Prediction Model Of Housing Mortgage Loan Prepayment Risk Based On LightGBM Algorithm

Since the rapid development of information technology in the past few decades,machine learning algorithms are now widely used in multiple fields,including the financial industry.Compared with traditional statistical modeling techniques,machine learning algorithms are more flexible,compatible,and much more powerful in dealing with big data,and has become increasingly popular in the industry.However,as the structures of machine learning models are relatively complex,it is hard for many users,such as clients lacking technical backgrounds or regulators emphasizing the interpretability of the models,to accept them,which restricts the application of these techniques in the industry.Prepayment risk is the main risk of the callable fixed-income securities,such as mortgages and MBS.Traditionally,Logistic Regression is the most widely used tool in prepayment risk modeling.Now,with the new resources of big data and the improvement of data-mining techniques,we can try to find a new method by implementing machine learning techniques,which can give consideration to both prediction accuracy and interpretability.Based on the data of the 30-year fixed-rate mortgages originated in North Carolina between 2004 to 2015,we first construct the objective variable,which represents whether or not the mortgager fully prepaid the loan,and a14-feature input dataset,including information of mortgage origination,monthly performance records,and macroeconomics.By using the LightGBM algorithm and Logistic Regression,we build two models to predict the prepayment activities and forecast the prepayment rate both on the loan-level and pool-level data.To enhance the interpretability of the LightGBM model,we also use the Accumulated Local Effects Plot,which depicts the margin contribution of the features.Empirical evidence shows that the LightGBM model has better prediction accuracy,compared with the LR model,by capturing both linear and non-linear relationships between the objective variable and the features.