Backtesting explained: why 80% of bot results don't work in practice

Pitfall 1: tick data versus 1-minute data

MT5 default backtest uses 1-minute candles. But during news or flash crashes the market moves 50 pips inside a minute. Your SL can be hit at a price not visible in candle data.

Realistic: use real tick data from your broker history, or use the MT5 'Every tick based on real ticks' modeling mode. Win rates typically drop 5-15% versus 1-minute backtests.

Pitfall 2: spread modeling is wrong

MT5 backtest often uses a fixed spread that is the current market spread. In reality spread varies per hour (Tokyo session wide, London open narrow, news event extreme).

Bots that are break-even or small-win on fixed spread often turn out losing on variable spread. Always test with spread-shock stress tests.

Pitfall 3: optimization (curve fitting)

Tester lets you test 50 parameter combinations, you pick the best, that one has 90% win rate over 2 years history. Live: 50% win rate.

Reason: you fitted parameters on the specific noise of that period, not on a real edge. When market regime changes, everything falls apart.

Anti-curve-fit: backtest on period A (2 years), validate on period B (6 months out-of-sample). If A and B are not comparable, your strategy is curve-fit.

Pitfall 4: survivorship bias

A seller shows you a bot with 1 year history. What you don't know: he tried 50 bots, stopped 49 because they lost, this one still works. Statistically: even random bots have a chance that 1 in 50 gives a great backtest.

Always ask: how many other strategies did you try before picking this one? No answer = survivorship bias.

How we do it

Our bot was built in 2023, tested in 2023-2024 on multiple broker feeds, run live on a €1,000 personal account for 8 months before any members were allowed to connect. The live numbers from that test match our backtest within 1.5% per month. That is what you want to see.