Why your live win rate is lower than your backtest win rate

The short answer

Four mechanisms account for most of the gap, and they are testable in order. Did you select this strategy because of its win rate? Were intrabar outcomes assumed in your favour? Were realistic costs included? Is the sample big enough that the difference is even meaningful?

Working through those four questions takes an afternoon and will usually locate the gap, because in most cases the gap is not mysterious. It is one of these four, and often more than one at once.

One: the number you selected on is the number that is inflated

If you tested many strategies and kept the one with the best win rate, you have selected the maximum of a set of noisy estimates. The maximum of many noisy numbers is biased upward, every time, even when no strategy has any edge at all.

This is why a strategy can go from 62% in testing to 50% live without anything breaking. The 62% was partly real and partly the luck of that particular sample; when you selected it, you selected the luck too. Live trading draws a fresh sample, and the luck does not come with it.

The defence is not to stop searching. It is to record how many things you tried, and to be more suspicious of a number precisely because you chose it. A win rate that survived selection is a weaker piece of evidence than the same win rate that fell out of a single pre-specified test.

Two: intrabar assumptions decide the outcome

On any bar where the price touches both your stop and your target, the backtest has to decide which happened first. It cannot know, because the bar only records the open, high, low and close. So it applies an assumption.

The common assumption is favourable to the strategy, which means some of the wins in your backtest were trades that live would have been losses. A strategy with tight targets and tight stops trades inside the bar constantly, so this affects a large share of its trades rather than an occasional one.

The test is straightforward: switch the modelling from open prices only to every tick, and re-run on the same window. If the win rate moves materially, your reported figure was partly an artefact of the fill assumption and the live number should be expected closer to the pessimistic version.

Three: costs are subtracted from the winners too

Win rate is usually reported before costs, because costs are usually configured separately and often left at zero. But spread and commission are charged on every trade. They do not change which trades are winners and losers by much in a low-frequency strategy; in a high-frequency strategy they change plenty.

Consider a strategy that targets a small profit and a small loss with a tight stop. The spread on entry and exit can be a significant fraction of the target. Trades that cleared the target by less than the spread in the backtest become losses live, and the win rate falls by exactly that fraction.

Add the costs explicitly and re-run before going live. If the win rate survives realistic spread, commission and swap, you have eliminated one of the four mechanisms.

Four: the difference may not even be significant

A win rate is an estimate, and estimates from small samples are uncertain. If your backtest won 62% of 40 trades and live you are at 50% after 25 trades, those two numbers are not yet distinguishable. At 40 trades, the standard error on a 62% win rate is about 7.7 percentage points; the 95% range around it spans roughly 47% to 77%.

That is the honest reading: a live win rate below the backtest number is not by itself evidence that anything is wrong. It is evidence that you have fewer observations than the question requires. Reporting both the win rate and the trade count, always together, is the only way to avoid this confusion.

What our own terminal checks showed

We run every strategy that survives our filters on a real MetaTrader terminal before publishing, comparing backtest results against what the terminal produced on the same instrument.

Of the 1,315 strategies in our verification record, 631 could not be run on that terminal and are reported separately rather than counted either way. Of the rest: 494 verified, 84 came out weak, 63 failed, 29 diverged and 14 mismatched.

Looking specifically at strategies whose backtest return was positive, 559 had comparable figures on both sides. Of those, 25 — about 4.5% — came out flat or negative on the real terminal. Read that number carefully, because the interesting part is the denominator: these are strategies that had already passed our filters, and still roughly one in twenty flipped sign when a program we did not write read a feed we do not control.

The remaining failure modes are not about the sign flipping. Weak, failed, divergent and mismatched are separate outcomes we report separately, because collapsing them into a single failure count would overstate the problem in one direction and hide it in the other.

How to diagnose your own gap

Start with the sample. If your live trade count is under 100, the gap may be noise and you should not act on it yet — check the arithmetic before concluding anything.

Then compare the trade lists, not the summary statistics. Put the backtest trades and the live trades side by side by entry time. The differences will cluster: entries that live took a bar later, exits that live filled worse, trades the backtest took that live never saw. The clustering tells you which mechanism is operating.

Then re-run the backtest with pessimistic assumptions — worst-case spread, every-tick modelling, full costs — and see whether the live number falls inside the range the pessimistic version produces. If it does, the strategy is behaving as expected and the original number was the optimistic tail.

What this does not mean

It does not mean the backtest was fraud or the platform is lying. A backtest is a simulation with stated assumptions; the gap is what happens when the assumptions meet a market that is not obliged to honour them.

It does not mean every strategy underperforms live by a fixed amount. There is no such constant. A patient strategy on a liquid instrument with wide stops will show a much smaller gap than a fast strategy with tight targets.

And it does not mean the live number is the truth either. Live results from a small number of trades carry their own error bar. The point is not that live is right and backtest is wrong; it is that neither is meaningful without the trade count that produced it.