What if Tomorrow is the World Cup Final? Counterfactual Time Series Forecasting with Textual Conditions

2026-05-14

AI Summary

This work addresses the challenge that existing time series forecasting methods struggle to model counterfactual scenarios influenced by future free-text events and are often restricted to structured conditions. It introduces a novel forecasting task that integrates counterfactual reasoning with free-text conditioning, proposing a text attribution mechanism to disentangle mutable and immutable factors. Leveraging a diffusion model architecture and a new conditional fusion strategy, the approach enables dynamic modeling of future scenarios described by complex textual narratives. The study further establishes an evaluation framework encompassing both factual and counterfactual settings, allowing effective assessment of predictive performance without access to actual future sequences. Experiments demonstrate that the method significantly improves accuracy across diverse real-world scenarios, exhibiting strong generalization and robustness to highly stochastic and fine-grained textual conditions.

Abstract

Time series forecasting has become increasingly critical in real-world scenarios, where future sequences are influenced not only by historical patterns but also by forthcoming events. In this context, forecasting must dynamically adapt to complex and stochastic future conditions, which introduces fundamental challenges in both forecasting and evaluation. Traditional methods typically rely on historical data or factual future conditions, while overlooking counterfactual scenarios. Furthermore, many existing approaches are restricted to simple structured conditions, limiting their ability to generalize to the real-world complexities. To address these gaps, we introduce the task of counterfactual time series forecasting with textual conditions, enabling more flexible and condition-aware forecasting. We propose a comprehensive evaluation framework that encompasses both factual and counterfactual settings, even in the absence of ground truth time series. Additionally, we present a novel text-attribution mechanism that distinguishes mutable from immutable factors, thereby improving forecast accuracy under sophisticated and stochastic textual conditions. The project page is at https://seqml.github.io/TADiff/

Problem

• counterfactual forecasting
• time series forecasting
• textual conditions
• stochastic events
• forecasting evaluation

Innovation

• counterfactual forecasting
• textual conditions
• time series forecasting
• text attribution
• evaluation framework