Quants ‘running into walls’ with AI interpretability

A number of hedge funds are “running into walls” in their efforts to apply artificial intelligence in investing – particularly when it comes to understanding what the algorithms they use are doing. 

“Some funds and firms are… stumbling and you don’t hear about it because they don’t open funds based on those [failed efforts],” said Joseph Simonian, director of quantitative research at Natixis Investment Managers.

Simonian was speaking on a panel at the Cayman Alternative Investment Summit on February 7 – a gathering of hedge fund managers including some of the biggest in the industry, such as Bridgewater Associates and Man Group, with a slant this year towards topics such as “The rise of the robot” and “The quantum shift: humans versus machines”.

Hedge funds, asset managers and banks are investigating the potential of machine learning in areas ranging from predicting mortgage prepayment rates to improving trading execution or helping glean insights from company analyst calls.

But one of the biggest obstacles, the speakers noted, was interpretability – the ability to explain why an AI algorithm came up with the output it did. This is easier for simple machine learning methods like decision trees, but becomes cumbersome for more complex methods like neural networks.

Michael Weinberg, chief investment officer of Mov37 – a spinoff from fund of funds Protégé Partners that works with start-up funds on machine learning strategies – said there was a split emerging between managers that require a strong economic rationale to deploy an AI algorithm and those who dismiss the need for one.

Blackrock in the past year shelved work on certain liquidity risk models based on neural networks because of the challenge of interpreting them. However, Dario Villani, former global head of portfolio strategy at Tudor Investment Corporation, who recently started a new machine learning hedge fund, the Duality Group, has argued that machine learning’s superiority lies in its ability to spot economic patterns that humans could not know intuitively.

“There’s a spectrum,” Weinberg said. “In terms of quants, some are using machine learning only when it’s entirely understood, and at the other extreme, some do not need to understand why they are generating alpha.”

Recently, senior industry figures have warned about the dangers of misusing the new technologies, such as allowing machine learning models to pick patterns from data that turn out to be statistical flukes.  

Simonian believes quants should do extensive pre-model work to determine a coherent rationale for how they apply machine learning before deploying a model based on false assumptions or bad data.

Today’s AI techniques are technically simpler than the methods that many quants grew up with in fields such as derivatives pricing and traditional econometrics, he said. But because the new techniques are more flexible, they can also be more risky.

“You can do a tree-based model like a random forest. You can throw any variable you want in there and get output, and depending on testing it will seem, on the surface reasonable,” he said. “With freedom comes responsibility.”

There are other problems too, he added. Firms might want to be seen using the newest tools, but that can clash with what has worked in the past. There is “a lot of groping for solutions based on AI and machine learning”. But some firms with well-founded investment strategies and a successful track record have found it difficult to integrate AI into their investment approach, he said.