This is a cookbook that I compiled (for my personal use) when I was working for a major US bank as a model validator. I used it as a field manual for the validation of cross-sectional regression models in general and whole sale credit risk models (PD, LGD, CCAR) in particular.

End-to-end model validation includes data validation, model methodology validation, model implementation validation (i.e. code replication), and risk scenario analysis (e.g. CCAR stress testing). This manual addresses the needs of model methodology validation.

Once methodological issues are identified, typically two sets of tools are employed to fix or mitigate the issues: for outliers in the data, robust statistics of some sort are used to mitigate the impact of noise; for model specification issues, machine learning techniques are often used to identify "the right model". Summaries of these tools are outside this book's scope.

For whole sale credit risk models, data are often scarce and each outlier requires thorough understanding. That's why regression-based small sample statistical inference is still relevant and ML techniques are not a panacea.

This summary has two main features. First, it's a practical guide for practitioners, yet it retains the flavor of a systematic exposition of highly technical materials. Second, I tried my best to explain the intuitions behind various econometric tests by presenting the key ideas.

When I was at the Model Validation Group (MVG) of the US bank aforementioned, I had the honor to work with a group of highly dedicated, highly competent, and super easy-going quants and financial economists. They have gone to different places and different directions ever since. This summary is a salute to my former MVG colleagues: I wish you all the best, my friends.