Friday, December 15, 2006

The speed of finance

The speed of light is not fast enough if you're in the business of vacuuming up micropennies (keeping bid ask spreads tight) using an automated trading system. You've got to have physical proximity to the exchange in order to keep lag times short! I guess the endpoint to this arms race is to have your servers co-located in the same facility as those of the exchange! Or, to be fair, the exchange could impose a random delay on all trades which is of order of the light travel time around the earth.

Is this kind of activity "making markets more efficient" -- the usual defense of financial shenanigans -- or just gaming the system as much as possible (the other standard description :-) ? If we're talking about milliseconds, I imagine that is about as efficient as we need. Also, I'm a bit confused as to why an outside system would be better at matching bids and asks than the internal algorithm of the exchange itself (unless it is taking on a little risk; see below). It's not surprising that these two types of entities -- trading automatons and exchanges themselves -- would eventually start stepping on each others' toes.

The next step is a little more AI in the algorithms. They probably already peek at futures and options prices to predict the direction of movement in the underlying, taking a bet rather than just locking in an arbitrage.

WSJ: NORTH KANSAS CITY, Mo. -- About four years ago, Dave Cummings moved his trading firm's computers from a storefront in this Kansas City suburb to buildings in New York and New Jersey that house central computers for two big electronic stock exchanges.

The move shaved a precious fraction of a second from the time it takes Mr. Cummings's firm, Tradebot Systems Inc., to buy or sell a stock on computer-based exchanges like Archipelago. It now takes Tradebot about 1/1000 of a second to trade a stock, compared with 20/1000 before the move -- a difference of about the time it takes a computer signal to zip at nearly the speed of light from Kansas City to New York and back.

That may not seem like a big difference, but in Mr. Cummings's obscure corner of the stock-trading universe, speed is critical and fractions of seconds loom large. Tradebot's computers are programmed to detect, among other things, tiny, fleeting differences between bid and offer prices of stocks, then to pounce, buying stocks at one price and almost immediately reselling them for a fraction more. If his firm hadn't moved its computers, says Mr. Cummings, "we'd be out of business."


Dozens of other firms, ranging from Citadel Derivatives Group to a brokerage unit of J.P. Morgan Chase & Co., also employ split-second trading strategies. That has set off an arms race to shave the time it takes for orders to reach the computers of electronic exchanges. In their quest for the choicest locations, at least 40 of Tradebot's competitors have carted their computers to the same buildings, a practice known as co-location.

The kind of trading practiced by Mr. Cummings is a particularly fast form of "algorithmic" or "black-box" trading, in which computer programs decide when to buy and sell securities. Hedge funds such as SAC Capital Advisors, D.E. Shaw & Co. and Renaissance Technologies have been using computers in their investment strategies for years. These days, a variety of new computer strategies rely on lightning-quick trading. For example, computers are being programmed to take news headlines into account when executing trades, and media companies including Dow Jones & Co., publisher of The Wall Street Journal, and Reuters Group have begun releasing news in computer-readable formats that cater to them.

Mr. Cummings's strategy -- which is shorter term than most and is highly reliant on speed -- was made possible by the growth of electronic-trading networks. These trading platforms -- Globex at the Chicago Mercantile Exchange, Archipelago at NYSE Group, INET at Nasdaq Stock Market and others -- account for more than half of all trading in household-name stocks and financial futures contracts.

Trading mostly Mr. Cummings's own money, privately held Tradebot, which has about 20 employees, makes between $30,000 and $150,000 a day and up to $20 million a year, people familiar with its finances say.

Electronic exchanges use computer systems to match buyers and sellers. They execute orders without involving floor traders known as specialists, who arrange transactions through auctions on the NYSE. For years, these specialists and the Wall Street dealers who traded Nasdaq stocks profited on gaps between bid and offer prices.

Opportunity for Profit

The electronic marketplaces bill themselves as a democratizing force -- a way to cut out specialists and dealers. Mr. Cummings realized that these electronic-trading pipelines also provided a new opportunity for profit: If he used fast enough computers and programmed them just right, he could harvest the same kind of trading profits that specialists and dealers long have gathered, albeit on a smaller scale for Mr. Cummings.


In Mr. Cummings's world, the fundamentals of a stock are of little consequence. His firm favors large stocks, such as Microsoft, because it can trade in and out quickly. On one recent afternoon, a computer screen in Tradebot headquarters indicated that the firm accounted for more than 10% of that day's trading in Microsoft. On many days, Tradebot's trading totals account for as much as 5% of all Nasdaq trading, on par with trading levels at such giant firms as Fidelity Investments.

"He's had a huge impact," says Jamie Selway, former chief economist of Archipelago and now head of White Cap Trading, a small New York brokerage firm. By constantly trolling to buy and sell, he says, Mr. Cummings enables others to trade more quickly. "You're probably trading with Dave Cummings, but don't know it," Mr. Selway says.

