The Effect of Imbalanced Data and Parameter Selection via Genetic Algorithm Long Short-Term Memory (LSTM) for Financial Distress Prediction.

Financial companies are grappling with a burning issue about bankruptcy prediction. There are many methods for bankruptcy prediction, including statistical models and machine learning. Real-life datasets are often imbalanced with high dimensionality. Therefore, it is challenging to train a robust model to predict bankruptcy. Thus, we first applied an oversampling technique known as the Synthetic Minority Oversampling Technique (SMOTE) to reduce the skewness of the data. The balanced data was trained with the baseline models, the ensemble classifiers using different combination methods and the long short-term memory (LSTM) model. In addition, we employed an optimization technique called a genetic algorithm (GA) to optimize and determine the learning parameters of an LSTM network. We further determine the effects of using different training/testing ratios on the developed models. An autoencoder long short-term memory (LSTM) model was developed to extract the best feature representation of the input data. A comparative analysis was carried out between the LSTM-GA and autoencoder-LSTM. The results show that the improved LSTM-GA model with an accuracy of 98.11% performs better than other models. Overall, the research work concluded that all models and LSTM have good performances, while the optimized LSTM model via genetic algorithm outperforms the classical machine learning models. [ABSTRACT FROM AUTHOR]