Essays on Economic Forecasting with Machine Learning

Abstract

[1] This cumulative dissertation includes four essays on economic time series forecasting. It broadly examines how we may meaningfully adopt and adapt machine learning algorithms in data-rich environments for more accurate predictions of (macro-)economic dynamics. All four chapters derive from a consistent data basis: The Federal Reserve Economic Data (FRED) provided by the Federal Reserve Bank of St. Louis. The first chapter, co-authored with Boriss Siliverstovs, is mainly concerned with the evaluation of macroeconomic forecasts for a wide range of macroeconomic indicators in case of possible instabilities.[2] This study systematically broadens the relevance of possible model performance asymmetries across business cycles in the spirit of the recent state-depen- dent forecast evaluation literature (e.g. Chauvet & Potter, 2013) to hundreds of macroeconomic indicators and deepens the forecast evaluation of the recent factor model literature on hundreds of target variables (e.g. Stock & Watson, 2012b) in a state-dependent manner. Our results are consistent with both strands of the literature and generalize the former to over 200 macroeconomic indicators and differentiate the latter across three levels of temporal granularity: We document systematic model performance differences in both absolute and relative terms across business cycles (longitudinal) as well as across variable groups (cross-sectional) and find these performance differences to be robust across several alternative specifications. The cross-sectional prevalence and robustness of state dependency shown in this article encourages economic forecasters to complement model performance assessments with a state-dependent evaluation of predictive ability. The second chapter focuses on predictive macroeconomic modeling with ensembles of trees. While the recent macroeconomic forecasting literature shows that double targeted forests (DTF) can enhance the predictive ability of single targeted forests (STF) (e.g. Borup, Christensen, Mühlbach, & Nielsen, 2020; Medeiros, Vasconcelos, Veiga, & Zilberman, 2021; Wochner, 2018), I study the merits of triple targeted forests (TTF) as a flexible and rich framework for non-linear tree-based ensemble predictions in macroeconomic forecasting when dealing with high-dimensional and mixed-frequency environments. TTFs offer interesting and novel modeling choices across the different steps of the framework, especially in combination with recent advances put forth in the literature (e.g. novel/traditional factors, supervised/unsupervised factor selection). An extensive out-of-sample forecasting experiment for short-term U.S. output growth shows that optimally tuned and suitably designed TTFs can offer considerable gains in predictive accuracy and may outperform well-established distributed lag and autoregressive models as well as single and double targeted forests. Nevertheless, factor-augmented linear models remain a competitive benchmark. In the third chapter, I elaborate on forest-based model estimation in the presence of possible state-dependent dynamics and thereby build upon the first two chapters. Specifically, I propose so-called Dynamic Factor Forests (DFF) for macroeconomic forecasting, which extend the model-based trees of Zeileis, Hothorn, and Hornik (2008) in the spirit of Garge, Bobashev, and Eggleston (2013) to model-based forests and synthesize the recent machine learning, business cycle and dynamic factor model literature within a unified statistical machine learning framework. DFFs are state-dependent, non-linear and smoothed forecasting models that allow us to embed theory-led factor models in powerful tree-based machine learning ensembles conditional on the business cycle state. DFFs incorporate in the spirit of the recent literature four key features: (i) Complex conditioning sets allow to account for generalized and potentially time-varying state-dependent dynamics while (ii) forest- based randomization decorrelates the dynamic factor trees in the ensemble. (iii) The inclusion of Lasso-based state-dependent regularization not only serves as a guard against overfitting but also enables switching mechanisms between structurally distinct state-dependent models. (iv) In addition to traditional principal component-based factors, novel factor transformations are examined. An extensive out-of-sample forecasting experiment for short-term U.S. GDP growth provides promising results in favor of DFFs. In the fourth chapter, I combine recent advances in the literature for more accurate density predictions of U.S. inflation. Recent work by Medeiros et al. (2021) shows that point forecasts of Breiman’s (2001a) random forest machine learning algorithm can systematically outperform well-established benchmarks at predicting U.S. inflation. This chapter extends their work from point to density forecasts: On the one hand, the predictive densities of well-established random walk, autoregressive and stochastic volatility benchmarks are examined. On the other, three recent forest-based machine learning models are assessed: Namely, Quantile Regression Forests (Meinshausen, 2006), Bayesian Additive Regression Trees (H. A. Chipman, George, & McCulloch, 2010) and Distributional Forests (Schlosser, Hothorn, Stauffer, & Zeileis, 2019). The chapter shows that the most recent Distributional Forest (DIFO) algorithm often achieves the best point and density predictions. DIFOs allow to simultaneously account for locally evolving means, nonlinearities, time-varying volatilities as well as heavy-tailed densities all of which constitute stylized characteristics of inflation dynamics highlighted in the recent literature, which is shown to be key for their predictive superiority. [1] The summaries provided herein are based on the abstracts of the corresponding essays. [2] The remainder of this paragraph corresponds to the abstract in Siliverstovs and Wochner (2021).