Predicting corporate financial distress- A deep neural network approach
Background. Predicting bankruptcy is of great importance for creditors, investors and other stakeholders. Early warning signs of financial distress allow stakeholders to take action to minimize the negative consequences of a bankruptcy. Altman Z-score is one of the most used credit scores for predicting financial distress and this study reassesses the Altman Z-score using a neural network approach to compare the accuracy of the models within a 6-year time horizon. Objective. The objective with this study is threefold. First, we test what hypertuning settings that yield the most accurate result for FDP using a neural network model. The second objective is to test if the accuracy of the Altman Z-score improves when the model is reassessed using the same variables but with a neural network approach. The last objective with this study is to test what happens to the accuracy of the assessed Altman Z-score when it is tested on a global data set. Method. The method applied on the problem of predicting bankruptcy is a Recurrent Neural network (RNN) with Long short-term memory (LSTM). This is a machine learning technique commonly used for forecasting on time-series but could potentially yield accurate prediction on categorial outputs as well. By feeding the RNN-LSTM with Altman's original financial ratios and applying previous time-steps, the model trains on large amount of data and tests on a perfectly balanced test-sample between bankrupt and non-bankrupt firms within 6 years. Result. The original Altman Z-score performed 48.91% accuracy for predicting bankruptcy in the manufacturing industry when tested on a domestic data set. Our reassessed Altman model performed an 87.45% accuracy on the same data set. When the data set was extended to a global data set, the Altman Z-score performed a 50.1% accuracy for predicting bankruptcy while our model performed an accuracy of 72.95%. Conclusion Our result shows that the reassessed Altman Z-score using a neural network method outperforms the original Altman Z-score in terms of accuracy. This finding strengthens the theory that intelligent models, like the neural network, can outperform statistical models, like MDA, by loosening some of the assumptions that must hold true in the statistical models. The accuracy of the reassessed Altman Z-score decreases in accuracy when it is tested on a global data set compared to a domestic one. This indicates one of two things. Either the increased data points on a global data set is not enough to compensate for the increased heterogeneity, or the variables in our models were originally chosen for a domestic data set and are not well suited for a global data set.
MSc in Accounting and Financial Management