Flash Boys: A Wall Street Revolt (17 page)

HE’D JOINED GOLDMAN
at an interesting moment in the history of both the firm and Wall Street. By mid-2007 Goldman’s bond trading department was aiding and abetting a global financial crisis, most infamously by helping the Greek government to rig its books and disguise its debt, and by designing subprime mortgage securities to fail, so that they might make money by betting against them. At the same time, Goldman’s equities department was adapting to radical changes in the U.S. stock market—just as that market was about to crash. A once sleepy oligopoly dominated by Nasdaq and the New York Stock Exchange was rapidly turning into something else. The thirteen public stock exchanges in New Jersey were all trading the same stocks. Within a few years there would be more than forty dark pools, two of them owned by Goldman Sachs, also trading the same stocks.

The fragmentation of the American stock market was fueled, in part, by Reg NMS, which had also stimulated a huge amount of stock market trading. Much of the new volume was generated not by old-fashioned investors but by the extremely fast computers controlled by the high-frequency trading firms. Essentially, the more places there were to trade stocks, the greater the opportunity there was for high-frequency traders to interpose themselves between buyers on one exchange and sellers on another. This was perverse. The initial promise of computer technology was to remove the intermediary from the financial market, or at least reduce the amount he could scalp from that market. The reality turned out to be a windfall for financial intermediaries—of somewhere between $10 billion and $22 billion a year, depending on whose estimates you wanted to believe. For Goldman Sachs, a financial intermediary, that was only good news.

The bad news was that Goldman Sachs wasn’t yet making much of the new money. At the end of 2008, they told their high-frequency trading computer programmers that their trading unit had netted roughly $300 million. That same year, the high-frequency trading division of a single hedge fund, Citadel, made $1.2 billion. The HFT guys were already known for hiding their profits, but a lawsuit between one of them, a Russian named Misha Malyshev, and his former employer, Citadel, revealed that, in 2008, Malyshev had been paid $75 million in cash. Rumors circulated—they turned out to be true—of two guys who had left Knight for Citadel and guarantees of $20 million a year each. A headhunter who sat in the middle of the market and saw what firms were paying for geek talent says, “Goldman had started to figure it out, but they really hadn’t figured it out. They weren’t top ten.”

The simple reason Goldman wasn’t making much of the big money now being made in the stock market was that the stock market had become a war of robots, and Goldman’s robots were slow. A lot of the moneymaking strategies were of the winner-take-all variety. When every player is trying to do the same thing, the player who gets all the money is the one whose computers can take in data and spit out the obvious response to it first. In the various races being run, Goldman was seldom first. That is why they had sought out Serge Aleynikov in the first place: to improve the speed of their system. There were many problems with that system, in Serge’s view. It wasn’t so much a system as an amalgamation. “The code development practices at IDT were much more organized and up-to-date than at Goldman,” he says. Goldman had bought the core of its system fifteen years earlier in the acquisition of one of the early electronic trading firms, Hull Trading. The massive amounts of old software (Serge guessed that the entire platform had as many as 60 million lines of code in it) and fifteen years of fixes to it had created the computer equivalent of a giant rubber-band ball. When one of the rubber bands popped, Serge was expected to find it and fix it.

Goldman Sachs often used complexity to advantage. The firm designed complex subprime mortgage securities that others did not understand, for instance, and then took advantage of the ignorance they had introduced into the marketplace. The automation of the stock market created a different sort of complexity, with lots of unintended consequences. One small example: Goldman’s trading on the Nasdaq exchange. In 2007, Goldman owned the (unmarked) building closest to Nasdaq. The building housed Goldman’s dark pool. When Serge arrived, tens of thousands of messages per second were flying back and forth between computers inside the two buildings. Proximity, he assumed, must offer Goldman Sachs some advantage—after all, why else buy the building closest to the exchange? But when he looked into it he found that, to cross the street from Goldman to Nasdaq, a signal took 5 milliseconds, or nearly as much time as it would take, a couple of years later, for a signal to travel on the fastest network from Chicago to New York. “The theoretical limit [of sending a signal] from Chicago to New York and back is something like seven milliseconds,” said Serge. “Everything more than that is the friction caused by man.” The friction could be caused by physical distance—say, if the signal moving across a street in Carteret traveled in something less direct than a straight line. It could be caused by computer hardware. But it could also be caused by slow, clunky software—and that was Goldman’s problem. Their high-frequency trading platform was designed, in typical Goldman style, as a centralized hub-and-spoke system. Every signal sent was required to pass through the mother ship in Manhattan before it went back out into the marketplace. “But the latency [the 5 milliseconds] wasn’t mainly due to the physical distance,” says Serge. “It was because the traffic was going through layers and layers of corporate switching equipment.”

