Click for a transcript.
[Jad:] Hey, I'm Jad Abumrad.
[Robert:] I'm Robert Krulwich. This is *Radio Lab*, and speed is our subject.
[Jad:] You beat me to it. Actually, that's what this whole next segment is about. See, I had it in my bones just to set it up. I got this idea from my friend Andrew Zolli, who is a fantastic writer, wrote the book *Resilience: Why Things Bounce Back.* We were at a diner, I was telling him about this show, and he says, "You should do something about the stock market."
[Robert:] And you were like...
[Jad:] I was like, "I'm the last person who should do something about the stock market." And he’s like, "No, no, no, forget everything you think you know about the stock market."
Most of us, when we think about stock markets, if you just close your eyes and think about the financial world, what you imagine is a bunch of people in a room wearing funny-colored jackets, shouting at each other, waving bits of paper. This kind of raucous scene—people screaming, trying to figure out what a price is. And we have this cultural iconography of how the financial system works, which is largely divorced from reality.
Because, here’s the first surprise: somewhere between 50% and 70% of all trades that happen on what we think of as Wall Street aren’t executed by human beings as a result of human decisions. They’re actually executed by algorithms at a speed, rate, and scale that is beyond our comprehension. So, I decided I would try to comprehend this new world he was describing.
And, since this is a subject matter that generally makes me, frankly, frightened, I decided to call up David Kestenbaum from *Planet Money.*
[David:] Hey, Jad.
[Jad:] Hello, David K. There could be more than one David, there probably are, on Twitter. In any case, it didn’t click for either of us just how fast, how inhumanly fast trading had gotten until we visited this firm called TradeWorks.
[David:] Nice to meet you, David.
[Jad:] So we go into this little building in New Jersey. It looks like a startup or something, and this guy says, "Hello, my name is Mike Beller. I’m the Chief Technology Officer of TradeWorks." Mike over here sits us down at this computer, opens up this little program that logs exactly what’s going on in the market, insanely specific.
[Mike:] You could pick a stock. We could look at Yahoo, for example. We can literally pick some time of day that we're interested in.
[Jad:] What time is this at?
[Mike:] This is at 11:35:26.97.
[Jad:] Seconds, really?
[Mike:] And in fact, that’s not enough precision for us because we deal in microseconds—millionths of a second. We have another way of measuring time, which is the number of microseconds since midnight of the previous day.
[David:] Can you read that 417 number?
[Mike:] Sure. 417,729,979,559 microseconds since midnight.
[Jad:] Wow. So, do you always have lunch at like 2,305,000?
[Mike:] [Laughter] No, that’d be really early.
[Jad:] How many trades do you do in a day?
[Mike:] It depends a lot. A high-frequency trader might do a thousand trades in a minute. It’s about that tempo, but it’s kind of very bursty.
[Jad:] Now, what happens during those bursts is a bit of a mystery.
[David:] It’s very hard to see what’s going on. Often, says Andrew, it’s the computers testing the market, testing to see if they can find a nibble on the other side. They’ll fire out a bunch of buy and sell orders, and when another computer bites, they’ll quickly cancel the ones that didn’t stick. Like, "Nope, sorry, didn’t want to do that." They’re doing this on a microsecond basis—buy, no sorry, sell, buy, sell, sell again, no, forget about that, buy, nah.
And they create huge volumes of transactions that just disappear into the ether. There are some computer algorithms, he says, whose whole job is to combat other algorithms, fake them out.
[Andrew:] For example, we just had a very good example happen about a month ago in Kraft.
[Eric Hunsader:] That’s Eric Hunsader. He tracks high-frequency trading for the firm Nanex.
[Andrew:] Kraft? Like Kraft cheese?
[Eric:] Yes. What we saw was this algorithm jump into the market, buy up a bunch of Kraft, which jammed the price up, allowing that algorithm to sell at much higher prices to other algorithms. And we calculated out—it cost them $200,000 to push the price up, but they were able to sell about $900,000 of stock, netting a gain of over half a million dollars in a matter of seconds.
