Odd Lots - Why One Of The Most Successful Quant Funds Decided To Create Its Own Video Game

Episode Date: February 6, 2018

Quantitative finance is red hot. These days, basically everyone (banks, hedge funds etc.) is hiring mathematicians and coders. So what differentiates one quant shop from any other? On this week's episode of the Odd Lots podcast, we speak to Alfred Spector, the CTO of Two Sigma Investments, which is one of the most successful quant firms in the world. Spector is a computer scientist who previously did long stints at both Google and IBM. He tells us about why Two Sigma spent resources to create its own video game, and what the firm does to ensure that technologists and mathematicians are eager to work there.See omnystudio.com/listener for privacy information.

The news doesn't stop on the weekends. Context changes constantly. And now Bloomberg is the place to stay on top of it all. Hi, I'm David Gurra. Join us every Saturday and Sunday for the new Bloomberg this weekend. I'm Christina Rafini. We'll bring you the latest headlines, in-depth analysis, and big interviews. All the stories that hit home on your days off.

And I'm Lisa Mateo. Watch and listen to Bloomberg this weekend for thoughtful, enlightening conversations about business, lifestyle, people, and culture. On Saturday mornings, we put the past week's events into context, examining what happened in the markets and the world. That on Sundays, we speak with journalists, columnists, and key political figures to prepare you for the week ahead. Join us as soon as you wake up and bring us with you wherever your weekend plans take you. Watch us on Bloomberg Television.

Listen on Bloomberg Radio, stream the show live on the Bloomberg business app, or listen to the podcast. That's Bloomberg this weekend. Saturdays and Sundays starting at 7 a.m. Eastern. Make us part of your weekend routine on Bloomberg Television, radio. and wherever you get your podcasts. And welcome to another episode of the Oddlots podcast. I'm Tracy Alloway.

And I'm Joe Wisenthal. So Joe, I think we have to come clean about this particular episode. We do have to come clean. Before we get into the discussion, there's a big, what is it, 800-pound gorilla in the room that we have to address? Yeah. So the gorilla is that last week we recorded an amazing podcast, all about technology and its role

in finance and the broader world. And then we were hit by our own technological snafu. It's right. So we recorded the greatest episode in the history of the entire podcast. It was amazing, one of a kind, the kind of conversation that you dream of. And then, unfortunately, the audio was bad and the entire thing was ruined. Yes. Something happened with the computer. The computer said no. We're still trying to figure out what the exact issue was, but we learned an important lesson about the pitfalls of technology, which gives us an excuse to have our guest come on and try to have the conversation all over again. So here we go.

And since this episode is kind of about the relationship between technology and finance, we can at least pretend that there's some lesson here and what happened to us that's relevant to the episode. But really, like, I wasn't really exaggerating when I said it was a great conversation, and it would have been so hard to, it would have been very hard to try to replicate that or to try to pretend we were just doing it again for the first time. So just in the spirit of honesty and recreating spontaneity, we wanted to get it out of the way and be honest with our listeners that this is a take two of that conversation. Who knows? Maybe it will be even better the second time around. The important thing is we learned a lesson about. backup systems and tech. All right.

So here goes. But we can't now pretend to do our schick where we don't know. We're like, what are we going to talk about this time? Because that would really be contrived after that interview. No. No, I wasn't going to. Okay, I'm going to just bring our guest on. Our guest for today, for the second time, is Alfred Specter.

He's the chief technology officer of 2 Sigma. He's also a former engineer at Google, and he was also at IBM for a very long time. He's an extremely well-known name in the realm of technology and also in quant-driven finance, and he's been nice enough to join us yet again on Oddlott. So thank you, Alfred. Really appreciate it. It's my pleasure to be here. And by the way, the probability that there's a failure and a technology system is somehow proportional

to the seniority of the person that's involved. So if we ever give a demo to like a really senior person, it's much more likely to fail. I'm afraid I engendered the failure. No, not at all. But, you know, Tracy introduced you as, you know, the CTO at 2 Sigma. You don't seem like a guy who's very busy or anything. So I'm sure it was very easy for you to reschedule your time and just come back in for a second day. Very easy indeed.

But no, seriously, thank you very much for coming back in and recreating last week. The way we started our conversation last week, and really the first thing we discussed is that, that your firm, 2 Sigma, it's a very well-known quantitative hedge fund, is known for having a game. You've created a video game and created a competition for people all around the world to come and design programs to master the game. So tell us what is this game that you have people do and why do you have people try to beat it? So a couple of years ago, we introduced a game, a programming competition game, where first we within the company and then eventually members of the general public got a chance to write computer programs that would try to win some strategy game.

