How to Backtest Prediction-Market Strategies on Polymarket & Kalshi

What makes prediction markets different to backtest

A prediction-market contract is not a stock or a perp. Each outcome is quoted 0 to 1 — its implied probability — and at settlement one side is worth exactly $1 and the other $0. Crypto up/down markets settle in 5 to 60 minutes against a reference price, spreads widen and depth thins into the close, and the whole life of a market can be shorter than a single candle on a normal chart.

That shape changes what a backtest needs. Sampling a last price once a minute misses most of a five-minute market; ignoring the book hides the spread that decides your P&L on a contract that pays at most a dollar; and mislabeling the settlement flips winners into losers. The rest of this guide is about getting each of those right.

Step 1 — Start from the real order book

A backtest can only be as honest as its data. You need the full order book — every resting bid and ask with its size, on both sides — captured on every change rather than on a fixed clock, because event-driven capture is the only way to see a short-dated market's whole arc. Neither Polymarket nor Kalshi serves its own historical order-book depth, so replaying the book as it was requires a provider that captured it continuously.

DepthFeed serves exactly that for both venues in one schema: Polymarket captured event-driven from the CLOB websocket, Kalshi polled continuously at full depth (up to 100 levels per side), each snapshot carrying epoch-millis timestamps and a joined underlying crypto reference price. One loader reads either venue.

Step 2 — Model fills against depth, not the mid

This is the step that separates a backtest you can trade on from one that lies. A marketable order walks the book — it fills the best level first, then the next, so its average price is worse than the touch whenever it consumes more than the top level. A resting order joins the queue and only fills if the market trades through its price with enough size behind it. Filling everything at the mid with zero slippage, the default of most casual backtests, systematically overstates edge.

Whichever route you take, insist on a depth-aware fill. On a four-cent round-trip spread, on a contract that pays at most a dollar, the difference between a mid fill and a real one is enormous — and it compounds over every trade.

Step 3 — Settle at the real outcome

Because the payoff is binary, settlement is not a rounding detail — it is most of the P&L. Polymarket up/down markets resolve against a 'price to beat' set at the open versus the final reference at the close; Kalshi contracts settle yes/no against a published reference price at expiry. Backtest against the real resolution, not an assumption, and sanity-check the split: across the settled crypto up/down markets we've captured, outcomes land near a coin-flip, so any backtest whose settlement skews far from that is probably mislabeling winners.

Step 4 — Pick your workflow: no-code or API

There are two ways to run all of the above on DepthFeed. The no-code path is the Backtest Lab: pick a coin and window, choose a preset or write a short rule, select a fill model, and it replays resolved markets and settles them for you in the browser — depth-aware VWAP fills included, no pipeline to build.

The API path is for custom research at scale: pull full-depth snapshots over the REST API, reconstruct the book in your own stack, and run whatever model you like. Because the historical REST API and the live WebSocket emit the identical JSON, the loader you write to replay history reads the live feed unchanged — you trade the exact code you backtested.

Step 5 — Forward-test before you risk money

A backtest is fit to the past, however lightly, so the last step is to run the survivor forward on markets whose outcome nobody knows yet. Deploy the rule to paper trading — or drive it from your own bot over a webhook — and it trades live books with virtual cash, filling at the real bid/ask and settling automatically at resolution. The equity curve it builds is the forward evidence a backtest can't give you.

That completes the loop: data → honest backtest → forward test. Kill ideas cheaply on history, graduate the ones that survive to live paper books, and only then consider real capital.

Common pitfalls

• Last-price or hourly data: hides the spread and misses the life of a 5–60 minute market.
• Mid fills: assuming zero slippage overstates edge, badly, on a contract that pays at most $1.
• Assumed settlement: label winners from the real reference price, not a guess — it's most of the P&L.
• Look-ahead bias: only react to data with a receive timestamp at or before your decision time.
• Skipping the forward test: an edge that only exists on settled history may not survive live markets.