Walk-Forward Analysis
Rolling train and test windows, and what efficiency actually measures.
How the windows roll
The engine divides your history into consecutive segments. It optimises parameters on the first train window, applies them unchanged to the test window that follows, then rolls both windows forward and repeats. Six windows is the default.
Because parameters are never fitted on the window being measured, the aggregate test performance is a genuine out-of-sample estimate rather than a rehearsal.
Interpreting efficiency
Efficiency is out-of-sample return divided by in-sample return. A value near 1.0 means the strategy performed as well outside the fitting window as inside it. Below 0.50 the edge is largely a fitting artefact and the gate fails.
Efficiency above 1.0 happens occasionally and is not automatically good — it usually means the test window caught a favourable regime rather than that the strategy improved.
Window count trade-offs
More windows give a more reliable estimate but leave less data in each train segment, which can starve parameter optimisation. Fewer windows do the opposite. Six windows over twelve months of hourly data is a reasonable default for most crypto strategies.