Mr. Cummings, 37 years old, learned to program as a teenager when he also worked at his father's computer-software store. After earning an engineering degree at Purdue University, he worked for three years at a health-care software company. In the mid-1990s, he moved to the Kansas City Board of Trade, donning a forest-green jacket and taking to the pits with other traders, buying and selling futures contracts tied to stocks and wheat prices.

In 1999, he quit his trading job, settled at a computer in a spare bedroom of his house and set out to develop software to replicate the actions of a floor trader in an arena where floor traders don't exist -- the electronic marketplace. Months later, he rented a small office in a bank building, furnished it with old furniture and a few computers, and launched Tradebot with $25,000 of personal savings.

Harold Bradley of American Century Investments, a Kansas City-based money-management firm, saw Mr. Cummings demonstrate the system while Tradebot was hunting for money to trade. Mr. Cummings told potential investors that he had created a robot to do what floor traders did -- hence, the name Tradebot. His firm would buy and sell based on signals from its computer program, collecting lots of nickels and dimes by being faster than other traders. Mr. Bradley says he declined to invest, figuring it would be difficult to make money on a large scale.

Mr. Cummings raised several hundred thousand dollars from a Chicago trading firm that wanted to use his programs. (Mr. Cummings bought this partner out in 2002.) Tradebot focused on stocks listed on the Nasdaq, which were easier to trade electronically than those at the New York Stock Exchange, where rules, at that time, limited electronic trading.

Tradebot took no long-term views on where stock prices were heading. Instead, it aimed to profit off tiny differences between what investors were willing to pay for heavily traded stocks and what others were willing to sell them for. In 2000, when stock pricing began shifting from increments of 1/16 of a dollar to one cent, many Wall Street trading desks gave up that business, figuring it was no longer profitable enough. That left Tradebot and competitors such as Citadel, Automated Trading Desk and Getco LLC to go after the profits.

Mr. Cummings's programs took into account fluctuations in stocks as well as in financial instruments such as stock-index futures. His computers, for example, would simultaneously try to buy Microsoft stock from investors at, say, $27.69 a share, and to sell it at $27.70, for a penny-per-share profit. Theoretically, human traders could spot the same opportunities, but Mr. Cummings's computers could usually get there first. His firm sometimes traded more than 1,000 times a minute. It ended most days owning no stock, cashing out all its positions.

Tradebot's employees constantly monitored the program trading. Mr. Cummings says he figured that once in a while the strategy wasn't going to work. It was often best, he concluded, to move out of the way when the market was moving sharply in one direction or another. He did not want to be stuck holding big blocks of stocks when their values unexpectedly plunged. If trading in a particular stock began losing money, an employee could switch off the program for that stock.

In the early 2000s, electronic marketplaces such as Island ECN and Archipelago courted professional traders. They offered rebates on certain kinds of complex trades such as "limit orders," in which traders offer to buy or sell shares within certain price parameters. The rebates allowed Tradebot to profit on some trades that merely broke even, according to several former employees.

By 2001, other firms were employing computer strategies similar to his, and some of them had an advantage: Their computers sat in Island's headquarters building in lower Manhattan. As a result, Tradebot's computers were a fraction of a second behind its rivals in trying to grab the best prices. "We were excluded because of the speed of light," Mr. Cummings says. "We had to move our computers."

Mr. Cummings discussed the problem with Island executive Matthew Andresen. "Our customers outside of New York are at a big disadvantage," Mr. Andresen recalls telling him. "We agree it should be a level playing field."

In addition to transaction fees Mr. Cummings's trades generate for Island, he agreed to pay a few thousand dollars per month to put Tradebot's computers in Island's facilities. The difference was significant: In the 1/50 of a second it used to take to execute an order from Kansas City, Tradebot could buy and sell at least 20 times.

Archipelago, an Island competitor now owned by the NYSE, was also after Tradebot's business. Mr. Cummings told Archipelago he was concerned that its system could take as long as a second or two to process his orders. "I can't manage that risk," he told Mr. Selway, then Archipelago's chief economist, Mr. Selway recalls. At times, Tradebot had to "throttle back" its Archipelago trading because that exchange's system couldn't handle the volume fast enough, says Mr. Cummings. Mr. Selway figures he spent a hundred hours on the phone with Mr. Cummings discussing how to make Archipelago better.

A few months after sealing the Island deal, Mr. Cummings visited Archipelago Chief Executive Gerald Putnam in Chicago and sized up the firm's computer room. Archipelago executives worried that letting Tradebot move its computers in would prompt complaints from other brokerage firms. They told Mr. Cummings to wait for Archipelago to complete a new data center in Weehawken, N.J., that could accommodate traders' computers.

Mr. Cummings didn't wait. He moved some Tradebot computers about a mile away from Archipelago's Chicago office. Then, when Archipelago finished its New Jersey data center, he moved the computers there. If a rival's computers "are in the exchange, and I'm across the street, then I lose," Mr. Cummings explains.

