Why we use it
If you pick the best parameters using the whole history, the backtest already knows the future and will look better than reality. Walk-forward optimization only ever uses information that would have been available at the time, which gives a more honest estimate of live performance.
How Quant Arena applies it
Each strategy is trained on a 3-year window and then traded for the next 6 months with the chosen parameters, rolling forward through the history. The walk-forward result is what we show. Until 3 November 2026 we kept the default parameters whenever they did better over the whole history, but that choice itself looked at the full sample and made about one in five results slightly optimistic, so we dropped it.
How much survives out of sample
Every strategy page now has an out-of-sample check. Across the 486 strategy-and-market pairs with parameters to tune (data through 3 November 2026), the settings chosen on each 3-year training window had a median Sharpe ratio of 0.32 there, but only 0.19 in the six months after each choice, when they were actually traded. In 357 of the 486 pairs (73%) the out-of-sample Sharpe ratio was lower than the training one.
Re-selecting the parameters was not a clear win either: over the same out-of-sample months it beat simply keeping the default settings in 248 of the 486 pairs, about half. That is why we treat walk-forward results as an honest estimate, not as proof that tuning adds value, and why each strategy page shows both numbers side by side.