So in fact, it isn't really a game of people, but it's a game of programming where you program something to try to win. The game was really successful internally and excited our engineers and got them to think really deeply about algorithms and about how to structure situations in game theoretic ways. And we decided to launch it, thinking that it would attract many programmers that would then hear about 2 Sigma. Some of them might actually decide they want to work with us. It would also educate people because it requires very sophisticated and clever programming to win these games, and we're really interested in educating more and more people in tech. It was sufficiently

successful. The first year that we did it again, and this year there were, I'm a lot of about 6,000 players that wrote bots, as we call them, to play from about 1,000 organizations, 100 countries. In the top 10, there were six nations represented. And in the top 10 winners of this, two of them were high school students, amazingly enough. One of them from Brooklyn and one from Argentina. So I'm trying to rethink all my questions from last week. No, no, new questions.

Okay, fresh questions. We hear a lot about the competition. for talent in technology. You obviously have all these financial firms that want programmers, coders, people like that, and they're competing with tech firms in Silicon Valley. How intense is that competition? And what's the benefit of trying to attract competition through something like this game versus more traditional enticements to the financial industry, like just offering people, say, a lot of money? Well, I think first and foremost, what we're seeing is technology playing a bigger and bigger role in almost every industry.

I refer to that as CS plus X for all X. So the innovation is occurring at that intersection of computing and X. It's certainly happening now in finance. But I think what comes first is technological excellence. So we see ourselves as having to play in exactly the same markets for technology. talent, then tech companies in many domains, and I think that will occur even beyond finance and health care and education, et cetera, in the future, this kind of a global technology community.

We try to appeal to that in having a culture internally that values technology, that values algorithms, it values careful thinking, values terrific engineering, and we try to portray that externally so that people know that's the kind of firm they're joining. So tell us about the game specifically. What kind of game is it? So the game is a turn-based strategy game. So this year, somewhere either two or four players on the game when the game starts. The players have three ships each in outer space.

And the goal is to have the ships take over a large number of planets and basically take over the galaxy that they're part of. It's really simple in a way that the ships can really do only three things. They can move a certain number of positions. They can land on a planet and they can take off from the planet. And there's some things that happen when they encounter other ships and when they get on the planet, how they gain strength and when more ships are created. But there are only three commands to do it. On the other hand, there are many, many possible positions in the galaxy,

and that's what makes the game interesting. there is a huge combinatorial explosion, as we say, of moves that you can make at any given time. So that it's extremely challenging to write a program to win in this galaxy. Compare the complexity of this game to a sort of move-based game like we would, like chess, for example. So in chess, the thing we think about, despite all the complexity of doing it, is that there's only one piece you move at a time. And that piece, depending upon the piece, can do different kinds of things. But we call it a branching factor of 35. At each move in the game, you can do about 35 things.

In HALight, the branching factor is 10 followed by 2,500 zeros. So a very, very large number of moves. So it's essentially impossible for a human to play. But a bot can play it really well because computers, as we know, are pretty fast. So people are playing this game, which bots have been most successful and what types of strategies have they been pursuing? This is a really interesting question. In the game, you might think that the approach should be that people should sit down, players should sit down, and think hard about should they go to a near planet that's very large, should they go to a distant planet that's smaller, should they hide out in a corner and wait for other players to interfere with each other? the like, that's an algorithmic approach to the game, or there's the question of, should we be

doing what, say, the deep mind people in that Google subsidiary in London are doing, and building AI programs that play the game against each other and learn the right approaches by essentially trial and error and by seeing which wins. Both approaches are used in the game. The top players, the top, say, 30 or 40 players used algorithmic approaches where they really thought things through. However, now this year, some of the top players in the top 50 or 60 actually built very simple bots with very small amounts of code that actually learned by playing the game millions and millions of times. And it's quite interesting that that actually is working in a world which is this difficult. And, of course, you mentioned the Google Deep Mind endeavor. It's important in the history of chess computers, this is the two different approaches.

So back in the 90s, when we think of Casparov versus Deep Blue, Deep Blue at the whole library of games and all these grandmasters training it. And the new generation just learns chess from day one. And it teaches itself without any GMs or anything. And these days, that new approach is what works. But what you're saying in this game, you've seen some success from both approaches. That's right. In the recent Alpha Go program that DeepMind did, they learned to be a world champion in chess in four hours of play without much background, really remarkable. This game is considerably harder.