On Oct. 9, 2002, Tradebot for the first time traded 100 million shares in a single day, most of them stocks listed on Nasdaq. At that time, the NYSE was still a private cooperative controlled by floor brokers and specialists who traded stocks the old-fashioned way. It maintained limits on electronic trading, including restrictions on certain big orders and on rapid in-and-out trading. With an eye toward trading more NYSE-listed stocks, Tradebot bought Philadelphia-based Bloom Staloff Corp., which traded NYSE stocks from its specialist business at the Philadelphia Stock Exchange. Tradebot installed faster computers, but the NYSE restrictions proved troublesome. Mr. Cummings shut down the operation at a loss within six months and continued focusing on Nasdaq stocks. (In October of this year, the NYSE started opening its 2,700 listed stocks to electronic trading.)

Analysts estimate that Mr. Cummings's net worth surpassed $10 million several years ago. His trading business is probably worth several times that, although as a private firm, its finances aren't disclosed.

Over the past three years, about 40 other firms have moved their computers closer to the action, with bigger brokerage firms following smaller trading firms. Merrill Lynch & Co., Goldman Sachs Group, Deutsche Bank AG and J.P. Morgan Chase all have computers sitting near marketplace mainframes, or at least in the same neighborhoods.

Some exchanges, such as the Chicago Board of Trade, have resisted co-location, mostly because they don't want to put at a disadvantage investors who trade through firms that haven't moved their computers.

'Question of Fairness'

"It begs the question of fairness," says Andrew Brooks, head of stock trading at mutual-fund firm T. Rowe Price in Baltimore. "You shouldn't win just because you have better access to speed." Some investors also complain that rapid traders often cancel and resubmit orders, clogging the trading systems of electronic exchanges when markets are busy.

The Securities and Exchange Commission allows co-location, as long as electronic marketplaces give firms equal access to prized computer locations. The SEC has looked into the fairness issue and intends to continue to monitor it, according to two people familiar with the agency's approach.

Last year, Mr. Cummings says, he grew worried about the future of his trading strategy. More firms were co-locating. Consolidation in the exchange business, he feared, would lead to higher fees. The NYSE had acquired Archipelago, and Nasdaq had bought Island's successor company, INET.

In 2005, Mr. Cummings turned over day-to-day management of Tradebot and shifted most of his focus to a new venture: creating an exchange-like competitor to the NYSE and Nasdaq that would match buyers and sellers for a small fee. He called it Better Alternative Trading System, or BATS.

That June, he emailed about 100 traders and other industry contacts about BATS. Within days, senior executives from both NYSE and Nasdaq called to ask why one of their biggest customers was setting up a rival marketplace.

"The market needs competition," Mr. Cummings recalls telling Robert Greifeld, Nasdaq's chief executive officer.

"I wish you luck, but it's harder than you realize," Mr. Greifeld recalls replying.

Mr. Cummings has already poured about $4 million into the venture, which now occupies most of his time. He now travels on monthly sales calls and sends potential BATS clients a wooden baseball bat with the company's logo.

A few months after starting in January, BATS drew minority investments from Lehman Brothers Holdings Inc., Credit Suisse Group and Morgan Stanley. BATS now regularly handles about 5% of Nasdaq trading, the firm says.

7 comments:

Anonymous said...

If Mr. Cummings is really taking out $20,000,000/year and algorithmic trading is a classic Zero Sum Game, who exactly is this "small player" taking the money away from... after expenses? Goldman? Merrill? NYSE Group? I can't see this strategy working in the long term.

Anonymous said...

One can envision some Joe Shmo filing a class action suit and as a result regulations requiring simultaneous release of quotes in NYC, Chicago and SF (just do a delay)

Anonymous said...

"It's not surprising that these two types of entities -- trading automatons and exchanges themselves -- would eventually start stepping on each others' toes."
Not surprising and a question of logic:exchanges were once owned by members who were making the rules and providing liquidity.Now they are owned by shareholders essentially running a toll-booth and NOT providing liquidity.The new shareholder owned exchange model is flawed,and logically some competition is stepping in albeit slowly since exchanges are protected by regulators allegedly for the public benefit.

Anonymous said...

"The next step is a little more AI in the algorithms."

Highly doubtful. AI has been hyped in finance for decades, with little practical impact.

Steve Hsu said...

David,

By AI I just meant looking for discrepancies or signals in derivatives prices related to the underlying, to perhaps guess at the direction of the next move.

Full blown AI is very far off, at least in my opinion :-)

http://infoproc.blogspot.com/2004/11/generalized-turing-test.html

Anonymous said...

According to the following very interesting (57 min) presentation, prof Dreyfus, says AI will never be capable of replicating human behavior...

http://webcast.ucsd.edu:8080/ramgen/UCSD_TV/11335.rm

Anonymous said...

I don't think AI could ever be considered the same as human intelligence, but with the learning capabilities it sure has come a long, long way.
Matt

Blog Archive

Labels