RockTS: Robust Time Series Forecasting based on Information Bottleneck and Optimal Transport

ICLR 2026 Conference Submission12891 Authors

Keywords: Time series forecasting

Abstract: Time series forecasting plays a crucial role in numerous real-world applications. Existing works mostly assume clean and regular historical sequences for predicting future ones. However, real-world time series data often contain anomalous subsequences that deviate from the normal patterns of the entire series, posing challenges to accurate forecasting. In this paper, we propose RockTS, a novel end-to-end framework for robust time series forecasting based on Information Bottleneck and Optimal Transport, which integrates the detection and imputation of anomalous subsequences into the forecasting task through a unified optimization objective. RockTS first introduces a detection process for anomalous patterns based on Information Bottleneck, which compresses representations of time series while retaining the information more relevant for effective forecasting. It then imputes the detected anomalous regions with normal patterns through a novel reconstruction strategy based on Optimal Transport for forecasting. Experiments on multiple real-world and synthetic datasets demonstrate that RockTS achieves superior robustness and forecasting performance.

Primary Area: learning on time series and dynamical systems

Submission Number: 12891