Information-Theoretic Benchmarking for Multimodal Time-Series Forecasting

This research introduces a synthetic benchmark to evaluate whether text annotations meaningfully contribute to multimodal forecasting. It validates six mutual information estimators, including MINE and InfoNCE, in measuring the predictive value of textual context.