A backtest showing a perfect equity curve is almost always a lie — not out of dishonesty, but out of over-optimization. That's the overfitting trap, and it's the first thing to understand before trusting an Expert Advisor, yours or a seller's.
What is overfitting?
Overfitting means tuning a strategy so finely to the past that it ends up "memorizing" history rather than capturing reproducible logic. The result is spectacular on test data… and collapses live, because the future never looks exactly like the past. The more parameters and filters you add to "improve" a backtest, the higher the overfitting risk.
Signs of a too-good backtest
- Too-smooth equity curve: reality is bumpy. A perfect rise is suspicious.
- Too many parameters: each extra "optimal" setting is a chance to fit noise.
- Single period: testing on one favourable window proves nothing.
- No realistic costs: ignoring spread, slippage and commissions inflates results artificially.
How to backtest honestly
A few principles that separate a credible test from an illusion:
- Realistic data: include real costs and stress periods (not just easy markets).
- Out-of-sample validation: reserve a part of history the strategy never "saw", and judge on that.
- Walk-forward: re-optimize then test on the next period, repeating — it simulates a system's real life.
- Simplicity: prefer robust logic with few parameters over a "perfect" machine on paper.
The link with past performance
A regulatory and common-sense reminder: past performance does not predict future results. A good backtest isn't there to promise a gain; it's there to eliminate bad ideas and understand a system's behaviour (notably its drawdown) before exposing capital.
Why it's central for AI
AI systems are, by design, excellent at memorizing the past. That's exactly what makes overfitting even more dangerous in reinforcement learning. The defence is always the same: out-of-sample validation, guardrails, humility. That's the spirit in which Adestto AI designs and evaluates its bots — read how our bots learn from their mistakes and the design framework.
For the practical foundations, see the MT5 bots guide. To test other common misconceptions about automated trading, see 10 myths true or false.
Educational content. Adestto AI is a software and educational-content publisher — not a broker or an investment adviser. Past performance does not predict future results; trading involves risk.