Applying AI Techniques to Credit Reporting and Risk Modeling

The presentation, delivered by a senior algorithm expert from Du Xiaoman Financial and organized by DataFunSummit, explores the application of AI techniques to credit reporting and risk modeling in the financial sector.

Credit data, issued by the People's Bank of China, includes personal basic information, loan transaction details, non‑loan credit information, and query records, forming a comprehensive and nationally unified credit system that is essential for most financial institutions.

Credit modeling heavily relies on this data; traditional scorecard models use expert‑crafted features for interpretability, while more complex models aim to extract richer features or employ end‑to‑end approaches, with the latter often achieving the best performance.

Model 1 addresses the semi‑structured nature of credit reports by integrating numerical and categorical features with self‑attention and cross‑attention mechanisms across loan and credit‑card sequences, and applies two‑layer multi‑head attention for textual fields, favoring a shallow transformer due to sparse text.

Model 2 enhances temporal trend capture by encoding loan and credit‑card sequences separately, then concatenating them into a unified sequence and applying session‑level sequence modeling to learn borrowing patterns over time.

Model 3 further incorporates repayment sub‑sequences nested within loan timelines, combining repayment trends with basic loan information to strengthen the model’s ability to predict repayment behavior.

Model 4 leverages graph neural networks to build association networks for textual entities (e.g., addresses, company names) in credit reports, enabling the model to infer risk signals from rare or long‑tail named entities.

For pre‑training, a BERT‑style masking approach is adapted to credit reports, but due to strong local correlations, a hierarchical softmax over discretized feature combinations is introduced, yielding substantial performance gains over non‑pre‑trained baselines.

The risk ranking model shifts the optimization target from probability estimation to ranking metrics such as AUC and KS, using sorting objectives and knowledge‑distillation techniques (e.g., pandemic‑related risk) to improve short‑term delinquency discrimination.

Interpretability of complex models is addressed with Integrated Gradients and SHAP methods, explaining feature contributions while acknowledging challenges like baseline selection and computational cost.

The Q&A session highlights practical encoding strategies for textual fields, categorical codes, sample definition (one credit report per user), and network designs for mixed state and behavior data.

The speaker concludes the session with thanks to the audience.