Artificial intelligence has often underdelivered on Wall Street. The Huttenlocher brothers aim to prove skeptics wrong with their robotic bond seller.
By Hal Lux
December 2000
Institutional Investor Magazine
Replacing a bond salesman with a machine that thinks seems like a no-brainer. The robo-salesman should be faster than a human. It would never call in sick. And it surely wouldn,t need to score courtside Knicks tickets to close a sale.
That’s the promise of artificial intelligence, the kind of computer program that infers and analyzes and learns from experience. In academic settings, computers have been taught to construct syntactically correct sentences, and robots have learned to complete manual tasks through trial-and-error patterning. There have even been successful commercial applications, such as voice recognition and text analysis systems that answer customers, telephone and e-mail inquiries without human intervention.
But it has turned out to be a lot more difficult to automate the judgment and reasoning that go into structuring a complex financial deal or loan, managing an investment portfolio, picking winning stocks , or selling bonds.
Just ask the Huttenlochers: Carl, a former fixed-income and equity derivatives trader for John Meriwether’s ill-fated Long-Term Capital Management hedge fund, and his older brother Daniel, a Cornell University computer science professor. The pair set out last year to build a computerized bond salesman, powered with artificial intelligence, that would bring a new level of methodical sales coverage to the fixed-income market. Their robo-salesman doesn,t yet have a catchy name à la Deep Blue, the IBM chess-playing machine that in 1997 defeated then,world champion Garry Kasparov. For now, think of the machine, which focuses on convertible bonds, as Deep Discount.
The $200 billion convertible bond market is a promising target for AI applications. Small, thinly traded and composed of fairly complicated securities, it’s a market where sophisticated intelligence gathering carries a premium. Unlike other bond markets, where dealers and large institutional investors do most of the trading, the convertibles market has lately come to be dominated by a multitude of small hedge funds. As a result, a sell-side team of, say, ten salespeople and traders would strain to track and make an active market in somewhere around 50 convertible issues.
Enter the Huttenlochers, Internet-based software program. Painstakingly recording and analyzing the preferences of institutional investors entering orders, Deep Discount constantly scans the market for order flow that matches an investor’s historical trading activity. When it spots a reasonable opportunity, the system spits out an e-mail suggesting a trade to investors , or to a human sell-side salesman covering institutional accounts.
The Huttenlochers, who aim through their San Francisco,based company, Intelligent Markets, to license their technology to large banks and exchanges, say they are out to assist, not replace, live salespeople. “I don,t see software like ours ever being better than a human salesman focusing on a client 100 percent of the time,” says Carl. “We,re not saying we,ve built something that’s smarter than any human being. We,re saying that there is a repetitive nature to a part of this business and technology can have a big impact.”
The nature of human intelligence is hard to grasp, and artificial intelligence can be even more maddeningly elusive. Teach a computer to play chess, comprehend and answer a few verbal questions, distinguish between a nuclear reactor and a rock, and most experts would say you have progressed from mere programming to something that mimics aspects of human intelligence. Software that executes a trade at a specified market level wouldn,t be considered intelligent, but a program that anticipated when an investor might be inclined to liquidate a position might be said to exhibit intelligence. Says Dan Huttenlocher, “Artificial intelligence is the use of machines to solve problems that people didn,t think machines could solve.”
In many respects, trading is a perfect market for AI. Stocks and bonds are basically expressed in numbers, the language of computer processing. Trading is increasingly moving to automated systems, making it easier to instantaneously process and analyze investor data. And bonds are relatively fungible products , an investor looking for a B-rated steel company junk bond may not care whether he buys National Steel Corp. or LTV Corp. By contrast, most book buyers at Amazon.com would not consider From Dawn to Decadence, a cultural history of the past 500 years, a suitable substitute for Harry Potter and the Goblet of Fire.
The Huttenlochers are not the first or the only innovators to develop and sell artificial intelligence. As with much of modern computing, AI’s first advances occurred after World War II, led by personalities such as British computer science pioneer and legendary wartime code breaker Alan Turing. Thinking machines conjured up all sorts of robotic visions , like that of HAL, the independent-minded talking computer that went haywire on a space voyage set by Arthur Clarke in a fictional 2001. The reality was much more mundane, confined to experiments in places like the Massachusetts Institute of Technology Artificial Intelligence Lab until some commercial applications began to emerge in the 1980s. Again, typical of computer industry trends in the second half of the century, nothing went as quickly as technologists and entrepreneurs had anticipated. “People’s expectations about some of these technologies were just too high,” says Louis Salkind, an expert in robotics who now manages money for quantitative investment shop D.E. Shaw & Co.
