Electronic Data Systems Corp has signed a profit-sharing deal with Sabre Fund Management to provide an asset allocation model based on a neural network. The firm has developed a network that will help investors to decide in what proportions their money should be invested, across stock, bonds and currencies. The system has to be trained on a single personal computer, accelerated with an Intel Corp 80860 RISC chip. The network is fed large amounts of historical data and macro-economic data from which it extracts predictive relationships to help it guess at asset performance. When trained, the network can be moved to an unaccelerated 80486 machine, but will continue to train itself by measuring its predictions against real-world results. The network and users will talk to each other via Microsoft Corp’s Excel spreadsheet. Electronic Data will be paid on a performance-related basis, meaning that if the system does not come up to scratch, neither will EDS’s return on the contract. The system took a year to develop, eight months of which was testing.
