Logical Designs has been engaged in a research and development project using neural networks to learn the trading strategy for stock futures. There are many claims made about the trading systems currently offered for sale today. The basic problem with most systems is that the authors of the systems test on the same data that was used to develop the system.

The basic approach being taken by Logical Designs is to apply neural network theory to the problem of developing a trading system. While there are many neural network paradigms currently available, none met the requirements for our design. Supervised methods seemed ineffective since there are a great number of buy/sell points that would yield a given return on investment. Tests attempting to train a supervised network to predict all points that could yield a profit proved to be ineffective. There is no way to know which set of points can be related to patterns in the data. Thus, the decision was made to develop a neural network algorithm whose goal was to maximize return on investment directly.

One result of this development effort was the patented Graded Learning Network. This algorithm allows a network to be trained to make trades in order to maximize return on investment without having to determine optimal trading points prior to training.

"Duane (Logical Designs) developed a trading model that actually did make money for some time, then slipped. His was the best of any and all models I have seen and I think I have looked at all the commercial 'canned' networks for sale." Larry Williams, Club 3000 News 94.07.

The current network has been trained on tick data for the SP500 future for five years of data (1986-1990), using a day trading strategy. The resulting network was tested on data from the first five months of 1991. The network achieved a profit of $30,000 for this test period, which represents a return on investment of over 100% per month. Additional research is being conducted to improve these results. Current efforts are being directed toward the development of an on line real-time trading system.

Network performance is directly related to the amount of data presented to the network. Too much information can unnecessarily increase the size of the network and make training more difficult. Too few network inputs may fail to provide sufficient information for the network to train effectively. The success of a neural network project depends on the proper analysis of the problem and correct choices related to input representation, network architecture, and training set development.

Logical Designs has developed neural algorithms not available from any other source. We can supply a custom software and hardware solution to your problem, or work with the tools you have in house. Logical Designs has the ability and experience to make your project and application a success.