Scenario Analysis with Multivariate Bayesian Machine Learning Models
Abstract
We present an econometric framework that adapts tools for scenario analysis, such as variants of conditional forecasts and generalized impulse responses, for use with dynamic nonparametric models. The proposed algorithms are based on predictive simulation and sequential Monte Carlo methods. Their utility is demonstrated with three applications: (1) conditional forecasts based on stress test scenarios, measuring (2) macroeconomic risk under varying financial stress, and estimating the (3) asymmetric effects of financial shocks in the US and their international spillovers. Our empirical results indicate the importance of nonlinearities and asymmetries in relationships between macroeconomic and financial variables.
Keywords
Cite
@article{arxiv.2502.08440,
title = {Scenario Analysis with Multivariate Bayesian Machine Learning Models},
author = {Michael Pfarrhofer and Anna Stelzer},
journal= {arXiv preprint arXiv:2502.08440},
year = {2025}
}
Comments
Keywords: conditional forecast, generalized impulse response function, Bayesian additive regression trees, nonlinearities, structural inference; JEL: C32, C53, E44