Forecasting and interpreting factors of labor productivity using ensemble machine learning methods

Abstract

This study explores the application of nonlinear machine learning models to forecasting regional labor productivity and identifying its key determinants. Several ensemble methods were employed, with the random forest model achieving the best predictive performance. To account for interregional heterogeneity, clustering of labor productivity time series was performed and a categorical label indicating each region’s cluster membership was introduced, which captured structural differences and enhanced the sustainability and interpretability of the forecasts. The forecasting accuracy was compared with that of traditional linear models commonly used in domestic research, revealing a consistent superiority of nonlinear approaches. Interpretation of the final random forest model using the Mean Decrease in Gini criterion indicated that the most influential factors affecting labor productivity are the growth rate of real wages and the dynamics of investment in fixed capital, which aligns with previous empirical findings. Furthermore, the use of SHAP values enabled a more detailed assessment of the direction and magnitude of each feature’s contribution, enhancing the interpretability and transparency of the model. The findings provide evidence supporting the transition from conventional econometric approaches to machine learning methods for forecasting macroeconomic indicators.