[David:] Now, to put that in context, back in the day, you know, 20 years ago when humans still ran the trading pits...
[Larry Tabb:] I’m Larry Tabb, founder and CEO of the Tabb Group. The average time it took to execute a trade was around 11 or 12 seconds back then.
[David:] And when you ask people how we got from 11 or 12 seconds to 417,729,979,559 microseconds since midnight, the answer is kind of surprising.
[Andrew:] But I'll start with the obvious part—at least it’s obvious to people who work in finance. A basic law of the market is that the fastest person usually wins. There’s always a benefit to getting information faster than the other guy.
[David:] This has been going on since Julius Reuter used carrier pigeons to send stock quotes faster than the guy on horseback.
[Jad:] That was in the 1850s.
[David:] Here’s a more modern example. Say the latest job numbers come out—U.S. employers added 227,000 jobs in February. If those numbers are good, the stock market is going to go up. So, if you can get the numbers and rush to the market before anyone else, you can buy the stock before it goes up and make a lot of money, right on the "buy low, sell high" principle.
[Jad:] But when the markets turned electronic, which began to happen in the early '90s, this basic law created a situation that was totally bananas.
[David:] What do you mean?
[Jad:] So imagine it’s the year 2000. You’ve got this market in New York—it’s electronic, basically just a building on Broad Street with a giant computer inside, matching buyers and sellers. And you have traders in different parts of the country connected to this market. Some are using automated trading bots. One day, this guy, Dave Cummings, in Kansas, notices his robot keeps getting beat. When it would send a trade to New York, like a buy order, some other robot would swoop in, get there first, and snatch up the trade. And it occurs to this guy, Dave—wait a second, is it because I’m in Kansas? If the other guy’s closer to New York, then his cable would be shorter, so I need to move closer to New York.
[Robert:] No, no, no, because we’re talking about the speed of light.
[Jad:] Well, close to the speed of light, still. Obviously, it’s because he’s in Kansas.
[Robert:] What do you mean "obviously"?
[Jad:] Because the speed of light is like a foot a nanosecond. You’re going to get your ass kicked if you’re in Kansas.
[Robert:] I don’t... how do you know this for a fact?
[Jad:] Yeah, it’s a foot a nanosecond. It takes a billionth of a second to go a foot. It’s 3 * 10 to the—
[Robert:] Why do you act like this is something everybody knows?
[Jad:] I know this because when I was in physics, if I needed to delay a signal by a nanosecond, by a billionth of a second, I just added an extra foot of cable.
[Robert:] Did you really do that?
[Jad:] Yeah, see, the proton-antiproton would collide, and it would create a muon that would go out, and you only wanted to measure—you wanted to filter all the junk so you knew when it was going to arrive roughly, so you had a little window it had to arrive in. But you had to get the timing of the window right, so it meant adding a delay. And we would just add cable—that was the easiest way.
[Robert:] You would literally go get some cable and just splice it in?
[Jad:] Not splice, like, there are LEMO connectors.
[Robert:] Oh, of course, LEMO connectors.
[Jad:] Here’s another way to think about it. Say the time it takes for information to get from Kansas to New York is something like this: [beep-beep sound].
[Robert:] Did you hear that?
[Jad:] Yeah. The first beep is when it leaves Kansas, the second beep is when it arrives in New York. We actually slowed that down just a bit so we can hear it better, but the point is, that is fast. There’s still a little space in there between the beeps, which is the travel time. Very, very little space, but even if these signals are traveling at millions of miles an hour, close to the speed of light, if somebody is a few hundred miles closer to New York than you, and they leave at the same time as you, well, then it’s going to be...
[[Beep-beep-beep sound]]
[Jad:] You hear that? That beep in the middle is some other dude beating you by a few milliseconds. These little differences matter because they’re trying to get in and out super fast, and maybe each trade they’re only making a fraction of a penny.