Broadly speaking, there were three problems Serge had been hired to solve. They corresponded to the three stages of an electronic trade. The first was to create the so-called ticker plant, or the software that translated the data from the thirteen public exchanges so that it could be viewed as a single stream. Reg NMS had imposed on the big banks a new obligation: to take in the information from all the exchanges in order to ensure that they were executing customers’ orders at the official best market price—the NBBO. If Goldman Sachs purchased 500 shares of Intel at $20 a share on the New York Stock Exchange on behalf of a customer without first taking the 100 shares of Intel offered at $19.99 on the BATS exchange, they’d have violated the regulation. The easiest and cheapest solution for the big banks to this problem was to use the combined data stream created by public exchanges—the SIP. Some of them did just that. But to assuage the concerns of their customers that the SIP was too slow and offered them a dated view of the market, a few banks promised to create a faster data stream—but nothing they created for customers’ orders was as fast as what they created for themselves.

Serge had nothing to do with anything used by Goldman’s customers. His job was to build the system that Goldman Sachs’s own proprietary traders would use in their activities—and it went without saying that it needed to be faster than anything used by the customers. The first and most obvious thing he did to make Goldman’s robots faster was exactly what he had done at IDT to enable millions of phone calls to find their cheapest route: He decentralized Goldman’s system. Rather than have signals travel from the various exchanges back to the Goldman hub, he set up separate mini–Goldman hubs inside each of the exchanges. To acquire the information for its private ticker plant, Goldman needed to place its computers as close as possible to the exchange’s matching engine. The software that took the output from the ticker plant and used it to figure out smart trades in the stock market was the second stage of the process: Serge rewrote a lot of that code to make it run faster. The third stage was called “order entry.” As it sounds, this was the software that sent those trades back out into the market to be executed. Serge worked on that, too. He didn’t think of it this way, but in effect he was building a high-frequency trading firm within Goldman Sachs. The speed he created for Goldman Sachs could be used for many purposes, of course. It could be used simply to execute Goldman’s prop traders’ smart strategies as quickly as possible. It could also be used by Goldman’s prop traders to trade the slow-moving customer orders in their own dark pool against the wider market. The speed Serge gave them could be used, for example, to sell Chipotle Mexican Grill to Rich Gates at a high price in the dark pool while buying it from him at a lower price on a public exchange.

Serge actually didn’t know what the speed was being used for by Goldman’s prop traders. As he worked, he became aware of a gulf in understanding between himself and his employer. The people at Goldman with whom he dealt understood the effects of what he did but not their deep causes. No one at Goldman had a global view of the firm’s computer software, for instance: He figured that out on the first day, when they asked him to look into the code base and figure out how the different components talked to each other. In doing so, he saw that there was shockingly little documentation left behind by the people who had written that code, and that no one at Goldman could explain it to him. He, in turn, was not privy to the commercial effects of his actions—in part, he sensed, because his superiors did not want him to know them. “I think it is done intentionally,” he said. “The less you know about how they make the money, the better it is for them.”

