Quantile graphical models: prediction and conditional independence with applications to systemic risk

Victor Chernozhukov, Alexandre Belloni, Mingli Chen

TL;DR

This paper introduces Quantile Graphical Models (QGMs), a powerful framework for mapping dependencies between variables at specific points in their distribution—most importantly, the extreme tails. By moving beyond standard Gaussian assumptions, QGMs allow researchers to visualize and quantify financial "contagion" and systemic risk in high-dimensional networks where traditional models fail.

What Problem Does This Paper Address?

Traditional statistical models, like Gaussian Graphical Models (GGMs), assume that data follows a normal distribution and that relationships between variables are constant. However, in fields like finance, we are rarely interested in the "average" day. We care about the "crash" days—the extreme tail events.

Real-world data is often non-Gaussian, asymmetric, and heavy-tailed. Two stocks might appear independent during normal market conditions but become highly correlated during a meltdown. Standard models miss this "tail interdependence." Furthermore, modern datasets are often "high-dimensional" (where the number of variables $d$ is much larger than the number of observations $n$), making traditional inference unreliable. The authors aim to provide a mathematically rigorous way to estimate these networks at any quantile ($\tau$) while maintaining valid statistical inference.

Methods

The authors propose two distinct types of QGMs:

  1. Conditional Independence QGMs (CIQGMs): These characterize whether two variables are independent given all others at a specific quantile. They are ideal for validating the structure of causal models.
  2. Prediction QGMs (PQGMs): These focus on the "best linear predictor" under asymmetric loss functions. These are more robust because they allow for "misspecification"—meaning the model still provides useful insights even if the underlying relationship isn't perfectly linear.

Key Technical Innovations: * $\ell_1$-Penalized Quantile Regression: To handle high-dimensional graphs, the authors use Lasso-style penalties to "thin out" the network, identifying only the most significant connections. * Neyman Orthogonality & Multiplier Bootstrap: They develop a de-biased inference method that allows for valid confidence intervals even after performing model selection. This ensures that the "edges" identified in the graph are statistically significant. * $\mathcal{W}$-Conditioning: An extension that allows the graph to change based on external events (e.g., a graph of stock dependencies specifically when the market index is down).

Datasets

The paper validates the models using both simulations and real-world financial data:

  • International Financial Contagion: Equity index returns for 45 countries (21 developed, 21 emerging, and 3 frontier markets) from the MSCI index (2009–2013).
  • U.S. Stock Interdependence: Daily returns for 86 S&P 500 stocks from the CRSP database (2003–2013).
  • Synthetic Benchmarks: Hub graphs and Erdős-Rényi random graphs used to prove that the algorithm can correctly recover known network structures in non-Gaussian settings.

Key Insights and Ideas

  • The "Aha" Moment on Tail Risk: The authors show that financial networks are not static. The "Up Tail" ($\tau = 0.9$, representing high volatility/distress) is significantly denser and more interconnected than the median network. This confirms that markets become "locked together" during crises.
  • Asymmetry is Feature, Not a Bug: Unlike Gaussian models where the relationship between A and B is symmetric, QGMs allow for directed edges. A shock in Germany might predict volatility in Greece, but the reverse might not be true.
  • Robustness to Non-Gaussianity: By using quantiles, the model is naturally robust to outliers and heavy tails, which are ubiquitous in financial time series.
  • Systemic Risk Contribution: The model introduces a "network-cooperated $\Delta CoVaR$," which measures how much a specific institution (or country) contributes to the risk of the entire system by accounting for spillover effects across the whole network.

Key Results

  • Network Density: In the application to international contagion, the authors found that interdependence increases dramatically at the up-tail. This suggests that "diversification" often disappears exactly when investors need it most.
  • Identifying "Systemic Shippers": The model identified Germany and France as the primary sources of volatility transmission in the Eurozone, while countries like Greece and Spain were net "receivers" of systemic risk.
  • Superiority over Glasso: In simulations with non-Gaussian data, the QGM approach correctly identified conditional dependencies that standard Graphical Lasso (Glasso) completely missed or misidentified.

Limitations and Future Work

  • Computational Intensity: Solving high-dimensional penalized quantile regressions for every node in a large graph is computationally more expensive than the OLS-based methods used in Gaussian models.
  • Linearity: While PQGMs handle misspecification, the core of the model still relies on linear combinations of variables. Exploring non-linear kernels within this quantile framework is a suggested path forward.
  • Temporal Dynamics: The current framework is largely cross-sectional. Integrating these models into a more formal time-series (VAR-style) framework to capture lead-lag relationships more explicitly is a potential area for expansion.

Summary updated

Abstract

The understanding of co-movements, dependence, and influence between variables of interest is key in many applications. Broadly speaking such understanding can lead to better predictions and decision making in many settings. We propose Quantile Graphical Models (QGMs) to characterize prediction and conditional independence relationships within a set of random variables of interest. Although those models are of interest in a variety of applications, we draw our motivation and contribute to the financial risk management literature. Importantly, the proposed framework is intended to be applied to non-Gaussian settings, which are ubiquitous in many real applications, and to handle a large number of variables and conditioning events. We propose two distinct QGMs. First, Condition Independence Quantile Graphical Models (CIQGMs) characterize conditional independence at each quantile index revealing the distributional dependence structure. Second, Prediction Quantile Graphical Models (PQGMs) characterize the best linear predictor under asymmetric loss functions. A key difference between those models is the (non-vanishing) misspecification between the best linear predictor and the conditional quantile functions. We also propose estimators for those QGMs. Due to high-dimensionality, the two distinct QGMs require different estimators. The estimators are based on high-dimensional techniques including (a continuum of) `1-penalized quantile regressions (and low biased equations), which allow us to handle the potential large number of variables. We build upon a recent literature to obtain new results for valid choice of the penalty parameters, rates of convergence, and confidence regions that are simultaneously valid. We illustrate how to use QGMs to quantify tail interdependence (instead of mean dependence) between a large set of variables which is relevant in applications concerning with extreme events. We show that the associated tail risk network can be used for measuring systemic risk contributions. We also apply the framework to study international financial contagion and the impact of market downside movement on the dependence structure of assets' returns.
BibTeX
@article{chernozhukov2018quantile,
  author = {Victor Chernozhukov and Alexandre Belloni and Mingli Chen},
  title = {{Quantile graphical models: prediction and conditional independence with applications to systemic risk}},
  journal = {Journal of Econometrics},
  year = {2018},
}
Figure