- 12 Mar 2019
- Managing the Future of Work
How Goldman Sachs is using technology to redefine banking
Welcome to the Managing the Future of Work podcast. I’m your host, Harvard Business School professor and visiting fellow at the American Enterprise Institute, Joe Fuller. Today I’m speaking with Marty Chavez, vice chairman, co-head of securities, and former CFO of Goldman, who oversaw much of that transformation. Marty will describe how modern technology is making some work more efficient—fundamentally changing the jobs of traders and bankers—and even spurring redesign of the bank’s business model. Marty will also share his insights for executives looking to navigate the transition to artificial intelligence [AI]. Thank you, Marty.
Marty Chavez: Pleasure to be here.
Fuller: You know, Marty, I’m interested; you were quoted recently as saying that trading and coding are the same thing. Was that a rhetorical flourish, or did you really mean it?
Chavez: I mean it. It is literally the truth. What I would say to amplify on that is that it’s an evolution. And as that old adage would have it, the future’s here—it’s just not evenly distributed. So, depending on the trading product, the trading desk, the trading region, that will be more or less the case. But definitively in, say, U.S. cash equities, that has happened. So, we’ve gone through an evolution. It’s gone on for decades, but in the last few years, with our own legal and compliance teams, and with our regulators, we have formalized the notion. It’s got many names for it, but we’ll just, for these purposes, call it “traders who code.” And right now, if I look out on the trading floor and I look at, say, our program trading business for U.S. equities, it is—there’s no test that I could perform to tell which one was a trader and which one was a data scientist or strategist in our terminology, because the skills have completely merged into human beings who are both.
Fuller: And over time, can you describe how that process took place? And how did the work processes change? And how did the culture change as you started introducing new types of talent and automating more of the business?
Chavez: So, the evolution began a long time ago, and really I was the beneficiary of this. Two gentlemen who are still with us—our retiring chairman, Lloyd Blankfein, and Mark Winkelman, who’s on our board of directors—had this notion that, to get the best result, they weren’t communicating effectively with the technologist. They would be talking past each other. Traders would say to the technologists, “Would you build a risk-management system?” And then the technologists would build something, and then they’d deliver it to the traders, and the traders would be unhappy and they’d say, “Well, that’s not what I meant.”
Fuller: I thought traders are always unhappy.
Chavez: Well, there might have been a little bit of that as well. But certainly, they were speaking different languages. And so they had the idea: Why don’t we find people with that math and software and engineering skillset and put them on the trading desk right where we can see them, right next to us, and then they will live and breathe the trading business. And then they will figure out what to build.
Fuller: So, create a bilingual …
Chavez: That is exactly it.
Fuller: … or ambidextrous workforce.
Chavez: That’s right. And so that began in the early ’90s. I joined the firm in ’93. I was roughly number 12 in that group that we called “strategists,” as I mentioned, later shortened to “strats.” And it continued from there. And in those days, we had a very strict—and appropriately strict—separation between traders on the one hand and strategists on the other, because, among other things, it was clear to us—and it’s still the case—that the people who are building the math and the software for trading and for measuring value and for measuring risk need to be separate, because you could imagine if the trader was doing the math that came up with the value of the derivative, then the trader could have the value of the book be whatever the trader wished it to be. So there were quite separate and strong boundaries around them. And so traders were not allowed to commit code to the production environment, and strategists were not allowed to commit capital. One group commits code, the other group commits capital. As time went on, of course, there was the advent of electronic or algorithmic trading. And so you would have engineers—strategists—writing software. And, as time went on some more, the software that the strategists wrote would put trades into the market. This was an evolution, and it was finally dawning on more and more of us that we had gotten to a point where it wasn’t useful or accurate to say that there was this division—these people commit code, and these people commit capital, and never the twain shall meet. So, then, we got into the immensely complicated journey of preserving all of the restrictions and constraints and boundaries that we’ve always had with the ones I just discussed, while recognizing that these engineers were trading—or they were traders who were coding. And because a trader is a very particular construct in the regulatory world, we went with this idea of traders who code. But, really, to get this done required immense amounts of thought, information barriers, controls, policies, procedures, conversations with regulators. But now we’re reached the point where we have many traders who code.