But even if they had wanted him to know how the money was made, it is unclear Serge would have cared to know. “I think the engineering problems are much more interesting than the business problems,” he says. “Finance is just who gets money. Does it wind up in the right pocket or the left pocket? It just so happens that the companies that make money are the companies like Goldman Sachs. You can’t really win in that game unless you are one of these people.” He understood that Goldman’s quants were forever dreaming up new trading strategies, in the form of algorithms, for his robots to execute, and that these traders were meant to be extremely shrewd. He grasped further that “all their algorithms are premised on some sort of prediction—predicting something one second into the future.” But you needed only to observe the 2008 stock market crash from inside of Goldman Sachs, as Serge had, to see that what seemed predictable often was not. Day after volatile day in September 2008, Goldman’s supposedly brilliant traders were losing tens of millions of dollars. “All of the expectations didn’t work,” recalls Serge. “They thought they controlled the market, but it was an illusion. Everyone would come into work and were blown away by the fact that they couldn’t control anything at all. . . . Finance is a gambling game for people who enjoy gambling.” He wasn’t a gambler by nature. He preferred the deterministic world of programming to the pseudo-deterministic world of speculation, and he never fully grasped the connection between his work and the Goldman traders’.

What Serge did know about Goldman’s business was that the firm’s position in the world of high-frequency trading was insecure. “The traders were always afraid of the small HFT shops,” as he put it. He was making Goldman’s bulky, inefficient system faster, but he could never make it as fast as a system built from scratch, without the burden of 60 million lines of old code underneath it. Or a system that, to change it in any major way, did not require six meetings and signed documents from informational security officers. Goldman hunted in the same jungle as the small HFT firms, but it could never be as quick or as nimble as those firms: No big Wall Street bank could. The only advantage a big bank enjoyed was its special relationship to the prey: its customers. (As the head of one high-frequency trading firm put it, “When one of these people from the banks interviews with us for a job, he always talks about how smart his algos are, but sooner or later he’ll tell you that without his customer he can’t make any money.”)

After a few months working on the forty-second floor at One New York Plaza, Serge came to the conclusion that the best thing they could do with Goldman’s high-frequency trading platform was to scrap it and build a new one from scratch. His bosses weren’t interested. “The business model of Goldman Sachs was, if there is an opportunity to make money right away, let’s do that,” he says. “But if there was something long-term, they weren’t that interested.” Something would change in the stock market—an exchange would introduce a new, complicated rule, for instance—and that change would create an immediate opportunity to make money. “They’d want to do it immediately,” says Serge. “But if you think about it, it’s just patching the existing system constantly. The existing code base becomes an elephant that’s difficult to maintain.”

That is how he spent the vast majority of his two years at Goldman, patching the elephant. For their patching material he and the other Goldman programmers resorted, every day, to open source software—software developed by collectives of programmers and made freely available on the Internet. The tools and components they used were not specifically designed for financial markets, but they could be adapted to repair Goldman’s plumbing. He discovered, to his surprise, that Goldman had a one-way relationship with open source. They took huge amounts of free software off the Web, but they did not return it after he had modified it, even when his modifications were very slight and of general, rather than financial, use. “Once I took some open source components, repackaged them to come up with a component that was not even used at Goldman Sachs,” he says. “It was basically a way to make two computers look like one, so if one went down the other could jump in and perform the task.” He’d created a neat way for one computer to behave as the stand-in for another. He described the pleasure of his innovation this way: “It created something out of chaos. When you create something out of chaos, essentially, you reduce the entropy in the world.” He went to his boss, a fellow named Adam Schlesinger, and asked if he could release it back into open source, as was his inclination. “He said it was now Goldman’s property,” recalls Serge. “He was quite tense.”

Open source was an idea that depended on collaboration and sharing, and Serge had a long history of contributing to it. He didn’t fully understand how Goldman could think it was okay to benefit so greatly from the work of others and then behave so selfishly toward them. “You don’t create intellectual property,” he said. “You create a program that does something.” But from then on, on instructions from Adam Schlesinger, he treated everything on Goldman Sachs’s servers, even if it had just been transferred there from open source, as Goldman Sachs’s property. (Later, at his trial, his lawyer flashed two pages of computer code: the original, with its open source license on top, and a replica, with the open source license stripped off and replaced by the Goldman Sachs license.)