Machine learning forecasting in the macroeconomic environment: the case of the US output gap

This paper aims to forecast deviations of the US output measured by the industrial production index (IPI), from its long-run potential output, known as output gaps. These gaps are important for policymakers when designing relevant economic policies, especially when a negative output gap may show economic slack or underperformance, often associated with higher unemployment and low inflation. We use a dataset that includes 32 explanatory economic and financial variables and 18 lags of the IPI, spanning the period from 2000:1 to 2022:12, resulting in 50 variables and 276 monthly observations. The dataset is fed to five well-established machine learning (ML) methods, namely decision trees, random forests, XGBoost, long short-term memory (LSTM) and support vector machines (SVMs), coupled with the linear, the RBF and the polynomial kernel. Moreover, we use the standard elastic net logit method from the area of econometrics as a benchmark. Our results indicate that the tree-based ML techniques perform better in-sample, and the best overall forecasting model is the XGBoost achieving an out-of-sample accuracy of 91.67%.