 4) + 1 +def page_of(title): + key = norm(html.unescape(re.sub(r"<[^>]+>", "", title)))[:40] + for i in range(toc_end, n): + if any(l.startswith(key) for l in lines[i]): + return i + 1 + return None +missing = [] +def fill(m): + title = m.group(2) + pg = page_of(title) + if pg is None: + missing.append(title); pg = "" + return f'{m.group(1)}{title}{pg}' +nav = re.search(r'