Expanding or rolling training windows

What trade-off determines whether you use expanding or fixed-length rolling training windows for a drifting financial dataset?

Expanding windows retain more observations and can lower estimation variance if old relationships remain relevant. Rolling windows discard stale data but estimate from fewer observations. Compare both as complete historical procedures, including retraining costs and chosen window lengths. Select the rule within development periods and evaluate its stability across regimes. A rolling window is not automatically superior merely because markets change.