V.11:6 (261-266): A Neural Network System For Reliable Trading Signals by Marlowe D. Cassetti

A Neural Network System For Reliable Trading Signals by Marlowe D. Cassetti

For those of you who have complained that you have to be a rocket scientist to understand what we run in Stocks & Commodities, here's a rocket scientist to explain it to you.

Artificial intelligence (AI) technology can be applied successfully to the art of trading, and to this end, cutting-edge AI technology has produced a high-performance trading system. Central to this system is a unique neural network that differs from the conventional, fully connected network. The genesis of this network is the neural network tutorial featured in Carol Halquist and George Schmoll's "Neural networks: A trading perspective" 1989 article in STOCKS & COMMODITIES, in which a simple neural network was programmed in the C language, using the code listing in that article as a starting point. This configuration is designed so the network recognizes specific input patterns and then associates those patterns with the desired output. The network also features data preprocessing, rapid training and good trading performance.

A NEURAL NETWORK CONFIGURATION

The neural network configuration identifies unique patterns in price momentum, price acceleration, volume movement and price volatility. Its configuration is a limited one. A limited connection configuration is specifically designed to recognize patterns, which in turn feed the feature detector hidden layers. (See Figure 1 for a schematic of a limited neural network.) Pattern detector nodes contained in the first hidden layer connect to feature detector nodes in the second hidden layer. That layer in turn feeds the output node that produces the single network be output. Because this network connects neurons logically, it forces the training to evaluate the inputs as distinct patterns and then associate the features of the different input patterns.