Polymarket Backtest Methodology

Mid-fills and minute aggregation are the two classic silent assumptions: the venue sits on the 0.99/0.995 grid 98.6% of the time and is one-sided 16.9% of 5-minute snapshots, so the gulf between "assumed mid" and "actual executable" is the difference between a strategy that proxies and a strategy that would have filled.

Rules 4 and 5 exist because the sample is the claim: 80 resolved markets support a hypothesis, 8,000 support a habit, and no backtest speaks beyond its window.

A bound-join strategy is backtested on 500 resolved 5m markets with the same fill model the fill simulation defines. The report separates strategy P&L, fill cost, and one-sided sample share; the conclusion names the window and the grid resolution.

That report is the deliverable — the methodology exists so the deliverable cannot hide the venue's texture.

The methodology prohibits the strongest claims on the weakest evidence: no "edge" from a 40-market mid-filled run, no annualized number from a month of windows. The honest ceiling is a claim that survives the sample, the fills, and the re-run.

Three biases account for most overfit polymarket backtests: look-ahead (a feature computed after the resolve second), survivorship (testing only markets that filled), and regime mixing (blending V1-era and V2-era fee rules, changed at the April 28, 2026 cutover). Each is checkable against the recorded archive.

The methodology on this page is built to surface all three before you report, because a backtest that cannot name its regime will not survive the next cutover.

• Backtest
• How-to backtest
• Fill simulation