Skip to content
Crypto.world

Research

The Anatomy of a Decentralized Prediction Market: Microstructure Evidence from the Polymarket Order Book

Prediction markets are usually studied through their prices (calibration, forecast accuracy, longshot bias), while the trading-cost side of these venues has been largely unobservable because early…

The Anatomy of a Decentralized Prediction Market: Microstructure Evidence from the Polymarket Order Book
arXiv
2604.24366
Published
2026-04-27
Authors
Philipp D. Dubach

AI summary

Overview

Research area: Market microstructure of decentralized prediction markets (quantitative finance, q-fin.TR), specifically the on-chain limit order book of Polymarket.

Technical level: Advanced. The paper assumes familiarity with limit order books, effective/realized spreads, Kyle's lambda, Glosten-Harris decomposition, Lee-Ready trade-direction inference, and panel econometrics with robust standard errors.

Scope: A tick-level, cross-sectional study of 600 pre-registered Polymarket markets, built by joining a 30.3-billion-event WebSocket order-book archive to 255 million authoritative on-chain OrderFilled trades, reporting eight stylized facts plus a measurement result on the limits of order-book-only trade-direction inference.

What This Paper Is About

Prediction markets are usually studied through their prices (calibration, forecast accuracy, longshot bias), while the trading-cost side of these venues has been largely unobservable because early venues used scoring rules or sparse parimutuel pools. Polymarket now runs a genuine limit-order-book exchange on Polygon, which makes microstructure measurement possible for the first time at scale. The paper asks two questions: what does Polymarket's microstructure actually look like across many markets at once, and can standard equity-microstructure tools—which need to know whether a trade was buyer- or seller-initiated—be trusted when the only input is the public order-book feed?

Key Contributions

  1. A tick-level, cross-sectional dataset and pre-registered panel. A continuous archive of Polymarket's public WebSocket order-book feed covering 52 calendar days (2026-02-21 to 2026-04-15), containing 30,287,264,368 events across 385,198 distinct markets, joined to 255,425,405 on-chain OrderFilled events over a 28-day overlap window (2026-02-28 to 2026-03-27) and analyzed on a pre-registered stratified panel of 600 markets (top-100 by on-chain USDC volume plus a random-500 stratum).

  2. A measurement result that constrains the whole literature. Trade direction inferred from the public order-book feed agrees with on-chain ground truth only about 59% of the time (volume-weighted), barely above the 50% chance baseline and far below the roughly 81% Lee-Ready accuracy documented on Nasdaq. Direction-dependent measures therefore change sign when the inferred sign is swapped for the on-chain record.

  3. Eight cross-sectional stylized facts about Polymarket (longshot spread premium, depth concentration, block-clock alignment, maker concentration, category-conditional spread, archive-ingestion latency, self-counterparty wash share, and cross-sectional depth structure), none of which requires the on-chain join.

  4. A replication package and pre-registration discipline. Code, panel artifacts and the on-chain scrape pipeline are released (GitHub philippdubach/polymarket-microstructure, DOI 10.5281/zenodo.19811426), with the panel's SHA-256 recorded in the pre-registration document before analysis ran.

