Can Artificial Intelligence really manage risk?

Source: Wikipedia

There’s a new application for Artificial Intelligence. It’s called regulatory technology, or “regtech.” It takes the rulebook of a supervisory agency, like the FDIC, and translates it into a computerized logic engine. The regulators evaluate the logical consistency of their rules, and the banks structure and label their data to be queried. This allows the regulator to look for broad patterns in the data that might not be immediately apparent. AI can identify deep, dynamic structures much more robustly than a person.

But are these programs looking for the right thing? A hundred years ago the economist Frank Knight outlined the difference between risk and uncertainty. Risk can be measured and quantified. It generates statistical distributions that can be optimized and controlled. It’s a known unknown. Uncertainty is outside our data-set. We know that it’s relevant but we can’t quantify it. A card game generates risk. The possibility of playing with a stacked deck creates uncertainty.

“The Card Game” by Theodor Rombouts. Source: Web Gallery of Art

AI can’t manage uncertainty because you can’t optimize an algorithm with unknown data. Machine learning is really good at processing what can be seen. It can’t evaluate something that’s never happened before.

The focus of risk management is risk, not uncertainty. If the stock market declines by $200 billion, we have a pretty good idea what will happen. It’s a known risk. But when sub-prime mortgages declined by $200 billion in 2008, the ensuing chain reaction brought the financial system to the brink. There were no observations on which to train a machine-learning system. While human risk managers can miss uncertainty, too, we at least have the benefit of inference, analogy, and creativity. We know that history rhymes, and we can evaluate new developments against other theoretical frameworks. When sub-prime defaults began to ripple through the system, some risk managers executed a “Big Short” to protect themselves and their investors.

AI is useful in preventing historical failures from repeating. But it may miss the most dangerous forms of risk-taking: the ones we haven’t seen, yet. The end result will likely be lower volatility, but fatter tails – less day-to-day risk, but more systemic risk.

Douglas R. Tengdin, CFA