Latent variable portfolio credit models
From FinMath
Stub.
This term describes dynamic models for default and default intensities of multiple assets characterized by Bayesian inference of unobserved "factors" computed for a plurality of time steps. Some latent variable models may be interpreted as attempts to filter dynamic evolution of conditional default probabilities models which are at face value static. Computations in these models are often special cases of inference algorithms on Bayesian networks including Gibbs sampling (as described in Duffie et al who use the term "Frailty Model") and more general Markov Chain Monte Carlo techniques. "Factor" structural and copula models may be cast as latent variable models, though the naive inference of default intensity dynamics may lead to questionable changes in default probabilities post default. Gaussian factors lead to non-trivial sampling problems given the difficulty of conditioning multivariate normal variables on inequality evidence - to borrow terminology from Bayesian inference literature - thus leaving plenty of room for cleverness and practical improvement on overaching technologies.
Latent variables may have economic significance as economic indicators or proxies for macro, industry or local influences on credit quality. Alternatively, the factors may be mathematical conveniences derived from a representation of dependence among variables and used only as intermediate steps in a calculation. There may be many equivalent representations, thus leaving open the approach to general criticisms of factor models, and superficially different models are often equivalent (Andersen - pdf).
