This paper proposes a rating methodology that is based on a non-linear classification method, a support vector machine, and a non-parametric isotonic regression for mapping rating scores into probabilities of default. We also propose a four data set model validation and training procedure that is more appropriate for credit rating data commonly characterised with cyclicality and panel features. Tests on representative data covering fifteen years of quarterly accounts and default events for 10,000 US listed companies confirm superiority of non-linear PD estimation. Our methodology demonstrates the ability to identify companies of diverse credit quality from Aaa to Caa-C.