State Street turns to a scientist to improve its trading and risk strategies.
Financial services isn’t normally something most of us consider “scientific” — but that’s about to change, thanks to the advent of analytics and machine learning. “Because it’s worked well with pattern recognition,” notes State Street Global Exchange’s Jeffrey Bohn, who is chief science officer for the company, “we think it’s possible to potentially train an algorithm to see the macroeconomic regime we’re currently in.” And the potential benefits of regime identification, adds Bohn, is that it “could vastly improve our risk forecasting.”
In a conversation with MIT SMR’s David Kiron and Sam Ransbotham, associate professor of information systems at the Carroll School of Management at Boston College and guest editor for the Data and Analytics Big Idea Initiative for the MIT Sloan Management Review, Bohn discusses how he is developing better trading and risk strategies for clients using State Street’s proprietary data and analytics.
Do general managers need to improve their data and analytics capabilities?
Senior executives don’t always come with deep experience in data and analytics and don’t take the time or may not have the interest or the capacity to process the analytics that may be quite crucial to their decision making in today’s rapidly changing markets. I think it’s better now than it has been, but there’s still a ways to go.
What has changed or what is changing that’s making it better?
The primary driver is a more proactive regulatory system. If you’re a commercial bank, you now have to pass certain stress tests and follow all sorts of regulations. If you don’t pass them, you can’t pay dividends, so some executives are forcing themselves to learn about analytics because it has real impact on their personal compensation.
The increased focus on risk and the fact that younger executives running financial institutions may be more likely than older executives to have some kind of analytics background are also factors, although there still are not a lot of quants running them.
What additional skills does someone without a quant background need to make good use of the analytics now expected at financial institutions?
One is a grounding in statistics. A lot of people take statistics in college or in their MBA program, but the average understanding of problems, like how you process uncertainty, is not very good. This is a broader problem in our society today. It’s very important to have a conceptual understanding of the details underlying a model: what are the assumptions, how confident are people about the data-generating process, how confident are they that they’ve got a representative sample of data. For better or worse, increasing market complexity and regulatory assertiveness will require a higher level of quantitative training for all financial executives.
Everybody understands that they need to up their game in analytics. When I talk about new people, I’m referring to a population of executives in their 30s and 40s who are not quite at that senior level, but who have grown up in an industry that’s required them to have a greater understanding of quantitative output than what was required of their predecessors.
The other problem is that there isn’t always enough effort made to understand the details behind the models. The whole subprime situation is a good example. The rating agencies opted for an overly simplistic model that did not capture the full risk because their incentives were based on quickly growing a new market rather than investing the resources to unpack the actual risk. More rigorous risk analysis would have materially reduced the size of this market. That behavior led to people making poor investment decisions and to profound consequences for financial institutions and the economy in general. Sadly, going forward, we don’t have the option to entirely avoid complex financial products. Creating products in the retirement and insurance arenas for an aging society will require a certain degree of quantitative complexity.
Is State Street doing anything specific to try to address these education issues?
We are, yes. I run something called GX Labs, which sits inside State Street Global Exchange, where we’re trying to pull together data, advisory services, and solutions in a central location.
My team focuses primarily on the intersection of data science and portfolio risk. One of things we’re doing is talking with clients about how to come up with better dashboards. People invest in data collection and in the modeling teams, but then they don’t design compelling dashboards. They actually have the right data, but they aren’t always using it effectively.
Is that a lack of data or is that a user interface problem?
I know institutions that had plenty of information about their subprime exposure pre-2008, but it wasn’t communicated well to senior management, so no action was taken. These communication difficulties persist.
Is that an issue with getting the right information to the right people at the right time in the right way?
Yes, but also educating those people so they fully understand the link between a particular set of analytical output and what that means for their decision making.
Prior to joining State Street, I worked on a project where I assisted a credit-portfolio management group in creating an optimizer. In the end, the portfolio manager did not use the tool, as he felt he had better intuition based on the other data he was receiving. We are in a transition period within financial services where we are trying to balance how much to rely on computer algorithms versus on human intuition. Other industries seem to be in the same transition, whether it be pilots relying entirely on computers to fly planes or riding in self-driving cars.
Your title at State Street is chief science officer. What is a bank doing with a chief science officer?
That’s a very good question. There are a lot of new technologies and new analytics emerging because of the increase in computational capacity. There are also new techniques for managing data, new platforms for doing distributed computing, and new kinds of databases. There are also new computational mechanisms which, historically, have been used for (video) gaming, but that are now being repurposed for use with some of our other applications. And there are new distributed ledger technologies like blockchain.
Managing all this seemed to fall under the umbrella of research, which relies heavily on managing academic partnerships. These partnerships are designed to do a couple of things. One is to stay abreast of trends in data and computer science as well as in statistics, economics, and math. We’re building and maintaining a consortium of partnerships with academia as well as with other non-financial groups. For example, we recently created the Consortium for Data Analytics in Risk (CDAR). State Street and other companies can work with universities in the context of this kind of consortium to share ideas and spark new research. It provides a platform for us to experiment and do research in new areas, and also allows for some interdisciplinary exchange with industries that may be a few years — or sometimes many — years ahead of where financial services are. That’s why we have a chief science officer.
How many people are in your group?
Right now we’ve got about 15 — 10 full-time staff and then four or five consultants, but the size of the group is really determined by the projects we’re working on.
Can you go into any detail about what one of your machine-learning projects involves?
Liquidity, or liquidity regime identification, is one area where it could potentially be applicable. There is potential to use some of these machine learning/deep learning algorithms to identify what we call fragile liquidity environments.
We’re looking, for example, at systemic risk identification and macroeconomic regime identification. So far, deep learning has been really successful in pattern recognition. Self-driving cars is a good example. Another example is training an algorithm on a picture and then having the algorithm go find every picture of that person. Because it’s worked well with pattern recognition, we think it’s possible to potentially train an algorithm to see the macroeconomic regime we’re currently in. This kind of regime identification could vastly improve our risk forecasting.
Another area where machine learning has potential is fraud detection. If you’re looking across trading activities in a large financial institution you may be able to train an algorithm to identify where there’s fraud.
Those are some of the more exciting areas we’re getting into.