Fuller: How far do you see this going? How far do you see AI and machine learning going inside Goldman, going inside the sector? Because we often talk about AI or machine learning optimizing UPS routes or helping price surplus inventory for a retailer. A lot of them, the memes out there about how technologies are going to affect the work of middle-skills workers, manual workers—not elite, white-collar, highly educated people. How do you see this unfolding in that population?
Chavez: Well, let’s step back for a moment and consider the breakthrough in machine learning that’s happened in roughly the past 10 years. We see the evidence of that breakthrough all around us. Just look at how good machine translation has become, for instance. And voice recognition and image detection, right? In your Google photos, it knows all the people. And you can find that beach you were on by just typing the name of it, right? This is all amazing. And I would submit that, if a problem can be posed in the following way, machines will be extremely effective at answering the problem. And the way is: You have a large number of training instances, and, one way or the other, by hook or by crook, usually with human beings, you’ve dichotomized all those training instances. So, there’s a million images, and these ones contain cats, and these ones don’t contain cats. If you can frame a problem that way, you can use all these techniques. And the techniques will work, and they will work at scale, and they will work extremely well. And they will be faster than people could be. And so, it’s amazing how many problems can actually be formulated that way. And so, for those problems, it is just the nature of the world that human beings, just as we would generally leave multiplication to our calculators, will leave those kinds of problems to the computers. And then the human beings will move on and do different things. At the same time that we’ve had these breakthroughs, there’s vast number of problems that, as far as I know, as far as most of us know right now, cannot be formulated in the way I just described. And it might take a long time—I’m not going to make any prediction when that break through will happen—it could happen in two years or it could happen in 50. I think it will happen, but I have no sense of the timeframe. And in all of those areas, those are things that human beings are just going to continue to do.
Fuller: How far can this go, in terms of work processes in an elite, white-collar institution? You’ve regularly referred to trading and equities and foreign exchange. Does this get into bonds? Does this get into markets that are between two parties and not broader markets? Does this get into other core work processes, M&A advisory, things like that, eventually?
Chavez: So the answer is: It gets everywhere eventually, just at different speeds. Let’s take the example you mentioned from corporate bonds. It’s often stated that, well, it makes sense that equities would be highly automated, because there are so many exchanges. It’s been electronic for such a long time. But then, of course, the counter example is foreign exchange, which is also heavily automated and electronified …
Fuller: … very fragmented …
Chavez: … but almost without exchanges. There are venues of various kinds, but it’s mostly bilateral trading on these venues, right? And then there are other markets, which are mostly voice, though increasingly those are relatively rare. There’s this sort of voice connected with …
Fuller: … by which you mean people are literally calling each other on the phones, right?
Chavez: Just literally calling each other on the phone, but often it’s calling each other on the phone, but also typing to each other on Bloomberg or on Symphony, right? So you see that variety as well. And so, really, what we’ve observed—the key—is not so much is it exchange traded, or is it even traded on a venue, but rather is there a lot of data? If there’s a rich data set—and that’s often correlated with something that’s exchange traded or something that’s highly liquid—but if there’s a rich data set, or if the rich data set can come into existence, then that’s going to be a very rich opportunity for machine learning and work-flow automation. And so, you can get very far away from the trading examples that are usually given, and get into, for instance, M&A. Now, I would say M&A is going to be a business that’s incredibly focused on people interacting with other people in every scenario I can see for as long as I can see. At the same time, there’s a lot of work that goes into figuring out, well, does this merger combination make sense? And on what terms? And we’re increasingly finding that the computers can take datasets, such as company financial information, and apply machine learning to them to make adjustments to them; do merger math on all kinds of pairs; do that at scale, out in the Amazon cloud; suggest to bankers this is something worth discussing with your client; generate the first version of the pitch book that you could use to motivate it. So it’s a mix.
Fuller: I’m sure there are a lot of investment banking analysts who can’t wait for the day that that’s widespread so they’re not working all night seven nights a week. But, in any case, there’s, of course, a big difference between a model that is reaching a conclusion it’s been trained to reach and a model that might later need have to explain how it reached that conclusion.
