Since we are in the era of ‘big data’ and fast computers, you may wonder whether machine learning has conquered finance, more specifically whether the computer can now replace humans in making accurate predictions of asset prices. If so, more power to robo-advisers! If not, we’ve got to downplay the hype.
There has indeed been big leaps in in machine learning in specific domains (by the way, I will use machine learning interchangeably with artificial intelligence or AI). In 2017, the world recently witnessed one particularly impressive feat when Deepmind’s AlphaGo program beat the world’s champion Lee Se-Dol in the complex game, Go. Deepmind went on to develop an even smarter version of the program called Zero that can learn on its own, given the basic rules. After three days of self-play, Zero was smart enough to defeat the version of itself that beat Lee Se-dol, winning handily 100 games to nil. We are told that there is more to come!
Despite this impressive feat, it is strange that we don’t hear much about the success of AI in finance. I can think of two reasons why this is. One, the tech geniuses are keeping their secrets close to their chest for obvious reasons. Two, maybe there isn’t much success to brag about in the first place. It’s possible that even today’s clever machine learning technologies are not clever enough to be able to predict stock and bond prices accurately. To understand why this might be the case, let’s consider the limits of AI according to some of the world’s top AI experts.
Dr. Dave Ferrucci, co-founder and chief executive of Elemental Cognition and a former AI expert at IBM stresses that machine learning is simply a statistical technique for finding patterns in large amounts of data. He adds that “having a computer spew out an answer is not sufficient in the long term. You want to say, ‘here’s why‘ “. But understanding the “whys” of financial asset prices is much harder than understanding the rules of a board game like Go. Despite its complexity, Go, like all board games, is actually quite easy for computers to figure out because the rules are finite – there is no hidden information and importantly, no element of luck (unlike finance). This means that researchers applying AI to Go can have access to a perfect simulation of the game. They can program their software to run millions of tests and be sure it’s not missing anything. Not many fields meet these criteria. Key examples that do include language translation, speech recognition and image recognition. Finance isn’t one of them.
It is not difficult to see why. As Stanford University’s Andrew Ng, one of the founders of Google Brain, Google’s deep learning project, emphasizes, AI works only for problems where clear inputs can be mapped or linked to a clear output. This criterion limits the applicability of AI to problems involving categorizations such as the fields mentioned above. Finance is orders of magnitude more complex than these problems. Take stock market predictions for example. It is incredibly hard to pin down the most important inputs or causal variables that drive stock prices. First, there are too many of variables to consider (GDP, inflation, interest rates, oil prices, exchange rates, and most slippery of all, market sentiments). As Warren Buffet once said: “investment is a game of a million inferences.”
Secondly, the financial environment is a dynamic one. Variables that are highly influential for stock prices during high inflation periods may be less so during low inflation periods and vice versa. Trying to “catch” the best set of predictors is like trying to catch a butterfly; the more you chase it, the more it will elude you.
Third, even if you think you can nail down have the relevant predictors variables, how do you put them together in a model to generate your predictions. A model is an equation that relates what you want to predict (say, stock returns) to a set of predictor variables. Coming up with an accurate prediction model is far from trivial. You can ask any finance professor; I guarantee you will walk away unsatisfied with the answer (I know; I was a finance professor myself).
What about computer scientists and mathematicians, the people most closely associated with pushing the limits of AI? Swetava Ganguli and Jared Dunnmon are two ‘quants’ at Stanford University who has done research in using AI to predict bond prices. They ask: how good different machine-learning techniques are in predicting the future price of corporate bonds? They use standard “shallow learning” techniques as well as more exotic neural-network techniques to answer this question.
Before I summarize their findings, some background on bonds. Corporate bonds are IOUs issued by firms or to raise money for the firm’s business. In return, bond investors receive interest and at the bond’s maturity, the bond’s principal value. As every bond investor knows, bond prices are interest sensitive. Moreover, they have to contend with default risk. The higher the default risk, the lower is the bond price. Yet, compared to stocks, bonds are relatively simple instruments and bond prices are generally less volatile.
Given these features of bonds, you would think that predicting bond prices is a piece of cake for machine learning programs. This is not what the researchers found! I won’t go into the details of their study (see the reference below). They found that while the best predictions came from the neural networks, these fancy techniques took several hours to work their magic – too slow to be useful for actual bond traders. Interestingly, simpler such as linear regressions aren’t so bad in terms of both accuracy and speed. This is a slap in the face for AI. At the same time, it also affirms that simpler methods may be more robust (read: less bad) for highly complex problems, a fact that data scientists have known for some time. That may be the reason why there are so few stories about the success of AI in financial predictions.
Swetava Ganguli and Jared Dunnmon (2017), “Machine learning for better models for predicting bond prices“, arXiv:1705.01143 (link)