Financial Risk Management: Business Requirements and Technical Solutions

Speaker: Su Xiaolin, Data Platform Architect, shared insights at DataFun AI+ Talk.

Business Perspective: Before any technical implementation, understanding the business is crucial. The talk highlighted risk‑control problems such as identity authentication (ID cards, on‑site photos, account systems, facial recognition, live verification) and fraud types (fake data, third‑party fraud, self‑fraud, account theft). These issues affect core financial services like wealth management, transfers, payments, consumption, and especially credit, where risk quantification is vital.

Risk‑Control Toolbox: The speaker described four main weapon categories: (1) Internal data organization – leveraging bank and internet data, cross‑departmental features, and company‑wide data sharing; (2) Feature mining – expert pattern expansion, feature engineering, and data‑point derivation; (3) External data integration – API connections and data partnerships; (4) Data expansion – enriching datasets for better modeling.

Technical Perspective: The presentation covered feature construction logic, emphasizing embedding to transform high‑dimensional sparse data into interpretable low‑dimensional features. Feature evaluation metrics were shown for both new product launches and model iterations. Non‑linear features are converted to linear ones using GBDT or Random Forest, producing decision‑tree leaf nodes that represent user portrait aspects. A reference to an ACM paper (https://dl.acm.org/citation.cfm?id=2648589) was provided.

Sample Accumulation Process: Three stages of sample definition were described: early stage (few samples, many features, short cycles), middle stage (moderate samples, many features), and later stage (abundant high‑quality samples). Corresponding methods include sample replacement, partitioning, cleaning, and using DPD trends or collection strategies. Sample partitioning strategies (time‑based, random, rule‑based) evolve from Leave‑One‑Out to 5‑fold and finally OOT (out‑of‑time) validation.

Model Optimization: Simpler models are recommended despite complex business contexts. Evaluation metrics such as AUC (overall performance), KS (cut‑off relevance), and GINI (equivalent to AUC) were discussed, along with score mapping examples.

Online Validation & Evaluation: Validation includes algorithmic metrics (feature importance, score distribution, business logic) and business metrics (approval rate vs. default rate, credit limit and utilization changes, GMV impact). Model review checks KS stability across Train/Test/OOT, PSI drift, and swap‑in/out behavior.

Deployment Pitfalls & Recommendations: Common issues include feature leakage, drift, interruption, and large‑scale real‑time computation. The speaker advised a three‑step rollout: offline verification, online comparison, and small‑traffic validation before full deployment.

Resources: A PPT download is available via the DataFun community; the author’s bio notes extensive experience in internet finance risk modeling and data science leadership.