A deep learning model-based approach to financial risk assessment and prediction

Abstract

This paper proposes a split-lending network model for bank credit risk, calculates whether a bank fails by simulating the changes in bank assets and liabilities over time, and adds the default rate calculation considering the characteristics of the default rate when a bank borrows. Meanwhile, the XGBoost-based classifier is used instead of random forest to improve the accuracy of classification, and the grcForest_XGB model is established. The activation function Sigmoid, the error function mean square error function, and the adam optimization function with the best effect at present are used to predict the accuracy of the grcForest_XGB model, and the different models are compared with the grcForest model for comparison experiments. The experiments show that the grcForest model has higher AUC and KS metrics of 0.8224 and 0.6368, respectively. The recall value is 0.8319, which ranks first among the six models. The Acc value is 0.9732, which is only 0.05 lower than LSTM, and is at a higher level. This study shows that the model is more accurate for risk assessment, can predict financial risks in advance, and make an effective assessment of financial risks.

Keywords

• financial risk
• unbundling network model
• grcforest model
• optimization function
• activation function
• 68t05