StatFuse: Bridging Statistical Inference and Neural Prediction for Interpretable Forecasting

Abstract

The integration of traditional statistical methods with modern deep learning architectures offers opportunities to develop prediction frameworks that balance accuracy and interpretability. This paper introduces StatFuse, a hybrid approach synthesizing statistical decomposition with neural prediction while maintaining rigorous uncertainty quantification. By combining time-series analysis principles with neural architectures, the framework achieves strong and competitive performance across benchmark datasets. The methodology incorporates conformal prediction intervals for distribution-free coverage guarantees and employs statistical diagnostics and perturbation-based attribution for feature importance. Experimental validation on economic forecasting and public health monitoring demonstrates that StatFuse improves performance on two of four benchmarks and remains close to strong baselines on the others, while offering enhanced interpretability.