Leading Edge R&D:

FINANCIAL NEURAL NET TECHNIQUES

As reported by Edward Rosenfeld, new pre-processing tricks are helping to make neural network (NN) approaches better at detecting patterns in financial markets. Developers are optimistic such tricks will help investors beat the market. Otis Port, writing in BusinessWeek magazines Developments to Watch column this month, noted two such developments. One, from the U of Kebangsaan in Malaysia, uses what it calls a modified returns function, trained NNs that then achieved a 91% hit rate when forecasting a rise or fall in a stock price the next day. The test were run using historical stock performance data.

In another test of a NN technique on historical financial data, Thomas Hellström of the Umea U in Sweden reported that his use of stochastic processes and NNs had resulted in impressive returns. His work is based on the development of what he calls ASTA: the automatic stock trading agent. He says the ASTA tool is for research in predictions of stock prices. It provides: a test bench for old and new trading strategies; an interactive development tool for trading rules; an non-interactive development tool for trading rules; and provides data generation for post processing. Hellström¹s recent publications are listed, with many available for download, from the following URL: www.cs.umu.se/~thomash/pub.htm.