Business Failure Prediction of Construction Contractors Using a LSTM RNN with Accounting, Construction Market, and Macroeconomic Variables

In the construction industry, predicting business failure and providing early warnings are critical challenges in the prevention of business failure chain reactions. Most relevant studies have developed models that predicted the probability of business failure within 1 year using financial ratios. Although a few studies have attempted to use nonfinancial information, they did not provide empirical evidence that this addition can improve the prediction performance of a model. To address these problems, this study proposed a model that used not only accounting variables but also construction market and macroeconomic variables to predict failure probability from 1 to 3 years. We examined the effects of combinations of these variables on the business failure prediction performance of construction contractors in the United States and compared the effects of combinations of these variables between three models that predict business failure within 1, 2, and 3 years. This study developed a prediction model using a long short-term memory (LSTM) recurrent neural network (RNN), which is a deep-learning algorithm. The results showed that the prediction model using both the construction market and macroeconomic variables had approximately 2%, 3%, and 4% higher prediction performance compared with that using only accounting variables when predicting within 1, 2, and 3 years, respectively. This means that the business failure prediction model had superior prediction performance from a long-term perspective when the construction market and macroeconomic variables were used in addition to accounting variables. The results of this study are expected to provide empirical evidence regarding the effect of input variable selection on the prediction performance for each prediction period and useful references for improving performance of predicting business failure of construction contractors.