Most EA buyers size their account from one number: the maximum drawdown in the backtest or the signal history. "Max DD 15%, so I'll be fine with a 20% stop." It feels conservative. It usually isn't, and the reason is simple arithmetic rather than anything wrong with the EA.
The maximum drawdown you see is the worst drawdown that happened in that particular sample of trades. Run the same strategy for longer, or on a different sequence of the same kind of trades, and the worst drawdown gets bigger. Not because the EA broke, but because a longer sample gives bad streaks more chances to line up.
This post puts numbers on that so you can size your account and your drawdown stop with your eyes open.
The model (illustrative)
- Fixed-fractional EA risking 1% of equity per trade
- 45% win rate, winners +1.5R, losers −1R (a positive edge of +0.125R per trade)
- Trades independent of each other
- 20,000 simulated trade sequences per row
This is a deliberately healthy EA. Every sequence comes from the same strategy with the same edge. The only thing that changes is how many trades we look at and which random order they arrive in.
1. Max drawdown grows with the number of trades
| Trades in sample | Median max DD | 95th percentile max DD | Median longest losing streak | 95th pct streak |
|---|---|---|---|---|
| 100 | 9.1% | 17.2% | 6 | 11 |
| 250 | 12.7% | 22.4% | 8 | 12 |
| 500 | 15.7% | 26.5% | 9 | 14 |
| 1,000 | 18.8% | 30.2% | 10 | 15 |
| 2,500 | 23.2% | 34.2% | 12 | 16 |
Same EA, same settings. A typical 500-trade history shows a max DD of around 16%. Run it for 2,500 trades and the typical worst drawdown is around 23%, and one sequence in twenty goes past 34%.
So "max DD 15%" on a 500-trade backtest describes the past 500 trades. It's a poor estimate of the worst drawdown over the next three years.
2. How often does live trading beat the backtest's worst?
Take a 500-trade backtest, note its max DD, then trade the same EA live. How often does the live max DD end up worse than the backtest's?
| Live trades after the backtest | P(live max DD > backtest max DD) | P(live max DD > 1.5 × backtest max DD) |
|---|---|---|
| 250 | 31.7% | 8.4% |
| 500 | 50.4% | 17.1% |
| 1,000 | 68.3% | 29.3% |
| 2,000 | 81.7% | 44.5% |
Over a live period as long as the backtest, it's a coin flip. Over four times as long, the backtest's worst drawdown gets beaten more than four times in five, and in almost half of sequences it's beaten by 50% or more.
And that's with an edge that never decays. Real EAs face changing spreads, slippage, regime shifts and over-fitted parameters, all of which push live drawdowns higher, not lower.
3. The lucky backtest problem
Backtests and signals that get shared, sold or chosen tend to be the ones that look best. In the model, if we keep only the backtests whose max DD landed in the best quarter (12.8% or lower), then trade 500 live trades:
the live max DD is worse than the backtest's in 87.2% of cases.
That's selection bias. If you picked an EA (or a set file, or an optimisation pass) because its drawdown was low, expect the live drawdown to be higher. The low number was partly luck, and luck doesn't carry forward.
4. Losing streaks: a quick formula
The longest run of losses in n trades with loss probability q is roughly:
Longest losing streak ≈ ln(n × q) / ln(1 / q)
With q = 0.55 (a 45% win rate):
| Trades (n) | Approximate longest losing streak |
|---|---|
| 100 | 6.7 |
| 500 | 9.4 |
| 1,000 | 10.6 |
| 2,500 | 12.1 |
This lines up with the simulated medians above. It grows slowly, but it always grows. At 1% risk, a 12-loss streak alone is about an 11.4% drawdown (0.99^12 ≈ 0.886), before counting any losses either side of it.
5. What to do with this
- Note the sample size behind any max DD. "15% over 300 trades" and "15% over 3,000 trades" are very different claims.
- Size for a multiple of the backtest's drawdown, not the drawdown itself. In this model, the 95th percentile drawdown over 2,500 trades is about 2.2× the median drawdown of a 500-trade backtest. Your EA's multiple will differ, so run your own Monte Carlo on its trade list (MT5 lets you export the trade history from the Strategy Tester report).
- Set your account or per-EA drawdown stop above the realistic future worst, not at the historical worst. A stop placed at the backtest's max DD will trigger on a perfectly healthy EA roughly half the time over a comparable live period.
- Distrust a drawdown you selected for. If you chose the EA or the set file because its drawdown was the lowest, add extra margin.
- Re-estimate as live trades come in. Live history is the only sample that includes your broker, your VPS and your spreads.
We track per-EA drawdown against limits like these at Virtual Risk Desk, but the method above needs nothing more than a trade list and a few lines of code.
If you would rather not write that code, VRD Quant runs 2,000 bootstrap resamples of each EA's closed trades on your account and shows a 95th-percentile drawdown and expectancy for each EA, from 10 closed trades upward. It reads live account history, which fits step 5. Like any resampled number it is an estimate, and live drawdown can still exceed it.
Tools mentioned
VRD Quant ($69): 95th-percentile drawdown and suggested operating capital per EA, from its own closed trades. Read-only.
https://www.mql5.com/en/market/product/204796
VRD Insight (free): per-EA results next to the account's maximum drawdown, so you can re-check as live trades come in.
https://www.mql5.com/en/market/product/204799
Virtual Risk Desk
All figures are illustrative outputs of the stated model (fixed 1% risk, 45% win rate, 1.5R winners, independent trades). They are not a backtest of any real EA, a forecast, or a performance claim.
