Irina Rish

LLMs outperform specialized TSFMs and regression baselines in intervention-aware time series forecasting, yet all models exhibit a stark performance degradation when predicting responses to novel, unseen covariate futures.