| This thesis models (i) corporate exits (bankruptcy and mergers and acquisition) and (ii) corporate credit rating changes. For corporate exits, extensions of the basic univariate discrete hazard regression model are made in two directions: (i) competing risks and (ii) stochastic frailty. Both extensions are motivated and supported by current empirical evidence. Under a classical proportional hazard specification the stochastic frailty can be viewed as a baseline hazard with random effects. For credit rating migration, we consider an extension of the basic multi-period ordinal logistic regression model with a latent macroeconomic process. The inclusion of an unobservable random process is especially important since the true intercorporate dependence structure may not be well described by using only observable covariates. A common random factor, or stochastic frailty, imposes a cross sectional dependence and induces a serial correlation, which could be a useful feature for modeling the intercorporate dependence.The idea and estimation methods are applied to comprehensive real data sets compiled from various sources for both corporate exits and credit rating transitions. Quarterly firm specific financial variables and macroeconomic variables of U.S. industrial firms spanning 1986 to 2006 are used. These data sets have been constructed from Compustat and CRSP and have been accessed through a license from the Ivy Business School Library at UWO. In-sample and out-of-sample prediction results are provided and compared. We find the random process model is more robust against missing macroeconomic covariates, a very useful feature in regression models.A measure of prediction accuracy is important to test the model performance. Instead of the ad-hoc prediction power association methods and percentage measures currently used in the literature, we propose a Poisson approximation and entropy related measure to test and measure prediction accuracy in the aggregate and disaggregate fashion respectively. These developments appear to be novel in this field.Keywords. credit risk, credit score, credit rating, bankruptcy, merger and acquisition, hazard function, ordinal regression, competing risks, frailty, prediction, accuracy, in-sample, out-of-sampleThe parameters could be estimated by an EM algorithm, but this requires intensive computations. As an alternative we propose a posterior mode estimation method which may be viewed as a penalized maximum likelihood estimation with roughness penalty. By assuming a multivariate mean reverting autoregressive process prior, the resulting posterior mode estimates are smooth. We examine the relationship between the effect of smoothing and inclusion of observable macroeconomic variables. |
