Forecasting Subseasonal Temperature Extremes with Machine Learning: Skill, Interpretability, and Paths Toward Uncertainty Quantification

Abstract: Subseasonal temperature extremes pose major challenges for energy systems, agriculture, and public health, yet their prediction remains difficult due to weak atmospheric persistence and limited skill of physics-based models beyond 10 days. In this talk, I will present recent work evaluating the ability of machine learning methods to forecast weekly mean temperature and temperature extremes at 1–4 week lead times across five regions of the contiguous United States. We compare a tree-based ensemble method (random forests, RF) and a recurrent neural network approach (ensemble echo state networks, EESN) against persistence and enhanced climatological baselines.

Across all lead times and regions, RF achieves the highest accuracy, particularly for extremes, reducing RMSE by up to 50% relative to climatology. EESN provides modest improvements over persistence at 1-week lead times but performs similarly to climatology at longer lead times. To improve interpretability and model parsimony, we apply a grouped permutation feature importance (PFI) framework to iteratively rank and select correlated predictor groups from an initial pool of over 1,100 atmospheric and land-surface variables. This approach reduces the effective input dimensionality to fewer than 10–75 features per model with no loss of predictive skill, while revealing region- and lead-time–dependent drivers of subseasonal predictability.

While the primary focus of this work is on point prediction skill, I will conclude by discussing ongoing efforts to extend this framework to probabilistic forecasting and uncertainty quantification. In particular, I will outline challenges associated with uncertainty estimation for subseasonal extremes and highlight how conformal prediction methods, combined with machine-learning–based forecasts, may provide a principled pathway toward calibrated, distribution-free uncertainty estimates for environmental forecasting applications.