Chavez: Well, certainly this is a huge area of emphasis for us. We are finding strong limitations in the current set of techniques, because we generally need to explain our reasoning. We need to—at least at the boundary where the machines are interacting with the rest of the world—we need to have strong and accurate explanations of what’s going on. And so machine-learning techniques that don’t give us that are not going to be particularly useful for this. And this is an important frontier of this work, and it is complicated. And the current set of techniques do not definitively give you the explanations while maintaining the accuracy.
Fuller: Right, right. One thing that I know you’ve been experimenting with is opening your platform to both customers and an open-app environment for development of additional capabilities.
Chavez: Yes.
Fuller: When I think back on my career dealing with Goldman and other banks, I would never had predicted the day that a bank would be essentially sharing the data and the tools by which it was reaching insights with third parties. That was the secret sauce, that was the formula of Coke. What’s led to that attitude shift? And is it something that’s being imposed on you, or is it something you’re using as a strategy to extend the bank’s reach and effectiveness?
Chavez: So, it’s absolutely a strategy that we’ve devised and embraced and are now propagating across the firm. It has had it’s controversies over time, just as you have described and as you would expect. If you start with the idea that, what is the mission of our trading business? So, we’ve actually defined it in six words and three verbs. It is: Serve clients, assess value, build scale.
Fuller: Okay.
Chavez: And so another way of thinking about it is, clients come to us because they have risks they don’t want; they want risks they don’t have. It’s our job to help them understand and analyze it and then transform it from what they have to what they want. And so, if that’s your mission, then you’re not thinking about the individual trade and the individual bid offer. You’re thinking about the relationship and what makes it more interesting and inspiring for the clients to keep coming back to us. And there are many analogies in Silicon Valley. When you look at eco systems that have been incredibly successful, you tend to be quite happy in your eco system. You don’t find the need to use that other website for buying stuff or that other search engine, because you just keep coming back to the one that you know and love. And that’s really the analogy. Another way of motivating it is to imagine a search company—an internet or web search company—that had treated its search engine as internal and proprietary. And so the client wants a search, you pick up the phone and you call the salesperson at the search engine company, and you read your search terms over the phone, and the salesperson types it into the engine and gets the results and reads them back to you. That’s a reasonably good description of the way Wall Street has traditionally worked. And so, when I frame it that way, you can see how there might be another way of doing it, which is to take a large number of activities and then make them available to your clients in really a self-serve format on a website. Or maybe the computer’s talking directly to other computers through application programming interfaces, as they’re called. And then you build on top of that the human aspect—the relationship aspect. And it’s all potentiated and enabled by all this work flow that’s automated.
Fuller: When you design that, that interface, are you designing them for the workforce the client has or has had? Or are you trying to lead the evolution of the client’s talent and have them go through some of the same transitions you have?
Chavez: So the answer is yes to both, and it is complicated. We have found, in multiple situations, when we took what I would call too much of a forward view of how this is all going to work, it just didn’t work for the clients, and we rolled it back. And so, generally you want to be a little bit ahead, because that’s going to be a differentiator, and it’s going to be interesting and attractive. Let’s say: You want to be on the cutting edge but not the bleeding edge, might be another way of putting it.
Fuller: Yep, yep. Let’s talk about how this has affected Goldman’s workforce—who you’re hiring, who you’re looking for. You alluded to your own case being a very early hire as the bank started to look for these ambidextrous type workers. Now, almost 30 percent of your workforce has a CS background.
Chavez: STEM background.
Fuller: STEM background. Talk a little bit about now just how you’ve pivoted that way, but also what it’s done culturally to the bank, and how do you see that evolving as the extent of AI and machine learning grows, and also as some of these people you’ve hired have matured and got more senior in the bank?
Chavez: Well, certainly I’ve been a beneficiary of this trend. So in my mind, I’ve been doing just about the same thing since I was 10, which was solving problems with math and software. It occurs to me is that I’ve just been doing the same thing and then the universe has really traveled in my direction. I don’t feel like I’ve changed that much, right? That’s certainly been an evolution. And one thing that we’re seeing at Goldman Sachs is that we used to see people with the math and software skillset—let’s call them engineers, writ broadly, and that’s an umbrella term that includes data scientists, and quants, and algorithmic trading engineers—all of those people. We have seen an evolution where they’ve gone from being support to the business, helpful to the business, valuable, to being one of the pillars of the business, alongside salespeople, and bankers, and traders, and all of the other professions—legal, compliance, controllers, and so on, to in many cases now being leaders of the business. At Goldman Sachs we look to have co-heads for almost all of our businesses, and so in many places across the firm, you’ll see examples of senior leaders at the firm who have this engineering skillset. And so that’s been an important evolution.
