2DkcTXceO

374 Statistics in a new erajoint default intensities that were used to price CDOs and mortgage-backedsecurities. These models also failed to predict well the “frailty” traits of latentmacroeconomic variables that underlie mortgages and mortgage-backedsecurities.For a multiname credit derivative such as CDO involving k firms, it isimportant to model not only the individual default intensity processes butalso the joint distribution of these processes. Finding tractable models thatcan capture the key features of the interrelationships of the firms’ default intensitieshas been an active area of research since intensity-based (also calledreduced-form) models have become a standard approach to pricing the defaultrisk of a corporate bond; see Duffie and Singleton (2003) and Lando (2004).Let Φ denote the standard Normal distribution function, and let G i be the distributionfunction of the default time τ i for the ith firm, where i ∈{1,...,M}.Then Z i =Φ −1 {G i (τ i )} is standard Normal. Li (2000) went on to assume that(Z 1 ,...,Z M ) is multivariate Normal and specifies its correlation matrix Γ byusing the correlations of the stock returns of the M firms. This is an exampleof a copula model; see, e.g., Genest and Favre (2007) or Genest and Nešlehová(2012).Because it provides a simple way to model default correlations, the Gaussiancopula model quickly became a widely used tool to price CDOs and othermulti-name credit derivatives that were previously too complex to price, despitethe lack of convincing argument to connect the stock return correlationsto the correlations of the Normally distributed transformed default times. Ina commentary on “the biggest financial meltdown since the Great Depression,”Salmon (2012) mentioned that the Gaussian copula approach, which“looked like an unambiguously positive breakthrough,” was used uncriticallyby “everybody from bond investors and Wall Street banks to rating agenciesand regulators” and “became so deeply entrenched — and was makingpeople so much money — that warnings about its limitations were largely ignored.”In the wake of the financial crisis, it was recognized that better albeitless tractable models of correlated default intensities are needed for pricingCDOs and risk management of credit portfolios. It was also recognized thatsuch models should include relevant firm-level and macroeconomic variablesfor default prediction and also incorporate frailty and contagion.The monograph by Lai and Xing (2013) reviews recent works on dynamicfrailty and contagion models in the finance literature and describes a new approachinvolving dynamic empirical Bayes and generalized linear mixed models(GLMM), which have been shown to compare favorably with the considerablymore complicated hidden Markov models for the latent frailty processes or theadditive intensity models for contagion. The empirical Bayes (EB) methodology,introduced by Robbins (1956) and Stein (1956), considers n independentand structurally similar problems of inference on the parameters θ i from observeddata Y 1 ,...,Y n ,whereY i has probability density f(y|θ i ). The θ i areassumed to have a common prior distribution G that has unspecified hyperparameters.Letting d G (y) betheBayesdecisionrule(withrespecttosome