What one paper documented about why naive Polymarket volume overcounts

What one paper documented

Public write-ups about Polymarket volume often skip the construction and keep the folklore: the site did billions, or on-chain volume is fake because someone added up the wrong log.

One working paper measured a version of that gap. It is Tsang and Yang, The Anatomy of a Blockchain Prediction Market: Polymarket in the 2024 U.S. Presidential Election (arXiv:2603.03136; current version on the opened pages is 11 August 2026). This note states what that paper documents about naive event-log totals versus exchange-equivalent turnover on one 2024 election tape. It is not a method for rebuilding a book. It is not a 15 August 2026 volume.

The same paper’s finding that the 2024 Trump book traded at a persistent positive Yes-plus-No gap is a separate live piece. This note does not re-try that identity.

The sample ends 6 November 2024

The authors study Polymarket’s Presidential Election Winner 2024 market. They write that the market contains 17 outcomes, and that Donald Trump, Joe Biden, and Kamala Harris account for the vast majority of trading volume. Those three outcomes are the ones they tabulate.

The sample is 2024-01-05 00:00:00 UTC through 2024-11-06 06:46:00 UTC, “when Fox News first projected that President Trump had won the election.” A footnote on the same page dates the actual market launch to late 4 January 2024, with orders flowing after midnight UTC on the 5th.

That window is the fact. It is not a statement about books on 15 August 2026.

Raw on-chain flows are not one kind of trade

Official docs retrieved 15 August 2026 still describe the machine the paper is measuring. Each market has two ERC-1155 outcome tokens. Split turns $1 of collateral into 1 Yes and 1 No. Merge burns that pair back into $1. Redeem, after resolution, pays $1 on the winner and $0 on the loser. The prices-and-orderbook page says a buy-Yes at $0.60 can match a buy-No at $0.40 because those bids add to $1.00: the dollar is locked and the pair is minted. That $0.60 / $0.40 pair is a documentation example, not a live fill. Matching is off-chain; settlement is on-chain.

The paper’s 2024 sample writes the collateral as USDC. The docs page retrieved this afternoon writes pUSD. This note does not collapse those names, and it does not invent a conversion date.

Tsang and Yang add the event-log problem. Completed matches settle on Polygon and emit OrderFilled (one row per fill) and OrdersMatched (one summary per matched transaction). Those logs record simple exchange of existing shares, minting of new pairs, burning of complete sets, and — in a single transaction — mixtures of exchange and mint or burn. Conversion exists as a fourth mechanism. They describe it and then exclude conversion from their turnover and inflow measures, because it is unilateral and, as they define those measures, does not change directional exposure.

Two naive totals miss in opposite directions

They walk a mixed trade from the 2024 tape. A buyer takes 238.095237 Trump Yes for 99.999999 USDC. Two hundred of those shares are a straight exchange (84 USDC). The remaining 38.095237 are minted against a complementary Trump No buy. Economically: 84 USDC of secondary exchange plus 38.095237 USDC of new pairs.

Read only OrdersMatched and the log assigns 99.999999 USDC to Trump Yes and drops the minted No side.

Add up OrderFilled and Trump Yes becomes 84 + 99.999999 = 184.999999 USDC, with Trump No at 22.095238 — the omitted No comes back, and the exchange leg is counted twice.

A simple three-fill exchange on the same pages is the clean case: volume, as they read it, is 123.9 USDC, not the sum of every fill.

Those are worked examples on one tape. They are not a recipe.

Three constructed measures, one window

From the fills they build three market-level objects for a window T.

Exchange-equivalent volume (V^E) is secondary-style turnover: the exchange component, plus offsetting mint and burn inside the window, counted once. They warn that V^E is defined on the window. Lengthen T and more mints become eligible to pair against later burns, so V^E generally rises. Monthly V^E is their cross-period measure. Daily V^E is not the same object.

Net inflow (F) is mint minus burn. Positive F is new collateral committed on that side.

Gross market activity (V^G) is V^E plus the absolute value of F — turnover plus the net capital that entered or left.

Conversion stays out of both V^E and F by design. A census that put conversion back in would be a different number. That number is UNKNOWN from the opened pages.

October 2024, as the tables print it

October is the most active month on their tables.

Trump-market, million USD (Table 4): V^E 391.030; F 176.417; V^G 567.448.

Harris-market, same month: V^E 191.928; F 106.229; V^G 298.157.

Biden-market, same month: V^E 0.017; F −0.002; V^G 0.019.

Table 5 puts the two naive constructions next to those objects. October Trump OrdersMatched is 477.246 million USD (15.90 percent below V^G). October Trump OrderFilled is 958.478 million USD (68.91 percent above V^G).

The abstract rounds that contrast to $958 million versus $391 million. The introduction writes $958.48 million and $391.03 million. This note keeps the table figures.

From July to November, they write, OrderFilled exceeds V^G by roughly 60 percent to 88 percent in the Harris and Trump markets, and is more than twice V^E. The decomposition is what lets them separate turnover among existing positions from fresh capital.

Those are 2024 monthly reconstructions. They are not a live 15 August 2026 frequency table.

The $3.6 billion figure is a venue-page total they cite

The introduction says Polymarket “processed over $3.6 billion in trading volume on the presidential race alone.” The footnote: that figure “comes from Polymarket’s own market page for Presidential Election Winner 2024, which displayed total volume of $3,686,335,059.”

This run did not open that market page. The $3,686,335,059 figure is paper-reported. The authors’ point on the same pages is that a headline total built from raw flows “should therefore not be read as conventional turnover.”

How any named live page computed “volume” on 15 August 2026 is UNKNOWN. This run did not retrieve a live volume line, and official docs retrieved this afternoon do not publish an OrderFilled-versus-V^E mapping.

What this paper does not say

It does not say a 15 August 2026 book uses OrderFilled, OrdersMatched, V^E, or something else. UNKNOWN.

It does not say the same overcount ratio holds on a later tape. The sample ends 6 November 2024. UNKNOWN.

It does not say displayed volume is wash, fake, or a fraud. The opened paper is an accounting reconstruction, not a wash-trading finding. A practitioner note they cite (Slivkoff, December 2025) was not opened on this run.

It does not say the complementary $1 identity failed, or that you should mint, merge, or trade a gap. That identity is a separate live piece.

It does not compare this venue to Kalshi.

If you came here because a thread treated a headline billion as stock-style turnover, or treated every on-chain total as fake, the useful move is smaller than picking up either story: keep the dated table, keep the event-log rule on the same line as the dollar figure, and refuse the live-volume claim until another primary replaces the 6 November 2024 end date.