Fuller: So it’s always dangerous to make predictions, especially about the future, but how do you see this, how do you see this shaking out? How do you see technology—obviously it’s going to continue to change—but how do you see the relationship between highly skilled humans and technology, generally, in the future of the bank?
Chavez: I’m firmly in the camp of people who say, of course predictions are risky, that human beings will always have something to do. Nothing that I’ve seen in 25 years suggests otherwise. One of the things that I’ve done and that my colleagues in our strategy group have all been doing for a long time, is the first few times you get asked a question by one of your colleagues—say a trader or salesperson or banker—you do a bunch of analysis, and then you show it.
Fuller: Socialize it.
Chavez: Socialize it, exactly. And then you see that that question is going to keep coming …
Fuller: … keep coming, right.
Chavez: And so then you package it up, so that you can look really smart, but really you’re just pressing F9. And then, right? And then the next thing you do is you package it up so that you just give it to your colleague as a tool, and kindly, politely, ask your colleague not to bother you with that request anymore. And I’ve seen that strategy be extremely successful, right?
Fuller: Profitable and successful.
Chavez: Profitable, right? So some people would say, “Well now, if I make it too easy, if I just give that over, well then I’m out of the loop, and maybe I’m out of a job. That’s my job security.”
Fuller: Right.
Chavez: I have not seen that strategy succeed. I’ve seen the strategy of automating oneself, or helping others automate a particular activity, in every case, allows you to go on and do more interesting, more valuable things.
Fuller: It’s interesting how we constantly hear this meme of, well, the company’s going to have to hire lifelong learners in the future. But a lot of companies aren’t very good at creating an environment that actually provides people a basis for learning, or even an insight into what they ought to be learning. As we think about the future of work, having companies realize that that type of move to the next question calls your talent to think in terms of how can they make themselves less necessary in the current, mature product and processes so they can move on to the next more interesting thing, more challenging thing, and higher value-added thing, is the way to go.
Chavez: Well, one of the things that we’ve been doing—and leaders that came well before me at Goldman Sachs have been doing for a long time—is learning through experimentation. Of course, there are many ways to learn—and certainly, one way to learn that I embrace and that I work to role model it, is I’m always taking Coursera classes, and I’m talking on our internal forums about the Coursera class that I’m taking right now. And it’s just a way of saying, I find this useful, and then other people do it, too. And so it’s become viral. But beyond that, something that the firm’s been doing a long time, which is experimentation of various kinds. And I’ve been the guinea pig or—well, in every case I can think of—the beneficiary of those experiments. So, as an example, in 2005, there was a thought: Why don’t we take a derivative quant—me—and drop this person in investment banking? And I remember saying, I don’t know the first thing about investment banking. And the thought was that I would be unlikely to harm, in any way, our marvelous investment banking franchise.
Fuller: Yes.
Chavez: And that it’s just barely remotely possible that I might think of something that you only think of when you put someone who’s out of context, right? And, indeed, that’s worked really well. One of the things that I participated in was studying a problem for corporate clients, which was buying back their own shares. And the way the business worked at that time is the clients weren’t too happy with the business because they paid a penny a share commission to buy back their own shares. And then we would think, well it’s only a penny a share, right? And so there was sort of a general unhappiness. And so the question was: Can we take derivatives technology and apply it to get a better result for the client? And certainly the received wisdom was: You couldn’t get there from here; it would never work in that scale. And then you would have to file regulatory filings, and you couldn’t delta hedge, and you’d have to deliver prospectuses. There were all kinds of received wisdom for why it couldn’t be done. And so, if you just drop someone in who doesn’t know all that and is going to ask a lot of questions and work with the team—that’s the hugely important part—you get something different. That’s something that I immensely admire about Goldman. We’re conducting these safe experiments. They’re circumscribed. We understand the downside, and we’re willing to say, “Well, maybe nothing will happen if we conduct this experiment. Let’s conduct it in a prudent and appropriate way with a lot of consultation with our clients and our regulators.” And then, that’s the only way you uncover the upside.
