Write the rule first
Specify the sample, data cutoff, comparison function and outcome measure before evaluating success. Keep a record of changes so a revised method is not confused with the original test.
Use walk-forward evaluation
At each historical cutoff, restrict the search to information available at that point. Evaluate the following period without letting it influence selection.
Choose a baseline
A research claim needs a comparison. Depending on the question, a baseline might be a simple return distribution or an unconditional historical sample. Explain why it is appropriate.
Account for dependencies
Overlapping windows, correlated instruments and repeated parameter trials can make evidence look stronger than it is. Count these dependencies and avoid treating every row as an independent experiment.
Report failure as well as success
Publish the evaluation period, missing data treatment, error distribution and sensitivity tests. If a strategy is tested, transaction costs and implementation assumptions also matter.