Modeling systemic risk with Markov Switching Graphical SUR models☆
Introduction
The financial crisis of 2008–2009 has shown that liquidity and valuation shocks may quickly propagate across the economic system and affect financial institutions operating in different markets, with different size and business structure, thus causing widespread losses and domino effects. Understanding the dynamics of cross-asset and cross-equity linkages is therefore of key importance to both systemic risk management purposes and to deal with contagion waves in times of crisis. Systemic risk shocks are conventionally referred to in the network literature as abrupt increases in the density of cross-firm connectivity (see, e.g. Billio et al., 2012, and references therein). Modeling firms’ connectedness through network analysis has been recently supported by a series of papers that have shown its in- and out-of-sample superior performance over traditional, correlation-based approaches.1
We extend the existing literature in a number of ways. First, we propose a system-wide inferential scheme based on a Markov-switching graphical model that allows us to simultaneously consider all the possible linkages among firms through constraints on the regime-specific conditional dependence structure. Second, we propose an identification scheme for different regimes of cross-firm connectivity based on a novel weighted eigenvector centrality measure, which is related to both the number and the strength of the connections between firms. Third, we provide an asset pricing application based on otherwise standard multi-factor pricing models in which the exposures of the assets to the risk factors (their regression betas) are allowed to change according to the regimes in cross-firm connectivity. This allows us to develop a unified framework where systematic and systemic risks are not mutually exclusive, in the sense that firm-specific exposures to sources of systematic risk may directly depend on the level of aggregate network connectivity.2 Finally, we provide a Metropolis-within-Gibbs sampling scheme which permits to jointly draw both the parameters of the factor pricing model, the latent states and the underlying regime-specific graphs.
Methodologically, we build upon the Gaussian Graphical model for multi-variate systems proposed by Whittaker (1990), Dawid and Lauritzen (1993), Lauritzen (1996), Carvalho et al. (2007b), Wang and West (2009), Wang (2010), Rodriguez et al. (2011), Wang et al. (2011) and Ahelegbey et al. (2016). In particular, Wang et al. (2011) developed a dynamic matrix-variate graphical model which allows to capture conditional dependencies under time-invariant graphs. We generalize and extend their framework by introducing Markov-switching dynamics in the graph structure within a Seemingly Unrelated Regression (SUR) model. More specifically, we propose a new Markov Switching Graphical SUR (MS-GSUR) which makes it possible to identify different regimes of network connectivity. The regime-switching identification problem is solved by exploiting the graph-theoretic properties of the state-specific conditional dependence structures of the error terms in the model. Thus, we provide a new weighted eigenvector centrality measure, which accounts not only for the number of adjacent nodes, but also for the weights of the edges and for the number of indirect connections between nodes, i.e. the number of walks between nodes. We formally show that our measure can be related to both (Bonacich, 1972) and communicability (see, e.g., Estrada and Hatano, 2008 and Estrada and Hatano, 2009), as well as to other measures used in the analysis of topological features of complex networks.
Our empirical application focuses on the cross-section of daily excess returns of the constituents of the S&P100 index over the period 1996–2014. The emphasis on stock returns is motivated by a widespread desire by policy makers and regulators to incorporate the most current information for the purpose of systemic risk measurement: stock prices of largely traded stocks reflect information more rapidly than other non-traded measures such as accounting variables. Informative cross-firm connectivity is estimated on the basis of the residual covariance structure of stock returns, conditioning on a set of tradable risk factors used in some of the most popular linear asset pricing models, namely the Capital Asset Pricing Model (CAPM), the three-factor model of Fama and French (1993), and Merton’s (1973) intertemporal CAPM (I-CAPM). The results are robust across model specifications.
Our main findings reveal that the dynamics of systemic risk can be captured by two regimes, in which a state of high connectedness characterizes the period 1999–2003 (marked by the passing of the Gramm–Leach–Bliley act, the inflating and bursting of the dot.com bubble, and the ensuing financial scandals) and subsequently, the great financial crisis of 2008–2009. We show that a few financial institutions turned out to heavily outweigh other firms in the network during these periods and that shocks to the Financial sector turned out to be the most systemically important. Finally, both a cross-sectional regression and rank-correlation analysis show that market capitalization does not significantly explain the relevance of a given firm within the network. However, firms which are more relevant within the network are more likely to suffer significant losses during periods of high systemic risk.
