Credit Scoring Based on Transaction Sequence Classification

The growing volume of credit card and personal loan applications has spurred the need for financial institutions to improve their credit scoring methods to make intelligent decisions. In this paper, we introduce a novel credit scoring framework based on transaction sequence classification. This framework involves a CNN-based structure for sequential feature extraction. It can be applied to both raw transaction sequences and feature matrixes aggregated by time window to capture short-term and long-term sequential features. Other non-sequential personal features are also integrated to the model to help making the final decision. In the experiments, evaluations on the performance of model-based, feature-based sequence classification and the whole framework are done respectively. Our sequence classification method as well as the credit scoring framework outperforms other state-of-art methods in the experiments on real-world labelled application and transaction data from a major payment organization.