May 05, 2018

What if Artificial Intelligence helps predict decently correct Portfolio Variant Bank Capital Requirements?

Sir, Tim Harford refers to “Prediction Machines by Ajay Agrawal, Joshua Gans and Avi Goldfarb [which] argues that we’re starting to enjoy the benefits of a new, low-cost service: predictions. Much of what we call artificial intelligence is best understood as a dirt-cheap prediction. “Cheap innovations often beat magical ones” May 5.

If a credit to a risky borrower is not excessively large, and carries a correct risk premium, it can provide more safety to a bank’s portfolio, than a credit to a borrower perceived as safe.

Unfortunately, and as was stated in “An Explanatory Note on the Basel II IRB (internal ratings-based) Risk Weight Functions”,“Taking into account the actual portfolio composition when determining capital for each loan - as is done in more advanced credit portfolio models - would have been a too complex task for most banks and supervisors alike.”

And so to make up for that difficulty the regulator decided: “In the context of regulatory capital allocation, portfolio invariant allocation schemes are also called ratings-based. This notion stems from the fact that, by portfolio invariance, obligor specific attributes like probability of default, loss given default and exposure at default suffice to determine the capital charges of credit instruments.”

And to justify it they argued that: “essentially only so-called Asymptotic Single Risk Factor (ASRF) models are portfolio invariant (Gordy, 2003).”

But, what if Artificial Intelligence had then allowed bank regulators to make their capital requirements portfolio variant? Many other bad things could of course have happened, but surely AI would have warned against too much exposure being built up with assets perceived (residential mortgages), decreed (sovereigns like Greece) or concocted (AAA rated securities) as safe. And also about too little exposures to what is perceived risky, like loans to entrepreneurs.

The danger is though that since we are clearly not capable to duly question human regulators’ expertize, we could end up questioning even less any Artificial Intelligence’s also quite possible mumbo jumbo.