So if we think about artificial intelligence, some artificial intelligence is just to try to duplicate what people do. So like an early problem in AI was digit recognition. Could you read, say, the numbers on a check automatically? That was AI just a few years back. That was a very hard problem. Now, then another problem in AI is to do something that humans do but do it better. So that's like self-driving cars. You can easily imagine that it should be possible, maybe it's hard, to build a self-driving car,

because we can do it pretty well. Then there are these questions of things which we can't even do. And that's a game like Haylight. Can we get AIs to do that? And there are implications, of course, in financial markets. We're all kind of challenged by predictions and optimization and financial markets. maybe it's very much the case that these AI systems in the fullness of time will do things we ourselves can't even think of doing today and making a better economic system.

So I'm always curious when it comes to these bots that are essentially self-learning the game, how good are they at dealing with spontaneity or the unpredictability of other people's decisions or say, you know, just a human playing the game who might make a mistake? Do they always assume that the other players are rational or can they react in some way to the unexpected, I guess? I think it's a really good question. I don't know the answer. And I think it's a subject of research now to understand that. Two things come to mind.

One is I saw some of the early newscasts on the early go-playing programs. And people thought they were really creative in doing things that hadn't been seen before. I'm not a go-officianado, but I believe that to be true. The second is it's certainly the case that many think that great creativity is kind of serendipity, or almost a kind of randomness that happens. And of course, if we think that and we think that great creativity comes out of kind of the random ideas that maybe one of our strange colleagues might have some days, that can be programmed. You can get the news whenever you want it with Bloomberg News Now.

I'm Amy Morris. And I'm Karen in Moscow here to tell you about. our new on-demand news report delivered right to your podcast feed. Bloomberg News Now is a short five-minute audio report on the day's top stories. Episodes are published throughout the day with the latest information and data to keep you informed. Yes, there are other products like this from a variety of news organizations, but they usually rerun their radio newscasts throughout the day. That's not what we do. We create customized episodes that can only be heard on Bloomberg News Now. We don't wait an hour to publish breaking news.

When news breaks, we'll have an episode up in your podcast feed within minutes. So you're always getting the latest stories and developments. Get the reporting and the context from Bloomberg's 3,000 journalists and analysts. We're all over the world. Listen to the latest from Bloomberg News Now on Apple, Spotify, or anywhere you listen. So obviously, humans can't play this game, Haley. It's way too complicated. Can humans appreciate the game, like the same way?

Like if you're watching two bots play against each other, is it understandable enough so that someone could look at the game and sort of grasp what they're doing? Absolutely. A couple of things about that. Number one is that if humans are going to want to program bots for the game, they have to find it entertaining. So it has to be an interesting objective that they're trying to achieve. And they have to be able to watch and understand what their bot is doing.

And it's quite exciting. So you need that for the game. And then secondly, we in fact saw that. In reality, many people have put up plays of the game on YouTube and other places where you can watch really interesting games and how they unfold. And you get to see the strategy. For example, what may happen is a player's bot, I have to be careful how I say this, a player's bot may realize that it has very little chance of winning.

But perhaps if it goes, hides in a corner, the other players may defeat each other and it might come in second. And that's a strategy that happened in the first round of the game. I feel really bad for that thought. Yeah. Well, in fact, we do tend to personify these things over time, which is another interesting aspect of how humans deal with computers. But we didn't see this behavior until the last week or so of Halite version 1.

And then all of a sudden, we call it an emergent behavior. It emerged from the game we never anticipated that that happened. And there are other kinds of strategies that also occur as well in the game. So you've been running two rounds of this game now, Haylight, right? Right. You're in your second iteration. Have you ever recruited anyone that was playing the game? Has it actually translated into tangible recruitment benefits for you?

Yes. Happy to say that. We have a tremendous employee that came out of Haylight 1, who's working with us in our London office. And we have many more people that remind us that they know about 2 Sigma because of HALight. So it's valuable from a marketing perspective as well. So I think it's something that will be around and helping us for a long term. In fact, I met a college intern who did HALight as a high school student and said that she knew about 2Sigma because she did it as a high school student. I want to turn to more just the, you know, talk about quantitative findings.

and some of the lessons you've learned. Before we do, though, and before we move off the game, when we, in our first attempt at recording this episode at the end, you said, oh, you wanted to talk a little bit more about some of the high school students who had done so well in the game. And so I don't want to forget to do that this time. Tell us a little bit more about how high school students

are who they are. How could they compete with the top computer scientists in programming a bot? So let's just start with one thing first. So you have to design a game, so that it's easy to get started with, right? That's a nice thing about checkers, right? For little kids, you can learn the rules quickly,

and yet it's pretty sophisticated to play it well. The same thing happens here. You want to build a game that's easy to get started with, but that has a really, really long path, maybe essentially an infinite path towards perfection. So maybe there can be no absolute perfection. You can play a very, very long time. Then it's a much better game.

