Stock ranking: neural networks vs multiple linear regression

The use of neural networks to replace classical statistical techniques for forecasting within the framework of the APT (Arbitrage Pricing Theory) model for stock ranking is examined. It is shown that neural networks outperform these statistical techniques in forecasting accuracy terms by an average of 36% and give better model fitness in-sample by one order of magnitude. Values are identified for network parameters for which these figures are statistically stable.<
>