Distributional Vector Autoregression: Eliciting Macro and Financial Dependence

Yunyun Wang, Tatsushi Oka, Dan Zhu

TL;DR

This paper introduces Distributional Vector Autoregression (DVAR), a framework that models how the entire probability distribution of multiple economic variables—not just their averages—evolves over time. By applying this to US data, the authors reveal that financial crises don't just lower growth; they can split the economic outlook into multiple "modes," creating a complex landscape of risk that traditional models miss.

What Problem Does This Paper Address?

In macroeconomics and finance, the standard tool for decades has been the Vector Autoregression (VAR). Traditional VARs are great at telling us how the average of one variable (like GDP) responds to a shock in another (like interest rates).

However, the "average" is a poor descriptor of reality during extreme events like the 2008 financial crisis or the COVID-19 pandemic. During these times, distributions become skewed, "fat-tailed" (extreme events become more likely), or even multimodal (where there are two or more distinct likely outcomes). Standard VARs, which often assume a simple bell-curve (Gaussian) error structure, are blind to these structural shifts in risk. While some newer models look at specific "at-risk" quantiles, they often struggle to model the joint relationship between multiple variables without falling into the "curse of dimensionality."

Methods

The authors propose a semiparametric approach that combines Distributional Regression (DR) with a technique called distribution factorization.

  1. Factorization: Instead of trying to estimate a complex multi-dimensional distribution all at once, they break it down into a sequence of simpler, conditional one-dimensional distributions.
  2. Semiparametric Estimation: For each step, they use a flexible regression model. Unlike a standard regression that finds one "best fit" line, DR fits a collection of models across the entire range of possible outcomes. This allows the impact of past events to vary depending on whether the economy is currently in a boom or a bust.
  3. Distributional Impulse Response Function (DIRF): This is the paper's most novel contribution. A traditional IRF shows how a shock moves the mean of a variable over time. A DIRF shows how a shock changes the entire shape of the future probability distribution—revealing if a shock makes the future more uncertain, more skewed, or more likely to collapse.

Datasets

The authors evaluate their model using two primary US indices from 1973:Q1 to 2019:Q1:

  • Real GDP Growth: Quarterly data from the Bureau of Economic Analysis.
  • National Financial Condition Index (NFCI): A weighted average of 105 measures of financial activity (risk, liquidity, and leverage) from the Federal Reserve Bank of Chicago.
  • Mixed-Frequency Extension: In the appendix, they also incorporate monthly NFCI data to see if higher-frequency signals improve quarterly forecasts.

Key Insights and Ideas

  • The Emergence of Multimodality: One of the most striking findings is that during the Great Recession (specifically 2008:Q4), the joint distribution of GDP and financial conditions became multimodal. The economy didn't just have a "lower average"; it had two distinct "likely" states—one bad, and one much worse.
  • Asymmetric Dependence: The relationship between finance and the macroeconomy isn't a two-way street. Shocks that tighten financial conditions have a massive, long-lasting impact on the "downside risk" of GDP. Conversely, improving GDP growth during a financial crisis has a surprisingly limited effect on fixing the underlying financial conditions.
  • The "Curse" Avoided: By using a semiparametric DR approach rather than a purely non-parametric kernel method, the authors can include more "lagged" variables (past data points) without the model becoming computationally impossible or losing accuracy.

Key Results

  • Superior Calibration: Using a "Probability Integral Transform" (PIT) test, the authors show their DVAR model is better calibrated than traditional kernel-based models, meaning its predicted probabilities actually match real-world frequencies.
  • Crisis Dynamics: In 2008:Q4, the model identified two distinct modes for the one-quarter-ahead forecast: one centered around -2% GDP growth and another even more extreme. As the forecast horizon extended to one year, these modes merged back into a single, more stable (unimodal) distribution.
  • Policy Simulation: The authors simulated a "distributional impulse"—essentially asking, "What if policy had successfully prevented financial tightening in late 2008?" The DIRF showed this would have significantly "thinned" the left tail of the GDP distribution, effectively eliminating the most catastrophic growth outcomes for 2009.

Limitations and Future Work

  • Variable Ordering: Like many VAR models, the results can be sensitive to the order in which variables are added to the factorization (e.g., does finance affect GDP first, or vice versa?). While the authors found their results were robust to alternative orderings in this specific case, it remains a theoretical concern for larger systems.
  • Dimensionality: While DVAR handles more variables than non-parametric methods, it still faces challenges as the number of variables grows. The authors suggest using Lasso regularization (a way to automatically ignore less important variables) as a promising path for scaling the model to dozens of variables.
  • Computational Intensity: Estimating parameters across every point of a distribution is more "expensive" than a simple mean regression, though the authors provide a straightforward estimation method to mitigate this.

Summary updated

Abstract

Vector autoregression is an essential tool in empirical macroeconomics and finance for understanding the dynamic interdependencies among multivariate time series. In this study, we expand the scope of vector autoregression by incorporating a multivariate distributional regression framework and introducing a distributional impulse response function, providing a comprehensive view of dynamic heterogeneity. We propose a straightforward yet flexible estimation method and establish its asymptotic properties under weak dependence assumptions. Our empirical analysis examines the conditional joint distribution of GDP growth and financial conditions in the United States, with a focus on the global financial crisis. Our results show that tight financial conditions lead to a multimodal conditional joint distribution of GDP growth and financial conditions, and easing financial conditions significantly impacts long-term GDP growth, while improving the GDP growth during the global financial crisis has limited effects on financial conditions.
BibTeX
@misc{wang2023distributional,
  author = {Yunyun Wang and Tatsushi Oka and Dan Zhu},
  title = {{Distributional Vector Autoregression: Eliciting Macro and Financial Dependence}},
  year = {2023},
  eprint = {2303.04994},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2303.04994},
}
Figure