What is a good win rate for a trading bot?

The number that matters is expectancy — win rate combined with average win and average loss. A 35% win rate at 3:1 gives 0.4R per trade; a 60% win rate at 0.5:1 gives −0.1R and loses money while winning most of its trades.

Win rate is nonetheless the most heavily advertised metric, because it is the most intuitive and the easiest to make impressive. A system taking tiny profits and letting losers run can show 90% and be one trade from disaster, which is exactly the profile of most martingale products.

Our own figures are 44.8% in-sample and 36.5% out-of-sample at a fixed 2:1 ratio. Both are above the 34% break-even for that ratio, and the gap between them is a measure of how much of the in-sample number was fitting rather than edge.

How many trades before the number means anything

Thirty trades is enough to notice a very large effect and nothing else. A hundred begins to be informative, and several hundred is where win rate stabilises enough to be treated as an estimate. A vendor quoting a precise win rate after a few weeks of live trading is quoting noise — see sample size.

This is also why our engine scores strategies on the Wilson lower bound rather than the raw rate. A strategy that won its only two trades has an observed 100% win rate and deserves no confidence at all.