Walk-forward validation evaluates a time-dependent model or strategy through a sequence of chronological windows. Each step uses only earlier data for development and then tests on the next period, after which the window moves forward.
This design better resembles repeated real deployment than a random train-test split. It can reveal whether performance survives changing conditions, although results still depend on window choices, data quality, realistic execution assumptions and disciplined parameter selection.
ELI5
Walk-forward validation tests a model or trading strategy in the same order that time actually passes. It develops the method using older data, tests it on the next unseen period, then moves both windows forward and repeats.
For example, a strategy could learn from January through June and be tested on July, then learn from February through July and be tested on August. This makes it harder to accidentally use future information, though unrealistic trading costs or repeated tuning can still mislead.
How is walk-forward validation different from one backtest?
It performs repeated chronological evaluations, so each test period remains later than the data used to develop or tune the strategy.
Does walk-forward validation prevent overfitting?
It reduces some forms of overfitting, but repeated tuning to walk-forward results can still bias the final performance estimate.
