Real‑Time Metrics Processing Technology for Financial Risk Control and Anti‑Fraud

Speaker: Yin Hang, Senior Technical Expert at Bangsheng Technology

Organizer: DataFun Community

This sharing session is divided into three main parts.

Part 01 – Risk‑Control Challenges in the Internet Era

1.1 Data leakage and rampant fraud: personal data such as passwords and bank card numbers are widely sold; black‑market actors exploit leaked data, making real‑time risk control essential.

1.2 Challenges faced by existing financial risk‑control systems: complex rule sets, reliance on post‑event processing, difficulty in real‑time parameter adjustment, and the need to integrate AI, machine learning, graph analytics, and device fingerprinting.

1.3 Opportunities and challenges of the big‑data era: shift from batch statistical analysis to real‑time analysis and decision‑making, requiring low‑latency processing.

1.4 Real‑time data processing (batch + streaming): distinguishing hot, warm, and cold data, and leveraging hot data for immediate decision‑making.

Part 02 – Real‑Time Metric Processing Technical System

2.1 Evolution and goals of real‑time computing: from simple low‑latency databases to Spark Streaming, Flink, and stateful stream processing achieving millisecond‑level response.

2.2 Indicator system for financial‑customer‑behavior management: defining and weighting various fraud‑related metrics for credit, transaction, login, and other behaviors.

2.3 Business scenarios and technical challenges of real‑time indicators: real‑time fraud detection within 100 ms, real‑time marketing, instant credit approval, and high‑throughput network bot detection.

2.4 Characteristics of real‑time indicator computation: long‑term, multi‑dimensional, massive data volumes requiring dynamic, time‑windowed calculations.

2.5 Computation models for real‑time indicators: distributed caching, patented time‑series handling, and big‑data model management achieving 13‑17 ms average latency.

2.6 Urgent need for AI models to be deployed in real time: rule‑based, statistical machine‑learning, and cyber‑netic approaches using random forests, GBDT, neural networks, etc.

2.7 Real‑time intelligent big‑data processing: end‑to‑end pipeline of real‑time collection, cleaning, analysis, and decision‑making.

2.8 Fast online of machine‑learning models: streaming cube (PipeACE) ensures real‑time feature availability for decision engines.

2.9 Full‑stack machine‑learning platform: integrated data collection, training, and application with multiple algorithms (GBDT, Random Forest, etc.).

2.10 Knowledge‑graph platform: real‑time definition of nodes, edges, and relationships.

Part 03 – Technical Features of Bangsheng’s Real‑Time Indicator Platform

3.1 Product and solution overview: focused on financial anti‑fraud and internet credit, covering strategy, consulting, modeling, data, and platform.

3.2 Understanding and practice of financial risk‑control system construction: device fingerprinting, black‑phone data, streaming cube, PipeACE, graph analytics, and machine‑learning modeling.

3.3 Position of the real‑time indicator platform within the overall architecture: small but critical component that stores incremental results and interacts with decision engines.

3.4 Interaction with related systems: storing intermediate results and feeding them to real‑time engines.

3.5 Real‑time indicator calculation in complex business scenarios: multi‑source data ingestion, time‑windowed rules (e.g., transaction count over the past 10 days), and dynamic threshold evaluation.

3.6 Core technology of the streaming cube: distributed cache, time‑series storage, and big‑data model management delivering sub‑millisecond latency.

3.7 Real‑time data processing capabilities: linear scalability, high concurrency, high availability, and loose coupling of modules.

3.8 Overall architecture of the real‑time indicator platform: proxy‑based data collection (Micro‑Probe), streaming processing, and integration with decision engines.

3.9 Functional architecture diagram (illustrated in the original slides).

3.10 Application case: transaction anti‑fraud comparison between systems with and without risk‑control.

Conclusion

Bangsheng’s real‑time indicator platform offers ultra‑low latency, rich built‑in algorithms, powerful data processing, and high reliability to meet demanding business requirements.