Walk-forward validation, and why one split can lie.

A backtest that chooses its parameters on the same data it reports results on is grading its own homework. Walk-forward validation splits the history in two: the earlier part is used to choose, the later part is used only to score, once.

The number from the later part is the only one that is evidence. The earlier one tells you how well the search fitted the past, which is a different and much less useful question.

1. Every combination in the strategy's grid is run on the first 70% of the period.
2. The one with the best Sharpe ratio there is picked.
3. A gap is cut out around the split: 1% of the period, at least 5 and at most 25 trading days, on each side. A trade open across the split would otherwise earn in both halves off the same price move and count twice.
4. The picked combination, unchanged, is run on the last 30% and scored.

The verdict compares the two Sharpe ratios. Keeping at least 70% of the in-sample figure is held up, 40–70% is weakened, below that is overfit, and a test-period Sharpe at or below zero is failed.

A single split is still one draw. On Apple, momentum's best in-sample cell scored 0.949 and then −0.303 on 2022–2023: failed. Moving the end date one year put 2024's rally in the test half, and the same strategy held up at 0.547. Bollinger bands went the other way, from held up to failed.

So the workspace also runs a rolling walk-forward, which slides the split through the period and shows the out-of-sample result at each step. A strategy that only passes at one cut point has not passed.

Walk-forward protects the final number from the search that produced it. It does not protect you from your own searching: if you rerun it with ten strategies and keep the one that passed, the test half has become part of the search. That is what the overfitting checks are for.