Polymarket backtest AI: describe a strategy in plain English and replay it on 250ms books

Backtest AI parses a strategy description into entry and exit rules — entries, take-profit, stop-loss, and window-end exits — then replays those rules over resolved crypto up/down markets from the archive: 118,251 of them across BTC, ETH, SOL, BNB, DOGE, HYPE, XRP, and ZEC at 5-minute, 15-minute, and 4-hour windows.

No Python and no CSV wrangling. Only markets that have already resolved are used, so every run is scored against what actually happened, and each run is read back in three parts: a Summary, Key Observations, and an Actionable Takeaway.

The engine walks the historical 250ms L2 bid/ask ladder at each tick: it reads the size available at each price level, models partial fills when a level runs out, and applies slippage when a larger order has to cross the book. On the published dataset, 5-minute markets are one-sided 16.9% of the time and 3.24% of snapshots are crossed — a mid-fill backtest would flatter most strategies.

Exits are tracked too, so the report can say why each run finished: take profit hit, stop loss hit, the window ended, or never got a fill. See how it replays before you buy.

• Standalone Backtest AI — [$29/mo](/pricing) — unlimited strategy backtests, ~500 AI backtests per month, 30-day market data for the AI, and unlimited Strategy Builder access. No REST data API.
• Add-on — +$19/mo on any data window — 500 backtests per month over the same resolved archive, read back with a Summary, Key Observations, and an Actionable Takeaway.
• [Starter](/signup) stays free — 1 AI backtest and 3 free strategy backtests on 250ms books with 3 days of history, enough to try the flow before paying.
• See how it measures against the alternatives.

The interaction is one sentence: "Buy DOWN on BTC 15m below 0.40, take profit at 0.55, stop loss at 0.30." Backtest AI parses that into entry, take-profit, stop-loss, and window-end exits, then replays it over the resolved crypto up/down archive — 118,251 contracts across BTC, ETH, SOL, BNB, DOGE, HYPE, XRP, and ZEC at 5-minute, 15-minute, and 4-hour windows.

Nothing is scored on an unresolved market. Every run is judged against what actually happened, on the 250ms ladder that really stood — the engine reads size at each price level, models partial fills when a level runs out, and applies slippage when an order must cross the book rather than transact at the mid.

The reread comes back in three parts: a summary of the run, key observations about where the rules won and lost, and an actionable takeaway naming the single adjustment with the largest measured effect. That structure is what makes the tool usable by someone who has never touched a DataFrame.

Where a scripted backtest traps beginners is exit attribution, and the report states it plainly: take profit hit, stop loss hit, the window ended, or the fill never arrived. A strategy whose reports fill with "window ended" has discovered that the price never paid its visit — the venue rebounds less than the charts imply.

Read the one-sided rate into every winner: 5-minute books are one-sided 16.9% of the time and 3.24% of snapshots are crossed, so a strategy that assumes you can always buy or sell a level is assuming a book that 16.9% of the time is only half there.

Use the exit-reason column as the diagnostics layer; it tells you whether you lost to the market direction or to the book structure, and those are different fixes.

The standalone plan at $29/mo folds the AI reads in; the add-on path attaches the same engine to a data subscription for teams that want both.

Reserve judgment until one run has been reconciled manually against a single stored snapshot — five minutes of checking beats a month of trusting a report engine.

• Backtest API
• Backtesting overview
• Resolved markets
• Crypto data coverage
• Plans & pricing