Working Paper “Machine Learning Algorithms for Short-Term Forecasting of Real GDP Growth Rates”

The Eurasian Fund for Stabilization and Development (ESFD) has released a working paper titled Machine Learning Algorithms for Short-Term Forecasting of Real GDP Growth Rates. The study evaluates the accuracy of short-term forecasts produced by machine learning methods and algorithms compared to conventional econometric models.

The research involved an experiment using quarterly data for Belarus from 2004 to 2024. The entire dataset was split into two parts: a training set (the first 70% of observations) and a testing set (the last 30% of observations). Indicators from the real, external, and financial sectors were used for evaluation and forecasting. Both conventional econometric models (such as AR, VAR, and BVAR) and machine learning methods (including ridge regression, LASSO regression, elastic net, boosting, bagging, random forest, and neural networks) were applied to the dataset. Python was used as the programming language for all computations.

The findings show that nearly all machine learning methods and algorithms outperformed traditional models in short-term forecasting of real GDP growth rates. LASSO regression, random forest, and neural networks delivered the best results. These models can be used both as an alternative and as a supplement to conventional models. Moreover, machine learning methods and algorithms avoid the lag shift effects typical of AR, VAR, and BVAR models. However, a statistical difference in accuracy was observed only when using a combined forecast approach based on LSM weights that produced the most accurate predictions.