How to Backtest a Trading Strategy Properly

A backtest is not proof that a strategy will make money. It is an experiment that asks a narrower question: how would a defined set of rules have behaved on a particular historical dataset under specified execution assumptions? The quality of the answer depends on the quality of the experiment.

Write the rules before looking at the result

Start with a clear hypothesis. Define entry conditions, exit rules, stop placement, target logic, position sizing, trading hours, spread assumptions and what happens when multiple signals occur. If the rules change repeatedly after every disappointing test, the process becomes curve fitting rather than research.

A strategy should be reproducible. Another person using the same data and rules should be able to obtain approximately the same result.

Use data that matches the strategy

A daily strategy can often be studied with daily bars, but an intraday EA that uses stop loss, breakeven and trailing logic needs much finer data. Bar-only testing can miss the order in which price touched the stop, target or trailing level within a candle.

For MetaTrader 5, MetaQuotes documents Every tick based on real ticks as the mode that most closely reflects recorded broker tick conditions. That does not make a test perfect, but it is more appropriate for execution-sensitive EAs than open-price-only testing.

Model real trading costs

Spread, commission, slippage and financing can change a strategy from profitable to unprofitable, especially when the average trade is small. Use realistic values from the intended broker where possible. Avoid zero-spread backtests unless the strategy will actually trade under a zero-spread structure with its corresponding commission.

For news-sensitive systems, remember that spreads can widen precisely when signals are most active. Testing only with an average spread can still be optimistic.

Separate development from evaluation

If the same data is used to invent the rules, optimize the parameters and judge the final strategy, the result is biased. A better process divides history into at least an in-sample development segment and an out-of-sample evaluation segment.

For example, parameters might be developed on earlier years and then frozen before testing on a later period. The out-of-sample result matters because the strategy has not been tuned to those specific observations.

Use walk-forward testing

Markets change. A single train/test split may depend heavily on the chosen dates. Walk-forward testing repeats the process through time: develop or calibrate on one rolling window, evaluate on the next unseen window, then move forward. The goal is to see whether an edge persists across multiple periods rather than one lucky segment.

Watch for look-ahead and survivorship bias

Look-ahead bias occurs when the strategy uses information that would not have been available at the time of the trade. Examples include using a candle's final close to enter earlier in that same candle, or accidentally referencing future indicator values.

Survivorship bias is especially relevant when testing baskets of securities because today's surviving instruments may not represent the historical universe. For major FX pairs the issue is different, but data-cleaning and broker-history changes can still matter.

Do not optimize only for net profit

A strategy can show an attractive final profit while relying on a few extreme winners, taking unacceptable drawdown or producing a long losing streak. Useful metrics include:

• Number of trades: tiny samples are unstable.
• Win rate: useful only together with average win and average loss.
• Expectancy: the average R or cash outcome per trade.
• Profit factor: gross profit divided by gross loss.
• Maximum drawdown: peak-to-trough decline.
• Average holding time: important for financing and exposure.
• Longest losing streak: important for psychological and capital planning.
• Results by market and regime: helps reveal hidden concentration.

Parameter stability matters

If an EMA length of 49 is highly profitable while 45, 47, 51 and 53 are poor, that narrow peak is a warning sign. Robust strategies often have a region of acceptable parameters rather than one magical value. Sensitivity testing helps distinguish a broad relationship from a historical accident.

Test different regimes

Trend-following and mean-reversion systems can behave very differently across trending, ranging, high-volatility and low-volatility periods. Results should be broken down by regime where possible. Gold, Bitcoin and major FX pairs also have different volatility structures, so a parameter set that works on one instrument should not automatically be applied to another.

How GodzillaBTC validates signal setups

The GodzillaBTC signal engine does not convert a model score directly into a probability. It compares similar historical setups and measures whether the defined target was reached before the stop within the evaluation horizon. The public tier is withheld when the sample is too small, quote data is stale or historical expectancy is weak. This is still research, not an audited performance record.

See the Validation & Performance Evidence page for what is currently measured and what is not yet claimed.

A practical backtest workflow

1. Write the strategy rules and assumptions.
2. Choose data quality appropriate to the execution logic.
3. Include spread, commission and realistic slippage.
4. Reserve unseen data before optimization.
5. Run sensitivity tests around chosen parameters.
6. Evaluate multiple markets and regimes.
7. Record drawdown, expectancy and losing streaks—not just profit.
8. Freeze the model and run out-of-sample or walk-forward tests.
9. Forward-test on a demo account before risking real money.
10. Publish limitations together with results.

Sources and further reading

MetaQuotes explains the behavior of Strategy Tester real-tick mode. The CFTC also warns that automated trading technology cannot consistently predict the future; see its forex fraud guidance.