Information Bottleneck Learning for Faithful Time Series Forecasting Explanations

The paper introduces IB-Forecast, an inherently interpretable framework for multivariate time-series forecasting. It decomposes forecasts into a learned periodic component and a residual component derived from explainable masks over input tokens. A budget-constrained information bottleneck lets users control explanation sparsity during end-to-end training. The authors report that IB-Forecast matches the forecasting error of leading black-box models while producing faithful explanations at no additional inference cost. With only 14–20% of observations, it delivers low-error predictions, and its native explanations outperform gradient-, occlusion-, and optimization-based baselines under matched sparsity across the evaluated datasets.