Walk-forward analysis: validating out-of-sample

In-sample (IS) is the slice of history on which you build and tune the strategy. Out-of-sample (OOS) is a slice kept aside, never used to decide the rules: it stands in for the future.

Performance that stays good OOS is an encouraging sign of a real edge. Performance that collapses OOS betrays overfitting: the strategy memorized the past instead of capturing a mechanism.

Rather than a single IS/OOS split, walk-forward chains rolling windows: you optimize on one window, test on the next, then shift and repeat across all history.

At the end, you stitch every OOS segment into a single curve: that's the performance you'd have achieved by periodically re-optimizing, without ever cheating with the future.

The gap between in-sample and out-of-sample performance is a thermometer for overfitting: the wider it is, the more your results owe to luck and fine-tuning, and the less to a real edge.

Walk-forward also exposes parameter stability: if the optimal settings change drastically from one window to the next, the strategy is fragile, even if the overall curve looks great.

Set aside a share of history from the outset that you never look at while building. Use it only at the very end, once, to judge — without re-touching the rules on it, or you'll contaminate the OOS.

Combine it with a Monte-Carlo (for risk of ruin) and a deflated Sharpe (for multiple testing). Together, these tests tell a robust edge apart from a curve fitted to measure.