This Japanese Robot Calls the Market Better Than a Human
This Japanese Robot Calls the Market Better Than a Human
By Tom Redmond & Toshiro Hasegawa
February 17, 2016 — 7:00 AM PST Updated on February 17,
2016 — 6:43 PM PST
Model has been right 68 percent of time over almost four
years
Former BOJ official holds a Ph.D. in artificial
intelligence
On the 10th day of every month, Junsuke Senoguchi has
just one thing on his mind -- the closing level of the Nikkei 225 Stock
Average.
That’s because Senoguchi, an unassuming man in his late
40s, has built a machine that’s been predicting the direction of Japanese
shares, and once a month he gets a progress report on its success. The model
makes a simple call -- whether the equity index will be higher or lower after
30 days -- and over almost four years it’s been right 68 percent of the time.
“I’m so happy” when it works, said Senoguchi, a senior
equity strategist at Mitsubishi UFJ Morgan Stanley Securities Co. in Tokyo.
“It’s because I feel I can predict the future.”
Algorithms have invaded global share markets, used by
everyone from high-frequency traders closing bets in fractions of a second, to
specialist asset managers whose strategies are determined by complex
quantitative analysis. Medallion, the fund created by James Simons at
quant-trading pioneer Renaissance Technologies, averaged a 71.8 percent annual
return, before fees, from 1994 through mid-2014. There’s also the growing field
known as robo-advice, which uses algorithms to suggest investments based on
clients’ goals and risk tolerances.
Science, Finance
Senoguchi has spent his career in finance, working as a
banking analyst for Lehman Brothers Holdings Inc. until it collapsed and then
moving to the Bank of Japan before taking his current job. But he also holds a
Ph.D. in artificial intelligence.
“Because my background lies in science, I was really
interested in seeing if I could use numbers to predict stock prices and the
economy,” he said. “Artificial intelligence gives you much better results than
conventional statistics.”
From March 2012 through Jan. 10, Senoguchi’s
robo-forecaster was right 32 times out of 47, he says. While the sample size is
small, the result goes beyond the 50 percent level expected from a coin toss. A
decades-long study by Philip Tetlock published in 2005 found expert human
forecasters on average do no better than chance.
“It’s hard to get over 60 percent without some
ingenuity,” said Akito Sakurai, a professor specializing in artificial
intelligence at Keio University in Tokyo. “You’d have to say that 68 percent is
pretty high.”
Investor Interest
Japan’s biggest institutional investors are taking note.
Senoguchi visited about 40 of them recently to field inquiries about his
creation.
Senoguchi’s machine operates like a chess computer,
poring over past data to identify patterns. He contrasts it with statistical
analysis, which he says works in terms of lines on graphs, normal distributions
and deviations from the mean. For him, that’s akin to forcing square pegs into
round holes. His model harnesses big data to see what happened and decide if it
will again.
Trading on Speed
Creating a Deep Blue -- the computer that beat chess
world champion Garry Kasparov in 1997 -- for the stock market is a complex
task. In short, Senoguchi’s machine develops hundreds of sets of rules by
combining 92 economic indicators and multiple timeframes, and then applies the
best one.
It uses elimination to decide which set has been most
predictive for stocks over the past 48 months. That becomes the model to
forecast month 49. It then deploys a decision-tree process from artificial
intelligence theory to make its guess.
Starting Afresh
Senoguchi says one strength is his model starts afresh
each time. When there are big changes in the environment, such as central bank
decisions, his model will compensate.
This is “definitely something extremely new,” said Keio
University’s Sakurai. “As far as I know, thinking up something like this and
actually using it for forecasting hasn’t happened before.”
The technique can be applied to other areas. Predicting
interest-rate moves isn’t hard, while currency rates sit at the other end of
the spectrum, Senoguchi said. Stocks are “fairly difficult,” but U.S. equities
would be easier because of the data available and lower volatility.
January Rout
Senoguchi’s model didn’t foresee Japanese shares’ worst
start to a year on record. Its prediction for the period ended Jan. 10 was for
the Nikkei 225 to advance. It wasn’t alone: almost half the forecasters in a
Nikkei survey saw their predicted lows for 2016 breached on the first day. The
Nikkei 225 added 2.5 percent on Thursday in Tokyo, trimming its decline for the
year to 15 percent.
Still, Senoguchi says his machine will get back on track.
“Sometimes the structure of the market changes greatly,”
he said. “The ability to change the model when this happens is a big difference
from previous approaches.”
Comments
Post a Comment