Start with a falsifiable hypothesis
An attractive indicator is not yet a hypothesis. State the mechanism, the scope, and the failure conditions before studying the result. A trend rule, for example, may capture persistence when markets move steadily, but lose money through repeated changes in sideways or reversing markets. Write that expectation down first.
Respect the direction of time
Time-series validation should preserve chronology. Separate development, validation, and final evaluation periods, and ensure each feature uses only information available at that point. Decide the validation process before tuning parameters. Repeated experimentation on the same final period can turn a test into another training set.
Choose a fair benchmark
A passive benchmark, cash, and a simpler rule help distinguish market exposure from strategy behavior. Use the same period and consistent cost assumptions, and compare risk as well as total return. Higher returns accompanied by much larger drawdowns may not answer the original research question.
Keep the whole experiment
Record the data version, parameters, time splits, cost assumptions, and random seed. Keep failures alongside successes. If small parameter changes reverse the result, investigate the sensitivity before building a stronger story. Reproducibility makes a conclusion inspectable; it does not prove that the conclusion will hold in the future.