Neural networks have been shown to be a powerful classification tool in financial applications. However, neural networks are basically black boxes that do not explain the classification procedure. The training results from neural networks, which are sets of connection weights expressed in numeric terms, hardly shed light on the importance of input attributes and their relationship for classification problems. To address this issue, researchers have developed different algorithms to extract classification rules from trained neural networks.
The purpose of this paper is to validate the prediction power of extracted rules from one algorithm GLARE (Generalized Analytic Rule Extraction). The input to the GLARE algorithm is a set of connection weights from a trained neural network, and the output is classification rules that can be used to predict new cases as well as to explain the classification procedure. We apply the conventional backpropagation and GLARE to a data set from the CompuStat database. The input to the prediction problem is a vector of financial statement variables, and the output is the rate of return on common shareholders' equity. To test the effect of the number of training epochs on rule extraction, we train the networks for 5 and 1000 epochs before rule extraction. To test the statistical significance of performance differences between conventional backpropagation and rules from neural networks, we perform paired t test for each pair of the average returns. The experimental results support the superiority of extracted rules to conventional backpropagation on selecting high return stocks.
Lam, Monica, "Extracting Rules from Neural Networks to Predict High Return Stocks" (2001). ICEB 2001 Proceedings. 159.