Predicting South African GDP Growth Rates Using Factor Models and Machine Learning Techniques

This paper investigates the usefulness of a combined factor estimation and shrinkage approach in forecasting South African GDP growth rates one-quarter ahead. For this purpose, 62 quarterly macroeconomic variables from Q1 1996 to Q1 2019 are examined in an empirical forecasting experiment in addition to a simulation study of the constructed model. It is found that the hybrid model, combining boosting with principal component analysis, leads to significantly lower forecast errors than standard autoregressive forecasting methods. However, this result is heavily dependent on the boosting parameters. Simulations of data with varying dimensions reveal that the gain in forecasting accuracy achieved by the combination method is larger when more latent factors exist and that datasets with higher dimensions increase performance. As an extension, the effectiveness of using a recurrent neural network with long-short term memory to produce the forecasts is evaluated, giving rise to similar findings. Hence, it is concluded that machine learning methods are valuable tools to predict quarterly South African GDP growth rates if the parameters are chosen properly.