Main Findings

  • Effective spreads are roughly an order of magnitude wider than in equities. On the central price decile the median quoted half-spread is around 200 bps, versus single-digit basis points for liquid US equities post-decimalization.
  • SF1 – Longshot spread premium. Median quoted spread is about 400 bps in the central [0.4, 0.6] mid-price range and rises to 1,300–1,800 bps for markets trading below 0.10. The lowest decile has a median of 1,818 bps. The pattern is asymmetric, with the low-probability side wider; the author argues the magnitude looks more like an inventory-risk constraint than the classical few-percent racetrack premium.
  • SF2 – Depth is layered, not top-heavy. Across 547 markets with non-null depth, top-of-book carries a median 0.136 of cumulative top-10 depth against a 0.10 uniform null, with no level capturing more than 14% at the median. Median KL divergence from uniform is 0.087 (IQR [0.024, 0.474]; p10–p90 [0.005, 0.985]), far from the log 10 ≈ 2.30 of a perfectly top-of-book book. Only 9% of markets are strongly top-heavy (L1 > 0.5); 57% have an L1 share within [0.05, 0.20].
  • SF3 – No meaningful block-clock alignment. The share of price_change events within ±100 ms of a 2,000 ms grid point has median 0.102 (IQR [0.087, 0.110]) against a 0.10 null; only 27 markets (4.5%) exceed 0.15. A binomial test rejects the exact point null in 351 of 600 markets (58.5%), which the author attributes to mechanical sensitivity from very large event counts rather than economic clustering.
  • SF4 – Broad maker diversity with a concentrated tail. Across 600 markets and 6.4 million trades, median volume-weighted maker HHI is 0.031 (about 32 effective makers), p90 = 0.119 (about 8 effective makers), maximum 0.40 (roughly 3 effective makers).
  • SF5 – Category-conditional spreads. Panel composition is Crypto 348/600 (58%), Sports 142/600 (24%), Other 75/600, Geopolitics 35/600. Median on-chain effective half-spreads are Sports 0.0075, Geopolitics 0.0001, Other −0.0004, Crypto −0.0393 probability points; medians lie within ±0.04 probability points of each other, but Crypto and Sports have within-category interquartile ranges of roughly 0.3 probability points.
  • SF6 – Archive-ingestion latency, not exchange latency. Median per-market delay is 41.5 ms (inter-market p25–p75 of 39–47 ms), with p90 = 166 ms and p99 = 6,108 ms. The author stresses this measures the collector pipeline, not trader speed.
  • SF7 – Low but nonzero self-counterparty wash share. Under a lower-bound detector (maker == taker, or a flipped pair within 128 blocks on the same market), median wash share is 0.97%, p90 = 4.5%, p99 = 10.6%, maximum 22.2%. The 25–70% range documented by Cong et al. (2023) on unregulated token exchanges is treated as a sanity bound, not a like-for-like benchmark, because the wash incentives differ.
  • SF8 – Depth is explained by duration, price level and volume, not time-to-close. The bivariate slope on log seconds-to-close is 0.8178 (HC3 SE 0.1133, R² = 0.1293, n = 322), falling to 0.5500 with category fixed effects and 0.3048 with log volume. Adding log market duration and log p(1−p) collapses it to +0.0084 (SE 0.1015, R² = 0.5702, n = 313); restricting to mid-price deciles 2–7 gives −0.2702 (SE 0.1687, n = 187). Duration enters at +0.222 (SE 0.073), log p(1−p) at −1.02 (SE 0.180), log volume at +0.41 (SE 0.046). The author calls the earlier positive slope "mechanical mediation."
  • Trade-direction inference is noise-dominated. Panel mean sign agreement is 0.615 (market-clustered bootstrap 95% CI [0.579, 0.653], 109 valid cells of 400, 55 markets); volume-weighted 0.592 (CI [0.542, 0.659], 125,080 matched buckets); panel median 0.591, IQR [0.526, 0.681], p10–p90 [0.464, 0.841]; per-window medians 0.586 / 0.591 / 0.618 / 0.645.
  • Sign noise flips the signs of headline measures. On the comparable top-100 subset, the effective half-spread changes sign between feed-inferred and on-chain directions on 67% of markets in the first 7-day window (50% in a second non-overlapping window), and Kyle's lambda on 60% (43% in the second window). The mechanism is structural: change_side marks which side of the book moved, not which side initiated the trade.
  • Spread decomposition is near-null once signs are correct. The Glosten-Harris decomposition converges on 97 of 100 top-100 markets; median effective half-spread is −0.0003 probability points, median transitory component 0.00001, median adverse-selection component 0.0—consistent with the earlier sign-flip calibration.
  • Tooling caveat. CLOB REST resolves all 385,198 archive market ids, while the Gamma metadata API indexes only 34,764 markets and is therefore unsuitable for the join.

Methodology in Plain English

The author collected two independent records of the same trading activity and compared them. The first is a continuous, event-by-event archive of Polymarket's public WebSocket order-book feed: every time a resting order is added, changed or removed, an event is logged, producing 30.3 billion rows across 52 days. The second is the blockchain itself: every settled trade emits an OrderFilled event on the CTF Exchange V1 contract, and that event carries a simple binary field (which asset id is zero) that identifies whether the taker was buying or selling. That field is treated as ground truth.

The two records are tied together through the conditional-token identifiers, using the CLOB REST endpoint to map market ids to YES/NO token pairs. Before any analysis, the author fixed the rules for choosing the 600 markets—volume ranking for the top-100 stratum, a seeded random draw (seed 20260424) from markets with at least 100 on-chain trades for the random-500 stratum—and recorded a hash of the resulting panel in the pre-registration document so no market could be added or removed afterward.

