I'm building walk-forward optimization into my process. The literature says everything from 4 to 50 windows.
Practically: how many in-sample / out-of-sample windows do you run before you trust a result, and what IS/OOS split do you use? Bonus question: anchored or rolling?
I use 6-10 windows with a 70/30 IS/OOS split. More windows means smaller samples, which means noisier estimates — there's a real trade-off.
What matters more than the window count is parameter stability across windows. If the optimal inputs jump around wildly from window to window, the strategy has no robust core regardless of average OOS performance.
Anchored vs rolling matters too. My filter: anchored WFO, 8 windows, and reject any strategy where more than 2 windows come in under PF 1.2 out-of-sample. Strict, but my live results got boring in a good way.
The parameter stability chart per window is exactly what QuantMogul's WFO report shows, and it's what convinced me my last "robust" strategy was actually noise. Rejected it before it cost me real money. Rejection is the feature nobody markets.