Polymarket market-making: a microstructure note

Q: Is market-making on Polymarket economically rational?

A: Yes, and emphatically. Every maker volume decile earns a positive volume-weighted markout at every horizon studied. Stratification is monotonic with an ~85× gap between the bottom and top decile at 10 minutes. There is a sharp discontinuity at D7 → D8 that looks like a step change in maker sophistication.

DATA

Source: SII-WANGZJ/Polymarket_data on Hugging Face — 568M on-chain OrderFilled events through March 2026. After filtering to markets with ≥ 5,000 fills (top ~3% by activity, ~75% of volume), the working dataset is 327,417,891 fills across 915,852 unique maker addresses, $13.99B USDC notional.

METHOD

For each fill, compute a markout against the nearest subsequent trade in the same market at horizons of 10 s, 1 min, and 10 min:

signed_yes = +yes_amount  if maker BUY_YES
             −yes_amount  if maker SELL_YES
markout_H  = signed_yes × (price_at_H − price_at_fill)

Forward-ASOF with strict inequality (future_ts > ts) — naive ASOF can self-match when two fills share a timestamp and contaminate short-horizon markouts. Volume-weighted mean across each maker's USDC volume, aggregated by decile (D1 = smallest 10% by volume, D10 = largest).

FINDINGS

Volume-weighted mean markout per (decile × horizon), units: USDC × YES-delta per fill.

• Every cell is positive. No maker tier systematically loses on adverse-selection markouts in this dataset.
• Monotonic D1 → D7 at all horizons.
• Discontinuity at D7 → D8. Markout roughly triples between the 70th and 80th percentile of makers — a threshold effect, not a smooth gradient.
• D10 dominates short horizons. 2.27 at 10 s vs 1.84 at 10 min — top makers capture the adverse-selection premium fast, before mean-reversion compresses it.

INTERPRETATION

The economic story is fair-value mean-reversion in a venue with a lot of uninformed flow. When a passive maker's quote is hit, either (a) the taker is informed and the price keeps moving against the maker, or (b) the taker is uninformed and the price reverts. On Polymarket — high retail participation, binary resolution events that produce mean-reverting micro-moves — case (b) dominates.

The D7 → D8 discontinuity is the most striking structural feature. The top 30% of makers by volume look like a qualitatively different strategy from the middle of the distribution: candidates include algorithmic cross-market arbitrage, latency-sensitive auto-cancellation, or programmatic liquidity provision.

CAVEATS

• On-chain only. Off-chain cancellations, partial fills, and bundled maker operations are invisible. This is the on-chain view of MM behavior, not the full venue.
• No resting book. Markouts are computed against subsequent trades, not against quoted mids. Standard practice when only trades are available, but it conflates spread capture and adverse-selection edge.
• ≥ 5,000-fill filter. The long tail of thin markets is excluded — markout estimator is too noisy below that density. Generalizing to the full set requires more compute (bigger scratch disk, longer wall time).
• PnL decomposition unreliable. ~51% of source rows have price × token_amount ≠ usd_amount by > 1%. Total PnL identity holds; the spread-capture / inventory-holding split does not.
• 1-hour horizon deferred. The 5th ASOF join exhausts local disk on a 16 GB / 66 GB-free laptop. Tracked for an overnight run on bigger hardware.

FOLLOW-UP

Two extensions identified.

1. Resting-book reconstruction to separate quoted spread from realized markout. Status: blocked by dataset. The Hugging Face source has fill events only — no OrderPlaced / OrderCancelled streams, no quote snapshots. Would require Polygon RPC log replay for the CTFExchange contract.

2. Market-classification stratification (politics / sports / crypto / macro / short-horizon price bets) and time-to-resolution at fill, to isolate which market types reward which strategies. Status: implemented. See finding 2 in the repo.

REPRODUCIBILITY

git clone https://github.com/avdheshcharjan/polymarket-market-maker
cd polymarket-market-maker
uv sync

uv run pmm download --file quant.parquet
uv run pmm metrics --source-file data/raw/quant.parquet \
                   --source-kind quant --engine duckdb \
                   --min-fills 5000
uv run python scripts/finding_01.py

Full pipeline ~46 min on a 16 GB Mac, peak scratch disk ~55 GB. 43 tests cover unit + property (hypothesis) + cross-engine equivalence (Polars vs DuckDB).

NOTE

This finding is less a novel claim than an empirical confirmation: on Polymarket's on-chain CLOB, adverse-selection markouts are systematically positive and maker-sophistication stratification is real, monotonic, and large enough to matter. The interesting work is downstream — explaining the D7 → D8 step and identifying which kinds of markets reward which strategies.