Justifying adverse actions with new scorecard technologies

It has been argued that flexible classification models such as neural networks, support vector machines, and random forests face resistance as credit scoring models because it is difficult to identify which characteristics contribute substantially to the overall scores. In fact, however, this is a misunderstanding arising from the fact that standard models are based on sums of transformations of the raw characteristics. We distinguish between the need to identify which characteristics contribute most to an individual‟s score and the need to identify which characteristics contribute to the performance of a scorecard. We describe solutions to these two problems, and illustrate by applying a range of scorecard approaches to some real credit card data.