Why Backtests Can Be Misleading

When the Strategy Tester runs an EA over five years of EUR/USD, it is not replaying the market you would have traded. It replays one broker’s stored price history, through a simplified execution model, with costs you configured yourself. The report at the end is exactly as honest as those three ingredients.

None of that makes backtesting useless — it is the right first filter for any rule-based idea, and how the MetaTrader tester works is covered separately. This guide is about the gap between the tester’s tidy world and a live account — a gap that is almost always in the strategy’s favor.

Historical prices contain some repeatable structure and a large amount of one-off noise. An optimizer cannot tell the difference. Give it four inputs with eight values each and it will run 4,096 separate backtests and hand you the best one. Some combination will look excellent on anydata — including random data — because picking the maximum of thousands of tries guarantees an impressive winner.

The symptom of curve fitting is a near-perfect historical equity curve that collapses when a parameter moves slightly or new data arrives. The more inputs a strategy has, and the harder they were tuned, the more of its backtest profit is noise that will not repeat.

A tester needs every tick, but stored history rarely has them all. MT4’s “Every tick” mode interpolatesticks inside one-minute bars: the platform invents the path price took between a bar’s open, high, low and close. A strategy that lives intrabar — tight stops, scalping targets, trailing exits — can be filled on price paths that never existed. The familiar 90% “modelling quality” figure describes how ticks were generated, not how accurate they are.

The history itself can also be wrong: missing days quietly bridged, spikes from a bad feed, symbols whose digits or contract specs changed years ago. MT5’s “every tick based on real ticks” mode is materially better, but it is only as good as the tick archive the broker provides — which often thins out a few years back, exactly where long backtests need it.

Most default backtests assume a fixed spread, often set near the lowest value ever seen. Real spreads breathe: a EUR/USD spread that averages 1.4 pips might sit at 0.8 during the London session and jump past 6 pips for a few seconds around a news release or at the daily rollover. A fixed 1.0-pip assumption books the calm hours and skips the expensive moments. Before trusting any result, check its cost assumptions line by line:

• Spread— variable or recorded spread, not the lowest fixed value; compare against what your own account actually averages per symbol.
• Slippage— testers fill at the exact requested price by default; real fills drift, so an allowance per side (tenths of a pip on majors, more on exotics) belongs in the model.
• Requotes and rejections— invisible to the tester, real in fast markets, especially on instant-execution accounts.
• Commission— per-lot round-turn charges if the account type has them; on EUR/USD, $7 per standard lot is roughly 0.7 pips.
• Swap— overnight financing on anything held past rollover, including the tripled midweek charge.

A historical price series carries no record of what executing in it felt like. Around a high-impact release the chart shows one long candle; it does not show the seconds when the spread was 8 pips, stop orders filled far through their levels, and new orders were requoted or rejected. The tester fills everything cleanly at chart prices with whatever spread you configured.

Broker differences compound this. Every broker’s feed has its own highs and lows — a stop that survives on one feed is hit on another — plus its own minimum stop distances, execution mode and server timezone, which shifts where daily candles open and close. The same EA over the same dates on two brokers’ data produces two different reports, and a live account adds a third.

A subtler distortion happens before any single report is read. Run one EA across 20 symbols and keep the three profitable ones, or sketch ten strategy ideas and develop the one with the best quick test — either way, the survivors were selected after the fact. With that many tries, a few standouts are expected from chance alone, so reporting only the winners makes luck look like edge. This is the retail cousin of survivorship bias, and no tester statistic flags it.

The standard defense is to deny the optimizer part of the data. Optimize on 2019–2023, then run the chosen settings once, unchanged, on 2024–2025. In-sample profit only proves the optimizer did its job; the out-of-sample result is the first real evidence about the strategy.

Walk-forward testingrepeats this on a rolling basis: optimize on a window, validate on the next slice, step forward, repeat. Settings that hold up across slices they were never tuned on are worth more than any single spotless run. Practical tester settings — real-tick data, variable spread, slippage allowances — are covered in the guide to improving MetaTrader backtests, and the free Monte Carlo Trading Simulator stresses a backtest’s trade sequence by reshuffling it thousands of times.

A good backtest is necessary, not sufficient. It cheaply rules out broken ideas, maps which parameter regions matter, and sets rough expectations for drawdown and trade frequency. What it cannot do is certify the future — it can only describe one version of the past, usually the friendliest one.