Measuring Sovereign Contagion in Europe Presented by Jingjing XIA Caporin, Pelizzon, Ravazzolo, and Rigobon (2013)
Abstract What they do? Analyze the sovereign risk contagion using CDS spread and bond premium data for 8 major European countries. How they do it? Nonlinear regression; quantile regression; Bayesian quantile with heteroskedasticity What they get? - the propagation of shocks is constant during risk spillover among countries is not affected by size of shock - with bond yield data intensity of propagation decreases after 2008 financial crisis
Outline 1. Introduction 1.1 Motivation and Challenges 1.2 Methodology 2. Data Description 3. Estimation Approach 3.1 Nonlinear Approach 3.2 Quantile Regression 3.3 Bayesian Quantile Regression with Heteroskedasticity 4. Robustness Checks 4.1 Stability of Parameters 4.2 Bond Spread Analysis 5. Conclusion
Introduction 1) Motivation 2010 European crisis gives rise to a series of literures Economists and policy makers are interested in topics such as measuring impact of crisis events in some country on other countries, and identifying channel of shock transmission 2) Challlenges It is empirically difficult to address these questions a) the definiton of contagion: the authors define contagion as the difference of shock propagation during normal time and crisis time. b) normal empirical techniques are problematic: omitted variable bias, heteroskedastic errors, simultaneous equation bias. c) structural estimation is constrained: it requires the specification the channel of contagion ex-ante.
Introduction 3) Methodology First explore nonlinear models but problematic The baseline model is a reduced form quantile regression model The authors compare coefficients at differnt quantiles. Use Bayesian approach to estimate QRM with heteroskedasticity Use DCC test for parameter stability and bond data for disadvantages in CDS data
Data Description 1. 5 year sovereign CDS spreads 2. Daily data from Datastream 3. 7 Euro zone countries and UK 4. Euro zone countries include Greece, Ireland, Portugal and Spain (crisis countries) as well as France, Germany and Italy (major ) 5. Data is from November 2008 to September 2011
Data Description
Estimation Approach 1) Rolling correlation and exceedence correlation Non-parametric approaches using 60 observation as rolling window. They suffer from heteroskedasticity. 2) Projection methods Contagion is reflected as a significant coefficient of nonlinear linkages The model takes into account heteroskedastic errors (small)
Estimation Approach 3. Quantile regression An estimation technique that aims at estimating conditional quantiles of dependent variables given control varibles instead of conditional mean like in least square methods. Usually we estimate 10%, 25%, 50% (median), 75% and 90% quantile coefficients. Advantages: 1) it can provide information about relationship between y and x at different points in the conditional distribution of y; 2) median regression estimates are more robust to outliers than mean regression; 3) it offers a richer characterization of the data; 4) it is suitable for heteroskedastic data.
Estimation Approach General form of quantile regression model Quantile regression estimates are the coefficients inside that minimizes the expected value of the check function
Estimation Approach Quantile regression model used in the paper is quantile dependent parameters we want to estimate. The coefficient of interest here is We define a normal scenario with and a bad scenario Compare coefficients under different scenarios and see if there is change in propagation mechanism (contagion)
Estimation Approach except for UK, most of the coefficients do not change drastically at median quantile and at 99% quantile.
Estimation Approach Where is heteroskedasticity? Volatility is the residual time varying standard deviation computed using and. The authors use Bayesian approach to estimate the parameters in the equation.
Estimation Approach The coefficients of Bayesian quantile regression with heteroskedasticity for French CDS is
Estimation Approach All countries have stable coefficients across quantiles except Italy
Robustness Checks 1. Parameter stability To show that the quantile regression parameters do not suffer from the problem of omitted variables and simultaneous equations like other techniques, the authors conduct DCC test (Determinant of the Change in the Covariance matrix). Intuitively speaking, DCC test compares the covariance matrices across two samples and take the determinant to express the statistic as a scalar Under null hypothesis that there is no change in the covariance structural across samples, DCC will be 0.
Robustness Checks In the paper, the authors define a series of thresholds which divide the data into high and low volatility regimes and then compute the covariance matrix using data from these two regimes and test whether or not DDC=0.
Robustness Checks 2. Bond spread analysis Disadvantage of using CDS: 2008 financial crisis Instead of comparing normal time and bad time, we compare bad time and really bad time Sample selection may be the reason for stability The bond spread data is obtained by differencing 5 year bond yields with 5 year interest rate swap rates. Data goes from 2003 to 2011 and is divided into 3 periods, pre- crisis ( ), post-crisis ( ) and full ( )
Robustness Checks
Conclusion 1. The propagation of shocks in CDS and bond spreads among European countries has been remarkably constant across quantiles during crisis period All the increases in correlation comes from larger shocks rather than similar shocks propagated with higher intensity (contagion). 3. Bond market data provides evidence for propagation instability supporting a reduced correlation during bad time. 4. Market views euro zone area as perfectly integrated before crisis and this view is broken after Strong euro zone countries do not need to worry about contagion but large country specific shocks can still cause trouble.