| Abstract [eng] |
This work introduces the deep learning models developed to predict firms’ bankruptcy. Two deep learning models were developed using data from Lithuanian companies: a multilayer perceptron and a convolutional neural network. Bayesian optimization method was used to find the hyperparameters of the models. To investigate the impact of class imbalance on models training, models were trained with imbalanced and balanced datasets. To assess the ability of models to recognize the bankruptcies in a data with high class imbalance, models were tested with imbalanced test sample. The results of deep learning models were compared with random forest, logistic regression, and support vector machine. The results showed, that in all cases, the convolutional neural network was the best model and deep learning models separated classes better than the random forest model, when models were trained with imbalanced dataset. |
