Credit rating is to differentiate the risk of different customers． However，most existing methods either fail to guarantee that the credit rating meets the criteria “the higher the credit rating，the lower the loss rate”，or fail to differentiate the customers with different default possibilities to the greatest possible degree．Therefore，the existing credit rating method cannot be used as an effective tool for loan decisions． This study gives credit grades according to the above criteria． The innovations are: firstly，credit grades are divided by taking the maximum difference between the cumulative frequency of non-defaulting customers and that of defaulted customers as the objective function and taking“the higher the credit rating，the lower loss rate”as the main constraint，so that the credit grades meet the above dual criteria; Secondly，a new algorithm for credit rating by setting the intervals of random segmentation points is proposed． The algorithm avoids that problem that a front division point may be designated to customers in the back，resulting in the failure of grading customers in the back． Finally，this study uses 3 045 small businesses from a Chinese bank for empirical study，and the results show that the credit grades thus derived meet the above dual criteria．