How many trades should you backtest?
385
trades needed
could measure 45% to 55%
Under 90%, a strategy with no edge clears the range. Lowering it buys a smaller number, not more certainty.
How to read this chart
Let's start with the default parameters.
- true win rate of 50%
- 95% confidence
- 385 trades needed
The whole chart reads as the following sentence.
If your trading strategy's edge has a win rate of 50% when backtested over 385 trades, then you can be confident that 95 runs of taking 100 trades will result in your actual win rate to fall between 45% and 55%.
The remaining 5 runs would land somewhere outside of that range. That is exactly what 95% confidence means, and why the number is not a guarantee.
If you drag the movable line to 30 trades, that same 50% win rate strategy can result in anything from 32% to 68%. So a trader who only backtests 30 trades and gets a 65% win rate has not necessarily found an edge.
They could have very likely found, through luck, the upper range bound of what is expected for a 50% win rate.
That is the question the range answers: how far from the truth can the number in front of you be? The number of backtested trades is the only thing that reduces the range.
How to use it
- Drag the white line across the chart to find out the correlation between the number of backtested trades and the range you can expect your win rate to fall into.
- Move the "precision wanted" slider and the white line auto jumps to the number of backtested trades needed for that precision level.
- "True win rate" is the edge you are assuming your trading strategy truly has. The chart has the widest range at 50% win rate.
- Confidence is how often the range should contain the truth. 95% is the standard. The range turns amber when under 10 precision points and green under 5.
The math, and what it assumes
The range is the normal approximation to a binomial proportion, the interval usually called Wald:
p ± z · √( p(1−p) / n )
Inverting it for a target margin E gives the trade count: n = z² · p(1−p) / E²
Why Wald and not Wilson
This chart runs forward: it starts from a win rate you assume is true and asks what a sample of n backtested trades could measure. That is a sampling interval, and the normal approximation is the right tool for it.
Wald's known weaknesses all belong to the backward question, which is estimating an unknown win rate from one observed result. That question needs the Wilson score interval instead, and it is a different tool.
Where it breaks
- It assumes independent trades and a win rate that never moves. Real trading and markets give you neither, because edges decay and losses can impact future trades. Read the total backtested trades number as a minimum amount.
- The range is widest at 50%, because p(1−p) peaks there. A strategy claiming 85% win rate needs fewer trades for the same level of precision. That is what the math says and not an error.
- Win rate on its own is not an edge. It knows nothing about your risk-to-reward ratio. So an accurate 55% win rate could still be a losing system.
- Under about 30 trades the approximation is not very reliable. Read the range as being inaccurate rather than it being precise.