For the eight stylized facts, the paper computes per-market distributions (spreads by price decile, depth shares by book level, block-alignment shares, maker Herfindahl indices, latencies, wash shares) and reports medians and percentiles across markets rather than pooling all events. For depth near resolution, it runs panel regressions of log mean depth on log seconds-to-close under five specifications, progressively adding category fixed effects, log volume, log market duration and log p(1−p), with HC3-robust standard errors.

For the measurement result, the author tries to do what equity researchers do: infer the aggressor's side from the order-book stream alone (a "LOOSE" rule treating any resting-size decrement as a trade), match the inferred trades to on-chain fills at 5-second and exact-price granularity across four disjoint 7-day windows, and measure how often the inferred direction matches. A second "STRICT" rule is used to propagate the noise into six direction-dependent measures and show what it does to their signs. Because the STRICT routine scales quadratically in the recent-event list, that propagation test is run on a single 7-day window (2026-03-07 to 2026-03-14) rather than the full 28 days; the supplied text is truncated at the start of that subsection's results.

Why This Matters

Impact on research. The paper's central methodological point is that a widely used shortcut—reading trade direction off a quote feed—does not work on Polymarket. Because a noisy sign does not merely attenuate a microstructure measure but can reverse its sign, a substantial body of direction-dependent results computed from Polymarket's feed alone could be pointing the wrong way. The 59% agreement rate and the 67% / 60% sign-flip rates give a concrete, quantified reason to source direction from on-chain OrderFilled events, and the released join code makes that practical. The eight stylized facts also establish Polymarket as an empirical object in its own right, at cross-sectional scale rather than the single-market, aggregated-data studies that preceded it.

Real-world applications:

  • Traders and market makers can use the spread-by-price-level profile (SF1) and the depth profile (SF2) to estimate realistic execution costs on Polymarket, especially in low-probability contracts where quoted spreads reach 1,300–1,800 bps.
  • Builders of prediction-market analytics can adopt the on-chain join as a design requirement, since the public feed alone cannot identify the aggressor.
  • Researchers benchmarking forecast accuracy or calibration can adjust for trading costs using the effective-spread figures, rather than assuming frictionless prices.
  • Compliance and market-integrity analysts get a direct, lower-bound detection method for self-counterparty wash trades (SF7) with a reported distribution across 600 markets.

Industry relevance. The findings matter to exchange operators and aggregators: the block-clock test (SF3) indicates quote updates do not meaningfully synchronize with Polygon block boundaries, and the maker-concentration measurement (SF4) documents a flatter liquidity-provider landscape than the handful of high-frequency firms that dominate US equities. The latency measurement (SF6) is explicitly framed as an archive-pipeline property, a useful caution against citing venue latency from collector-side timestamps.

Future Directions

  • Evaluate the last_trade_price event. Polymarket's market channel emits this event separately and it was excluded from the archive; testing its agreement with on-chain ground truth is flagged as an obvious follow-up that could recover some direction signal from the feed.
  • Within-market depth trajectories. SF8 is cross-sectional; whether an individual market's depth falls as it approaches its own resolution requires per-market depth time series that the panel does not materialize, and is explicitly deferred.
  • Graph-based wash-trade classification. The current detector is an explicit lower bound covering only direct and immediate-roundtrip self-matching; extending to multi-counterparty network classifiers would quantify the gap to the 25–70% token-exchange benchmark.
  • Decomposing the longshot spread premium. Distinguishing inventory risk from a behavioral longshot premium would require maker realized inventory and time-on-book data that the study does not collect.
  • Full-window propagation analysis. The direction-noise propagation test was limited to a 7-day window by the quadratic cost of the STRICT inference routine, leaving a full 28-day version as an engineering open question.

Target Audience

Quantitative researchers in market microstructure and prediction markets; empirical finance and econometrics practitioners working with on-chain data; builders and analysts of decentralized exchanges who need to know what the public WebSocket feed can and cannot reveal; and data engineers constructing tick-level archives who will benefit from the latency, join and metadata-source lessons (CLOB REST over Gamma, the 256-block reorg buffer, the split-fill filter). Readers primarily interested in forecast accuracy or calibration will find the spread and depth facts more useful than the inference section.

Authors’ abstract

We study the microstructure of Polymarket, the largest on-chain prediction market, using a continuous tick-level archive of the public order-book feed (30 billion events over 52 days) joined to the authoritative on-chain trade record. On a pre-registered stratified panel of 600 markets we report eight stylized facts: a longshot spread premium; a depth profile closer to uniform than to top-of-book; a null block-clock alignment effect; broad maker-wallet diversity with a…

The full abstract on arXiv
Read the original paper