Nowcasting Russia’s key macroeconomic variables using machine learning

Abstract

The article developed a methodology for nowcasting and short-term forecasting key Russian macroeconomic aggregates: real GDP, consumption, investment, export, import, using machine learning methods: boosting, elastic net, and random forest. The set of predictors included indicators of the stock market, money market, surveys, world prices for resources, price indices, and other statistical indicators of different frequency, from daily to quarterly. Our approach makes available a detailed examination of the changes in forecasts with the flow of new information. For most of the considered variables, a monotonic non-deterioration of the forecast quality was obtained with an expansion of available information. Furthermore, machine learning methods have shown significant superiority in predictive performance over naive prediction. The considered methods within the framework of the pseudo-experiment quickly showed a strong drop in real GDP, household consumption, and other variables in the context of the spread of the COVID-19 pandemic in the 2nd and 3rd quarters of 2020.