TimeSeed: Effective Time Series Forecasting with Sparse Endogenous Variables

ICLR 2026 Conference Submission755 Authors

Keywords: Time series forecasting

Abstract: Time series forecasting is widely applied across various domains. In real-world applications, there are many scenarios where endogenous variables are missing. Recent studies show that incorporating exogenous variables can significantly enhance the predictive accuracy of endogenous variables. However, the lack of a complete historical context introduces significant uncertainty in temporal dependence capture, particularly in systems characterized by non-stationary behavior. To address these challenges, we propose TimeSeed, specifically designed for scenarios with sparsely observed endogenous variables. Technically, TimeSeed reconstructs l sufficient endogenous series from both complete exogenous series and sparsely observed endogenous series, utilizing two types of data to extract stable information. Building on this foundation, we effectively transforming the challenging original prediction task into a sequence-based prediction task. Moreover, TimeSeed is built entirely upon linear layers, which significantly reduces computational costs. Experiments conduct on seven real-world datasets demonstrate that TimeSeed consistently outperforms state-of-the-art models in forecasting accuracy, achieving an average reduction of 13.01\% in MSE and 7.54\% in MAE, with a model size of only 0.19M parameters. Code is available at this repository: \url{https://anonymous.4open.science/r/Alistair-7}.

Primary Area: learning on time series and dynamical systems

Submission Number: 755