Volume 55

Exploring the Effectiveness of SHAP for Feature Selection in National Demand Forecasting Corné van Zyl, Xianming Ye, Raj Naidoo

Abstract

This study evaluates the effectiveness of SHAP (SHapley Additive exPlanations) as a feature selection tool for national demand forecasting, addressing the need for a comprehensive understanding of its strengths and limitations. We benchmark SHAP against common selection methods, including variance-based selection, XGBoost, and NSGA-II. Using a 1D Convolutional Neural Network to forecast hourly national energy demand in Panama, we compare the performance of features selected by each method. Our results indicate that SHAP is generally outperformed by XGBoost and may not be the most effective initial approach for feature selection, nor does it consistently rank features by their predictive value. However, SHAP proves more effective in refining smaller sets of features, maintaining model accuracy by eliminating non-contributive features

Keywords feature selection, SHAP, national demand forecasting, explainable artificial intelligence