Your EA Wins 80% of Trades. So Why Is It Losing Money?

Your EA Wins 80% of Trades. So Why Is It Losing Money?

20 November 2026, 00:00
Lee Kai Lun
0
8

Imagine an Expert Advisor with a dashboard that proudly announces an 80% win rate. Eight trades out of ten finish green. The statistics look rather pleased with themselves.

Then you open the account history and discover the net result is negative.

The EA isn't necessarily broken. The arithmetic might be doing exactly what you told it to do.

Win rate tells you how often you win. Trading expectancy tells you what those wins and losses are worth together. For an automated strategy, that distinction can separate a useful test from a dangerously flattering screenshot.

The number your win rate forgot to mention

The basic expectancy formula is:

Expectancy = (Probability of win × Average winning trade) − (Probability of loss × Average losing trade)

Here, average loss is expressed as a positive magnitude. Use the same units for both averages. If you calculate from net trade outcomes, account for commissions, swap and execution costs consistently.

Let's invent a purely illustrative 100-trade EA test:

  • 80 winning trades, averaging $20 each.
  • 20 losing trades, averaging $100 each.

The calculation is (0.80 × $20) − (0.20 × $100) = −$4 per trade.

Across 100 trades, that's −$400, before any costs not already included in those figures.

An 80% win rate has just lost its invitation to the victory party.

Now imagine a different hypothetical test: 40 wins averaging $150 and 60 losses averaging $75.

Its expectancy is (0.40 × $150) − (0.60 × $75) = +$15 per trade, or +$1,500 across the illustrative 100 trades, again before uncounted costs.

That second strategy loses more trades than it wins. Neither example proves a strategy will work in live markets. They simply show why win rate alone is a poor judge.

How to calculate expectancy from MT5 history

Export or review a defined period of completed trade results. Include the full set, not just the photogenic weeks.

Count winning, losing and zero-result observations. Divide each count by the total number of observations, including zero outcomes. Then calculate average net win and average net loss. A flat trade contributes zero; it should not be quietly promoted to a winner.

For a straightforward system where one completed position represents one strategy decision, you can compare the result with:

Total net outcome ÷ Number of completed trades

Both methods should agree if your classifications, units and costs match.

But be careful with multi-position EAs. A grid or basket system may generate several closing deals from one underlying trading decision. Partial closes can also create multiple deals. MT5 deals, orders and positions are not interchangeable units of analysis.

For basket strategies, consider calculating expectancy per completed basket or strategy cycle as well as per closed deal. Otherwise, one difficult market episode can appear as dozens of independent observations when the risks were actually linked.

Trading costs: the tiny bill that keeps arriving

Suppose an EA's gross expectancy is +$3 per trade, but average additional execution costs are $4. Its net expectancy becomes −$1.

That sounds trivial until the system trades frequently.

Spread is already reflected in actual Bid/Ask execution prices, so don't subtract it a second time from net historical results. Commissions and swap may be reported separately, depending on the data you export. Slippage is reflected in realized fills, but an idealized backtest may underestimate its effect.

When reviewing MT5 Strategy Tester results, ask whether the test's symbol specifications, spread assumptions, commissions, execution model and tick data resemble the conditions you intend to trade. A profitable-looking model with unrealistic costs is a demonstration of the model, not evidence of a live edge.

Why positive expectancy can still hurt

Even a strategy with positive estimated expectancy can suffer a long losing streak, severe drawdown or a ruinous oversized position.

Consider two systems with the same average outcome. One makes modest gains and losses. The other depends on a rare enormous winner while absorbing many small losses. Their average can match, but their capital requirements and psychological demands may be wildly different.

The sequence matters, too. Ten losses before a recovery can trigger margin pressure or a risk limit long before the eventual winning trade arrives.

And one optimized backtest isn't enough. Parameter hunting can manufacture a handsome historical average. Test unseen periods, different market conditions and realistic costs. Look at the distribution of outcomes, not just the final average.

A practical five-check EA audit

Before increasing lot size because the win rate looks impressive, check:

  1. Observation unit: Are you measuring deals, completed positions or complete baskets?
  2. Net results: Have commissions, swap and realized execution effects been handled without double counting?
  3. Outliers: Does one exceptional winner account for most of the positive expectancy?
  4. Stability: Does the result survive different periods and genuinely out-of-sample tests?
  5. Survival: Can the account withstand plausible drawdown, adverse fills and losing sequences?

For comparisons across different position sizes, expressing outcomes in multiples of initial planned risk (R) can help. Just remember that basket systems with changing exposure need a clearly defined risk unit. Calling every trade “1R” does not make its risk fixed.

The better question

A win rate is useful. It's just not a verdict.

The next time an EA reports 80% winners, don't immediately ask how to make it trade more often. Ask what the average winner earns, what the average loser costs, whether the sample is meaningful and how much capital the system needs to survive.

Your EA can be right most of the time and still be expensive to own. The calculator, unlike the dashboard, has no reason to flatter you.