Troubled firms exhibit different symptoms. A prediction model that incorporates non-monotonic relationships between distress risk and predictors, as well as complicated correlations among predictors is required. This paper investigates the advantages and limitations of using deep neural network models in corporate distress prediction, relative to traditional methods. I propose a deep neural network model with a hybrid layer containing predictors with the most recent values and the abstraction of their historical patterns. When predicting financial failures in the following month, the average pseudo R-squared of this deep neural network model are 47.32% and 44.56% on the training and test sets, respectively, compared to 32.92% and 32.19% for the logistic model proposed by Campbel...[ Read more ]

Troubled firms exhibit different symptoms. A prediction model that incorporates non-monotonic relationships between distress risk and predictors, as well as complicated correlations among predictors is required. This paper investigates the advantages and limitations of using deep neural network models in corporate distress prediction, relative to traditional methods. I propose a deep neural network model with a hybrid layer containing predictors with the most recent values and the abstraction of their historical patterns. When predicting financial failures in the following month, the average pseudo R-squared of this deep neural network model are 47.32% and 44.56% on the training and test sets, respectively, compared to 32.92% and 32.19% for the logistic model proposed by Campbell et al. (2008). This indicates that the traditional models fail to process more complicated information. Financial failures predicted by deep neural networks but not by the logistic model on average have much larger asset sizes, lower market leverage ratios, higher cash holdings, higher stock prices, less volatile stock returns, and higher past stock returns. Besides, these missed cases tend to occur during normal times when the economy is generally trending up.