Introduction Of Reject Inference Small Business Credit Scoring Models

As a creative financial strategy, small business credit scoring is widly used in US, but receives far less attention in our country. It has been indicated that small business credit scoring is helpful to improve the credit availability for small businesses and contributes to resolve their financing difficulty.Compared with traditional lending techniques, credit scoring realizes handling loans automatically to a large degree and thus improves the efficiency of approving and screening loan applicants, which helps the bank price these loans more exactly based on credit risk involved. Large quantities of studies indicate that credit scoring works well in small business lending and can improve small business's credit availability.This paper first defines and elaborates on the three key concepts in this study, which are small business, credit scoring and reject inference. After the comparison between traditional credit evaluating method, modern credit risk quantization model and credit scoring, deep analysis is extended on the usability of credit scoring in evaluating credit risk of small business. We establish the variable system and extend the main factor analysis on the financial variables. Thereafter, we build a credit scoring model based on logistic regression by use of the first sample. Then, we apply this model in the second sample and simulate credit screening just like a real bank. Thus we get a new sample with missing data, and based on this new sample we establish three new models, the first is Heckman two-step sample selection model, which is one kind of reject inference method, the second is censored model which ignores all the missing data, and the last is ideal model with complete data. By comparing these three models'performances through ROC curve, we conclude that the Heckman two-step sample selection model outperforms the censored model and gets much closer with the ideal model, which indicates that reject inference is of great importance in the credit scoring modeling to relieve the sample selection problem due to missing data.