Suppose a model experiences a large residual in a given period. Does this necessarily mean that the model is no good?
Of course not.
Models are not built to explain everything that can possibly happen. If something truly unusual occurs – Covid certainly qualifies – we may need to be fatalistic about it and accept that the phenomenon was simply unmodellable.
Suppose you had a credit model, in late 2019, that was capable of predicting what just happened with coronavirus.
Back then I would have argued that your model was badly overfitted.
Despite everything that’s happened this year, I’d make precisely the same assertion right now.
One big fear is that regulators and validators will now demand credit models specifically capable of capturing the impact of Covid. That they’ll shut the gate after the horse has bolted.
Ideally, the lessons drawn from Covid will be more general in nature. Our models should adapt so they help with the next crisis, whatever form it happens to take.