Genetically Optimized Neural Network Classifiers for Bankruptcy Prediction - An Empirical Study

The use of financial statement data to predict the future jlnancial health of an economic entity is generally considered a complex problem where non-linear pattern recognition methods such as neural networks (NN) can provide a performance advantage. In fact, bankruptcy prediction has emerged as a popular benchmark for neural network performance. However, the use of neural networks in bankruptcy prediction as well as in business applications in general has been hindered by the fact that large numbers of parameters have to be @e-tuned before NN can be used successfully while no analytical solution for this optimization problem exists. Sequential Genetic Algorithms (GA) are guided random search methods that are based on the biological principles of evolution. GA may be used as powerful function optimizers that manipulate a population of bitstrings so as to retain and recombine their most promising features (“survival of the fittest’?. In our empirical study, different parameters of NN such as topology, connection weights and input variable selection are encoded in the genetic material. In each generation, the corresponding NN were then trained and evaluated to determine the quality of the solution. Our study uses a large sample of 6667 real-world @uncial statements from German corporations. The performance of the neural networks is compared on the basis of the beta-error (misclassification of solvent companies), while the (more costly) alpha-error (misclassification of insolvent companies) is kept constant. Early results porn the combination of GA and NN appeared promising as the beta-error on the test set was reduced from around 40% to about 35%. However, the optimized neural networks did not outperform earlier solutions on the validation set which was not used at all during the optimization process. We termed this inability to generalize “overselection effect”, a problem similar to the “overlearning effect” well known from NN applications, The paper explores one approach for reducing the ‘overselection effect” and concludes with issues remaining for further research. This paper constitutes but one part of a larger research effort in which the results for the bankruptcy prediction problem obtained from multiple Soft Computing methods were compared with each other and with previously published experiments [1;2].