How Many Trades Before an EA Has Proved Anything?

18 November 2026, 10:20
Michael Thomas Harty
0
17

A 70%-win-rate EA with a small, real edge needs about 726 trades before its results clear the usual 95% bar. A zero-edge copy of the same EA has a 55% chance of being in profit after 100 trades.

Both numbers come from the same arithmetic, and together they explain why so many EAs look proven at 100 trades and broken at 400. This post works through that arithmetic on three illustrative EAs, so you can run it on your own.

The one formula

Measure every trade in R, where 1R is your average loss. Then an EA has two numbers that matter:

  • Expectancy (μ): the average result per trade, in R.
  • Spread (σ): the standard deviation of a single trade's result, in R.

The noise in your average shrinks with the square root of the number of trades, so the trades you need before the average is clearly above zero is:

n ≈ (1.96 × σ / μ)²

1.96 is the usual two-sided 95% level. Use 1.645 if you only care about one direction ("is it above zero?"). The ratio σ/μ does all the work. A small edge buried in big swings needs a lot of trades, and because the ratio is squared, halving the edge quadruples the trades.

Three illustrative EAs

Made-up profiles. Every win is the same size and every loss is 1R, which is simpler than real trading but enough to show the shape.

Profile (illustrative) Win rate Avg win Expectancy Spread σ Profit factor Trades for 95% (two-sided) One-sided
A: scalper 70% 0.5R +0.050R 0.687R 1.17 726 511
B: balanced 50% 1.2R +0.100R 1.100R 1.20 465 327
C: trend 35% 2.5R +0.225R 1.669R 1.35 211 149

The trend EA, with its ugly 35% win rate, proves itself fastest, because its edge is large next to its swings. The scalper, which feels the most reliable trade by trade, is the slowest to prove anything. A high win rate tells you how often you win, not how sure you can be that the EA has an edge.

What luck can do in 50, 100 and 200 trades

To see what that means in practice, we gave each EA a zero-edge twin: the same win rate, with the average win shrunk until expectancy is exactly zero. Then we simulated 200,000 runs of each and asked how often the twin looks good.

Profile Trades Zero-edge twin in profit Zero-edge twin at or above the real EA's expected result Zero-edge twin shows PF ≥ 1.3 Real EA at or below zero
A: scalper 50 57.0% 32.8% 22.3% 31.6%
A: scalper 100 55.1% 22.5% 11.4% 22.1%
A: scalper 200 53.5% 15.8% 5.1% 15.8%
B: balanced 100 46.1% 18.5% 9.7% 18.2%
B: balanced 200 47.1% 8.9% 2.8% 9.1%
C: trend 100 54.3% 5.9% 8.8% 8.3%
C: trend 200 52.5% 1.2% 3.3% 3.1%

Read across the scalper's 100-trade row. One time in nine, an EA with no edge at all shows a profit factor of 1.3 or better, which is better than the real scalper's long-run 1.17. And more than one time in five, the real EA is flat or losing after 100 trades. At that sample size, a good EA and a worthless one are hard to tell apart in either direction.

(Twin "in profit" sits near 50% rather than exactly on it because win counts come in whole numbers. It doesn't change the point.)

The same maths runs backwards: spotting an EA that has stopped working

Markets change, and an edge can fade. The formula tells you how long it takes to notice.

If the edge disappears completely, the shortfall against what the EA used to make becomes statistically visible (one-sided 95%) at roughly the one-sided numbers above: about 511 trades for the scalper, 327 for the balanced EA and 149 for the trend EA. Even then you only have about a coin-flip chance of catching it by that point.

If the edge only halves, those numbers quadruple: about 2,046, 1,310 and 596 trades. A scalper doing five trades a day could be running at half its old edge for well over a year before its results prove it.

That's the uncomfortable part. Results alone are a slow alarm. You'll usually see that something has changed in an EA's behaviour (trade frequency, average loss, holding time, drawdown depth against its own history) long before the profit and loss can prove it.

How to use this on your own EAs

  1. Put every EA in R. Divide each closed trade by that EA's average loss. Do it per EA, by magic number, never on the blended account.
  2. Compute μ and σ from the closed trades, then n = (1.96 σ / μ)². Write that number next to the EA.
  3. Below n, treat results as noise in both directions. Don't scale up because a fresh EA looks great, and don't kill a sound one because it's flat. Size it small enough that being wrong about it is cheap.
  4. Watch behaviour as well as results. A change in how the EA trades is an earlier signal than its P&L.
  5. Re-run the number when the EA changes. New settings, a new symbol or a new broker restarts the count.

Steps 1 to 4 need clean per-EA history. VRD Insight is free and splits an MT5 account into one row per magic number with each EA's drawdown and its usual gap between trades, so you can see which EA did what; it doesn't compute expectancy or σ for you, so step 2 is still your spreadsheet.

Tools mentioned

VRD Insight (free): per-EA results by magic number, drawdown and quiet-EA markers, so each EA's history stands on its own.
https://www.mql5.com/en/market/product/204799

VRD Quant ($69, free demo in the Strategy Tester): bootstraps each EA's closed trades into a p95 drawdown and suggested capital, from 10 closed trades up; it sizes the EA while the sample is still small, it doesn't prove the edge.
https://www.mql5.com/en/market/product/204796

Virtual Risk Desk

Educational. The three EA profiles are made-up illustrative examples with fixed win and loss sizes; probabilities come from a 200,000-run simulation and standard sample-size arithmetic on those assumptions. Not a forecast, not a performance claim, and not financial advice. Trading leveraged products carries a high risk of loss.