covariates

Timestral is a family of four time series foundation models (5M to 209M parameters) achieving state-of-the-art probabilistic forecasting on GIFT-Eval, TIME, and fev-bench via decoupled horizon forecasting, native future covariate support, and synthetic pretraining followed by mixed post-training.

Proposes exogenous dropout, a simple training intervention that randomly zeros entire exogenous channels, to improve robustness of time series forecasting models against corrupted covariates. The method matches or exceeds more complex architectures across multiple corruption types and domains.

This paper introduces conditional hypothesis generation, a framework that incorporates researcher-specified covariates to steer LLM-based text analysis toward discovering meaningful subgroup differences while addressing confounds like stratum imbalance and sign reversal.