How to Backtest a Polymarket Strategy (and Not Fool Yourself)

Most Polymarket strategies start the same way: someone notices a wallet that keeps winning, or a pattern in prices, and asks whether following it would have made money. A backtest is how you ask that question properly. It is also the easiest place in trading to fool yourself, because a backtest always returns a number, and a wrong method returns a flattering one.

What a backtest is, and what it is not

A backtest takes a rule ("copy these wallets when two of them buy the same side", or "buy when the price falls 10¢ in an hour and the spread is under 2¢") and replays it over a past period, trade by trade, as if you had been following it. It adds up what you would have bought, at what price, and how those positions resolved.

It is a description of the past under stated assumptions. It is not a forecast, and it is not a record of trading: no order was placed. Everything below is about making those assumptions honest.

Step 1: Use an entry price you could have paid

The most common mistake is to copy a trader's own fill price. You could not have had that price. You see the trade after it happens, and by then the best offers are often gone.

An honest backtest enters after the signal, at a price a follower could still have paid, and never above the limit you set. If a strategy only works at the leader's price, it does not work.

The same rule covers information. A rule may only use what was known at the moment it fires: a ranking as it stood that day, not today's ranking applied to last month. Using later information is called look-ahead, and it turns any backtest into a hindsight exercise.

Step 2: Model the fills, and say how optimistic the model is

On a prediction market the order book is often thin. A $10 copy usually fills; a $2,000 copy often does not, or fills at a much worse average price. A backtest that assumes every order fills in full is measuring a strategy nobody could have run.

Two things to require of any backtest:

• A fill model: a stated rule for how much of an order would have filled, and at what price.
• Its measured optimism: how much better than real fills that model has turned out to be when checked. Every model is optimistic somewhere; the honest ones tell you by how much.

In the Strategy Lab, each report states the run's own fill model in a sentence, with its measured optimism beside the simulated-results notice.

Step 3: Compare with a fair baseline, not with zero

"It made money" is not a result. In a period when the markets a strategy trades went the way most buyers expected, almost any strategy in those markets made money.

The useful question is whether your selection added anything. So set the result beside what you would have made doing the obvious thing in the same markets under the same rules:

• For a copy strategy, the baseline is copying everyone who traded those markets, with your limits and caps.
• For a strategy built on your own rules, it is the same price at a random moment, and the markets you chose without the rule.

A strategy that barely beats its baseline is telling you that the markets did the work.

Step 4: Check that luck is not carrying the result

Take the result apart before you believe it:

• Without the best wallet. If the profit disappears when one wallet is removed, you have found one wallet, not a strategy.
• Without the best three days. If it disappears without its best stretch, you have found a lucky week.
• Count the copies. Ten copies are an anecdote. The Lab's verdict needs at least 30 copies and a t-statistic of at least 1 to say Promising, not proven, and at least 100 copies and a t-statistic of at least 2 to say Strong evidence. Anything less reads Weak evidence.
• Count your attempts. If you tried fifty variants and kept the best, the best one looks good by chance. Write down how many you tried.

Step 5: Keep one period you have never looked at

Every time you look at a result and adjust the rule, the rule learns the past a little more. The defence is a period you do not touch until the end.

For strategies built on your own rules, the Lab does this for you: the report covers the first three quarters of the window, and the last quarter stays sealed. You can run the final test on it once per strategy. Tune first, decide, then look.

Step 6: Test it forward, on live order books

A backtest looks back. The next test looks forward: paper trading runs the strategy on live data and scores each signal against the real order book at that moment, without placing an order. It is the closest thing to trading the strategy without trading it, and it is where fill problems show up. See Paper trading on Polymarket.

A worked example

Here is the whole path for one idea, as you would run it in the Strategy Lab. It shows what to do and what to read at each step; it quotes no result, because a result from one period is exactly what this guide tells you not to generalise from.

The idea. "Copy politics trades only when two well-ranked forecasters buy the same side."

1. Ask. Type the idea in plain words. The Lab's assistant turns it into a draft strategy and tells you which wallets it screened and why.
2. Read the settings before you run anything. The price limit per copy, the caps (money open, copies a day, a daily loss stop), and whether a copy is skipped when the leader has already sold.
3. Backtest 90 days. That is the default window; you can choose 30 or 60 days, or your own dates up to 366 days.
4. Read the verdict first, then the baselines. Is it Weak, Promising or Strong, and how does it compare with copying everyone in the same markets?
5. Read the funnel. Signals → Within the rules → Copied. If the book let in only a small share of the signals, the strategy is smaller in practice than it looks.
6. Read "without the best wallet" and "without the best 3 days". This is where most ideas end, and ending them here costs nothing.
7. Read "What a backtest cannot know". The report lists it.
8. If it survives, paper-trade it. A paper run lasts 10 days unless you choose another length, and counts toward 50 resolved copies before it says anything firm.

If the verdict is weak, the report offers Change the rules. Change one thing at a time and keep count of how many versions you tried.

What no backtest can tell you

• The future. Markets, fees and the traders you follow change.
• Your own execution. Delays, mistakes, and the difference between a modelled fill and a real one.
• Crowding. If many people copy the same wallets, the price moves before any of them fills.
• Why it worked. A backtest shows that a rule was profitable in a period, not the reason.

Where to go next

• The Strategy Lab shows real, dated answers before you sign up.
• The docs describe every part of a report: Backtesting and Paper trading.
• The Lab is part of the paid plans; the pricing page lists what each plan includes.

FAQ

How do you backtest a Polymarket strategy?+

Replay the rule over past trades using an entry price a follower could have paid after each signal, model how much of each order the order book would have filled, compare the result with a fair baseline in the same markets, check it again without the best wallet and the best three days, keep one period unseen for a final test, and then paper-trade it on live order books.

How much history should a Polymarket backtest use?+

Enough to produce a meaningful number of trades rather than a fixed number of days. In Polyrank's Strategy Lab a window is 30, 60 or 90 days, or custom dates up to 366 days, and the report's verdict depends on the number of copies: at least 30 for "Promising, not proven" and at least 100 for "Strong evidence".

What is look-ahead bias in a backtest?+

Using information that was not available at the moment of the trade, such as today's trader ranking applied to last month's trades, or the leader's own fill price. It makes a backtest look better than anything that could have been traded.

Is a good backtest enough to trade a strategy?+

No. A backtest is designed with hindsight and its fills are modelled, not executed. Test the strategy forward on live order books with paper trading first, and treat both as simulated results rather than forecasts.