Neural Network-based Personal Credit Risk Assessment Model

In recent years,the threshold of financial services has been lowered,and personal microfinance business prevails.This poses a huge challenge to the risk control capabilities of financial institutions.The emergence of bad debts will bring irretrievable damage to financial institutions,and the large number of personal loan applications makes the manual review almost impossible.The personal credit risk assessment model studied in this paper can judge the future overdue risk of users before lending,thus assisting financial institutions in lending decisions.Since China's credit information system is not yet perfect,most of the credit risk assessment models can only rely on user-related data accumulated within financial institutions.The model of this paper is mainly based on the user's transaction records and attribute characteristics.For the transaction sequence,this paper explores two representation methods,which are chronological sequence of events and statistical feature matrix organized by time window.The former is the direct representation of the original features of the transaction,while the latter is filled with aggregated features.Then,the corresponding neural networkbased feature extraction structures are designed for the two transaction representations.Bidirectional Gated Recurrent Units(GRU)is used to extract features from the event sequence while convolution neural network(CNN)with multi-size kernels is utilized to do the feature extraction of the transaction matrix.Finally,the neural network is used to combine the features extracted from both representations with the overall statistical characteristics of the transaction and the user attribute characteristics to establish the final personal credit risk assessment model.This paper also conducts an empirical study of the proposed model on a real data set provided by a payment company.For the two representations of the transaction,the sequence classification experiments of different models are carried out respectively,which shows the advantage of feature extraction of bidirectional GRU and multi-size convolution kernel CNN on two representations.For the model based on the proposed feature fusion method,the F1 score reaches 0.71 and the AUC reaches 0.90,which is much better than random forest.