Document Type : Original Paper
Author
Hakim Sabzevari University, Sabzevar, Iran
Abstract
This research examines different feature selection methods aimed at enhancing the predictive accuracy of macroeconomic forecasting models, focusing on Iran’s economic indicators derived from World Bank data. Fourteen feature selection techniques, classified into Filter, Wrapper, Embedded, and Similarity-based categories, were thoroughly compared. The evaluation utilized Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) metrics under a 10-fold cross-validation scheme. The findings highlight that Stepwise Selection, Tree-based approaches, and Similarity-based methods, especially those employing Hausdorff and Euclidean distances, consistently outperformed others with average MAE values of 32.03 for Stepwise Selection and 62.69 for Hausdorff Distance. Conversely, Recursive Feature Elimination and Variance Thresholding exhibited weaker results, yielding significantly higher average MAE scores. Similarity-based approaches achieved an average rank of 9.125 across datasets, demonstrating their robustness in managing high-dimensional macroeconomic data. These outcomes underscore the value of integrating similarity measures with traditional feature selection techniques to improve the efficiency and reliability of predictive models, offering meaningful insights for researchers and policymakers in economic forecasting.
Keywords
- Feature Selection
- Predictive Accuracy
- World Bank Indicators
- Macroeconomic Analysis
- Similarity Methods
Main Subjects
)