Information Bottleneck Learning for Faithful Time Series Forecasting Explanations

As time-series forecasting becomes critical for decision-making in sectors like energy and healthcare, the need for transparency has grown. However, existing interpretable models often lack rigorous guarantees that their internal structures actually drive their predictions, while post-hoc explanation methods are typically designed for classification rather than the temporal dynamics of forecasting. This paper asks: can we design a forecasting framework that is inherently interpretable, provides faithful explanations by construction, and maintains competitive accuracy without additional inference costs?

To address this, the authors propose IB-Forecast, which decomposes the forecasting process into two distinct parts: a disclosed structural context and a gated deviation readout. The structural context captures recurring patterns (seasonality, level, and scale), while the deviation readout focuses on instance-specific residuals.

The framework employs an Information Bottleneck (IB) principle to learn a sparse binary mask over input tokens. This mask acts as a gate, selecting only the most relevant historical information to pass to the readout. By training this gate end-to-end with a budget constraint, the model forces the system to prioritize essential historical evidence. Because the forecast is computed directly from these masked tokens, the explanation is inherently faithful—if a token is masked out, it is structurally impossible for it to influence the final prediction.

Experiments across multiple multivariate time-series benchmarks demonstrate that IB-Forecast achieves predictive accuracy comparable to leading black-box models. Crucially, the model provides faithful explanations that consistently outperform existing gradient-based, occlusion-based, and optimization-based baselines under matched sparsity budgets. The authors show that the model can deliver low-error predictions while utilizing only 14–20% of the historical input, effectively demonstrating that accurate forecasting often requires only a small fraction of the available data.