AI applications did proliferate in the financial world, though not to the extent that visionaries in the field once predicted. They are evident in the straightforward voice-recognition systems that deliver quotes and account information to brokerage customers, and in unsupervised order-routing systems that can select among exchanges for the best execution option on a stock trade. In consumer and small-business lending, AI systems have enhanced credit-scoring techniques, making possible almost instantaneous credit approvals , at least on standardized products such as mortgages and credit cards.
But there have been plenty of disappointments, too. In the 1980s and 1990s, numerous attempts to apply artificial intelligence to trading and investment management came a cropper. One highly touted AI technology for stock picking was neural networks , collections of mathematical functions organized in computer processors like neurons in the brain and thought to be useful at finding subtle patterns in massive amounts of data. “Fifteen years ago AI was not just a technology in finance, it was an industry,” says Philip Berber, an artificial intelligence expert who in March sold his online trading firm CyBerCorp to Charles Schwab Corp. “It has been a long time since you,ve heard people talk about it.”
The Internet is helping to drive the current wave of financial AI experiments. The online revolution has vastly increased the daily interactions between consumers and computers; technologies such as voice recognition or personalized shopping agents can thus become feasible, customer-friendly alternatives to typing queries into standard search engines. And the ongoing automation of financial transactions and services, from trading to research distribution, creates new opportunities for applications that perform best when large volumes of information are aggregated in computer code and operating rules are clearly defined.
“Artificial intelligence really underdelivered in its first generation,” says Paul Amoruso, CEO of Palo Alto, California,based ShockMarket, a start-up that plans to release a new series of AI-powered investment services early next year. “But the Internet is a channel for capturing data and a source of data. We now have massive quantities of behavioral data to which you can apply modeling techniques from previous AI systems, and that allows you to come up with practical business models.”
Mutual fund company Pioneer Investment Management, for example, has just rolled out Keith, Logan and Janet , three online animated computer characters that can answer basic spoken questions from brokers and retail customers. UBS Warburg, along with a Massachusetts technology company called Artificial Life, recently announced plans to develop intelligent software for risk profiling, asset allocation and portfolio optimization. Deutsche Bank and D.E. Shaw chairman David Shaw are among the venture capital backers of ShockMarket.
“Any mutual fund site has a tremendous amount of data for people to go through,” says Pioneer e-commerce head Iang Jeon. “I think you,re going to see a flourishing of this type of technology.”
In practice, artificial intelligence is now made up of several independent disciplines. SciComp, an Austin, Texas, company that has sold AI technology for writing risk management software to Merrill Lynch & Co. and Bear, Stearns & Co., is a specialist in “expert systems.” This specialty aims to incorporate in computer code the collective knowledge and typical problem-solving approaches of human experts in a particular discipline. The Huttenlochers, Intelligent Markets focuses on the more esoteric field of “computer learning,” which uses software to spot patterns and predict behavior by processing and analyzing vast amounts of data. (The Huttenlochers, like many practitioners in their specific discipline these days, generally choose to market their technology as “computer learning” rather than use the undifferentiated “artificial intelligence” label.)
The increasing volume and complexity of financial customers, interactions over the Internet should play to the strengths of AI. The first successful Internet financial services picked off the low-hanging fruit, using robust but relatively dumb technology. One reason for the rapid takeoff of online stock trading was that it doesn,t take very sophisticated technology to match buyers and sellers in the highly liquid equities markets. But as financial firms target more esoteric markets and instruments, from bonds to derivatives, they will likely require the more subtle touch of AI software. “These technologies are still not deployed as widely as they should be on Wall Street,” says D.E. Shaw’s Salkind. “Do they make sense for a large group of customers? Definitely.”
Like many people trying to crack a financial problem with artificial intelligence, Dan Huttenlocher, 42, began his scientific research a long way from the stock market.
A Ph.D. in computer science from MIT and a former senior scientist at Xerox Corp.'s famed Palo Alto Research Center, Huttenlocher holds 17 patents for machine technology that mimics human vision. He once designed a computer program for the Defense Department that could spot tank movements on surveillance plane videotape.
“Dan Huttenlocher is a very, very smart fellow,” says Stanford University visiting professor Andreas Weigend, an expert in the financial applications of AI and computer learning who has consulted for Goldman, Sachs & Co. and Morgan Stanley Dean Witter. “Anything he builds, you can be sure will be good.”
On his own, Dan wouldn,t have paid too much attention to finance. His brother Carl, 28, had come up with the idea for a computerized salesman while trading convertibles at Long-Term Capital Management, where he was bombarded with calls from salesmen looking for business. “The idea was to build Carl’s ideal salesman,” says Jason Topaz, formerly LTCM’s chief technologist, who became a co-founder of Intelligent Markets. Soon Carl called in his brother to build the system.
