Selective reuse of prior ensemble data improves the latest air temperature forecast over North America

Accurate subseasonal to seasonal (S2S) weather forecasts are critical for managing risks to society, yet improving forecast skill remains challenging. Ensemble forecasting mitigates atmospheric chaos but is limited by computational cost and by declining accuracy at longer lead times. Previous attempts to incorporate previous ensemble forecasts have yielded little improvement in the accuracy of the latest forecast because members from earlier initializations tend to degrade forecast quality. Here, we introduce a simple yet powerful postprocessing approach, lagged ensemble analog subselection (LEAS), which selectively chooses previous ensemble members that best predicted the most recent conditions. Using hindcasts of daily maximum 2-m air temperature over North America from four state-of-the-art S2S weather forecast models, we show that LEAS enhances both deterministic and probabilistic skill across multiple weeks, including for extreme heat events. The method reduces systematic bias as well as variance error, outperforming conventional lagged ensembles without requiring additional simulations or changes to model initialization. The improvement arises from filtering out poorly performing members and effectively emulating enhanced initialization of both atmospheric and land–surface states. LEAS combines principles of analog forecasting with lagged ensembles, extending their impact from short-term to multiweek predictions. Its simplicity and generality suggest broad applicability, not only to machine-learning–based weather forecasting but also to other predictive systems that rely on repeated initialization, such as hydrological, climate, and Earth system models. By extracting more value from existing forecast data, LEAS advances toward the upper limit of forecast skill achievable within current model frameworks while avoiding added computational burden.