22. Machine Learning in Macroeconomics: Application to DSGE Models

Machine learning—a tool primarily belonging to the statistical, physical, and technological streams—has found its way into the economic sciences too. This paper attempts to conduct a brief review of literature with respect to some of the machine learning tools being used in macroeconomic forecasting, in particular, in DSGE models. The DSGE models have found widespread applicability owing to their firm theoretical micro-foundations in the Real Business Cycle school. Despite their practicality and applicability, these models drew criticism following the global financial crisis broadly due to structural drawbacks of (i) actual economic agents rarely exhibit rational expectations (ii) structure of markets rarely following the efficient market hypothesis, and (iii) aggregation of the behavior of the representative agent to derive macro-processes being true only under restrictive assumptions. This was further augmented by econometric issues of overfitting, simplifying yet false and non-testable identification restrictions, structural (and not practical) specification of the model, etc. The machine learning techniques are applied to attain the objective of reducing dimensionality, and better predictive performance. We then exposit the detailed specifications of two machine learning techniques utilized in studies which implemented the DSGE using machine learning tools, viz., random forests and support vector machines (SVMs), which are used in the classification of the data used in DSGE models. The machine learning techniques applied to classify the data in a DSGE model across studies include the random forests, SVMs, neural networks, logistic regressions, etc. The paper concludes with the implications that the introduction of the machine learning techniques has on DSGE models, and the policy implications thereof.