Intelligent Risk Control Platform: Design Principles, Strategy and Model Lifecycle Management, and Architecture

The talk introduces an intelligent risk control platform that serves business needs by building a risk‑control middle‑office centered on real‑time, fine‑grained, joint, agile, intelligent, and unified capabilities.

It proposes the "five‑full double‑core" model: full coverage across subsidiaries, business lines, channels, and processes, combined with a dual core of rule‑engine and AI‑engine decision flow, supported by atomic data and modular architecture.

Strategy full‑cycle management includes configuration (rules, scorecards, decision tables), version control, AB testing, and continuous monitoring of business and statistical metrics to ensure effectiveness.

Model full‑cycle management covers data ingestion, feature engineering, AutoML (automatic table joining, feature generation, algorithm selection, hyper‑parameter tuning), and both batch and online learning to address model drift.

The business architecture is divided into three layers: data layer (integrating internal transaction data with external credit data using various databases), platform layer (metric center, decision service center, strategy and model centers, graph computation), and application layer (supporting anti‑fraud and other use cases).

Capability atomization is further broken down into infrastructure (storage, container scheduling), service (data engines, offline/online services, AutoML model management, data quality), and business modules (data center, strategy studio, scenario center, monitoring dashboards).

Two case studies illustrate the platform's impact: a hard‑real‑time transaction fraud detection system achieving sub‑10 ms latency with a hybrid rule‑AI engine, and a nationwide bank’s omni‑channel anti‑fraud solution handling over 500 rules and 2,000 real‑time metrics, intercepting nearly 10,000 high‑risk transactions per month.

The presentation concludes with thanks and a call for audience engagement.