Research on Dynamic Credit Risk Prediction in Commercial Banks Based on a Hybrid Scorecard-LSTM-Attention Model

Full abstract

Read the full abstract

With the increasing fluctuations in the global economy and the development of digital finance in China, commercial banks are facing two major challenges: the declining quality of credit assets and the growth of bad loans. Traditional credit scoring models based on Logistic Regression are easy to explain, but they struggle to handle complex data that is high-dimensional or changes over time. To improve the accuracy of credit risk prediction, this study develops a hybrid model called Scorecard-LSTM. This model combines traditional credit scoring methods with LSTM networks, also added an Attention Mechanism to help the model focus more on key customer behaviors. The study uses data from a commercial bank in Taiwan, including basic customer information, financial status, and repayment history. By using methods like WOE encoding, feature binning, and joint training, the model connects the clear logic of traditional scorecards with the powerful learning ability of deep learning. The experimental results show that this hybrid model is much better than traditional models in both accuracy and stability. With an AUC value of 0.7741, the model performs very well in identifying bad loans and providing early warnings for potential defaults.