Backtesting Checklist: 7 Numbers to Check Before Trusting an EA

Backtesting Checklist: 7 Numbers to Check Before Trusting an EA

18 十一月 2026, 06:20
Xiang Jun Jiang
0
2
Most EAs die between the tester and a live account. Not because the tester lies, but because people read only one number from it: the final profit. I have reviewed a few hundred tester reports over the last two years, and the same seven numbers decide whether a system is worth real money. Here they are, in the order I check them.

1. Total trades

If a report shows 300 percent growth on 40 trades, that is not an edge, it is a coin flip with leverage attached. I want at least 200 trades on the tested period before I take any result seriously, and I prefer 500. Fewer trades means the sample cannot separate skill from luck.

Two practical notes. First, check "short trades" and "long trades" separately; a system that only works on longs in a trending market is a long-only system no matter what the code says. Second, if the trade count collapses when you shorten the period, the EA is probably fitting to a handful of big moves.

2. Profit factor

Profit factor is gross profit divided by gross loss. Below 1.0 the system loses money. Between 1.0 and 1.3 it is fragile: a slightly worse spread wipes it out. I want 1.4 or higher before I will even consider live testing, and I treat anything above 3.0 on a short sample as a warning sign rather than a bonus, because it usually means the sample contains one outlier trade.

3. Maximal drawdown, and how it was reached

The percentage matters, but the shape matters more. A 15 percent drawdown spread over eight months is survivable. The same 15 percent reached in two days usually means the position sizing is too aggressive for the account, and the next event will produce 40 percent.

Open the equity curve and look for vertical drops. Every vertical drop is a moment when the EA had too much exposure at once. Find those moments in the trade list and ask: would I have kept the EA running through that?

4. Recovery factor

Recovery factor is net profit divided by maximal drawdown. It answers a simple question: how much do I earn for each unit of pain? I consider 3 or better acceptable on a one-year test, 5 or better good. Below 2, the return does not justify the stress.

5. Average trade duration versus timeframe

Compare the average trade duration with the chart timeframe the EA was designed for. If a scalper built for M1 shows an average trade duration of nine hours, it is not scalping, it is holding, and it is exposed to overnight risk that the author never mentioned.

6. Spread and commission sensitivity

This is the number most people skip, and it kills more systems than anything else. Re-run the same test with a wider spread: 10 points wider for gold, 3 points wider for majors. If the profit halves, the system is trading noise. If the profit becomes negative, the EA is a spread-collection device for your broker.

Also check whether the report used tick data or OHLC. A one-minute OHLC model with tight stops will happily fill you at prices that never existed. If the author only shows an OHLC test on a scalper, that is the whole answer.

7. Equity versus balance curve divergence

When the equity line diverges from the balance line and stays diverged, the EA is carrying open positions for a long time. That is not automatically bad, but it changes the risk profile: your free margin is partially locked, and a margin call becomes possible even if the balance looks fine.

I also want to see the losing streak. Take the longest run of consecutive losses in the report and multiply it by the average loss per trade. If that number is more than 25 percent of the account, the sizing is wrong for a real account, regardless of what the profit factor says.

A fifth-minute sanity check

Before any of the above, look at the tester report header. Check the deposit, leverage, currency and the "modelling quality" line. Twice this year I found systems that looked great because the test was run on a 100,000 deposit with 1:500 leverage and 0.01 lot, which is not the configuration anyone would trade.

Then check whether the EA uses a fixed lot or a percentage of balance. A percentage-based system that grows from 1,000 to 50,000 in the tester will look spectacular, but the drawdown percentage stays constant only because the lots grew with the balance. That is fine, but you must accept the same growth in losses.

One more thing worth doing

Re-run the test on a different period, and on a different broker's data. An EA that only works on one broker's feed in one year is a research project, not a product. I test at least two periods: a trending one and a ranging one. If the EA only survives the trending period, write that down; it is not a flaw in the test, it is information about when to switch it off.

Summary

Seven numbers decide most of it: total trades, profit factor, maximal drawdown and its shape, recovery factor, average trade duration against timeframe, spread sensitivity, and the equity versus balance divergence. Add the losing streak calculation and a second period, and you have a filter that rejects most of what looks good on the first screen.

None of these numbers predict the future. What they do is reject systems whose past is a coincidence, and that is worth more than any additional indicator.

If you want a practical way to run this checklist, keep the tester report next to the trade list and go through the seven items in order, marking each one pass or fail before you look at the profit number again.