Fuller: Actually, just having the insight that nothing happening is a result that has value.
Chavez: Yes.
Fuller: You tested a hypothesis, and it was disconfirmed.
Chavez: Exactly.
Fuller: That does speak also to a cultural and organizational attribute where, if everyone’s standard is to put forward their best effort in the interest of the enterprise and are willing to have that best effort not yield necessarily the ideal result—but the standard is the best effort in the interest of the enterprise—then you can have that type of experimentation.
Chavez: And you do it with the right time horizon, knowing that, “Well, maybe this particular experiment didn’t yield something, but these individuals who are moving around the firm into different roles are going to be gathering and accumulating something. And you never know when that might result in something useful.
Fuller: Also, as we think of the future of work as well, the notion that we all know there’s ample research that having diverse teams is collinear with better performance. That diversity could be from an intellectual perspective—having the data scientist with the veteran investment and banker, and you’re going to look at those questions orthogonally and see what happens and see if there’s value created.
Chavez: That is a crucial and important aspect of diversity. I’ve seen it myself. If I were just sitting around imagining what a trading business would look like without actually being on a trading desk, I would never get to a useful result.
Fuller: I’m quite sure that some of our listeners are interested in what Coursera course is the vice chairman of Goldman Sachs taking?
Chavez: I do love programming languages. I don’t write a lot of code in my day job these days, but I still find it important. There’s a huge amount of innovation going on in programming languages. I’m always asking my colleagues, “Okay, if I’ve got space in my brain for one more programming language, what should it be right now?” I asked the colleague whose view I trust immensely, and he said, “Well, I’d recommend you learn Scala.” I went off and I took the courses on Coursera on Scala. They were difficult. I thought, “Well, a lot of things that I studied a million years ago—like functional vs. imperative programming—we really made some huge strides.” So that’s an example. There’s a fantastic data-science and machine-learning certificate series on Coursera as well that I’m also doing. Even though I have a PhD in machine learning from 1990—interestingly, the techniques are still the ones that we were using back then—everything that makes them work at scale is relatively new. If I want to get up to speed on the new tools and the new programming interfaces and the new paradigms, Coursera’s a fantastic way to do it.
Fuller: If you’re talking to our listeners, young people, students at our schools and other schools, or mid-career people, maybe, who are trying to advise their own kids or friends about, what do they need to know and learn to be prepared in the future? I’m regularly asked, despite having no basis for expressing a judgment on this ... My students here will ask, “Shouldn’t I start taking some computer science courses? Don’t I have to learn how to code?” We have a lot of our students here that take our famous Computer Science 50 class at Harvard College, which we also like to point out, is now one of the largest courses at Yale University taught by our faculty ... a little dig there. What would you say to somebody about what do they need to know and understand to be functional in this area without dedicating themselves and becoming deeply expert in it?
Chavez: Fantastic question. I don’t think that many people—and the future is an important skillset—need to know how to code. I do think that the algorithmic modular structured approach to problem solving is something that absolutely everybody needs to learn—the data-driven techniques to solving complex problems, something that everyone can learn and everyone really needs to learn.
Fuller: What the techniques are, what they’re good for, what to watch out for, or actually be able to use the techniques.
Chavez: And to understand, as you say, what they’re good for. What they’re not good for is equally important. And to be able to converse about them and to be able to lead people who are applying different skillsets to solve them. I think if you’re just treating it as an abstraction—“Oh, well, it’s this coding thing”—but you really have no idea how it works or what it is and what it isn’t, I don’t think that’s particularly effective. I would just draw an analogy to my own education. I showed up here at Harvard in 1981. I was in the science center during the first week. The professors from various departments were recruiting. Steve Harrison, from the biochemistry department, was recruiting. I told him I was a computer geek, a computer scientist, and he says, “Well, biochemistry wants you.” I said, “Well, I’m not interested in biochemistry.” I wonder if he remembers that. He said, “The future of biochemistry and the future of molecular biology is computational.” Now, 1981—that was a prescient thing to say. His lab was working on X-ray crystallography and envisioning proteins and applying machine learning. He was really early on this. He said, “I’ll work with you, and we’ll set up a biochemistry, a molecular biology concentration where you’d take a couple of biochem classes and a lot of other science classes, and a lot of other computer science classes, because we need, right now, to be educating people who are converse in both of these languages.” I would say to anybody do two things, and have computation embedded in it. One conversation that interests me immensely is, rather than advise Harvard students, “Why don’t you all go major in computer science?”—while I think that’s a beautiful thing—at the same time, I think it’s extremely interesting to look at all of our traditional concentrations and say, “What would the computational track within that concentration be? What would data science applied to it be?” and do that.
