Models for real-time nowcasting and forecasting of GDP components by expenditure

Abstract

This paper compares the forecasting performance of advanced machine learning and econometric models: two types of gradient boosting on decision trees (classical and natural), TabNet, one of the most effective neural network architectures for tabular data, a dynamic factor model (DFM), and a Bayesian mixed-frequency vector autoregression (MFBVAR). The paper asks which family of models and which particular models are the most accurate for relatively short time series, which approaches should be developed further, and why machine learning still yields no substantial gain over advanced econometric methods. Accuracy is compared using historical data within a fair experimental setup. Each model is required to forecast non-seasonally adjusted series and to use monthly predictors as they become available rather than quarterly regressors alone, i.e., to forecast in real time. Each model is also required to select the relevant predictors itself rather than rely on a factor set fixed by theory and must therefore be robust to the curse of dimensionality and include a built-in feature-selection mechanism. The data set contains 63 monthly predictors used to forecast key Russian macroeconomic indicators: GDP, household consumption, gross capital formation, and gross fixed capital formation in constant prices. The paper also proposes and tests the method for constructing predictors that use moving averages to capture the dynamics of several months when forecasting quarterly values. Gradient boosting models slightly outperform MFBVAR, whereas TabNet and the dynamic factor model demonstrate relatively weak performance. The proposed predictors yield accurate longer-horizon forecasts for GDP and gross fixed capital formation.