So we even wrote a paper about how to design these games called the design and implementation of modern online programming competitions. So again, going back to the ease of starting, we realized that since they're easy to start playing, they're accessible to high school students. So we went out and did a bunch of hackathons around the New York City area and some other places and had quite a bit of acceptance. We had almost 1,000 high school students doing this worldwide. And we learned about it because a teacher in Texas initially wrote to us and said that

it was a great opportunity for members of his class to start programming. And we think that early outreach is very important. It's also a core value of the firm because the co-chairs of the firm are very involved in mathematics education for young kids and also for programming educations via the MIT Scratch initiative for middle schoolers and high school students. So just one last thing. On that one, I've got to just mention it. So this kid in Brooklyn actually had an article.

written about him in the Brooklyn newspaper. So that was very exciting. I was called the Brooklyn High Schooler Takes on the World. We'll have to check that one out. We'll link to it when we post this. Yeah. So widening the conversation out to finance and tech, we were referring to 2 Sigma earlier as a very well-known quant fund.

I'm wondering what makes a quant fund a quant fund, given that nowadays it feels like pretty much every fund has some sort of systematic or programmatic trading actually happening? Right. So 2 Sigma is a tech firm that looks at many places where we can apply technology to optimize outcomes and finance. So we're also in insurance and we're in venture capital, et cetera. But certainly one of the things we do is investment management, as you mentioned. I think what differentiates us is, number one, the deep and long-term technical talent that we've had. After all, we were started by an MIT PhD and AI about 15 or more years ago, David Siegel, and John Overdeck, the other co-chair,

is a real expert, a mathematician, a silver math Olympiad, and a statistician. So the two of them really brought this to the firm quite a while back, and it's everywhere in the firm. Second is we do have scale in this. We've been doing it a long time, and I think that scale is really something that differentiates us from many of our competitors. Right, because as we know, we've all heard every bank CEO these days or at times they say, oh, we're really a software company that does banking or really a tech company. But you have a long experience with companies that are undisputably tech companies, Google and IBM.

What are the biggest differences in terms of culture that you see at a place like 2 Sigma versus your experience at Google? I think probably if you could name one, It's that technology is viewed at least as the equal, if not the driver of the core business. So at our firm, there's no question that those of us that do computer science, mathematics, and statistics, are viewed by almost everyone as the basis of the firm's success. Now, of course, we need and we're very happy to have the folks that do compliance and legal and all the other activities that are needed in the firm. but it's really a technology and math and statistics first operation.

I think the same thing is true at the really successful tech companies as well and became frankly less true at the tech companies that didn't do so well. It is kind of interesting that if you think at places where algorithms and programmatic strategies might be really interesting to do, the finance companies should theoretically be really, really intriguing because banks and insurers have these reams and reams of data that should be interesting for anyone with the technology background. But it almost feels like it's taken a little bit of time for people to catch on to that. And it's only now that a lot of the financial firms are making this really big push. Why do you think it's taken a bit of time? So one is, of course,

finance used technology very early on, right? It was among the earliest users just to computerize account records and transfers and such. So perhaps it's the case that because finance used a lot of technology, there became kind of a installed base of old technology that actually acted as an impediment to modernization. So I think that is one fact. So those of us that are newer in the business have an advantage. An example, of course, if you look at, say, online advertising, it didn't exist more than a couple decades ago, that's when it all began, so there can't be an installed base from the 1960s.

So I think we didn't have, if you will, negative inertia in Newfields, and we did have some of that in finance. The second is, I think it's important to understand what we should be doing in finance, and that's making financial systems, economic systems work better. So all of us like capitalism, we like decentralization, we like optimization of the firm, and we all hope that it will lead to Pareto-optimality and an efficient operation of society that produces lots of goods and services for all. But we all know that if we're not careful, inventories build up or prices get out of whack or people have irrational exuberance and the like, I believe with the proper application of data, the proper application of mathematics and statistics, we can do a better

job of running these economic systems. It's not easy, but I think that's really exciting. and I have a lot of success in attracting technical people to the firm because that's what I think we're doing. Do you proactively think about exactly what you said about building up some sort of legacy code base or some sort of legacy set of systems that 10 years from now you'll still be hewing to even if it's not the state of the art? I worry about it all the time. All of us in technology worry or should be worrying about the legacy that we will create. And it's a very difficult problem. If you think in the United States, there are literally, you know, millions and millions of programmers writing computer code all the time. All of that code will someday get old.

