XForecast: Evaluating Natural Language Explanations for Time Series Forecasting

Natural language explanations make AI forecasts more accessible: Traditional explainability methods require expert knowledge to interpret, but natural language explanations allow non-experts and stakeholders to understand time series forecasts, supporting more informed decision-making.

A new way to measure explanation quality has been developed: The paper introduces simulatability-based metrics that evaluate whether humans can accurately predict model forecasts using the provided explanations, offering a practical way to assess whether explanations are actually helpful rather than just plausible-sounding.

Explanation quality depends on reasoning ability, not model size: The research found that when large language models generate time series explanations, their numerical reasoning capabilities—not their overall size—determine explanation quality, suggesting that bigger models don't automatically produce better explanations.

Current evaluation methods for explanations were inadequate: The paper addresses a significant gap in responsible AI by acknowledging that evaluating forecast explanations is "difficult due to the complex causal relationships in time series data," and proposes metrics that align with human judgment to fill this gap.

Human evaluation remains important for validating AI explanations: The simulatability metrics were validated against human judgments, emphasizing that responsible AI development requires grounding automated evaluation methods in how actual people understand and use explanations.