Forecasts of the Tanzanian gross domestic product including factor models and the machine learning technique boosting

There has been extensive research on how ‘big data’ and machine learning techniques are useful for modeling low frequency macroeconomic variables. In this paper, it is analyzed if machine learning, variable selection, and shrinkage method boosting within the factor-augmented autoregression are helpful in forecasting the gross domestic product (GDP) of Tanzania. In addition, it is tested whether this forecast method outperforms the ‘simple’ auto-regressive (AR) with the lag order selected by the Bayesian information criterion. The estimation is based on a combination of the lagged GDP growth of Tanzania and the dynamical factors. The latter are based on principal component analysis (PCA) and contain the GDP growth of 51 other African countries from 1963 up to 2016. The evaluation of the effectiveness of the forecasts methods is based on simulations that include a data generating process. It was found that this method proved to generate reliable results, given that the input values are accurately described. The five forecasts of the GDP of Tanzania result all in a lag order of 2 selected by the Bayesian information criterion. The MSFE of the boosting method that includes principal component analysis is lower than the MSFE of the benchmark AR(2) method. Therefore, it was found that the factor models and the machine learning technique boosting are indeed reliable means for modeling and forecasting the gross domestic product of Tanzania, and they outperform the ‘simple’ AR(2) model.