[Back]

M. Frühwirth, P Schneider, L. Sögner:
"Markov Chain Monte Carlo estimation of issuer-specific and bond-specific components of credit and liquidity risk";
Talk: International Workshop on COMPUTATIONAL AND FINANCIAL ECONOMETRICS, Genf; 2007-04-20 - 2007-04-22.

Credit risk literature and industry represent the difference between risky bonds and risk-free bonds in the form of spreads. These
spreads, in general, include credit risk, liquidity risk and any market microstructure related differences. The main objective of this
article is to model interest rate risk, issuer-specific and bond-specific risk and to develop an econometric methodology to separate
and analyze these three types of risk. One of the standard procedures in the preceding literature to separate between issuer-specific
and bond-specific components assumes for each issuer the observation of a benchmark bond, where for this particular bond no
bond specific component is included. We refine this approach by an estimation technique that allows identification and estimation
of issuer-specific components on the one hand and bond-specific components on the other hand without having to make this strong
assumption. The model we are going to apply is based on the Duffie/Singleton credit risk framework. It consists of three building
blocks: the first for the risk-free term structure, the second for the issuer specific component of one particular issuer and the third is
a bond specific factor. By the model assumptions the number of observable time series (yields and/or bond prices) is smaller than
the number of latent processes, such that direct maximum likelihood estimation becomes impossible. We solve this problem by
means of data augmentation and estimate the model parameters along with the latent processes by means of Markov Chain Monte
Carlo methods. Since the autocorrelations of the sample paths are high and each simulation step of the sampler also requires to
solve a system of ordinary differential equations, the computational effort of our analysis becomes very high. One the other hand
side , this methodology neither requires a benchmark bond nor any discretization schemes, such that an exact Bayesian analysis
can be performed. MCMC runs with simulated data show that the parameters of our model can be estimated with high precession.
Furthermore, we apply our methodology to empirical data. We infer a risk-free term structure process from liquid swapmarket
data. Based on these estimates, issuer-specific and bond-specific risk are estimated from corporate bond data from the German
corporate bond market.