Effectiveness of Integrating Textual Context into Forecasting Models

Determine to what extent current time series forecasting models—including statistical methods, time series foundation models, and large language model–based forecasters—can effectively integrate relevant natural language textual context alongside numerical historical data to improve the accuracy of probabilistic forecasts of future observations.

Background

The paper introduces the Context is Key (CiK) benchmark to evaluate whether forecasting models can leverage essential natural language context in addition to numerical time series data. Traditional approaches typically operate on numerical inputs alone, whereas newer LLM-based methods promise to use textual information to guide forecasting.

Despite this promise, the authors emphasize that there has been no systematic evaluation of models’ ability to incorporate natural language context effectively, motivating the need to determine whether and how such integration genuinely improves predictive performance.