Dynamic Factor Trees and Forests - A Theory-led Machine Learning Framework for Non-Linear and State-Dependent Short-Term U.S. GDP Growth Predictions

Abstract

Machine Learning models are often considered to be “black boxes” that provide only little room for the incorporation of theory (cf. e.g. Mukherjee, 2017; Veltri, 2017). This article proposes so-called Dynamic Factor Trees (DFT) and Dynamic Factor Forests (DFF) for macroeconomic forecasting, which synthesize the recent machine learning, dynamic factor model and business cycle literature within a unified statistical machine learning framework for model-based recursive partitioning proposed in Zeileis, Hothorn and Hornik (2008). DFTs and DFFs are non-linear and state-dependent forecasting models, which reduce to the standard Dynamic Factor Model (DFM) as a special case and allow us to embed theory-led factor models in powerful tree-based machine learning ensembles conditional on the state of the business cycle. The out-of-sample forecasting experiment for short-term U.S. GDP growth predictions combines three distinct FRED-datasets, yielding a balanced panel with over 375 indicators from 1967 to 2018 (FRED, 2019; McCracken & Ng, 2016, 2019a, 2019b). Our results provide strong empirical evidence in favor of the proposed DFTs and DFFs and show that they significantly improve the predictive performance of DFMs by almost 20% in terms of MSFE. Interestingly, the improvements materialize in both expansionary and recessionary periods, suggesting that DFTs and DFFs tend to perform not only sporadically but systematically better than DFMs. Our findings are fairly robust to a number of sensitivity tests and hold exciting avenues for future research.