The Anatomy of Polymarket: Evidence from the 2024 Presidential Election

Abstract: This paper provides a comprehensive transaction-level analysis of Polymarket's 2024 U.S. Presidential Election market using complete on-chain data from the Polygon blockchain. Because blockchain-based prediction markets involve heterogeneous trade mechanisms: share minting, burning, and conversion alongside conventional exchange, naive aggregation of on-chain flows can misrepresent actual trading volume. To address this, we develop a volume decomposition that yields three complementary measures of market activity: exchange-equivalent trading volume, net inflow, and gross market activity. Applying this framework, we document three key episodes that shaped the market: Biden's withdrawal, the September presidential debate, and the emergence of whale traders in October. As trading volume grew, arbitrage deviations narrowed, Kyle's $λ$ declined by more than an order of magnitude, and cross-market participation broadened, painting a consistent picture of a market that matured over its ten-month life.

• The paper introduces a transaction-level volume decomposition that separates exchange-equivalent turnover, net capital inflows, and gross activity, avoiding double-counting from Polymarket’s minting and burning mechanisms.
• The paper finds that disagreement surged after Biden’s withdrawal and during October whale activity, while Trump-market inflows increasingly signaled a Trump victory before prices fully adjusted.
• The paper shows liquidity improved dramatically: Kyle’s lambda fell from 0.518 early in the market to about 0.01 by October, reducing the estimated impact of a $1 million order near 50% odds from roughly 13 percentage points to 0.25 points.

Overview

Tsang and Yang provide a transaction-level analysis of Polymarket's 2024 U.S. Presidential Election market, using the complete on-chain settlement record on Polygon from January 5 to November 6, 2024 (2603.03136). The paper makes two methodological and four empirical contributions: a volume decomposition that corrects for the platform's heterogeneous trade mechanisms; documentation of trading activity and cross-market disagreement around three episodes (Biden's withdrawal, the September debate, and the October whale activity); an analysis of arbitrage deviations from the $1.00 pricing constraint; a characterization of trader behavior; and a rolling estimation of Kyle's$\lambda$ as a measure of liquidity and manipulation risk. The unifying finding is a market that matured over its roughly ten-month life: liquidity deepened, arbitrage deviations narrowed, and price impact fell by more than an order of magnitude.

Institutional setting and trade mechanisms

Polymarket operates a hybrid-decentralized architecture: an off-chain central limit order book managed by a centralized operator, with atomic settlement through smart contracts on Polygon (CTFExchange for binary markets, NegRiskCtfExchange for categorical markets). Outcome shares are ERC-1155 conditional tokens built on the Gnosis Conditional Token Framework, collateralized in USDC.

The paper emphasizes that unlike equities, the supply of outcome shares is endogenous. Three non-exchange mechanisms matter:

• Share minting (splitPosition): matched opposing orders lock 1 USDC and mint a full (YES, NO) set; it can also be invoked unilaterally to enforce the price ceiling when $P_{YES} + P_{NO} > 1$.
• Share burning (mergePositions): the inverse operation, redeeming a full set for 1 USDC and enforcing the price floor.
• Position conversion (convertPositions): in categorical markets, $Q$ units of NO shares across $M$ outcomes convert into YES shares of the remaining outcomes plus $(M-1)Q$ USDC, enforcing cross-outcome consistency ($\sum_k P_{YES,k} = 1$).

The practical consequence is that naive aggregation of on-chain flows double-counts volume—a problem previously flagged by practitioners but not formalized academically.

Volume decomposition

Each transaction is decomposed into six components—trade, mint, and burn volumes for YES and NO tokens—using OrderFilled events. The classification hinges on comparing aggregate buy-side and sell-side USDC flows: equality implies pure exchange; buy exceeding sell implies minting (pure or mixed, distinguished by the number of distinct maker asset IDs); sell exceeding buy implies burning. Three market-level measures follow:

• Exchange-equivalent volume ($V^E$): trade volume plus offsetting mints against burns within the interval, approximating counterfactual secondary-market turnover.
• Net inflow ($F$): mints minus burns, capturing fresh capital committed to or withdrawn from a state of the world.
• Gross market activity ($V^G = V^E + |F|$): total economic activity including net capital movement.

The authors acknowledge two assumptions here: temporal pairing of mints and burns within an interval may involve different traders at different times, and the split between exchange and minting is partly an artifact of the matching engine rather than trader intent. These caveats are inherent to any reconstruction from settlement data alone.

