Research On Loan Forecasting Algorithms Based On Expected Loss Cost Sensitive Optimization

In recent years,credit risk is gradually rising with the lending's growing market share and the loan products have been enriched,which have penetrated into every aspect of daily consumption and operation.The extreme large,complex and unbalanced characteristics of data makes it difficult to improve the accuracy of lending forecasting.Some features have the phenomenon of distribution drift.Therefore,this paper made use of loan data from Lending Club to reduce the default ratio through solving the unbalance problem and eliminating distribution of drift,which on the other hand help to pick good loans and keep lending market sustainable.Data preprocessing and feature construction are carried out on the loan data from the four dimensions of personal information,loan information,bank account information and credit information to enrich the description ability of the model.This paper introduces dynamic cost sensitive item into the ensemble model to reduce the influence of unbalanced training samples on the model.Innovatively,the expected loss of loan is proposed as the cost sensitive item.The core is adding the loss object into each iterative learning process of the XGBoost and LightGBM algorithms to improve the prediction accuracy and enhance the detection of default categories.The optimized algorithm is named ES-XGB and ES-LGB.This paper introduces the PCE macroeconomic data as the time recession coefficient to eliminate the temporal trend.Finally,this paper gives the loan forecast system flow chart to standardize the lending forecasting process.Numerical experiments show that the algorithm models ES-XGB and ES-LGB proposed in this paper have the best performance in loan prediction compared with other integrated models.Monthly PCE data can effectively eliminate the tendency of features and improve the prediction ability of loan prediction model.It is proved that the sensitivity of expected loss cost and the optimization of macroeconomic statistics are applicable to the lending forecasting scenario.