Credit Risk Strategies: Data, Rules, and Model Development for Consumer Lending

The speaker, senior risk‑control director at Baorong Cloud, introduces the topic of credit‑lending risk‑control strategies, outlining the evolution from simple rule‑based decisions to data‑driven and model‑based approaches.

Industry Background : The consumer credit market has shifted from rapid growth to a period of consumption downgrade, slower retail sales, higher household debt ratios, and rising non‑performing loan rates, making cost‑effective risk control essential.

Traditional Scoring‑Card Development Process :

Goal definition – clarify good vs. bad customers.

Data integration and preprocessing – collect personal identifiers, transaction records, clean errors, handle missing/outlier values, and engineer features.

Feature selection, model tuning, and evaluation – use statistical significance, IV, clustering, binning, KS, VIF, PSI, etc.

Scoring and deployment – assign scores per bin to build an interpretable scoring card.

Machine‑Learning Model Development Process : Compared with scoring cards, ML reduces manual intervention, has lower interpretability, and focuses on parameter tuning to avoid over‑fitting. The workflow includes data preparation, model training, validation (KS, PSI), and deployment.

Pre‑loan Risk‑Control Flow and Strategies :

Risk‑control flow design – identify fraud, high‑risk users, reduce risk while lowering cost and improving efficiency.

Rate strategy – calculate expected return r = A·(1‑p) where A is credit limit, p is default rate; typically a fixed rate is applied with a score threshold for rejection.

Limit strategy – determine credit‑limit intervals [A₁, A₂] based on risk‑return trade‑offs; use sigmoid functions to replace step functions for smoother risk‑limit mapping.

Rule Effectiveness Diagnosis : Score rejected customers, compare score distributions of approved vs. rejected groups, identify abnormal rules, group rules for comparison, and adjust accordingly.

Model Optimization :

Address sample‑masking by incorporating rejected samples into training via proportionate allocation, simple enhancement, or parcelling methods.

Iteratively build model 1 on approved data, score all customers, remove high‑score rejected samples, apply parcelling on low‑score rejected samples, and train model 2; repeat to improve performance.

The presentation concludes with a summary of the discussed risk‑control workflow, rate and limit strategies, rule diagnostics, and model‑optimization techniques, emphasizing the importance of aligning risk strategies with business goals and data constraints.