| |
Investing with Neural Networks
Sophisticated
computer programs take the human element out of picking winners on
Wall Street
William Peter Hamilton, former editor of the Wall Street Journal,
was a market timer extraordinaire. Hamilton's investment instincts
beat the market by nearly 3 percentage points a year between 1930
and 1997. There's just one hitch--Hamilton died in 1929. His results
are real, but he is not--at least not any longer.
Those sparkling returns were produced by a VirtualHamilton
neural network--a branch of artificial intelligence whereby software
programs "learn" through trial and experience--created by a team
from New York University and Yale. The real Hamilton ran the Journal
in the early 1900s, but the academics mined his writings to
replicate the pundit's mind. They then fed the VirtualHamilton seven
decades of market data to see how it performed.
Techies joke that AI is a technology that is supposed to make
real computers act as they do in movies such as 2001: A Space
Odyssey and last summer's A.I. Wall Street's AI can't yet match
Hollywood's version of thinking, self-aware computers, but as much
as $250 billion is currently being managed using sophisticated
computer tools. These include neural nets, expert systems
(investment acumen distilled into rules of thumb), and genetic
algorithms (stock strategies digitally converted into cyberspace
creatures that mutate and evolve like human DNA).
"We all have the same data, and the question is what the hell
are we going to do with it," says Doug Case, chief investment
officer at Advanced Investment Technology in Clearwater, Fla. Case
sees AI as the key to decrypting high-velocity,
information-saturated financial markets. "AI can deal with that data
and handle these disorderly global markets," says Case, whose $1
billion firm is majority owned by State Street Global Advisors.
There's even a chance that as AI filters down to amateur stock
pickers (box, Page 24), the result may be warp-speed markets where
using this technology will be a must. "In this escalating arms race,
the humans with better information and more powerful AI tools will
be able to fight the more competitive battle," says John Moody, a
professor of computer science at the Oregon Graduate Institute and a
hedge-fund manager.
"Skunk Works." At PanAgora Asset Management in Boston,
researchers in the firm's advanced products division (nicknamed the
"Skunk Works" as homage to the secretive Lockheed Martin unit that
developed the Stealth fighter) have created a hedge fund where
stocks are traded by a "virtual securities analyst." It uses an
expert system to try to mimic human analysts by converting their
smarts into a string of programmed if/then statements. (If a
company's cash flow is less than the sector average, then the
quality of that cash flow is low.) A summary of these statements
then produces a buy/sell decision. "Real analysts think what they do
is some sort of art, but it can really be reduced to rules," says
Edgar Peters, the firm's chief investment officer. Why not just hire
a real analyst? A human analyst can analyze only a small number of
stocks. An AI analyst can cover them all--and without a fat expense
account or million-dollar salary.
Neural networks function more like the human brain. They can
compare existing stock-trading patterns with previous situations and
eventually "learn" what works and what doesn't as the program
digests more data. Unlike traditional financial models, neural nets
capture interconnections among financial variables. At Case's AIT,
neural nets search out linkages between stock performance and
variables such as price momentum, free cash flow, and the state of
the overall economy.
AIT's neural nets have discovered, for instance, that with
some stocks, the price-earnings ratio is a key indicator of its
future return during good economic times. But when the economy is
slowing, the stock's price momentum becomes more critical. Gaming
company Aztar is one of AIT's largest positions. With low inflation
and a steepening yield curve (a widening gap between short- and
long-term interest rates), AIT's models show valuation and price
patterns for the stock similar to those that have been bullish in
the past. AIT's AllCap large-stock portfolio has beaten the overall
market by an average of 3 percentage points a year since 1999. Those
strong, though not otherworldly, results back up Case's cautionary
contention that while AI is a formidable investing tool, "it's not
some holy grail."
Still, the results can sometimes be astounding. Standard
& Poor's uses a neural net to compile its Neural Fair Value 20
portfolio--available in its Outlook newsletter for $19.50 a
month--which gained 29 percent last year, compared with a 13 percent
loss for the S&P 500. The network constantly looks back six
months to find the factors that seem to affect stock prices to
predict the best performers over the next six months. Among the
stocks in the portfolio are Computer Associates, PacifiCare Health
Systems, and Tommy Hilfiger.
The VirtualHamilton presents a more tantalizing use of the
technology. Why not also a VirtualBuffett or VirtualLynch? These
digital doppelgangers might beat the originals by quantifying the
unconscious intuition of these fabled investors. Just as an Ichiro
Suzuki doesn't run trajectory and velocity calculations before
catching a fly ball, many managers probably don't fully understand
how they analyze stocks. Digitize a superstar manager's moves, and
you might be able to hack his financial mind. "That's called reverse
engineering," says Yale finance professor William Goetzmann. "And I
suspect it is scaring some managers away from using a single broker
who can view all of their trades." Using available information,
Goetzmann himself has been attempting to reverse engineer the
decisions made by managers of some unnamed mutual funds. "The idea
being to see what makes managers trade, what signals they use, and
if there is a magic formula," he says.
Math whiz. If there's a wild card in this investing arms
race, it may be FatKat. The company may sound like a villain in a
James Bond flick, but it's really a fledgling investment firm in
Wellesley, Mass., founded by inventor and AI evangelist Ray
Kurzweil. Although he's currently mum on FatKat, Kurzweil has
written about the potential of mathematical formulas known as
genetic algorithms to beat the market. The Darwinist process would
begin with software randomly generating a million sets of rules for
buying and selling stocks. Each set is a financial organism with the
rules constituting its DNA. The ones that can't beat the market are
killed, while the stock-savvy survivors mutate and breed until the
population is back to a million. Rinse and repeat 100,000 times.
"The surviving software creatures should be darn smart investors,"
he writes in The Age of Spiritual Machines.
How smart might AI programs get? By the year 2050, perhaps,
investment software programs may be able to "come up with their own
investment hypotheses, test them out, and implement them," says
Andrew Lo, director of MIT's Laboratory for Financial Engineering.
For now, though, humans still have a big role to play in the AI
investment process. While the numbers are being crunched, the world
keeps spinning and you need humans to keep track of it. At AIT, it
takes all weekend to download data and update investment models. You
also need humans to monitor the world for events that aren't
reflected immediately in the data, such as terrorist attacks. And
what happens if supersmart computers eventually get so good at the
prediction game that all investors are made of silicon rather than
carbon? Then the computers, as Kurzweil puts it, "will be trying to
outpredict each other." |
|
 |
 |
| Sign up to
receive our weekly tip sheet to what's new at
usnews.com. Subscribe
| |
|
 |
 |
|
| Subscribe
to U.S.News & World Report magazine and receive 4
FREE ISSUES!. | |
|
| |
Education
College
Graduate school