Fuller: One last question to you. You’ve been very public and direct about the need for institutions, starting with Goldman, to find ways to engage a broader population of perspective workers to remove anything that dissuades people from being interested in working for the firm and working in the industry. How does the type of change you’re seeing and we’ve been discussing ... how do we think about it through the lens of broadening accessibility to career paths? What do we need to be doing to enable more young people to be getting on a trajectory so they can take advantage of these various trends you’re talking about?
Chavez: It’s a complicated challenge. It’s like boiling the ocean, so where do you start? You start by doing a bunch of different things. This is really the argument for diversity, which we’ve been making with many of the companies for a very long time. When it really began to resonate is when the case for diversity moved beyond doing the right thing, a kind of altruism …
Fuller: … to doing the smart thing.
Chavez: ... to doing something that was not only right, but also smart. It’s the uncorrelated thinking, it’s something that we talked about. Take someone with this skillset, put her over there, mix and match, see what happens, run a lot of experiments. That’s worked incredibly well for us. We’re constantly asking ourselves, “How do we do this at scale? How do we cast a broader net? How do we interview students from more schools, more backgrounds, more parts of the world, in addition to all the traditional ways that we’ve recruited people? What do we do in the world to create more of these educational opportunities?” Something that I’m passionate about and something that our leader started a few years ago, we call Goldman Sachs Gives. A portion of partners’ compensation goes into a donor-advised fund subject to a number of constraints that are important. The partners are asked to direct those funds to various organizations. We’re all doing different things. Many times, many of our partners will gather together and emphasize a particular one. One that I’m particularly fond of is Girls Who Code. That’s something I get excited about. I see what it’s doing in the world, and there are hundreds of other examples like that.
Fuller: You hear two schools of thought on how AI helps or hurts this issue. You hear the school of thought that says: AI is going to allow me to efficiently change my paradigm of how I search for and evaluate talent. I’m not going to rely on the “We go to five campuses with alumni of those schools and they call up their buddies in the faculty and their coaches and their kids’ friends and say, ‘Okay, who should I want to talk to?’” You also hear that you can have bias built into AI; that, if you’ve taught the AI, that this is what success will look like: white males with advanced degrees. That’s how you’ve trained it. Where do you come down on that, and how do we avoid falling into that pitfall?
Chavez: That’s a tough one. Yes, we’re seeing the benefits, and you’ve described them beautifully, of being able to, for instance, actually read and evaluate in a serious and detailed way 150,000 CVs. That’s something that’s challenging. At the same time, who is equipped to ask the questions? What kinds of controls are you going to apply? What kinds of back-testing are you going to apply? And how critical are you going to be? We have a long tradition of thinking about all of these techniques. I think AI and machine-learning are amazing and almost miraculous. At the same time, to me, it’s also just like HP 12c calculators were some years ago. It’s an incredibly useful tool, and the tools are always evolving. That is the nature of our civilization. One thing that was certainly drilled into me as a junior Goldman person is, just because you press DEF-9 and the screen said this is the price of something, doesn’t mean that’s the price. This critical thinking—seeing it on the screen, seeing it as an output of a model—is interesting, probably useful, not necessarily right, not necessarily the truth, not necessarily seeing around all the corners, and imagining all the things that might arise. That kind of skillset, I think, is something that human beings are amazingly good at, and are going to continue to be good at. It’s going to be critical.
Fuller: Thank you, Marty, for sharing the story of the remarkable technological transformation at Goldman Sachs and the lessons you’ve learned along the way.
Chavez: It’s been a pleasure to be here. Thank you.
Fuller: From Harvard Business School, I’m Professor Joe Fuller. Thanks for joining us for the Managing the Future of Work podcast.