How to Backtest Polymarket Strategies

The recorded 250ms frames define the backtest: for each signal, the strategy transacts against that frame's ladder — the same ladder the venue carried at that second, with one-sided and crossed states intact. Minute-aggregated data cannot honor this model because the states are gone.

The measured reason the model matters: one-sided books at 16.9% and a median 42,254-share bound rest mean mid-fill results overstate returns structurally.

Define the signal, slice a resolved sample, and run it on the Backtest AI tooling or the REST pipeline: entry at executable side, slippage from consumed size, settlement on the recorded outcome.

The output report should separate strategy P&L, fill cost, and the one-sided share of the sample — if it does not, it has hidden the venue's real texture.

A backtest decides a habit, not a month's P&L: the honest interpretation names the sample, the model, and the assumptions, exactly as the backtest methodology page requires before any live dollar.

The honest backtest is walk-forward: tune on one resolved window, validate on the next, and never let a snapshot from the test period leak into the signal. On a venue whose final minutes are one-sided 76.2% of the time, leakage is the easiest mistake to make and the hardest to see.

The archive's reconciled structure makes the discipline mechanical — signal on the executable side, fill from the ladder, settle on the recorded outcome — so the backtest tooling merely enforces what a careful study would do anyway.

• Backtest
• Backtest API
• Fill simulation