A validation method that fits a strategy on one window of history, tests it on the window that follows, then rolls both forward, so that every tested period was unseen at the moment its parameters were chosen.
Fit on two years, test on the following six months, slide both windows forward, repeat. The reported result is the concatenation of the test segments only. The fitted segments are never scored, and a parameter changed after seeing a test segment invalidates that segment.
loop: fit window · test window · step size · aggregate the test segments only
Walk-forward validation answers a sharper question than a single train-test split: not whether one lucky parameterisation held up once, but whether the procedure that picks parameters keeps working as the market moves. It tests the whole method, re-fits included, which is what you would actually be running live.
The discipline is in what you are not allowed to do. The window lengths and the step size are chosen once, before any test segment is seen. Nothing learned from a test segment may flow back into an earlier fit. And the failures stay in the record: a walk-forward report that only shows the segments that worked is a curve fit with extra steps.
Its blind spot is regime scarcity. Rolling windows inside one calm regime produce a reassuring sequence that says nothing about the first shock, so the sample still needs to span at least one regime change. Walk-forward controls how you use the history you have; it cannot add history you do not have.
Covered in depth in the Strategy research fundamentals pillar hub.