Early detection of corporate failure using neural network

摘要

Bankruptcy prediction is important in today's turbulent business climate. Traditional statistical methods like multivariate discriminant analysis have achieved some success in this area. In this study, a branch of artificial intelligence: neural network is used to predict bankruptcy. In particular, a feedforward back-propagation neural network model is applied to a sample of U.S. bankrupt and healthy companies from manufacturing and services sector. Various topologies and learning parameters are experimented to find the highest possible overall accuracy rate for predicting bankruptcy. The experiments are conducted using software Neural Planner version 4.5. Results indicate that the optimal 5 hidden neuron three-layer network achieves acceptable, over 90% overall accuracy rate on testing samples. As early as three years prior to bankruptcy, neural network models correctly classify 83.3% and 82.5% of testing samples. This study also analyses the comparison between multivariate discriminant statistical model and neural network models. Results show that the latter outperformed the former.