The virtual assistant on your smart speaker and smartphone draws on artificial intelligence; so do every Google search and every Netflix recommendation. And while the bulk of the investment in AI and machine learning—the subset of AI in which computers look for patterns in large data sets—is coming from the tech sector, it is spilling into other areas. Deloitte points to growing use in manufacturing quality control, cybersecurity, customer service, risk management, and sales optimization.
Which uses of AI will come to ubiquitous fruition? Bloomberg offers a simple test: “Ultimately, the future of AI will depend on its ability to make money.”
Financial services seems like fertile ground for AI to generate profit. For decades, quantitative investing firms have used computers to trade based on complex algorithms. The next step is for the computers to ingest data on the history and current state of the market and generate their own trading strategies.
Early efforts haven’t set the world on fire, but their results so far should put hedge funds on notice. “Today’s small group of fully automated AI strategies are off to a middling start. Their performance beats the broader hedge fund industry but not the stock market.”
Yale Insights talked with Charles Elkan, global head of machine learning at Goldman Sachs, about how financial services firms can capitalize on the new era of data.
Q: How new is machine learning? How new is its application in finance?
Machine learning as a research area is not new, and even as an application area in other industries it’s not so new. Amazon has been using machine learning for recommendations for close to 20 years. But in finance the big firms and the big asset managers have really recognized its importance only in recent years as the technology has become advanced enough and widespread enough.
All the major firms are very aware of the opportunities; some of them are placing bigger bets than others. But sometimes it’s rational to be a fast follower rather than a true innovator. Every firm definitely needs a machine learning strategy to make sure that it’s not missing big opportunities.
In Goldman Sachs’ case, the firm has an overall strategy of increasing revenue in certain areas and that’s really driven by business. But within every business area there are opportunities to apply machine learning and the firm is investing and enthusiastic about those opportunities.
Q: What are the compelling opportunities for machine learning in finance?
One of the ways in which it’s helping is making existing models more sophisticated. A lot of the mathematical models used in finance are really linear regression. Now, with neural networks or other non-linear methods, we can extract more signal from the same data because we’re not just assuming that the signal is linear. Perhaps that allows incrementally improved accuracy.
Q: Does that make it harder for people to beat the market?
Yes. Some people are always going to be beating the market. Some are getting lucky. Some have figured out a method that other people haven’t yet figured out, but I don’t think any informational or methodological edge lasts forever.
Q: What other aspects of finance will use machine learning? Are some areas harder than others?
Some areas of the business are very high touch compared to other areas. At the strategic level, there may not be much opportunity to use machine learning to deepen the relationship between investment bankers and CEOs. But there’s certainly opportunity to use machine learning, for example, to understand better the entire universe of M&A opportunities.
I think some of the biggest opportunities for creativity is figuring out how to use machine learning or other AI technologies in areas of the business where at first sight it might not be applicable because that might be where the opportunity is to do something that other firms are not doing.
There are opportunities to use machine learning, especially via natural language processing or computer vision, for analyzing types of data that quantitative methods just were not touching before.
Every time that a financial firm sells a derivative or makes a loan, that’s a contract—a lengthy legal document. Natural language processing is really useful for understanding the content of those documents. A large firm is party to thousands, possibly hundreds of thousands, of contracts. It’s not realistic to know in real time the content of all those contracts down to which clauses might be activated in the near future based on outside events.
With automated natural language processing and automated event detection tracking news flows, firms are going to have closer to real-time understanding of their full portfolio. That’s going to be good for risk management. In the sense that reducing surprises reduces the possibility of crises, that’s going to be good for society.
Q: Is the technology going to replace humans?
Humans are continually being augmented by technology and technology is continually being used to do things that perhaps humans don’t really enjoy doing or humans are not very good at doing.
There are a lot of people in investment banks nowadays who essentially compile information and put it into spreadsheets. The more we can automate that, the greater the productivity we have and the more the humans are freed up to do work which is more interesting and a more appropriate use of their education.
Q: Will we reach a point where financial advisors will all be robo-advisors?
It’s not difficult to come up with formulas—60/40, 70/30 if you’re more aggressive, 50/50 if you’re less aggressive. A lot of the role of a financial advisor actually is to be a counselor and provide the human touch. The human financial advisor is talking to the client and really understanding the client. If the client is panicking in a market downturn, a financial advisor is talking that through with the client.
The financial advisor is an example of where natural language processing will be automating some of the work, but there’ll still be a lot of room for the human side.
Q: Are data-rich industries most likely to benefit from these technologies?
There’s a saying that data is the natural resource of the 21st century. Data is really valuable. But data by itself is a cost, not a benefit. Even if it’s cheap to store data, it’s still not free. One of the reasons machine learning is so important is that it’s a way to gain benefit from data.
Generally, the industries that have more data are going to be able gain more benefit from machine learning. However, it’s also important to understand how relevant historical information is to a given industry.
If we have data about lung cancer from 20 years ago, the disease hasn’t changed much in 20 years. The relevant timescale is that of evolution. Since it proceeds many orders of magnitude slower than scientific research, the data from two decades ago is of great value.
Whereas financial markets evolve quickly. A lot of finance involves adversarial situations where people are trying to predict what others are going to do. If one person is making a profit, someone else is making a loss; that naturally drives faster change. So data from 20 years ago may just not be useful or relevant nowadays.