How many trades do you need to trust a backtest?

By John, builder of Wise Apple · Updated August 27, 2026

Sixty-eight percent winners over 25 trades. It feels like a lot of evidence. It is roughly the evidence you would have about a coin after 25 flips — which is to say, not enough to tell a fair coin from a slightly rigged one.

Every number in a backtest report is an estimate with an error bar, and the report does not print the error bar. It prints 68%, in a large font, and lets you supply the confidence. This guide supplies the arithmetic instead: how wide the error bars actually are at different trade counts, why the count that matters is per regime rather than in total, and why the number of ideas you tested changes what any one of them can prove.

None of it needs more than a square root. All of it is the reason "30 trades" is a floor and not a target.

Where 30 comes from, and what it does not mean

Thirty is the number statistics teachers use for when a sample mean starts behaving like a normal distribution. It is a threshold for when the maths starts working, not for when the answer is reliable. At 30 trades the maths will happily tell you that your win rate is 68% plus or minus 17 points — which is a precise statement of not knowing.

The error bar on a win rate

For a win rate p measured over N trades, the standard error is √(p(1−p)/N), and a rough 95% interval is about twice that on either side. Here it is for a strategy whose true win rate is 55% — a good, realistic number:

Read the first row again. At 30 trades, a measured 55% is statistically compatible with a true 40% — a losing strategy at most payoff ratios. The report showed you 55%. The uncertainty was the other twenty points, and it was invisible.

The error bar on expectancy

Win rate is the easy case. What you actually care about is expectancy — average profit per trade — and its error bar depends on how spread out your trade results are. Measure results in R-multiples (profit divided by initial risk). A typical strategy's trade outcomes have a standard deviation somewhere around 1 to 2R, so take 1.5R. The standard error of the mean is then 1.5 / √N:

• At 30 trades: ±0.27R. An expectancy of +0.2R is within one error bar of zero. Unmeasurable.
• At 100 trades: ±0.15R. +0.2R is just over one error bar from zero. Suggestive.
• At 225 trades: ±0.10R. +0.2R is two error bars from zero. This is roughly where a modest edge becomes distinguishable from luck.
• At 900 trades: ±0.05R. Now you can start to ask whether it is +0.2R or +0.15R.

The general rule: to distinguish an expectancy E from zero at about two standard errors, you need roughly N ≈ (2σ / E)² trades, where σ is the spread of your trade results. Small edges with wild outcomes need enormous samples. A +0.1R edge with a 2R spread needs about 1,600 trades. That is not a strategy most people can test on daily bars in a lifetime of data, which tells you something about how many daily-bar strategies with tiny edges have ever been genuinely validated.

Trades per regime, not trades in total

The arithmetic above assumes each trade is an independent draw from the same distribution. Markets do not cooperate. Two hundred trades taken during one three-year bull run are two hundred draws from that market, and they tell you very little about a bear market, a sideways one, or a volatility spike.

So split the trade list by condition and count each bucket. Rising market, falling, sideways. High volatility, low. If any bucket has fewer than a few dozen trades, the strategy has not been tested in that condition, whatever the total says — and the condition with the empty bucket is usually the one that ends the strategy live.

This is why ten years of data is a reasonable minimum on daily bars. It is not the number of bars that matters. It is that ten years almost always contains at least one stretch you would have hated, and a strategy that has not met that stretch is untested in the only condition that matters.

The multiple-testing multiplier

Here is the part that undoes most sample-size arithmetic. Suppose you need 200 trades to trust one strategy. You test twenty strategies and keep the best. The best of twenty random strategies will look good — that is what "best of twenty" means — so the evidence required to believe it is much higher than for a strategy you tested once.

Harvey, Liu and Zhu's 2016 paper "…and the Cross-Section of Expected Returns" made this argument about the academic literature: after hundreds of published factors, the bar for a new one to be believed should be roughly a t-statistic of 3, not the traditional 2. The same logic applies at your desk. Every idea you tried and discarded raises the bar for the one you kept. Bailey and López de Prado's "deflated Sharpe ratio" (2014) is the formal version: it discounts a strategy's Sharpe by the number of trials that produced it.

What to do with all this

1. Print the error bar. For any win rate, compute √(p(1−p)/N) and double it. If the interval includes a losing number, you do not know the strategy wins.
2. Count per regime. Split trades by market condition. Empty buckets are untested conditions.
3. Get more trades honestly. More history, more instruments tested one at a time, or a shorter timeframe if the idea genuinely applies there. Do not get them by loosening the entry rule; that changes the strategy.
4. Count your attempts. Divide your enthusiasm by the number of things you tried.
5. Let time add samples. Freeze the rules and run them forward. Every new bar is a trade the optimiser never saw.

Sample size sits alongside overfitting as the reason the win rate in a report means less than it looks, and the report-reading guide puts trade count first for exactly this reason: it is the number that tells you whether any of the other numbers are numbers.

How Wise Apple reports sample size

Every Wise Apple report leads with N — the trade count — and the classification metrics beside it (precision, recall, MCC) are computed on the out-of-sample windows only, so the count that matters is the one shown. The Walk-Forward Window grades the strategy across several successive periods, which is a cheap way to see whether the numbers hold up regime by regime rather than in aggregate. And because ten-plus years of history are loaded for each instrument, the empty-bucket problem is usually a choice rather than a limitation.