Evaluating Feature Selection Methods for Macro-Economic Forecasting, Applied for Inflation Indicator of Iran

Abstract

This study explores various feature selection techniques applied to macro-economic forecasting, using Iran's World Bank Development Indicators. Employing a comprehensive evaluation framework that includes Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) within a 10-fold cross-validation setup, this research systematically analyzes and ranks different feature selection methodologies. The study distinctly highlights the efficiency of Stepwise Selection, Tree-based methods, Hausdorff distance, Euclidean distance, and Mutual Information (MI) Score, noting their superior performance in reducing predictive errors. In contrast, methods like Recursive Feature Elimination with Cross-Validation (RFECV) and Variance Thresholding showed relatively lower effectiveness. The results underline the robustness of similarity-based approaches, particularly Hausdorff and Euclidean distances, which consistently performed well across various datasets, achieving an average rank of 9.125 out of a range of tested methods. This paper provides crucial insights into the effectiveness of different feature selection methods, offering significant implications for enhancing the predictive accuracy of models used in economic analysis and planning. The findings advocate for the prioritization of stepwise and tree-based methods alongside similarity-based techniques for researchers and practitioners working with complex economic datasets.