Asset managers are leaning on AI to gauge risk. But should they?

Faye Kilburn expresses doubt that robots can learn to manage risk, at least in the short run. Her article was published by

Most attempts to apply AI so far have been in stock price forecasting. But risk managers are asking how the technology can be harnessed in their domain also. One area of exploration is the use of machine learning to replace traditional approaches to risk modelling.

Conventional risk models often treat markets as webs of essentially linear relationships. Each factor that contributes to risk gets a weighting – and those weightings don’t change. That’s a problem, as it tends to miss tail risks, according to Gareth Shepherd, managing partner at G Squared Capital, a London-based discretionary firm using machine learning to better understand idiosyncratic risk.

“The traditional approach equity research analysts take of using linear regression and bell curves to model idiosyncratic risk is a fairly antiquated tool. It’s like putting a horizontal ruler on a spherical Earth and trying to measure it. It’s just a weird thing to do,” he says.

Dario Villani, chief executive of machine learning hedge fund Duality Group, says current risk models fly in the face of the fact that assets are driven by elusive, shifting relationships, rather than fixed laws of risk and return. Villani is one of a group of quants who see machine learning as transformative, potentially unlocking the secrets of this non-linear behaviour.

In a non-linear machine learning model, the weighting would change over time, depending on a multitude of factors. For example, non-linear prepayment models for agency mortgage-backed securities built by MSCI depend on 30–100 variables that interact with each other differently depending on whether the loan is in or out of the money.

But do the new techniques harbour risks of their own? Yes, they do. Models in general can go wrong by picking up on false patterns in data, or simply through being too hard to understand. Both faults are amplified many-fold by AI because the datasets are so much bigger and the algorithms themselves so much more complicated. BlackRock shelved some liquidity risk models built using neutral networks because it couldn’t understand their inner workings. These risks will require careful handling.