Section snippets
A Markov switching graphical SUR model
Let be the stock returns of the firm in excess of the risk-free rate at time , and the -dimensional vector of systematic risk factors. In our baseline formulation, each time series of returns is modeled as a dynamic multi-factor linear model where the matrix includes an intercept plus covariates, is a -vector of time-varying regression coefficients, and is an error term that can be
Inference on networks and parameters
Let , be a sequence of decomposable graphs, the set of complete prime components of and the set of separators of . If the joint distribution of excess stock returns is Markov with respect to a decomposable graph , the joint density of given factorizes as For each subgraph or , , is the corresponding matrix of covariates,
Empirical analysis
Our application focuses on all the constituents stocks of the S&P100 index for which we have at least fifteen years of continuous trading days as of the end of our sample, leaving us with firms. The sample period is May 1996–October 2014. The S&P 100 represents about 63% of the market capitalization of the S&P 500 and about half of the total market capitalization of the U.S. equity markets as of January 2017. These stocks tend to be the largest and most liquid companies in the U.S.
We
Conclusions
In the aftermath of the great financial crisis, one of the main questions for economists and market participants has concerned the extent to which the economy is robust to unexpected shocks. In the language of network analysis, this translates into a desire to understand the nature and patterns of cross-firm connectivity. We address this question by developing a novel Markov Switching Graphical Seemingly Unrelated Regression (MS-GSUR) model, which allows us to jointly estimate standard SUR-type
References (57)
- et al.
Econometric measures of connectedness and systemic risk in the finance and insurance sectors
J. Financ. Econ.
(2012) Some unique properties of eigenvector centrality
Social Networks
(2007)- et al.
Some contagion, some interdependence: More pitfalls in tests of financial contagion
J. Int. Money Finance
(2005) - et al.
On the network topology of variance decompositions: Measuring the connectedness of financial firms
J. Econometrics
(2014) - et al.
Communicability graph and community structures in complex networks
Appl. Math. Comput.
(2009) - et al.
Common risk factors in the returns on stocks and bonds
J. Financ. Econ.
(1993) Community detection in graphs
Phys. Rep.
(2010)- et al.
Graph characterizations from Von Neumann entropy
Pattern Recognit. Lett.
(2012) - et al.
A novel centrality method for weighted networks based on the kirchhoff polynomial
Pattern Recognit. Lett.
(2015) - et al.
Laplacian centrality: A new centrality measure for weighted networks
Inform. Sci.
(2012)
Complex network measures of brain connectivity: Uses and interpretations
NeuroImage
Sparse seemingly unrelated regression modelling: Applications in finance and econometrics
Comput. Statist. Data Anal.
The network origins of aggregate fluctuations
Econometrica
Measuring systemic risk
Rev. Financ. Stud.
Bayesian graphical models for structural vector autoregressive processes
J. Appl. Econometrics
The architecture of complex weighted networks
Proc. Natl. Acad. Sci. USA
Modern Graph Theory
Random Graphs
Factoring and weighting approaches to status scores and clique identification
J. Math. Sociol.
Understanding risk and return
J. Polit. Econ.
On gibbs sampling for state-space models
Biometrika
Simulation of hyper-inverse wishart distributions in graphical models
Biometrika
Dynamic matrix-variate graphical models
Bayesian Anal.
Monte Carlo Statistical Methods
Marginal likelihood from the gibbs output
J. Amer. Statist. Assoc.
Cited by (0)
- ☆
We are grateful to Sylvia Kaufmann (guest editor) and three anonymous referees for their insightful comments and suggestions. Also, we thank Luc Bauwens, Matthias Büechner, Guido Consonni, Fulvio Corsi, Francis X. Diebold, Sylvia Frühwirth-Schnatter, Siam J. Koopman, Fabrizio Lillo, Chris Sims, Mike West, Kamil Yilmaz, and seminar participants at the Warwick Business School, Scuola Normale Superiore of Pisa, HSE of Moscow, the NBER Summer Institute 2015, the NBER-NSF Time Series conference 2015, the 8th Annual SoFiE conference, the SYRTO Conference on Systemic Risk, the 2nd Vienna Workshop on High-Dimensional Time Series in Macroeconomics and Finance, the European Seminar on Bayesian Econometrics ESOBE 2015, and the FMA Orlando 2015, for their helpful comments and advices. Monica Billio and Roberto Casarin acknowledge financial support from the European Union, Seventh Framework Programme FP7/2007-2013 under grant agreement SYRTO-SSH-2012-320270, from the Institut Europlace de Finance under the Systemic Risk grant, and from the Italian Ministry of Education, University and Research (MIUR) PRIN 2010-11 grant MISURA. This research used the SCSCF multiprocessor cluster system at University Ca’ Foscari of Venice. A previous version of the paper was circulating under the title “Modeling Contagion and Systemic Risk”.