Trading activity and disagreement

Monthly figures show explosive growth in the final months: in October, Trump-market volume reached $391.030 million versus $191.928 million for Harris, with October net inflows of $P_{YES} + P_{NO} > 1$0106.229 million (Harris). A notable signal visible only in net inflows is that in the final days, new money flowed disproportionately toward Trump YES while Harris NO inflows exceeded Harris YES—the direction of fresh capital consistently pointed toward a Trump victory even before prices fully reflected it.</p> <p>Using the daily net-inflow correlation between the Trump market and a combined Democrat market (Biden before July 21, Harris after), the paper identifies three regimes. The correlation jumps at Biden&#39;s withdrawal, reflecting simultaneous bilateral capital injections into both sides ($P_{YES} + P_{NO} > 1$15.495 million into Harris in July)—the signature of heterogeneous beliefs. It declines after the September 10 debate as traders converged toward consensus near 50–50 odds. It then surges from roughly 0.17 to 0.76 in early October, coinciding with whale accounts placing approximately $P_{YES} + P_{NO} > 1$2\lambda$ had by then fallen to 0.01–0.04, the market absorbed this disagreement through volume rather than extreme price dislocation—an empirical instance of the rational-countertrading mechanism predicted by Hanson, Oprea, and Porter's manipulation experiments.

Price efficiency

The deviation $P_{YES} + P_{NO} > 1$3 measures arbitrage efficiency under the smart-contract-enforced $P_{YES} + P_{NO} > 1$4\delta$P_{YES} + P_{NO} > 1$5\delta$ over time independently corroborates the liquidity improvement documented via declining price impact, consistent with a limits-to-arbitrage interpretation in which early mispricings persist because funding and execution constraints bind.

Trader behavior

Participation grew substantially across quarters, and intraday activity shifted from roughly uniform to strongly seasonal, peaking between 09:00 and 20:00 UTC—overlapping European and U.S. business hours. This concentration is sharper among the top 10% of traders by frequency and volume, suggesting sophisticated participants operate on schedules aligned with traditional financial markets.

Cross-market participation is heavily concentrated: about 40% of all traders participated exclusively in the Trump market, versus 18.4% exclusively in Harris. At the token level, 71.8% of traders touched Trump YES, with the largest subgroup (30.7%) trading exclusively in that token—directional bettors without hedging. The second-largest group (19.4%) traded all four major token markets, consistent with market-making or hedging. Only roughly 0.7% traded across more than two candidates, a degree of specialization consistent with information-based attention allocation.

Liquidity and manipulation risk

The paper estimates Kyle's $P_{YES} + P_{NO} > 1$6 for the Trump YES token using a tick-rule classifier for trade direction, hourly aggregation of signed flow, a log-odds transformation of prices to handle bounded-support heteroskedasticity, and a rolling 720-hour OLS regression of log-odds changes on net order flow. The results are stark: $P_{YES} + P_{NO} > 1$7 spikes to 0.518 in early, thin periods—implying a $P_{YES} + P_{NO} > 1$8p=0.5$P_{YES} + P_{NO} > 1$9-0.073$Q$0\hat{\lambda}_t$Q$1R^2 = 0.278$ over 277 observations.

The authors are careful about interpretation: $Q$2 measures how easily order flow moves prices, not why. High $Q$3 signals mechanical vulnerability to manipulation in low-liquidity regimes, whether or not observed moves were manipulative. By October, whales could still sway odds, but faced a deeper book and a counter-trading crowd.

Limitations and open questions

The analysis is confined to a single platform and a single event category; cross-platform comparisons with Kalshi, PredictIt, or Robinhood remain outside its scope. Whale identification relies on public reporting rather than direct on-chain attribution of trader identity. The temporal mint-burn pairing assumption underlying $Q$4 introduces approximation error that scales with interval length. Finally, welfare questions—whether such markets improve public forecasting, distort information, or create channels for political influence—are left open, though the paper's data infrastructure could support them.

Conclusion

This paper formalizes the measurement problem posed by tokenized contingent claims and applies the resulting framework to the largest prediction market event observed to date. Its decomposition separates churn from net capital commitment, revealing flow-direction signals invisible in volume alone. The empirical record shows a market whose arbitrage efficiency, depth, and breadth of participation improved markedly as the election approached, absorbing both informational shocks and large directional bets primarily through volume rather than price dislocation.