Forecasting GDP growth : a comprehensive comparison of employing machine learning algorithms and time series regression models

Abstract

In this paper, we do a comprehensive comparison of forecasting Gross Domestic Product (GDP) growth using Machine Learning algorithms and traditional time series regression models on the following economies: Australia, Canada, Euro Area, Germany, Spain, France, Japan, Sweden, Great Britain and USA. The ML algorithms we employ are Bayesian Additive Trees Regression Trees (BART), Elastic-Net Regularized Generalized Linear Models (GLMNET), Stochastic Gradient Boosting (GBM) and eXtreme Gradient Boosting (XGBoost), while Autoregressive (AR) models, Autoregressive Integrated Moving Average (ARIMA) models and Vector Autoregressive (VAR) models represents the traditional time series regression methods. The results assert that the multivariate VAR models are superior, indicating the chosen variables’ and the models’ suitability of forecasting GDP growth. Furthermore, we also do an assessment of the top three variables that drives the best performing Machine Learning algorithm of XGBoost to investigate whether it suggests the same variables in forecasting GDP growth as macroeconomic theory. In general we do see some evidence, but in many cases the algorithm emphasizes other variables than what macroeconomic theory suggests. Keywords – Time Series, Machine Learning, Econometric, GDP, Forecast