Backtesting with the assumptions in view.

Freeze the question before the evaluation

State the universe, timing, rule, parameters, objective, and evaluation procedure before reviewing results. Keep development observations separate from a later test. A change made after inspecting that later test consumes information and belongs in the research log.

A final report should show the scope of attempted variations, not only the most appealing version. The long-form backtesting article explains how selection changes the meaning of a reported result.

Audit data availability

Check whether the historical universe includes assets that later disappeared. Verify adjustment conventions, revisions, timestamp meaning, and availability delays. A modern dataset may describe the past without reproducing what was knowable at the time.

Write a data contract specifying fields, missing-value behavior, session rules, and version identifiers. Re-running the same rule against a changed input is a new result, not necessarily a reproduction.

Model execution and costs explicitly

Separate signal generation from executable orders and fills. A closing observation cannot automatically be traded at that already-finished close. Include a stated policy for spread, transaction fees, slippage, incomplete fills, and unavailable liquidity where relevant.

A convenient cost assumption is still an assumption. Examine how the conclusion changes under a plausible range rather than hiding the sensitivity behind a precise headline metric.

Read the full result

Review the return path, concentration, number of opportunities, exposure, drawdown, and conditions under which the rule failed. A few exceptional observations may explain much of an average. Results sampled at a coarse frequency can miss intervening losses.

Continue with risk and performance and the systems workflow. Download the research checklist to make data, execution, and selection questions part of each review.