How many trades does a backtest need before you can trust it?

Ask a statistician how big a sample you need and they'll answer with a question: how big is the effect you're trying to detect, and how noisy is the data? Trading answers both questions unkindly. The effects are small — a real edge might shift your average trade by a fraction of a percent — and the noise is enormous. That combination means backtests need more trades than intuition suggests, and far more than most screenshots have.

The rule-of-thumb ladder

There is no magic threshold, but the working ranges are well established:

• **Under 30 trades:** an anecdote. The outcome is dominated by a handful of lucky or unlucky fills; a single outlier trade can flip the verdict.
• **30–100 trades:** a hypothesis. Patterns start to separate from noise, but confidence intervals around your win rate and average trade are still embarrassingly wide.
• **100–300 trades:** a result worth validating. Statistical tests start to have real power here — this is where out-of-sample splits and walk-forward analysis become meaningful rather than theatrical.
• **300+ trades:** a result worth arguing about. At this size, degradation between in-sample and out-of-sample performance is signal, not noise.

To feel why, take a strategy with a true 55% win rate — a genuine edge. Over 20 trades, plain binomial math gives it roughly a one-in-three chance of showing 50% or worse. Your honestly profitable strategy fails the eyeball test a third of the time at that sample size — and an honestly *worthless* strategy passes it just as often.

Why calendar time doesn't substitute

"Backtested over five years" sounds rigorous, but a position-trading strategy that signals twice a quarter produces forty trades in those five years — an anecdote with a long runway. The reverse error also exists: a thousand trades generated inside six months of a single roaring bull market is a large sample *of one regime*.

That's the second requirement, independent of count: the trades should span different market personalities — trend, chop, collapse. A strategy meeting its first bear market in production is being validated with real money. Our market regimes explainer covers why strategies are regime-dependent almost by definition; the practical consequence is that a good backtest window includes at least one regime the strategy should *dislike*.

The quiet sample-shrinkers

Two things make your effective sample smaller than the trade count suggests:

• **Correlated trades.** Forty trades that are mostly the same breakout on correlated alts, taken the same week, are closer to five independent observations than forty. Independence is what statistics actually counts.
• **Multiple testing.** If you tried 50 parameter combinations and kept the best, your best backtest's trade count overstates its evidence — some configuration was always going to look good. This is the selection-bias problem the [deflated Sharpe ratio](/blog/deflated-sharpe-ratio) exists to correct, and it's why our engine asks how many variants you tried, not just how the winner performed.

What to do with a small sample

If your strategy simply doesn't trade often, you have honest options. Test it across several uncorrelated markets with identical parameters — ten coins with the same rules multiplies observations without multiplying tuning. Drop to a finer timeframe only if the edge plausibly survives the extra costs (fee drag grows brutally as trade frequency rises — see slippage and trading costs). Or accept the uncertainty explicitly: size the strategy as the hypothesis it is, not the certainty the equity curve pretends to be.

And whatever the count, split the data. A hundred trades where the last thirty were never touched during development beats three hundred trades of pure in-sample optimism. Our free backtester runs that split automatically — it takes about a minute to learn whether your sample is telling you something or just talking.