Businessweek Archives

Where Neural Networks Are Already At Work


Special Report

WHERE NEURAL NETWORKS ARE ALREADY AT WORK

In the world of finance, anything that provides even a slight edge over rivals can mean millions in extra profits. So investment pros are turning to gurus offering exotic computer technologies such as neural networks and genetic algorithms. The following stories show how some big Wall Street firms and small-time investors are exploring and sometimes adopting cutting-edge artificial-intelligence techniques, many of which have proven their mettle in such fields as aerospace and gambling.

FROM AEROSPACE TO ROCKET SCIENCE

Through much of the 1980s, Troy Nolen helped create top-secret military software. Hired by a major defense contractor, he worked on Air Force programs for guiding the flight and battle patterns of the YF-22 fighter. The software, running on five on-board computers, had to make split-second decisions based on data from ground stations, radar, and other sources. Plus it had to predict what the enemy planes would do, guiding the jet's actions accordingly.

These days, Nolen, 48, applies what he did for the military to the world of finance. The head of his own consulting firm, New Hampshire-based Norad Inc., he spends weekdays commuting from a houseboat in New Jersey to the Wall Street offices of Merrill Lynch & Co. His assignment: Develop a system to accurately set daily prices for more than 1 million corporate bonds. Merrill needs to determine current trading values for all of those bonds every business day between 4 p.m. and 5 p.m. Like the YF-22's software, the bond-pricing system must sift through reams of data from many sources and make swift judgments. "It's the same type of problem," Nolen says. "Only instead of pilots, we have bond traders."

Software has long played a big role on Wall Street. Until recently, however, computers were used mainly to crunch numbers. At the big brokerage houses, the chief technologists, known as rocket scientists or quants, are crack mathematicians. They know how to derive equations for setting bond prices or predicting market performance. But using those techniques in isolation is "the old way," says Tom Murphy, the Merrill vice-president who hired Nolen. A new breed of rocket scientists is adding artificial-intelligence methods for much smarter and more precise ways of predicting the market.

Perhaps the most accepted of these new techniques is so-called neural networks. Nolen is building a neural net that runs on a parallel-processing computer--one that uses many microprocessor "brains" to attack a problem. Modeled after the complex pathways of the human nervous system, such nets search for patterns in vast streams of data. For instance, the software can comb through information on how a certain bond has performed historically, then take a look at economic indicators plus the fluctuations of barometer bonds such as U.S. Treasury bills. Over time, it learns how different combinations are likely to affect each bond's value.

SURVIVAL. Nolen is most excited about a technique called genetic algorithms. As in a Darwinian universe where only the fit survive, such software procedures reject formulas that don't work and steer the system in the right direction. When used in bond-pricing models, Nolen compares such algorithms to survival mechanisms in the body. "Only the good cells live," he says. Merrill's Murphy is counting on Nolen to add this technique to the neural net. "If you don't use all these techniques together," he says, "you don't get the right answers."

Merrill is not alone. Shearson Lehman Brothers Inc. has been training its own neural net to help its traders forecast market patterns. A decade's worth of historical data was used in the software. For three years now, the program has been managing its own small portfolio. Since the software can "learn" from its mistakes, it gets better over time: The first year, it lost money, but it broke even in the second and turned a profit in its third, says Stephen P. Gott, Lehman's chief technology officer.

The next challenge at Lehman is to get nets to understand not just numbers but also the significance of news stories. Using an English-language-recognition system, neural nets are being trained to read financial news stories and decide how they affect different investments. For instance, if the President makes a statement on the budget deficit or the Fed cuts interest rates, what will that mean to the price of IBM? "We haven't got that licked yet," Gott admits.

In fact, these software techniques are still experimental. At Merrill, the new bond-pricing system isn't scheduled to go live until November. The only telltale sign that any of this software is beginning to rake in big money will come when the people creating and using it clam up. It's kind of like military secrets, Nolen says. "People who are having success don't want anyone else to know about it."

