FOCUS
40
FORTUNE.COM // NOVEMBER 2019
A.I. VS. THE WOLVES
OF WALL STREE T
INSIDE A .I.
AN INVESTOR SUBMITS a mountain of online orders
to sell shares in a grocery chain. It may look like
another routine trade, but it’s a scam.
What he’s actually done is spook others into selling their stock
in the company, causing its value to tumble. After canceling his
own sales, the scammer quickly swoops in to buy the thousands
of shares flooding the market at a bargain price.
This illegal scheme, known as spoofing, is well known. But it’s
difficult to police because shady traders can use algorithms to
place and swiftly cancel orders, over and over again.
In October, stock exchange Nasdaq intro-
duced a tool that uses deep learning, a subset
of artificial intelligence, to flag such suspicious
trades. The technology is supposed to make
it easier to identify fraud amid the billions of
trades that investors make annually.
Nasdaq has put the new software to work
alongside its more conventional systems for
spotting stock market crimes like insider
trading. What sets deep learning apart, says
Nasdaq’s machine-learning chief Michael
O’Rourke, is that it’s “good at finding things
that are very hard to describe.”
To train its deep learning, Nasdaq fed it
order and trade data from its exchange as
well as nonpublic information. After a year of
testing, the team that created the technology
judged it reliable enough to use more broadly.
One potential risk with using A.I. for spot-
ting scams is a flood of false positives—over-
whelming the staff who review flagged trades.
“You can’t miss, but you don’t want to over-
shoot,” says O’Rourke. However, it’s too soon
to calculate a meaningful error rate, he adds,
and he thinks a limited number of false posi-
tives is acceptable.
In the end, it’s up to federal regulators to
pursue wrongdoers—a challenge considering
the Securities and Exchange Commission has
limited resources. Still, Nasdaq’s team says
the new software will help by providing the
agency with better and timelier information.
Nasdaq isn’t alone in using machine learn-
ing to identify stock fraud. The University
of Michigan, for example, is also developing
A.I.-powered statistical techniques to counter
fraudsters. The problem, however, is that it
lacks access to the market data that is so criti-
cal to training its system.
“There are some daunting obstacles to
building algorithms to detect manipulation
when you don’t have a realistic toy market
to play with and when you don’t have access
to real-time data,” says Gabriel Rauterberg,
an assistant law professor at the university.
( Nasdaq rival New York Stock Exchange
may also be well positioned to develop deep
learning for detecting fraud, but it declined to
comment to Fortune.)
Eventually, Nasdaq says, it plans to sell
versions of its new technology to other ex-
changes, something it already does with other
software it has developed.
Nasdaq is tapping deep learning to police its marketplace
for certain kinds of hard-to-spot fraud. By Jonathan Vanian
ILLUSTRATION BY SÉBASTIEN THIBAULT
