Deciphering Long-Term Economic Growth: An Exploration With Leading Machine Learning Techniques

Abstract

Existing studies mainly focus on short-term economic forecasts, but research on long-term projections, particularly for periods spanning 6–10 years, remains insufficient, despite its importance. This gap may arise from the limitations of traditional linear methods in prediction tasks and pattern recognition, whereas machine learning techniques may help overcome these challenges. To address this, we employ five widely used machine learning models—artificial neural networks (ANN), random forest regression (RF), gradient boosting regression (GBR), extreme gradient boosting (XGBoost), and support vector regression (SVR)—using cross-country data from 109 countries between 1961 and 2019. To ensure robustness, we employ two distinct sampling methods for model validation. Our findings reveal that the ANN model outperforms others, particularly in long-term predictions (6–10 years), with an average out-of-sample prediction $R$-squared of 0.89. Furthermore, analyses using permutation feature importance (PFI) and SHapley Additive exPlanations (SHAP) methods indicate that while current growth rates are critical for short-term forecasts (1–3 years), two primary variables representing a country's foundational characteristics—real GDP per capita and “country-feature,” akin to a country dummy variable—are crucial for long-term predictions (4–10 years). This outcome demonstrates the ANN model's capacity to capture each country's unique characteristics and, through its highly non-linear nature, successfully execute complex, long-range forecasts. These results unveil the remarkable potential of machine learning in the realm of long-term economic forecasting.