FROM OTB TO OTC, A WINNING RECORD

Murray Ruggiero claims to have put himself through college by betting the horses. As a student at Southern Connecticut State University, he put his money down at off-track betting parlors four nights a week. His goal was not to win big on long shots but to devise ways to consistently spot safe bets that would provide modest returns. His all-time favorite: a horse named Doc Fella, which ran in the money for 50 races in a row. Instead of betting to win, Ruggiero would bet the horse to show. For a year, it was close to a sure thing, and Ruggiero would collect about $1.20 for every dollar he wagered, for a return of 20%.

As a computer science major, Ruggiero used what he was learning in school to figure out how to improve his odds of winning. He developed a crude expert-system program that ruled out horses that couldn't win, based on past performance and other data. The program included rules such as: "If a trotter worth less than $20,000 doesn't race in more than a month, it cannot win." Later on, he devised a neural network program called Dr. Trot that was able to predict winners in 38% of races, he claims, as compared with 30% for most professional handicappers.

BLOODHOUND. In 1988, Ruggiero co-founded a New Haven software company called Promised Land Technologies Inc. With his days at the races behind him, the ambitious 29-year-old's tiny startup is now focusing on more upscale forms of gambling: futures, options, bonds, and commodities trading. The five-employee company's first commercial program, called Braincel, has been purchased by about 1,000 people, many of them small-time investors who like the fact that the $249 software works along with popular spreadsheet programs including Microsoft Excel and Lotus 1-2-3.

While Braincel may be relatively simple to learn, it won't accomplish much--until it is trained. Like a bloodhound, it must be shown what to look for. As it is fed data, it spots patterns. "It can find subtle relationships in data," says Dan Williams, a spreadsheet product manager at Microsoft who reviewed the product.

Braincel, for instance, endeavors to predict when the Standard & Poor's 500-stock index is about to reach a weekly or monthly high. It does this first by checking each day's opening and closing prices: If the prices are getting closer to each other, and several other less obvious market patterns occur at the same time, a high may be imminent. The neural net makes and reconciles these calculations based on pastexperience.

NO NOISE. One customer, a Springfield (Mo.) business-school student named John Deatherage, uses Braincel for trading option contracts on the S&P index. Along with a new Promised Land companion program called FuturesBuilder, Deatherage created a neural network to model the index' fluctuations. He says the software reduces his risk and boosts his confidence. "It's not a crystal ball," he says. But he adds that "if the neural net says to get out of the investment, I get out."

The trick, says Ruggiero, is that the neural net knows how to "filter out the noise." In other words, it can learn when a drop in the market is just a random blip--mere "noise"--or whether it signals a real downward trend. Many investors rely on the market's moving averages to filter out the noise. But five-day averages, for instance, are misleading. If you are looking at the week's average on Friday afternoon, your freshest data point--Friday's closing point--is diluted by the older data in the average. Braincel, Ruggiero says, can help eliminate the problem by predicting what will happen on the following Monday or Tuesday, creating a predictive average.

So far, Ruggiero has had a difficult time selling to the big-time investment houses. "It's hard to show auditable results," says Microsoft's Williams. Ruggiero claims that Braincel has proven itself in trading futures for 30-year Treasury bonds. From March, 1991, until July, 1992, the software's suggestions would have given an investor an annual return of 292%, he says.

But as in horse racing, there's no such thing as a sure thing. Ruggiero's software can only offer enlightened advice based on a detailed understanding of the past. "If anyone really had something that could predict the S&P a month from today, they wouldn't be selling it," Ruggiero says. "They would be retired and living in Tahiti."Evan I. Schwartz in New York


We Almost Lost the Nasdaq
LIMITED-TIME OFFER SUBSCRIBE NOW

(enter your email)
(enter up to 5 email addresses, separated by commas)

Max 250 characters

 
blog comments powered by Disqus