Common Crypto Backtesting Mistakes and How to Detect Them

Prevent look-ahead and timing bias

Look-ahead bias occurs when a decision uses information that was not available at the simulated decision time. Common examples include trading on the same candle close that confirms the signal, using a completed higher-timeframe candle too early, or allowing a fill based on the favorable order of high and low inside one bar.[4][2]

Count how many ideas and parameters were tried

If hundreds of variants are tested and only the best is shown, its apparent performance includes selection luck. Keep a research log, limit degrees of freedom, and reserve data that did not influence the chosen rules. Nearby parameter values should produce a coherent region rather than one isolated peak.[3][4]

Count more than formal optimizer runs. Changing the symbol list, date range, indicator family, exit type, benchmark, and performance threshold after viewing results are all trials. Record them even when no report was saved. Define a search budget before exploration and group related variations under the hypothesis they test. Once a holdout result has influenced a choice, relabel it as development data and obtain genuinely later evidence. The apparent winner should face a higher evidentiary bar when it emerged from a larger search.

Inspect a parameter surface rather than a leaderboard. Choose two important parameters, test a small grid around the selected values, and look for a broad plateau of similar behavior across periods. An isolated spike suggests sensitivity to a few historical observations. Prefer a simpler point inside a stable region over the maximum. Repeat the comparison using net performance, drawdown, and trade count together, because a smooth return surface can still hide collapsing sample size or unacceptable risk.

Audit data coverage and universe selection

Using only assets that exist today can exclude failed or delisted instruments. Missing candles, insufficient indicator warm-up, changing contract specifications, and incomplete funding or open-interest snapshots can silently alter signals. Preserve warnings and state exactly which symbols were eligible at each time.

Construct a coverage report before the performance report. For each symbol show first and last usable timestamp, missing and duplicated intervals, warm-up consumed, required external series, and reasons for exclusion. Compare the eligible universe at each historical point with the universe the strategy claims to trade. If current listings are used as a proxy, name the resulting survivorship limitation and avoid conclusions about a dynamic market-wide portfolio. Never replace unknown external observations with zero unless zero is the documented meaning of the source.

Challenge fills, costs, leverage, and liquidation

Filling every limit order when price merely touches it, ignoring bid-ask spread, charging one fee instead of every fill, or omitting funding can create profit that execution cannot keep. Leverage must affect margin and liquidation exposure; it is not a multiplier that can be applied only to winning PnL.

Review execution assumptions with adversarial examples. Use a candle that touches both stop and target, a gap beyond the stop, a limit touched only at the extreme, an order below minimum notional, and a leveraged move near liquidation. Write the model's outcome before running the examples. Conservative deterministic behavior is preferable to selectively choosing the favorable path, and unresolved intrabar ambiguity should appear as a limitation or a sensitivity range rather than a precise fictional fill.

Avoid metrics and comparisons that hide fragility

Win rate without payoff size, return without drawdown, and annualization from a short period can mislead. Check trade count, distribution, recovery time, concentration by symbol and month, and performance against a simple benchmark. A strategy should not pass because one extraordinary trade covers persistent losses.