Forecasting Economic Recessions Using Machine Learning:An Empirical Study in Six Countries

This paper proposes a methodology for forecasting economic recessions using Machine Learning algorithms. Among the methods examined are Artificial Neural Networks (ANN), Support Vector Machines (SVM) and Random Forests. The datasets analysed refer to six countries (Australia, Germany, Japan, Mexico, UK, USA) and cover a time span of more than 40 years. All methods are compared against each other in terms of six evaluation metrics on their out-of-sample performance. In contrast to most similar empirical studies, the methodology developed focuses on the timepoints of the last four quarters before a recession begins rather than on those of a recession per se. It has been found that the SVM method tends to out-perform the others, as it classified correctly at least 75% of the pre-recessionary periods for half of the countries, with mean overall classification accuracy around 90% in these cases. Moreover, for all the countries under study, the traditional Logit and Probit models are always inferior to at least one Machine Learning-based model. Additionally, it turns out that macroeconomic variables representing a kind of debt - such as, household debt - are most frequently considered as important across the six datasets, in terms of the Mean Decrease Gini measure.