A convergence instability analysis of neural networks applications in financial data sets

The investigative route taken here is that of a well proven technique of 'learning', albeit by a machine, using Artificial Neural Networks (ANN). A proprietary software package, was used in the study, of charting the Share Price (SPRCE) and Earnings per Share (EpS) of two Australian banks. Key performance data was obtained for several companies over a period of five years and this information was 'trained' to result in the listed SPRCE and EpS of these companies over the stated rime period. The intention being, with the knowledge of anticipated key financial and performance information, some indication of the SPRCE or EpS could be inferred and thus investment decisions can be made. With certain settings of the computational process, some convergence problems were experienced with the data set training exercise. These unstable runs naturally resulted in less than satisfactory predictions. The use of a limited version of this proprietary package however, resulted in some success in 'learning'. The question remains however, of the credibility or the robustness of such decision-making aids. It is argued here that while some credibility can be given to results within certain marker types, such as a Bear, steady or Bull markets, it is virtually impossible to generate near accurate predictive rr ends on markets as a whole. Some solutions to this dilemma are presented in this paper.