Filter before Plug: One-for-All Framework for Covariate-Aware Forecasting with Time Series Foundation Models

ICLR 2026 Conference Submission18630 Authors

Keywords: Time series forecasting

TL;DR: We propose a a One-for-All framework where independently trained modules complement TSFMs.

Abstract: Time series forecasting plays a critical role in numerous real-world applications. Recent advances in Time Series Foundation Models (TSFMs) have achieved strong performance by modeling historical dependencies; however, they frequently neglect the impact of exogenous covariates. Existing methods either train from scratch, losing the advantages of TSFMs, or design plugin modules that are tightly coupled with specific architectures. To address these limitations, we propose FLUG, a One-for-All framework where independently trained modules complement TSFMs. We design an Endogenous Series Filter (EFit) module guided by the Hurst Exponent to separate exogenous components from the time series, thereby enabling TSFMs to focus on modeling and forecasting endogenous patterns. In parallel, we introduce a Covariate Plugin (CPin) module that employs Multi-Scale Patchify fusion and a Causal-Aware Masking strategy based on Gradient Reversal Layer to capture the exogenous information of the target variable. By decomposing endogenous and exogenous dependencies, FLUG enables integration of covariate information across a variety of TSFMs. To supplement existing publicly available covariate time series data, we curate and release four additional datasets. Extensive experiments on real-world business and supplementary data demonstrate the framework’s effectiveness and scalability.

Primary Area: learning on time series and dynamical systems

Submission Number: 18630