[Robert:] That’s it?
[Jad:] Says Andrew. But if you’re making a fraction of a penny millisecond after millisecond after millisecond, it can add up, right? But you have to be able to react really fast. So, when this guy in Kansas decided to move his robot to New York to get closer to the big market computer, it started a kind of land grab. There was a real estate bubble around some of these buildings because people were trying to buy physical real estate next to the exchanges so that the cables they would run into the exchanges would be just a few feet shorter than the other guy.
[Robert:] Wait a second, so does this mean, like, if I’m one stop up on the elevator and you’re two stops up, that I have the second-floor advantage?
[Jad:] Theoretically, yeah, that’s what it means. But I don’t know how far this real estate jockeying got because pretty early on, the people who run the market stepped in, and they were like, "Okay, this could get crazy." So they told the machine traders, "Okay, you want to be close to us? Fine, pay us some money; we’ll let you come inside."
[Robert:] Inside?
[Jad:] Inside our box, inside the mother ship.
[Robert:] Is there like some room where all these computers are keeping each other company now?
[Jad:] Oh, yes, there is. If you visit the New York Stock Exchange now—which we did, after going through months of security checks—what you see is the match itself, where the trades actually happen.
[Robert:] Amazing.
[Ian Jack:] Wow. So this is what, a 20,000-foot room?
[Jad:] This is Ian Jack; he’s head of infrastructure at the New York Stock Exchange. He showed us around.
[Ian Jack:] It has a number of rows of racks for customer equipment. In 2006, the New York Stock Exchange opened up this room; it’s the size of three football fields, filled with nothing but rows and rows of servers.
[Jad:] Banks, hedge funds, brokers...
[Ian Jack:] Yeah, a whole number of financial institutions.
[Robert:] Are these things trading right now?
[Ian Jack:] Absolutely. Each of these computers—there were close to 10,000 in the room, give or take—was at that moment analyzing the market, making a decision as to whether to buy or sell, and sending that decision over a cable into an adjacent room, where it gets bought or sold.
[Jad:] No people involved. If you stood still for a few seconds, the lights would go out. They automatically shut off if nothing moved because the assumption was there wouldn’t be people there.
[Robert:] And the whole idea of this place, says Ian, the whole premise is a level playing field.
[Ian Jack:] So any firm can come in here, and they’ll have the same access as anyone else.
[Jad:] And to make sure of that—this is my favorite part—every single rack within this facility has the same length of cabling to get to the network points at the end.
[Robert:] Exactly the same length?
[Ian Jack:] Exactly the same. Everybody gets the same length cabling. Whether you’re one foot away from the network hub or a thousand feet away, you get the same length.
[Robert:] I’m sure they send synchronized test pulses from both your trading computer and Jay-Z’s trading computer, and they make sure they arrive at exactly the same moment.
[Jad:] I like to imagine they have a guy with a tape measurer—that’s the guy you bribe.
[Robert:] Anyhow, you would think that since all machines can now be inside the exchange, literally inside the market building, the speed race would be over, right?
[Jad:] Yep.
[Robert:] No, actually, it only gets worse because the place we visited, the New York Stock Exchange, that’s just one market of many. I didn’t know this, but apparently when all trading went electronic, the markets fragmented.
[Larry Tabb:] It used to be that to trade stocks, you had the New York Stock Exchange, and then there was NASDAQ—really just those two markets.
[Robert:] Says Larry.
[Larry Tabb:] Now, there are 13 regulated exchanges. There are roughly 50 what they call "dark pools" in the marketplace.
[Jad:] Those are non-public?
[Larry Tabb:] Yeah, basically.
[Jad:] So, you’ve got these 60-some-odd different markets, and that’s created all these different speed races between them.
