Evaluating Ensemble Learning Techniques for Economic Forecasting: A Comprehensive Analysis of Model Performance

Abstract

Economic forecasting accuracy remains critical for policy formulation and investment decisions, particularly in emerging markets where economic volatility and data outliers pose significant challenges. Despite the widespread adoption of machine learning approaches, there has been limited research that systematically compares the effectiveness of ensemble learning techniques in handling outlier-rich economic datasets from developing economies. To investigate the comparative performance of Bagging, Boosting, and Stacking ensemble methods in predicting economic indicators using Indonesian historical data containing outliers, this study analyzed monthly economic data from 2010 to May 2024, encompassing inflation rates, economic growth, interest rates, and stock market indicators sourced from the International Monetary Fund (IMF) and Indonesia's Central Bureau of Statistics (BPS), with GridSearchCV employed for hyperparameter optimization across all ensemble models. Results demonstrated that Boosting achieved superior performance with the lowest Mean Squared Error (0.043), lowest Mean Absolute Error (0.147), and highest R² coefficient (0.732), while Bagging provided moderate but stable results with an R² of 0.589, and stacking underperformed significantly with the highest MSE (0.134) and lowest R² (0.171). Boosting's superior capability in capturing complex data patterns makes it the most suitable ensemble technique for economic forecasting in Indonesia. The study offers critical insights for model selection in outlier-prone economic datasets, providing practical guidance for economic analysts and policymakers in developing countries who seek robust forecasting methodologies.