We perform a thorough comparative analysis of factor models and machine learning to forecast Japanese macroeconomic time series. Our main results can be summarized as follows. First, factor models and machine learning perform better than the conventional AR model in many cases. Second, predictions made by machine learning methods perform particularly well for medium to long forecast horizons. Third, the success of machine learning mainly comes from the nonlinearity and interaction of variables, suggesting the importance of nonlinear structure in predicting the Japanese macroeconomic series. Fourth, while neural networks are helpful in forecasting, simply adding many hidden layers does not necessarily enhance its forecast accuracy. Fifth, the composite forecast of factor models and machine learning performs better than factor models or machine learning alone, and machine learning methods applied to principal components are found to be useful in the composite forecast.

No abstract