Tech Report CS-93-36
Feedforward and Recurrent Neural Networks and Genetic Programs for Stock and Time Series Forecasting
Peter C. McCluskey
September 1993
Abstract:
Adding recurrence to neural networks improves their time series forecasts. Well chosen inputs such as a window of time-delayed inputs, or intelligently preprocessed inputs, are more important than recurrence. Neural networks do well on moderately noisy and chaotic time series, such as sunspot data. A single neural network or genetic program generalizes poorly on weekly stock market indices due to the low signal-to-noise ratio. When the responses of a number of networks are averaged, the resulting forecast shows substantial profits on historical data.
(complete text in pdf or gzipped postscript)