[Jad:] Yeah, here’s a super basic example I talked about with Andrew. In Chicago, you've got this thing called the Commodities Market. Commodities are basic goods like corn, oil, soybeans, zinc, pork—that’s what they do in Chicago. Here in New York, we do equities, and equity is a share of a company. So you have basic goods in Chicago, stocks of companies in New York. Those are different kinds of things, but they’re connected to each other, you know?
[Robert:] Cuz, like, take oil, which is traded in Chicago.
[Jad:] Exactly. A lot of companies depend on oil, and they’re traded in New York. So, say oil goes up in Chicago—you can pretty much bet that right after that, a company like Exxon is going to go up in New York. But it won’t be instantaneous, right? Because information has a speed. Back in the days of the telegraph, as we learned, it took a quarter of a second—about that long—to get from New York to Chicago. Now, with fiber optic cables, it’s about 15 milliseconds.
[Robert:] I love that. I had no idea you could actually hear the time difference.
[Jad:] Yeah, that one, I think, is pretty accurate—15 milliseconds. But say you’re in Chicago, oil goes up, you know it, and you can get to New York in 14 milliseconds. Well, you’ve got one millisecond where you know the future, you know exactly what’s going to happen. You’re not even betting at this point; this is easy money.
So what happened over time was a race of people to provide the straightest fiber line between Chicago and New York.
[Robert:] That’s Mike Beller again from TradeWorks.
[Jad:] He’s part of this race. A couple of years ago, a company came along—not his, unfortunately—and spent some eight-figure sum to cut a straighter fiber line between those two points. And, according to some reports, they blew through a mountain to do it. They did a lot. Where the state-of-the-art for communication lines at the time between the two locations was about 15 milliseconds, they came along and made that state-of-the-art 13.3 milliseconds—a savings of about 1 millisecond each way.
[Robert:] Which is just... an eon.
[Jad:] It’s a thousandth of a second.
[Robert:] That’s not an eon.
[Jad:] Well, it’s an eon when your computer system can make a decision in 10 microseconds, which ours can—that’s 10 times faster. So your computer’s like, “I can do this so fast; I’m just waiting, waiting, waiting, waiting for the news from Chicago.”
So a lot of us were sitting around thinking, what can we do about this? Turns out, there was a way to get from Chicago to New York a little faster because the speed of light through air is a little faster than through a fiber optic cable. So what they’re doing now is building a series of towers to beam the signal through the air from one tower to the next, all the way from Chicago to New York.
[Robert:] So that would bring the travel time down to about... in the neighborhood of around 8 and a half milliseconds?
[Jad:] Yeah, that would be going from this [beep-beep] to this [beep]. I mean, come on, that’s a lot of potential savings.
[Robert:] I can totally hear the difference. Is it helping? Are we fast enough now? Can we... stop?
[Jad:] Um, here’s the thing. That’s Mino Narang, the CEO of TradeWorks. He joined us for part of the interview, and he told us:
[Mino Narang:] Actually, we would love to stop this arms race. Yeah, absolutely. The arms race is a huge drain on resources.
[Jad:] But he says they just can’t.
[Mino Narang:] As it stands, when a new technology comes out that makes it possible to be faster, if I don’t adopt it and my competitors do, I will lose out to them. I have to do it.
[Jad:] And looking at Mino, you could tell this part of the job is just like the plumbing. It just kind of makes him weary.
[Mino Narang:] Yeah, no, couldn’t care less.
[Robert:] Why not just call a truce? Everyone says, "We’re not going to try and go faster. We’re already way faster than any human can think. It’s fast enough. We’re going to stop." Why not call a truce?
[Mino Narang:] Because there’s a thing in game theory called the Prisoner’s Dilemma.
[Robert:] So someone will cheat, you’re saying, basically?
[Mino Narang:] Yeah, you can’t put a gun to everyone’s head and force them to abide by this truce, even though we’d all be better off if you could.
[Jad:] Well, who would be better off?
