Veriton Investment

AQR’s ‘Hard to Believe’ Study Spurs Clash Over AI Use for Quants

(Bloomberg) -- Wall Street quants and leading financial academics are clashing over whether artificial intelligence has upended one of the core principles of systematic investing.

Quant traders, who use rules-based strategies derived from data analysis, have long believed their models get less effective when they become too complicated. That’s because they suck in too much of the distortive noise that makes predicting markets such a challenge in the first place.

Most Read from Bloomberg

Why New York City Has a Fleet of New EVs From a Dead Carmaker

Chicago Schools Seeks $1 Billion of Short-Term Debt as Cash Gone

Trump Takes Second Swing at Cutting Housing Assistance for Immigrants

A Photographer’s Pipe Dream: Capturing New York’s Vast Water System

A London Apartment Tower With Echoes of Victorian Rail and Ancient Rome

But a researcher at AQR Capital Management has sparked a backlash with a study claiming the opposite — that rather than being a liability, bigger and more complex models might offer advantages in finance. The paper, titled , showed that a US stock market trading strategy trained on more than 10,000 parameters and just a year of data beat a simple buy-and-hold benchmark.

“This idea of preferring small, parsimonious models is a learned bias,” said Bryan Kelly, head of machine learning at AQR and one of the paper’s three authors. “All of us are on a day-to-day basis using these large language models that were revolutionary in their success because of this push toward extraordinarily large parameterizations.”

The research has triggered a heated debate since it was published in the prestigious last year, among both peers in the quant industry and those in related academic circles.

At least six papers, including from scholars at Oxford University and Stanford University, have now challenged its findings. Some argue the study has a questionable design that renders it irrelevant for live trading. Others say it’s less cutting-edge than it appears anyway. (Kelly has subsequently written a defense.)

Among the most notable critics is Stefan Nagel, a finance professor at the University of Chicago — the very school where two of AQR’s founders met and where the firm’s original investment philosophy took shape. His first reaction? “I found the empirical results hard to believe,” he said.

After digging into the details of the study, Nagel concluded that because the model was dissecting just 12 months of data, it was simply copying signals that had worked more recently. In other words, it was following a momentum strategy — a well-established trading approach.

“It’s not because the approach learned from the data that this effect is there,” Nagel said. “It’s because they did something mechanical implicitly, and this mechanical thing happened to work well by luck.”

Jonathan Berk, a Stanford economist who was among the first and fiercest critics of the paper, called it “virtually useless” for aiming at predictions that tell you nothing about what drives asset returns. Daniel Buncic at the Stockholm Business School said the study makes some obviously wrong design choices to reach its conclusions.

Co-written with Semyon Malamud at EPFL in Switzerland and Kangying Zhou at Yale University, the paper has provoked this response because it challenges a long-held assumption about forecasting financial markets.

While modern AI can perform remarkable tasks like telling cats from dogs in an image, that’s because it can learn from a massive supply of photos, and because animals have defined and unchanging features. In contrast, stocks provide an inherently limited amount of data (especially for slower-moving strategies that may only trade once a month), and each can be swayed by countless different forces.

The fear has always been overfitting — that complex models will learn from all the noise in historical data, much of which may not apply in future trading. So quants have traditionally relied on relatively simple insights, like the famous Fama-French three-factor model (which analyzes returns based on each company’s size, valuation and relationship with the broader market).

AQR itself was built on such so-called factors, which aim to outperform over long stretches of time. It is only in recent years that the $146 billion money manager has raised capital for machine-learning strategies and said not all trading signals have to be backed by economic theory. Kelly’s main contention is that traditional quant models are so simple they under-fit, producing inferior forecasts, while sufficiently complex models actually learn not to overfit too much.

To be sure, the critics don’t argue that machine learning has nothing to offer finance. They mainly view the paper’s results as too good to be true.

“The methods have a role and can be used,” said John Campbell, an economics professor at Harvard University who co-founded Arrowstreet Capital, a quant firm. “But some of the most eye-catching results have successfully been called into question.”

Even Ben Recht at the University of California, Berkeley — a renowned computer scientist who back in 2007 developed the method used in the paper — weighed in in his blog, saying “the hype cycle gets everyone confused.” The method in the paper was far from cutting-edge AI, he said, and anyhow didn’t seem necessary for the task at hand.

To Kelly, who teaches at Yale alongside his AQR gig, criticisms of the paper are “a little bit hollow” for focusing on the narrow aspects of what was ultimately proof of concept research.

“The practitioner world understands that these conceptual methods, when implemented in a more sophisticated manner, are going to be beneficial,” he said. “The exact ideal combination of how much of frontier machine learning methods to use versus more traditional economically oriented methods — that’s still something we’re trying to understand.”