Machine learning has done magic, such as beating human chess champions. But in finance, expectations for the technology may need to come down a notch or two, according to quantitative firm AQR.
In a report published Monday, AQR argues that the benefits of machine learning will likely apply to problems involving optimizing portfolio construction, such as risk management, transaction cost analysis, and factor construction — at least at first. That’s because markets are different from other areas where machine learning has come to offer up breakthrough research, according to “Can Machines Learn Finance?”
[II Deep Dive: Younger, Cheaper, Smaller, Smarter: Meet the New Breed of Quants]
Machine learning changes the way problems are solved. Traditional computer programmers define all of the rules or parameters of a game. Machine-learning applications, in contrast, are fed data so they can then determine the rules and relationships. Sorting pictures of dogs from cats is a well-known example of machine learning in action. A traditional programmer couldn’t program in the infinite number of variations that exist between dogs and cats.
But cats and dogs are not markets.
In financial markets, the so-called signal-to-noise ratio is low, meaning outcomes aren’t particularly predictable. If they were active, managers would consistently outperform their benchmarks.
“The low signal-to-noise is not some unfortunate coincidence of markets,” AQR said in the report. “On the contrary it is a feature ensured, and constantly reinforced, by simple economic forces of profit maximization and competition.” In other words, hedge fund managers and other active traders start making decisions based on information they have and push up stock prices. And they don’t stop until they’ve made as much money as possible. Stock prices reflect that information and the only force that moves markets is unforeseen events or unexpected data.
Machine learning doesn’t do well in that sort of environment. On top of that, markets are adaptive.
“If a researcher identifies a new signal that captures a particular form of asset mispricing useful for predicting prices, then as the signal becomes more widely known, more traders act on it, correcting prices more quickly,” wrote the authors of the report. “The market eventually absorbs the information, and the data generating process changes due to the very actions of agents in the market.”
In other disciplines, as AQR wrote, “cats don’t begin morphing into dogs once the algorithm becomes good at cat recognition.”
Financial markets are also different from other sectors in that asset managers need to explain their models to investors — not always easy in the machine learning world.
Although big data can be useful, AQR also argued that finance is a time-series discipline. New data on returns, for example, are only generated with the passage of time. Big data from social media or other platforms have a short history, which is of limited use in investments.
“The limited time series presents a challenge for meaningful backtesting. With a short history, it’s even more difficult to form a precise estimate of strategy performance which ultimately means that even very strong signals might prudently receive only small weights in a portfolio,” according to AQR.
Still, AQR argues that other areas of investing, such as risk, have a “high signal-to-noise” ratio — they can be predicted pretty well, as can the price-impact of transactions. Investors’ behavior doesn’t overrun these signals like their behavior squashes signals on returns.
Alpha is scarce and has remained so in the face of advancements over the past few decades.
AQR acknowledges this.
“Using new data and machine learning to build alpha (i.e., to find new unique sources of return predictability) heads straight into the most competitive aspect of financial markets,” according to the paper. “As more investors enter the market with similar tools, the mispricing corrects and that alpha compresses to zero.”