Let the Tree Decide: FABART. A Non-Parametric Factor Model for Nonlinear Oil Shock Transmission

Abstract

The question of how oil supply news shocks transmit to real activity, financial conditions and regional labor markets is back at the center of the macroeconomic research agenda. To answer this question, we introduce the Factor Bayesian Additive Regression Tree (FABART) model, a nonlinear factor-augmented vector autoregression model, and apply it to a large U.S. macro-financial dataset with externally identified oil supply news shocks. The framework incorporates a flexible nonparametric measurement equation that allows nonlinear transmission to emerge from the data without imposing a specific functional form on the asymmetry. We find that adverse oil supply news shocks generate stronger and more persistent contractions in real activity than the expansions associated with favorable shocks of comparable magnitude, with especially pronounced differences in industrial production, interest rates and equity prices. Employment responses are highly heterogeneous across U.S. states, with substantially stronger contractions in manufacturing-intensive regions than in energy-producing states. Across shock magnitudes, nonlinearities arise mainly between very small and moderate oil-price movements: small shocks generate weak and imprecisely estimated responses, while even moderate shocks produce economically meaningful effects on industrial production and regional employment. Larger shocks, however, do not systematically generate proportionally stronger responses across variables and shock signs.