Comparison of Forecasting Power of Statistical Models for GDP Growth Under Conditions of Permanent Crises for Application in Strategic Risk Controlling

Abstract

The study evaluates the effectiveness of combining different forecasting models to predict Russia's GDP growth rates for the upcoming quarter. The ensemble model utilized in this study consists of a dynamic factor model (DFM) and a neural network with long- and short-term memory (LSTM). The research compared the root-mean-squared errors (RMSE) of the ensemble model with other popular models such as ARIMA, VAR, SVR, and CatBoost, and found that the proposed ensemble model performed better than the LSTM and competitor models but did not improve upon the DFM forecasts. Additionally, the study identified key indicators with high predictive power for the Russian economy by analyzing the DFM eigenvectors and LSTM integrated gradient coefficients.