Working Paper “GDP nowcasting: from traditional econometric models to machine learning algorithms”

EFSD ANALYSTS HAVE ASSESSED THE POTENTIAL OF MACHINE LEARNING ALGORITHMS TO ENHANCE THE ACCURACY OF SHORT-TERM MACROECONOMIC FORECASTING

Analysts from the Eurasian Fund for Stabilization and Development (EFSD) have prepared a working paper entitled GDP Nowcasting: From Traditional Econometric Models to Machine Learning Algorithms. The report explores the potential of modern machine learning (ML) algorithms to improve the accuracy of short-term macroeconomic forecasts.

The authors set two main objectives:

• to determine whether ML algorithms can outperform conventional econometric models in terms of forecast accuracy; and
• to assess whether these methods can not only complement but also replace conventional models.

A key feature of the study is the use of high-frequency indicators to produce early estimates of low-frequency macroeconomic indicators, such as GDP, which are released quarterly with a significant time lag.

The analysis uses data from 2002 to 2024 for Armenia and Belarus. The dataset is divided into training and test samples to estimate model parameters and compare forecast accuracy.

The study provides a comparative assessment of three conventional econometric models and nine ML methods and algorithms. The performance of these methods was evaluated by the accuracy of predictions on the test sample. The experiments showed that the most effective ML algorithms were LASSO regression, Boosting, Random Forest, SVM and RNN. Moreover, combined nowcasting based on different methods further enhanced the accuracy of forecast compared with using individual models.

The findings suggest that ML methods and algorithms can serve both as an effective complement to conventional econometric models and as their alternative, particularly for nowcasting macroeconomic indicators.