MOSES: Macroeconomic Forecasting with Models and Sentiment Synthesis

Abstract

This paper applies Bayesian predictive synthesis to nowcasting gross domestic product (GDP) and forecasting inflation in Russia. The novelty of the research lies in the following: 1) Bayesian predictive synthesis is used to combine a wide range of machine learning (ML) methods, as well as Bayesian vector autoregressions with mixed data frequencies; 2) sentiment indices from the Bank of Russia, the European Central Bank, and the US Federal Reserve System are applied in the forecasting; 3) the impact of seasonal adjustment on the forecasting errors of gradient boosting models and neural networks is assessed. The results show that the most accurate GDP nowcasts for Russia are provided by cubic Bayesian vector autoregression. In forecasting inflation in Russia, the proposed method of combining models outperforms first-order autoregression, Bayesian model averaging, ensemble, and neural network methods. Seasonal adjustment leads to an underestimation of the root mean squared errors of neural network forecasts. The results are robust on data vintages published within three months after the test period and are statistically significant according to the Diebold–Mariano test.