Reproduce “Improving Accuracy Without Losing Interpretability: A ML Approach for Time Series Forecasting” on OpenResearch

About this paper

Yiqi Sun, Zhengxin Shi, Jianshen Zhang, Yongzhi Qi, Hao Hu, Zuojun Max Shen

A new hybrid forecasting algorithm, W-R, enhances time series prediction accuracy by an average of 78% in P50 Quantile Loss on diverse datasets while maintaining interpretability through a novel quantitative definition and a weighted combination function for decomposed components. The method applies an ML module to dynamically adjust component weights and residuals, outperforming both pure ML and traditional decomposition benchmarks.

Key insights

• Interpretability can be quantitatively defined by the nature of inputs and the transparency of the mapping function, enabling ML modules to be integrated for accuracy without sacrificing overall model understanding.
• Replacing simple additive component combination with a dynamic, ML-driven weighted sum, augmented by residuals, allows for capturing complex interdependencies between components, leading to higher accuracy.
• Restricting the learned weights to a moderate interval around 1 via a normalization function contributes significantly to improved accuracy by enhancing the estimation quality of individual components.

Summary by alphaXiv.