Financial Forecasting Through Data Mining : A Comparative Evaluation of Probabilistic Neural Networks and Other Models

In recent years, neural networks have been applied extensively to the task of predicting financial variables. Even among neural network techniques, backpropagation algorithm has been the most popular methodology. However, the advantages of other learning techniques such as the swift response of the probabilistic neural network (PNN) suggests the desirability of adapting other models to the predictive function. Unfortunately, the conventional architecture for probabilistic neural networks yields only a bipolar output corresponding to Yes or No; Up or Down. This limitation may be circumvented in part by using a graded forecast of multiple discrete values. More specifically, the approach involves a bipolar architecture comprising an array of elementary PNNs. This paper explores a number of interrelated topics: (1) presentation of a new architecture for graded forecasting using an arrayed probabilistic neural network (APN); (2) use of a "mistake chart" to compare the accuracy of learning systems against default performance based on a constant prediction; and (3) evaluation of several backpropagation models against a recursive neural network (RNN) as well as PNN, APN, and case based reasoning. These concepts are investigated against the backdrop of a practical application involving the prediction of a stock market index.