Counterfactual Time Series Forecasting with Textual Conditions

Shuqi Gu, Yongxiang Zhao, Baoyu Jing, Kan Ren

Keywords: counterfactual; time series forecasting; multimodal

TL;DR: We introduce counterfactual time series forecasting with textual conditions, providing an evaluation framework and an attribution-forecast paradigm to improve forecasting under complex and stochastic scenarios.

Abstract:

Time series forecasting plays an increasingly important role in real-world scenarios, where future trajectories are shaped not only by historical patterns but also by forthcoming events, whose subtle and complex influences pose significant forecasting challenges.
Two key aspects emerge in this context. First, forecasting must adapt dynamically under stochastic counterfactual conditions, raising fundamental difficulties in both conditional forecasting and evaluation. Second, the conditions themselves are often complex, and accurately modeling their influence remains non-trivial. Traditional methods typically rely solely on historical information or address only factual future conditions, while neglecting counterfactual scenarios. Moreover, most existing approaches are limited to simple structured conditions, resulting in poor generalization to real-world complexities. To address these gaps, we introduce the task of counterfactual time series forecasting with textual conditions, which leverages unstructured text to enable flexible, condition-aware forecasting. We propose a comprehensive evaluation framework capable of assessing models under both observed data and counterfactual settings, even in the absence of ground truth time series. Furthermore, we present a novel attribution-forecast paradigm that separates mutable from immutable factors, leading to more precise forecasts under sophisticated and stochastic textual conditions.

Primary Area: learning on time series and dynamical systems

Submission Number: 10103