PROGRAM AIDS STOCK MARKET PREDICTIONS

Computers, in general, are dumb. And when they can learn -- using a neural network program -- they are slow.

But an LSU electrical engineering professor, Subhash Kak, has developed a neural network program that can learn instantly.

Kak has already licensed one version of his program, named Anvish -- the Sanskrit word for "search" -- to Neural Technologies out of Kansas City. Neural Technologies will use it as the basis for a smart search engine.

But Kak expects his work to have more wide-ranging applications. One field in which he is particularly interested is stock market predictions.

"This program does 30-60 percent better than other techniques at predicting highs and lows," Kak said. One of Kak's graduate students, Asheesh Mehta, investigated the value of his neural network program by feeding it information about Ford and IBM stock prices. The program outperformed the "buy and hold" method by more than 22 percent for IBM and more than 18 percent for Ford; it outperformed the "moving averages" method by 34 percent for IBM and more than 54 percent for Ford.

Traditional neural network programs, which learn from previous history, take minutes or hours to learn a procedure. Kak's program differs in that it can learn instantly. The value of this "instant learning" capability shows up in an application like stock market prediction, where a large number of variables is required for accurate results. Kak's program can analyze vast quantities of complex data and identify patterns from which predictions can be made.

"Predicting a financial timing series is much harder than predicting a random timing series," Kak said. That is because of the huge number of variables affecting stock prices. Some variables, such as the state of the world economy, will affect a local economy sometimes but not at other times. A variable such as the overall strength of a sector, such as energy, may affect some stocks in the sector more strongly than others. Furthermore, in a market the prices are a result of the decisions of millions of investors, whose behavior cannot be predicted, he said.

Kak said his program should be good at predictions for chaotic systems, such as weather. The key is the quantity and quality of the data the program has to work with. The program can also be used in the prediciton of business and engineering time series.

The program, which has fewer than 200 lines of code, can also be used to interpret data to determine what is relevant and what is not, Kak said. "If a search engine returns 3,000 hits, that's not very useful. This program gives us the possibility of creating an intelligent system, which can make possible the 'mining' of information. What we have is just the beginning of it."