Applying External Data in Consumer Credit Risk Management: Framework, Evaluation, and Joint Modeling

The presentation introduces the practical application of external data in consumer credit risk management, emphasizing three key points: understanding where and why data is used (strategy/model), knowing data characteristics and suitable scenarios, and evaluating and applying the data effectively.

It is organized into four parts: an overview of the consumer credit risk operating framework and required data types; challenges faced when introducing external data; evaluation methods for external data in risk operations; and how joint modeling can enhance the effectiveness of external data.

The risk operating framework aligns credit risk control with business goals across acquisition, pre‑loan, loan‑in‑progress, and post‑loan stages, each aiming to identify and retain high‑value, low‑risk customers while filtering out high‑risk, low‑value ones. Data is categorized into internal and external sources, with the focus on external data.

External data types are divided into three categories: demand‑related data (frequency, urgency, amount, purpose), performance risk data (basic attributes, financial behavior, non‑financial behavior, public records), and performance ability data (basic attributes, income, assets, liabilities, business information). These data support strategies in acquisition (channel, blacklist, crowd identification, bidding, recall) and pre‑loan (admission, anti‑fraud, quota, pricing).

Key challenges in integrating external data include selection criteria (compliance, stability, interpretability, effectiveness, business benefit), a six‑step integration process (scenario definition, sample preparation, third‑party data retrieval, offline evaluation, online integration, monitoring), and issues such as sample selection, data drift, and effect decay.

Evaluation methods involve assessing data quality, stability, variable discrimination (IV, Lift, KS, AUC), and ROI, followed by small‑scale online experiments before full rollout.

Joint modeling is advocated to improve the performance of external data, especially when only scores are available or sample sizes are limited. Challenges include compliance constraints, limited Y‑labelled samples, early‑stage business data scarcity, and multi‑channel segmentation. Solutions include transfer learning (using source domain weighting, MMD, EM) and clustering modeling to handle heterogeneous customer groups.

The presentation concludes with a discussion on the necessity of segmentation modeling, balancing model management costs against ROI, and practical guidelines for deciding when to adopt joint or segmented models.