Direct answer
Responsible backtesting of MT5 Expert Advisors means you test a well-defined EA against historical data using explicit assumptions for execution, costs, and risk handling—then verify the result with out-of-sample testing and bias controls. Historical performance is not proof of future results, so the goal is to measure robustness and identify failure modes rather than to predict.
Mechanism and definition
An MT5 Expert Advisor (EA) is an automated trading program that turns defined rules into actions. In a backtest, the platform simulates how those rules would have behaved over past price data. Because the simulation depends on inputs and assumptions, “backtesting responsibly” starts with defining what you are actually testing:
- Data definition (input reality): specify the historical timeframe(s), symbol(s), and whether the test uses only bar data or more granular ticks. If you use bar data, intrabar price movement and order timing are approximated.
- Assumptions (simulation mechanics): define how orders are filled (especially for market vs. limit orders), what “spread” you assume, and how slippage is modeled.
- Costs (kostensoorten): include recurring and event-like costs such as commissions, swaps/rollover effects, and realistic spread assumptions. If you leave out a kostensoort, the backtest can become systematically optimistic.
- Variabele factoren (what changes in real markets): liquidity, volatility, and execution quality can change over time, so you must treat them as variable rather than fixed.
A simple example helps: if your EA uses stop-loss and take-profit levels, then the backtest outcome depends on whether the simulated price path touches those levels and in what order. If your data resolution cannot represent that path, results can misstate how often exits would occur.
Evidence and example checks
A responsible workflow uses multiple evaluations designed to expose fragility:
- Out-of-sample testing: Split history into at least two parts. Use one segment to evaluate design choices and another segment only to assess performance. If results collapse on the second segment, the EA may be overfitted.
- Walk-forward testing: Instead of one split, repeatedly train/evaluate on rolling windows. This reduces the chance that you accidentally “fit” one period’s peculiarities.
- Bias controls (controlebron): watch for common failure modes:
- Overfitting: tuning many parameters to one dataset.
- Look-ahead bias: using information that would not have been known at the time.
- Survivorship bias: only evaluating instruments or history that remain relevant.
- Stress under different regimes: evaluate how the EA behaves when volatility rises, correlations change, or trend conditions reverse. This targets robustness instead of a single “average” result.
When you document these steps, include the assumptions you used for costs and execution. Outcomes should be judged relative to those assumptions, not as absolute truth.
Limitations and risks (including at least one failure mode)
Even with careful methodology, backtesting can fail in material ways:
- Data limitations: If the data resolution is too coarse (for example, only OHLC bars), order timing and intrabar level crossings can be inaccurately represented.
- Execution mismatch: Real trading can differ from the simulator in fill quality. Slippage and spread variations can change risk and reward.
- Structural regime shifts: Markets can change permanently. Historical relationships do not establish future results.
- Stop/exit modeling error: The sequence of price movements around stop-loss and take-profit is critical; small modeling differences can materially change outcomes.
Verification and next question
To verify whether your backtest is credible, you should be able to answer—using your own documented assumptions—how the EA would execute under alternative spreads/slippage assumptions and how performance changes across out-of-sample periods. If you cannot clearly explain (1) the data resolution, (2) the kostensoorten included, (3) the aannames for fills, and (4) the out-of-sample and walk-forward results, then the backtest is not yet “responsible.”
If you share your backtest setup at a high level (data type, timeframe, order type used, and cost assumptions), you can also compare whether the results are robust to changes in those aannames and variabele factoren.