And I'm afraid it will look like the substructure underneath Lexington Avenue out here sometime and make it very difficult to build the next subway. But in banking, you still hear stories about some of the banks having, how do you say it, cobble or cobble, this programming language that stemmed from, I think it was World War II. basically invented in the 1940s and 1950s. And if you're one of the programmers who can still actually code in this ancient, ancient software language, apparently you can earn big money. So it does seem to be something of an issue. So common business-oriented language, cobal, yeah, I think it comes probably from the late 50s and 60s, not World War II, but you're in the right track there.

And, yeah, there's a lot of cobal code around. and some of it was written by employees who retired, maintained by the employees they trained, who have now retired, and the next generation is maintaining that. And you can just think of the engineering challenge. Do you rewrite it all, but do you even know what it does? It's a real challenge for organizations to deal with that. I don't believe we have any cobal. In fact, I'm certain we have no cobal and two signal. Cobol free. One of the things you hear a lot Silicon Valley people talk about is, the importance of culture as the enduring moat or the enduring sustainable advantage and that with whatever else that goes on as long as they have a superior culture, that that allows them to beat the competition. How do you guarantee that that's in place at 2 Sigma? And when you think

about all of these new funds or legacy funds that sort of want to do quant unit or banks trying to get into quant stuff, how much do you see that as an advantage towards competitors who would otherwise want to commodify what you're doing? I think in all of our organizations, talent is the first and most important thing. So the talent today is possessing of many opportunities because there's so many applications of advanced computer science and machine learning and AI and the like.

So we really feel that that culture is really important. And the culture is, it's hard to pin down exactly what it is. certainly it's clear objectives for the business. Certainly it's a clear understanding of what we do for our clients, and we have to understand what to do and feel good about doing that really well. But it's also soft and other things. Just if you think about the boards where people were talking about Haylight, you know, you read them if you're an employee and you feel good about working at the firm.

One of them said it's an absolute blast discussing strategies, sharing replays, and getting excited about the games with friends. That's a great place to work when you're doing that for the world. Last question, unless Joe has more, what's your top tip when it comes to avoiding technological errors such as the one we experienced last week? Well, my original career as a professor at Carnegie Mellon was in reliable distributed systems. And that means that you have to have duplication at many levels of a system. So how do you make sure you have two of everything in the challenge? That's important in financial markets so that we have capacity to keep operating.

That's probably important in games. We had many servers that could run Haylight, so if one of them, God forbid, had a problem, another one would keep running. In fact, we ran maybe tens or hundreds of servers simultaneously to deal with the load. It's probably important in radio and podcast, too. On that note, a perfect tip for all of us to remember in all endeavors of our lives. Alfred Spector, thank you very much for joining us. It's my pleasure.

or I enjoyed doing it again. So I really hope we don't have to bring Alfred in for a third time. But it was really... I disagree. I was going to say, wait, wait, wait. I was going to say it was really enjoyable speaking to him for another 30 minutes. Agree. And if we have to do it a third time, I'm looking forward to that as well.

But in all seriousness, I think we did a pretty good job sort of recreating the magic of that first. No, I really, I love that. And I love like, you know, we talk a lot about quantitative. finance in our work and we'll talk about various well-known strategies, momentum strategies, and other uses of alternative data. We talk about that all the time and in our reporting, but we don't talk about the sort of what needs to happen for people to come up with that stuff and the idea of that this stuff has to happen through recruitment and culture and academic study.

So I feel like this is an interesting, unexplored facet of all this. Yeah, I absolutely agree. And I have to say some of the machine learning that we're We were talking about this notion that bots, once they realized that they were probably not going to win or they didn't have a good chance of winning, they went and they hid in some obscure corner of the haylight galaxy. That kind of strategy is just really fascinating. And it's amazing to think that high schoolers potentially are coding that kind of learning

into the system. And the fact that in the early rounds of the game, they weren't doing that and that they learned that sort of adaptive approach over time is really fascinating. Fascinating. Does it make you think of Skynet? Makes me think of SkyNet a little bit. Definitely. Okay.

All right. This has been another edition of the Oddlots podcast. I'm Tracy Alloway. You can follow me on Twitter at Tracy Allaway. And I'm Joe Wisenthal. You can follow me on Twitter at the stalwart. And be sure to follow our hardworking producer, Tofer Forges, at Forhes T, as well as the head of podcast at Bloomberg, Francesca Levy, at Francesca today.

Thanks for listening. San Francisco. On April 4th, 2023, around two in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16.