Using non-statistical induction based techniques for financial forecasting

In this paper, we used data consisting of attributes containing financial performance information on failed and non-failed banks. We developed and tested several models using three induction-based machine learning techniques (C4.5, a backpropagation neural network and SX-WEB) and linear discriminant analysis. All models showed test set classification correctness under 74% when trained and tested with a data set containing all attribute values for the year prior to failure. We analyzed individual attribute predictiveness and developed models by using different combinations of the most predictive attributes. C4.5 and discriminant analysis showed higher test set classification correctness when trained with the most individually predictive attributes. We conclude this paper with directions for future work.