Predicting surprises to GDP : a comparison of econometric and machine learning techniques

Abstract

This study takes its inspiration from the practice of nowcasting, which involves making short horizon forecasts of specific data items, typically GDP growth in the context of economics. We alter this approach by targeting surprises to GDP growth, where the expectation is defined as the consensus estimate of economists and a surprise is a deviation of the realized value from the expectation. We seek to determine if surprises are predictable at a better than random rate through the use of four statistical techniques: OLS, logit, random forest, and neural network. In addition to evaluating predictability we also seek to compare the four techniques, the former two of which are common in econometric literature and the latter two of which are machine learning algorithms most commonly seen in engineering settings. We find that the neural network technique predicts surprises at an encouraging rate, and while the results are not overwhelmingly positive they do suggest that the model may identify relationships in the data that elude the consensus.