Updated 2026-07-25
Walk-forward validation is a testing method where a model is fitted on one period and evaluated only on the period that follows, then the window rolls forward and the process repeats. Every measurement is therefore made on data the model had never seen, in the order it actually arrived.
The point is to remove look-ahead bias, the quiet killer of backtests. If parameters are chosen using the same data used to score them, the result measures fitting skill rather than predictive skill, and it will not survive contact with live markets. See overfitting and out of sample. Walk-forward is the method we use to measure our own liquidation-level work, alongside sample sizes and confidence intervals rather than single headline figures. The method and the current numbers are published on the proof page.