Real-world time series forecasting is often challenging due to non-stationarity and distribution shifts, where the optimal forecasting model varies across different temporal regimes and horizons. In this work, we introduce a method called Adaptive Future-Guided Ensemble Learning (AFG-EL), a two-stage framework that performs drift-aware, sample-level routing over a heterogeneous model zoo. AFG-EL learns dynamic fusion weights from meta-features of the historical window and incorporates a future-guided training signal from a relative-future teacher or scorer, emphasizing learning on regime transitions and drift-sensitive segments. Crucially, the inference process remains strictly causal, requiring only historical data and extracted meta-features. We further use sparse routing with an entropy-based fallback mechanism to enhance stability when routing confidence is low. Our experiments on several commonly used forecasting datasets demonstrate that AFG-EL consistently outperforms strong single-model baselines, uniform averaging, and adaptive fusion baselines.
Jing et al. (Thu,) studied this question.