Capturing Invariance on Multi-Party Data for Decentralized Credit Scoring

With the rapid development of financial markets, diverse institutions can collaboratively develop credit scoring models. This collaboration can significantly increase the number of available observations for training the model, which can improve the performance of credit scoring model. However, this collaboration faces the problem of privacy leakage and distribution shift. To this end, we propose a novel decentralized method called invariant and collaborative credit scoring (ICS) based on federated learning. Specifically, we design an invariant features extractor to capture invariant features from multi-party data based on invariant risk minimization. Furthermore, we propose a hybrid modeling strategy to predict whether the borrower will default, and lessen the loss discrepancy to enhance the performance of the invariant features extractor. Empirical evaluation shows that the proposed method significantly outperforms the benchmarked methods in terms of both discrimination performance and granting performance. Ablation study further reveals the core performance-enhancing effect of every design artifact underlying ICS. The proposed approach can leverage a larger number of observations from multiple parties securely and effectively to train the model.