Identifying impact of variables in deep learning models on bankruptcy prediction of construction contractors

Purpose

The study seeks to identify the impact of variables in a deep learning-based bankruptcy prediction model, which has achieved superior performance to other prediction models but cannot easily interpret hidden processes.

Design/methodology/approach

This study developed three LSTM-RNN–based models that predicted the probability of bankruptcy before 1, 2 and 3 years using financial, the construction market and macroeconomic variables as input variables. Then, the impacts of the input variables that affected prediction accuracy in each model were identified by using Shapley value and compared among the three models. This study also investigated the prediction accuracy using variants of input variables grouped sequentially by high-impact ranking.

Findings

The results showed that the prediction accuracies were largely impacted by “housing starts” in all models. As the prediction period increased, the effects of macroeconomic variables on prediction accuracy increased, whereas the impact of “return on assets” on prediction accuracy decreased. It also found that the “current ratio” and “debt ratio” significantly influenced the prediction accuracies in all models. Also, the results revealed that similar prediction accuracies could be achieved using only 8, 10, and 10 variables out of a total of 18 variables for the 1-, 2-, and 3-year prediction models, respectively.

Originality/value

This study provides a Shapley value-based approach to identify how each input variable in a deep-learning bankruptcy prediction model. The findings of this study can not only assist in obtaining better insights into the underlying concept of bankruptcy but also use to select variables by removing those identified as less significant.