Downloaded 39 times
Uploaded byNeo4j
Banking Circle: Money Laundering Beware: A Modern Approach to AML with Machine Learning and Graphs
The document discusses Banking Circle's use of graph technology and a data-driven approach to improve its anti-money laundering efforts. It represents payment data as a network to extract features for machine learning models that detect suspicious activity. This approach generates fewer false alarms than rules-based systems while identifying more high-risk payments and accounts. Network-based investigations also help analysts explore connections more efficiently. The new system screens over 1 million payments daily and has increased alerts leading to compliance actions by 1300% while reducing total alerts by 30%.
More Related Content
Similar to Banking Circle: Money Laundering Beware: A Modern Approach to AML with Machine Learning and Graphs
![[DSC Europe 22] Anti-Money Laundering ML Modeling approach - Gizem Akar](https://cdn.slidesharecdn.com/ss_thumbnails/gizemakarantimoneylaunderingmlmodel-221130080718-afd93383-thumbnail.jpg?width=640&height=640&fit=bounds)
Banking Circle: Money Laundering Beware: A Modern Approach to AML with Machine Learning and Graphs
- 1. BANKING CIRCLE Advanced Analytics andgraphs in AML Banking Circle Financial Crime and Graph Technology
- 2. 2 Confidential Our vision isto build a global account-to-account payments network and related financial infrastructure that is simple, fast and low-cost, with world-class levels of security and compliance, for all major businesses 2 Executive summary • Banking Circle is a payments bank, delivering payments and banking services by connecting to the world’s clearing systems. • This requires a more performant approach to Anti money laundering (AML) than customer facing banks. • A data driven approach to AML has significantly improved BC’s ability to detect potentially suspicious flow • A network representation of our payment data is a core part of this data driven approach • Embeddings improve machine learning models • Visualisations aid manual investigations
- 3. 3 Confidential Banking Circle Financial Crimeand Graph Technology
- 4. 4 Confidential Scale Financial Crime andGraph Technology
- 5. 5 Confidential What does thismean for AML at Banking Circle • Lots of payments (data) • Different data to other banks / firms in the system • Traditional approach of throwing bodies at the problem won’t work • We need to use the data to make better decisions.
- 6. 6 A data drivenapproach in AML can help mitigate some of the issues in “traditional” systems • Traditional AML approach ‒ Based on rules: Manually tuned decision tree - > forest of very shallow trees • E.g: Catch certain words, amounts or locations. Attempts at catching patterns/behaviours ‒ (Very) high false positive rates • Data driven approach: • Kill/adjust non-performing models based on continuous testing and evaluation • Complex ensemble ML models ‒ Gradient boosted trees, neural networks, natural language processing Enrich with new data representations Augment with statistics based features: ‒ Anomaly detection, pattern descriptions Incorporate external data sets
- 7. 7 SCAM – Systemfor Catching Atempted Moneylaundering – an ensemble of different machine learning models New Payments Static Rules + NLP Supervised Models Entered into Hitlist DB w. Explanation Diversification Models Manual Evaluation Feedback from Monitoring Team Hit 1 Hit 2 Hit 3 Hit 4 Hit 5 Hit 6 Entity A Hit 1 Hit 2 Hit 3 Hit 4 Hit 5 Hit 6 Entity A Hit 1 Payment 2 Hit 3 Hit 4 Payment 5 Hit 6 Entity A Hit 1 Payment 2 Hit 3 Hit 4 Payment 5 Payment 6 Entity A Hit 1 Payment 2 Payment 3 Payment 4 Payment 5 Payment 6 Payment 1 Payment 2 Payment 3 Payment 4 Payment 5 Payment 6 Unsupervised models – do we have payments coming through that we can learn from ‒ Decision tree based models ‒ Statistical measures of similarity ‒ NLP Will generate new hits and reduce bias Will help label payments not found by rules or supervised methods Supervised models – what can we learn from already evaluated payments ‒ Decision tree based models ‒ Deep neural network based models Will score each payment (and account/client) and constantly improve performance as more payments are manually evaluated
- 8. 8 Core risk scoringmodel Payment Not suspicious Potentially suspicious 0.8 0.2 The “best” model minimizes a combination of error rates and complexity on a historical data set Probability Payment details Payment and monitoring history Data related to payment Payment patterns and connections
- 9. 9 SCAM relies ona network representation to capture connections Query • Entities (owners, directors etc) • Accounts • Countries related to • Payment • Account • Entity within 1,2,…, n links Graph interaction used for feature extraction
- 10. 10 Network based featuregeneration Basic properties that are difficult to get from “traditional” DBs ‒ Number of connections for an account or entity ‒ Names of all beneficiaries, remitters and their locations ‒ Distance to tax havens or known crooks. ‒ Number of addresses used or associated directors ‒ etc Advanced feature generation (WIP) ‒ Graph properties generate features • Connectedness and centrality vary between business types of customers • Color propagation – diffuse risk through a network ‒ Embedded subgraph to generate features ‒ Community detection ‒ ML Anomaly detection as feature input into ML models Results in an AUC improvement of 0.02 points ‒ 10% -25% reduction in false negatives, no change in false positives ‒ Halving of overall alerts generated
- 11. 11 Confidential AI on connecteddata has led to better AML decisions while reducing operational costs • AML system based on: • Rules (prevent egg in face, compliance with regulation) • machine learning running on top of network data to mitigate the deficiencies of a purely rule based system and catch new behaviour rather than known typologies fitting predefined static rules. • Daily AML screening of all processed payments through same system. • AI is the only way to future proof an AML model in real time whilst constantly increasing the value of all data in the ecosystem. Operational scaling
- 12. 12 Confidential AI on connecteddata has led to better AML decisions while reducing operational costs Expose known unknowns and increase holistic detection based on connections and enriched data • 50 % accounts closed or escalated to compliance now due to pure AI related AML findings. • 10+ % AI of alerts escalated vs few % for traditional rules. • External requests does not lead to additional tp’s In-house development by cross-functional team has allowed BC to • Learn from already evaluated transaction history. • Be adaptable and continuously identify payment patterns not covered by a static rule set. A data driven approach to AML has allowed BC to highlight fewer non-risky payments, generate less false positives, thus more time to investigate suspicious payments: • Number of payments has increased ~200% ; the number of generated AML alerts has decreased ~30%. • Number of AML alerts leading to compliance handling or account closure have increased ~1300%. Currently screening over 1.000.000 payments / day
- 13. 13 Confidential A closer lookat the implementation • A streaming setup based on Azure service bus and Kubernetes • Some data is cached on a daily basis, but most is computed live • Graph nodes and connections are added for each payment • Features are computed on the graph • All computation done in cypher • Limited to methods available Neo4j
- 14. 14 Graphs in antimoney laundering investigations Empowering investigations Investigation of known associates and connections ‒ New view onto the same data ‒ Linking diverse data sources in one view, complements “tabular” reports ‒ Natural representation for criminal networks / organized crime ‒ Very effective – some analysis is cut from days to minutes
- 15. 15 Investigation tool allowseasy access to full data model Empowering investigations Investigate individual entities, accounts or transfers Resulted in open source neographviz python package
- 16. 16 Investigation tool allowseasy access to full data model Empowering investigations Investigation of connections
- 17. 17 Take away Adata driven approach to AML has significantly improved BC’s ability to detect potentially suspicious flow A network representation of our payment data is a core part of this data driven approach ‒ Machine learning models ‒ Manual investigations
- 18. 18 Thank you foryour attention Contact: Ruben Menke rum@bankingcircle.com