Methodology

This page documents how every number on this site is produced: where the data comes from, how the backtest runs, how the A–F grade is calculated, and how we verify that a generated EA behaves the same as the engine that tested it. Where something is a limitation, we label it as one.

1. Historical data

• Data source is Dukascopy minute-level (M1) history, timestamped in UTC.
• Coverage is currently 14 instruments, starting January 2022, and is updated daily through the previous day's close.
• When no date range is given, backtests use a rolling 24-month window.
• Higher timeframes (M5 through D1) are aggregated from the M1 series, not downloaded separately.

2. How the backtest runs

• Bar-by-bar simulation. Signals are evaluated once per bar close; no future data is read.
• Stops and targets are checked at M1 precision — not merely against the high and low of the strategy-timeframe bar.
• If both stop and target are touched inside one bar, the unfavourable side fills (a deliberately conservative assumption).
• A fixed per-instrument spread is charged; entry prices include the configured slippage. Commission and swap are not modelled (see section 6).
• In multi-timeframe strategies, a higher-timeframe indicator only ever reads closed higher-timeframe bars.

3. Out-of-sample split

By default the final 20% of the period is held out as an out-of-sample segment. It takes no part in parameter tuning, and the report scores its trades, net profit, profit factor and win rate separately. A large gap between the two columns usually means the parameters are overfit.

4. How the A–F grade is calculated

Out of 100 points, weighted across five components:

Letter bands: A ≥ 80 · B ≥ 65 · C ≥ 50 · D ≥ 35 · F < 35. One hard rule overrides the rest: if net profit is negative the score is capped at 40 — a losing strategy can never grade above D.

5. Does the generated EA match the backtest?

This is the step most tools skip and the one that matters most: an accurate backtest is worthless if the MQL5 code it hands you behaves differently. We run the same strategy on both sides and diff the fills:

• Dukascopy data is imported into MT5 as a custom symbol and the generated EA runs in the MetaTrader 5 Strategy Tester;
• its fills are diffed against our engine's over the same period.
• Measured result: on single-timeframe strategies, 48 of 48 trades matched, entry prices identical in 48 of 48, exit reasons identical in 48 of 48. On a multi-timeframe (triple screen) strategy, 17 of 17 matched.
• Known benign difference: the tester's tick generation can overshoot a stop fill by up to roughly 1.2 pips.

Every generated EA is also compiled in MetaEditor before delivery and must report zero errors.

6. Limitations (known and stated)

• Commission and swap are not modelled. Strategies holding positions for long periods will underperform their backtest in reality.
• Spread is simulated as a fixed value, so widening around major news releases is not reproduced.
• Partial fills, rejections and slippage distributions are not simulated; orders are assumed to fill at the expected price (stops excepted).
• Your broker's data differs from ours — different feed, different server timezone (usually EET), different spread model. Re-testing in your own MT5 will produce different numbers. Judge a strategy by whether the trend and statistical profile agree, not trade-for-trade (more in the school).
• Past performance does not indicate future results. A backtest filters out what clearly doesn't work; it does not forecast.

7. What we hold ourselves to

• Every figure on this site comes from the engine described above. We don't publish results we can't reproduce.
• Backtest reports are never paywalled and never softened — a strategy that deserves an F gets one.
• If the methodology changes, this page is updated.