Insolvency prediction by neural networks

The paper aims to develop an insolvency prediction model for companies by using neural network methodology. Insolvency prediction models are important for company owners, managers, investors and creditors who want to predict the financial health of the company in order to activate certain measures before it is too late. Data sample for the research consisted of 1500 Croatian companies. Financial ratios based on the companies' balance sheets and income statements had been calculated. The output of the model consisted of a binary variable indicating whether the company will be insolvent in the next period of observation or not. Three different neural network algorithms were tested, and the stability of the model results was evaluated by a cross-validation procedure. The best model was selected on the basis of the highest average hit rate of all samples. Since the main purpose of this paper was to extract important financial ratios in predicting whether the company will be insolvent or not, a sensitivity analysis is performed on the best model. The results indicate that the financial ratios of insolvent firms differ significantly from ratios of firms that are not insolvent, and that neural networks are an efficient tool in insolvency prediction.