(Bloomberg) – Wall Street Quantz and leading financial scholars are clashing over whether artificial intelligence has overturned one of the core principles of systematic investment.
Using rule-based strategies derived from data analysis, Quant Traders has long believed that if a model becomes too complex it will be less effective. That's because they inhaled the market that predicts the market in the first place.
However, researchers at AQR Capital Management sparked backlash in a study claiming opposition. It is possible that larger and more complex models offer financial benefits, rather than responsibility. The paper, entitled Virtue of Complexity in Proof of Return, showed that US stock market trading strategies were trained with over 10,000 parameters, and that data for just one year has broken the simple buying and holding benchmarks.
“This idea of preferring small, parolitic models is a bias that we've learned,” says Brian Kelly, head of machine learning at AQR and one of the three authors of the paper. “We all exist on a daily basis using these large-scale language models that have been successful for this push towards very large parameterization.”
Since it was published in the prestigious Journal of Finance last year, the study has sparked fierce debate among fellow Quant industry and associated academics.
At least six papers from academics at Oxford and Stanford University are currently challenging their findings. Some argue that the virtue of complexity studies has a questionable design that is unrelated to live trading. Others say they are less cutting edge than they can see anyway. (Kelly wrote about the defense afterwards.)
Among the most notable critics is Stephen Nagel, a professor of finance at the University of Chicago. This is the school where the two founders of AQR met and the company's original investment philosophy was formed. His first reaction? “I found it difficult to believe in empirical results,” he said.
After delving into the details of the virtues of complexity research, Nagel concluded that the model analyzes data for just 12 months, so it only copies the recently worked signal. In other words, it followed a momentum strategy, an established trading approach.
“It's not because the approach learned this effect from the data,” Nagel said. “It's because they did something implicitly, and this mechanical thing happened to be lucky.”
Jonathan Burke was one of the first and fierce critics of the virtue of complexity, a Stanford Economist who called it “virtually useless” with the aim of predicting that doesn't convey anything about what drives the return of assets. Daniel Bunsic of Stockholm Business School said the study clearly made the wrong design choice to reach a conclusion.
The virtue of the complex paper co-authored with Swiss EPFL and Semyon Malamud of Kanzy Zhou at Yale University, sparked this response as it challenges long-standing assumptions about financial market forecasting.
Modern AI can perform amazing tasks such as conveying dogs to dog cats in images, but that's because animals have defined, unchanging features, as they can learn from the massive supply of photographs. In contrast, inventory essentially provides a limited amount of data (particularly due to slow movement strategies that can only trade once a month).
Fear is always overfitted. Complex models learn from all the noise in historical data, many of which may not be applicable to future transactions. So Quento has traditionally relied on relatively simple insights, like the famous Fama France three-factor model (analyzing returns based on the size of each company, its valuation, and its relationship with the broader market).
The AQR itself is built on such so-called factors and aims to be superior over the long term. It has been recently that a $146 billion money manager has raised capital for machine learning strategies and said that it is not necessary to support all transaction signals with economic theory. Kelly's main argument is that traditional quantum models are very simple and unfit, generating inferior predictions, but in reality they learn that sufficiently complex models do not depend on too much overload.
Certainly, the virtue of complexity critics does not argue that machine learning has no funding. They primarily view the results of the paper as the less true, the better.
“This method has a role and can be used,” says John Campbell, a professor of economics at Harvard University, co-founded Arrowstreet Capital, a Quant Firm. “But some of the most eye-catching results are being questioned.”
Even Ben Retch of the University of California, Berkeley, a well-known computer scientist who developed the method used in the virtues of complex papers in 2007, said on his blog that “the hype cycle confuses everyone.” The paper method is far from cutting-edge AI, and it seems that it is not necessary for the task at hand.
For Kelly, who teaches alongside AQR gigs at Yale, the criticism of the virtue of complexity is “a little hollow” to ultimately focus on the narrow side of proof-of-concept research.
“The world of practitioners understand that these conceptual methods are beneficial when implemented in a more sophisticated way,” he said. “An ideal combination of the amount of frontier machine learning methods used with more traditional, economically oriented methods. This is what we're trying to understand.”
More stories like this are available at bloomberg.com
