Application Of Multi-task Learning In Credit Risk Modeling

Basel ? incorporates Loss Given Default(LGD)into measurement framework for calculating bank regulatory capital,so that LGD becomes an important research topic in the field of credit risk management.At the same time,the accurate prediction of the LGD is important to the calculation of regulatory capital for the bank to ensure capital adequacy and maintain a healthy financial system.This paper adopts the Multi-Task Learning(MTL)method,hoping to improve the accuracy of LGD prediction.This article selects the 2013-2016 loan data provided by the official website of the well-known online lending platform Lending Club in the United States,including the identity information,financial information,and loan information of the borrower,to learn the Loss Given Default and the Probability of Default(PD).Only the data of the Charged Off and Fully Paid loan status is retained,and samples of intermediate transition states such as Current and Late are excluded.The sample of Charged Off is defined as the positive sample and Fully Paid is defined as the negative sample for the PD prediction task.And the Loss Given Default is calculated as the target variable of the LGD prediction task.This paper focuses on the problem of data scarcity and limited existing measurement models in the prediction of the LGD task,and proposes to use MultiTask Learning methods for the prediction of LGD and PD simultaneously,hoping to improve the prediction accuracy of the LGD task.This paper establishes an MTL model for hard parameter sharing of neural networks,and compares the effect of the MTL model on the prediction of LGD with other single-task models such as linear regression,logistic regression,and single-task neural network.This paper uses L1 regularization to complete the preliminary feature selection,and explores the influence of the weight ratio of the two loss functions on the model evaluation results in the MTL neural network.The result shows that the introduction of the PD auxiliary task and using MTL neural network model can improve the prediction accuracy of the LGD task.The two tasks are trained at the same time,so that the LGD task can use more information.One is the large data of the Fully Paid samples,which cannot be used in a single task.The second is the correlation that contains a wealth of information between the two tasks.