In some ways, convertibles are a salesman’s perfect market. A convertible is a bond with an imbedded call option on the issuing company’s stock. When the underlying stock is selling below the bond’s strike price, the convertible acts more like a bond; when the convertible is in-the-money it looks more like a stock.
Analytical models are used to price convertibles, but tidbits of information remain very valuable to clients trying to predict price moves and gauge liquidity in securities that often trade infrequently. “A lot of the time on a convertible bond desk is spent in idea generation. It’s really challenging to come up with ideas that your clients are interested in,” says one former convertible bond analyst. “A sales guy will come into a meeting and say X hedge fund is getting ready to bite. He’s going to sell this high-tech, in-the-money bond because he has had redemptions and he needs the money.”
The salesman is one of the mythic figures in American business, but trying to explain where human intelligence fits into the art of the sell is no easy task, especially on Wall Street.
At investment banks, salesmen are generally thought to have cunning, not raw brainpower. Structured-products specialists who build innovative derivatives are certainly considered big thinkers in the securities industry. Swaps traders who can crunch cash flows in their heads also have a brainy reputation. Even equity research analysts who know the difference between Ebitda and FIFO can seem relatively intellectual by Wall Street standards. Salesmen are usually credited with more dubious mental skills. While they can,t quite describe what constitutes intelligence in a bond salesman, some Wall Street professionals insist they,ve seen it. “Are they rocket scientists?” asks D.E. Shaw’s Salkind, a Ph.D. in computer science who once built a large convertible bond market-making operation at the firm. “No. But I have great respect for the abilities and creative talents of the top salesmen. It’s not a simple thing to determine the proper level of interaction to have with a client. There is definitely an art to what they do.”
To be sure, gossiping on the phone may not be considered the highest achievement of Western civilization, but it happens to be an activity that even the smartest computer can,t do. And even the fiercest advocates of computerized finance doubt that the human relationships that run the bond market are going away anytime soon. Still, Carl Huttenlocher saw plenty of room for improving the sales function. When he banned salesmen from calling, they used Bloomberg to blast him with 400 to 500 electronic messages a day with trade suggestions. “You do business with people who help you,” says Huttenlocher, who became convinced that technology could cut down the number of useless calls as well as vastly expand the number of clients that a sell-side firm could cover.
One year later, with the help of a few million dollars in venture capital from Robert Bass, a member of the famous Texas-based investor clan, and thousands of research hours by a dozen computer scientists, the convertible bond system is complete. The system uses the discipline of computer learning to predict what investors will do when faced with different trading opportunities.
Technically speaking, the system relies on neural networks and a much newer computer learning technology called support vector machines. First developed in the early 1990s and based on advances in statistical theory that reduce the amount of data necessary to run certain programs, SVMs have been used successfully in applications ranging from handwriting recognition to machine vision.
By covering more ground than humans possibly could, the Huttenlochers say, their system will add sales coverage that simply doesn,t exist on convertible desks. Instead of covering clients on a hit-or-miss basis, the AI salesman considers every order entered into the system and every piece of flow that hits the market. “A computer can do a better job in some respects because it’s much more methodical than any human,” says Dan Huttenlocher. “But the best salesman is responding to nuances that no system can capture.”
Now Intelligent Markets its deep into its own sales effort. Despite the collapse in Internet and technology stocks, the Huttenlochers succeeded in closing a $12 million round of venture financing in October. And they have big plans: In addition to building sales systems for other types of bonds, they are making an end run around potential customers by launching their own alternative trading markets for convertibles. The market, called BuySideDirect, combines an electronic crossing market , investors can trade directly without the cost of dealers , with the intelligent salesman, which essentially provides free coverage. Potential sell-side customers have objected to the firm’s also running a market that’s trying to disintermediate them. “We recognize the conflict,” says Carl Huttenlocher, who reveals that BuySideDirect may get spun off, perhaps even to a group of sell-side firms, in the future.
Yet Deep Discount itself is too new to have a clear future on Wall Street. Institutional investors and dealers are reserving judgment until they see the system in action. Only then will the Huttenlochers really know the value of what they have unleashed. To date, they have not announced any partnerships.
“In theory, it would be great,” Jonathan Sandelman, who runs Banc of America Securities, convertible bond and equity derivatives operation, says of the AI program. “Carl’s a smart guy; we tried to hire him. But you just have to see if it lives up to its billing.”
For artificial intelligence, that’s a familiar hurdle.