[Mino Narang:] Look, even though this speed race sucks for us, it’s actually helping you. Because, on a basic level, anytime you replace a human with a computer, things are going to get faster, they’re going to get cheaper. And now that the machines are competing, it’s getting cheaper still. In 1992, it would have cost you about $100 to trade a thousand shares. Now? 10 bucks.
So yes, humans have been completely supplanted when it comes to short-term trading, and humans who complain about that are being disingenuous, okay? They have not been displaced by anything other than the fact that they can’t compete.
[Robert:] You seem like... you’ve had to... you seem defensive.
[Mino Narang:] Well, just because I can explain the economics of the business doesn’t make me defensive.
[Jad:] That also sounded... defensive.
[Narrator:] If Mino did sound defensive, it’s only because he, Mike, and everyone in their industry have had to answer a lot of questions over the past few years about where all this speed is taking us. And those questions always come back to one particular day: May 6, 2010, when things got a little... fruity.
[Eric Hunsader:] We hadn’t had a down day in a long while. The market had been slowly creeping up for quite a while.
[Jad:] And that’s Eric Hunsader again, the analyst who’s been tracking high-frequency trading.
[Eric Hunsader:] He says that day, even though things had been going really well, that day had started off down pretty hard, which made some sense because there was bad news coming out of Athens. People were nervous. But then, at a very specific moment, 2:42 in the afternoon, 14:42 and 44 seconds, all hell breaks loose.
[News Anchor:] Neil, let me just—let me just interrupt for a second because this market is dropping precipitously. It just went -500, it is now... 560 even offer, seven even offer, six half are trading here now, six even trading, see it on the screen. The Dow is losing about... 653 points now, Dow is down 707 points, 81 even are trading here, the 79 trading. Boom, there it goes. Look at this market, it continues to slide. Look at it—835. This is the widest we have seen us in years, now it’s down 900—wow, almost 1,000 points. This will blow people out in a big way like you won’t believe. Cancel all orders! Down 1,000 points! Cancel all orders!
[Narrator:] At 2:45 and 27 seconds, an emergency circuit breaker shuts off for 5 seconds. And that was the end of the slide. When it went out and stopped for 5 seconds, that was the bottom of the market—1,000 points down. Several hundred billion dollars vanished in two and a half minutes.
Equally weird, when trading started again, the market bounced right back up. About two and a half minutes later, it was 600 points higher than the bottom. It was like, *boing.*
These kinds of swings had happened before, but never that fast. And speed is one thing. Arguably, what’s more troubling is that we still, two and a half years later, don’t really know what happened. I mean, the SEC investigated for months, released this giant 84-page report where they essentially blamed the whole thing on one bad algorithm—that this guy in New York was trying to sell a bunch of stocks, told his computer to do it, and his computer just did it a little too aggressively.
[Eric Hunsader:] No, that’s not how it went down at all.
[Narrator:] Eric doesn’t agree. He thinks what happened is that all the high-frequency computers just clogged the network.
[Eric Hunsader:] Really, the cause of the flash crash was system overload.
[Jad:] Cuz he says a basic feature of these computer algorithms is when they detect that the network is slow, they pull out. You know, one of the maxims on the street is, “When in doubt, stay out, or pull out.” And so if you’ve got this one computer selling a ton of stock and no computers left to buy, that creates a vacuum. Now, there were people who argued that high-frequency trading had actually made the situation better. Cuz, you know, Andrew says the markets did bounce back right up to the top. The computers self-corrected.
[Robert:] Perhaps.
[Jad:] But the point is, nobody had any idea. And that’s what gets him—that we’re in a situation now where, when things go wrong, they go wrong in the blink of an eye, and then it takes us years to figure out what happened.
[Robert:] The question that comes up is, have we crossed some kind of Rubicon? Have we passed into a realm where the complexity, the speed, the volume of all this stuff makes it no longer human-readable?
[Jad:] We just don’t know what the system is doing and can’t, in principle, find out when things go wrong.
[Music fades]