Ai in Business

Watch out, stock pickers.

Researchers have been working on an artificial-intelligence computer program designed to mimic the way an analyst uses financial news. In simulated trading, the program beat the S&P 500, and when combined with quantitative stock-picking techniques, it saw a return on trades of more than 20%.

European Pressphoto Agency

The New York Stock Exchange

To make stock predictions, the program does what is known as “text mining” — scanning large volumes of content and analyzing the words in it. Computer-aided quantitative funds already are plentiful, but they analyze numerical data rather than text. The new program is different because it attempts to simulate what has traditionally been a human activity.

“Our approach is more like the analyst approach, simulated by a program,” said Hsinchun Chen, director of the University of Arizona’s Artificial Intelligence Lab, in an interview with Digits. “You have an analyst reading papers, looking for clues that others have not observed.”

The program, which was first reported by MIT’s Technology Review, scans stock prices and financial news and buys or shorts stocks it believes will move more than 1% in the next 20 minutes. The system sells the stocks after 20 minutes.

“When you do long-term predictions, there are many variables,” Dr. Chen said. “But … you can have an advantage if you look at five minutes, 10 minutes.”

In evaluating the meaning of a single news item, “the system won’t be as accurate as an individual analyst,” Dr. Chen said. “The computer is maybe 80 to 85% accurate when analyzing text, but it can read maybe 100,000 times the amount of data.” He said the amount of data being analyzed is one thing that would make it difficult to game the system.

The system’s creators tested it using data from five weeks in the fall of 2005 — more than 9,000 news articles and 10 million stock quotes. During that test, the system, called AZFinText, had an 8.5% return on trades, beating the S&P...