Coming from a market risk background, I am wondering how to validate / backtest a credit risk model in practice. Here, I am specifically not asking about the PD/migration validation, but about the default aggregation parameters (, say sector correlations), i.e. those components of the model that operate on the credit risk dependence structure.

In Market Risk applications, I would usually look for time series of my risk factors (say, rates obtained from Bloomberg/ICAP); combine them with my valuation model(s), potentially add another layer of distributional assumptions (or not) and estimate desired quantities, say a VaR. At the same time, the availability of market data (seems to) allows me to perform a daily backtest of my model and offer insights into the quality of my assumptions on the joint distribution of the risk factors.

From a credit risk perspective, how do practitioners proceed with this topic, especially if your portfolio is not retail heavy (homogeneuos), but consists only of a small number of debtors, say between 1k and 10k names? In this scenario, I would assume that there simply is not a sufficient number of defaults in my portfolio so that I can do proper backtesting ... Would you then simply buy default data (as I would do with with market data from Reuters, Bloomberg / ICAP...) from somewhere else (a default data / credit quality data vendor?) and try to make sense of that data relative to your portfolio? Or do you approach your aggregate model validation from a more theoretical standpoint, say using competitor models, stressed parameters etc.?

I hope my question is not too general - effectively I am wondering how to do credit risk model backtesting - and whether that really is a thing. Thanks for any pointers.


1 Answer 1


Back-testing VaR models makes sense given the right holding period. Back-testing daily any credit portfolio (by credit portfolio here we assume non-listed instruments, but more like loans and other bilateral agreements, even securitization investments) I could argue that perhaps doesn't make much sense. I assume that your model is somehow close to the traditional RiskMetrics models whereby you have a transition probability matrix and then you work out expected and unexpected loss of your credit portfolio. This transition matrix at best consists of annually combined data and attempts to predict the percentage of credit deteriorating or improving. These can be scaled of course down to a quarter or a month if you want to.

Like all back-testing routines, you want to compare the predicted loss vs the realized loss. In the event of default or partial default (delinquencies) you portfolio's value should be marked down accordingly. This number is then compared against the expected loss (prediction). I wouldn't run this more frequently than monthly.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.