Predicting survival years of Japanese listed firms using neural network

Predicting the number of survival years of firms is the extension of predicting bankruptcy of firms. In this research the number of survival years to be predicted was taken to be 1 to 5 because of data availability. Prediction models for survival years were built using a three layered neural network and financial data of 32 failed manufacturers in Japan and 32 non-failed pair-sampled counterparts. Financial data were condensed from the original 277 indexes, that is, ratios and non-ratios, to five sets of indexes. This was done using the results of statistical tests between average of failure and average of non-failure firms. The five sets of indexes, respectively, consist of 5, 9, 13, 20, and 21 indexes. To learn financial data patterns of 24 out of 32 pairs of firms mentioned above 126 models were used and four models succeeded in learning; two models each for two and three survival years. The effectiveness of the four models was determined with holdout sample data of eight pairs of firms. As a result the most effective model was a model to predict survival of two years using 21 indexes and 50 hidden units with 0% type I misclassification and 12.5% type II misclassification