Backtests are research, not receipts

Understand replay coverage and the difference between simulated and executed outcomes.

Avoid hindsight

Replay should expose only information available at the time of each decision. A perfect retrospective drawing is not a prediction.

Read the coverage

A requested three-month window can return less. Missing frames, delisted symbols and new listings affect the sample.

Check costs and ambiguity

A candle can touch both target and stop without telling you the event order. Conservatively handle ambiguity and execution costs.

Separate the ledgers

Historical generated candidates, current card levels and actual filled trades answer different questions. Never publish them as one win-rate number.