Credit Scoring and Default Prediction Using Pufferfish Optimization Algorithm

This work develops a novel hybrid deep learning model for credit scoring and default prediction. Initially, the dataset is directed into data cleaning and scaling techniques to eliminate noise and artifacts inherent in data. The resultant pre-processed dataset is fed to feature fusion phase, which is executed by a Multi-Layer Perceptron (MLP), trained through the application of the Pufferfish Optimization Algorithm (POA). Ultimately, the selected features are processed in the credit scoring and default prediction phase. During the prediction phase, the forecasting is conducted utilizing the proposed SDBFN-GRU, which represents an integration of Sparse Deep Belief Network with Fuzzy Neural Network (SDBFNN) and Gated Recurrent Unit (GRU). Finally, the results of experiments demonstrate that the proposed SDBFN-GRU methodology demonstrated superior performance, which attained an accuracy of 95.8%, a sensitivity of 96.2%, and a specificity of 94.6%, compared to conventional models.