Skip to content
RegTech Codex
Directory

Fatf AML KYC Transaction Monitoring

AML Transaction Monitoring Model Governance: SR 11-7 Validation, ATL/BTL Tuning & AI Explainability

Supervisory consent orders routinely strike down "arbitrary threshold adjustments" tuned solely to shrink analyst headcount—requiring independent SR 11-7 conceptual soundness validation, Above/Below-the-Line sampling, and SHAP explainability.

By Victoria Ashworth, LL.M, CAMSReview not recorded
AML Transaction Monitoring Model Governance: SR 11-7 Validation, ATL/BTL Tuning & AI Explainability — RegTech Codex 2026 quantitative & supervisory data card
RegTech Codex (FATF Recommendations, CDD/EDD Beneficial Ownership & SAR Typologies): Independent model validation, ATL/BTL stratified sampling, data-lineage reconciliation, and SHAP ML.Source: RegTech Codex Institutional Research DeskLicense: CC-BY-4.0

Reviewing major BSA/AML enforcement actions issued by the Federal Reserve, OCC, FDIC, NYDFS (Part 504), and the UK FCA reveals two recurring multi-hundred-million-dollar failures: unmapped upstream transaction codes silently omitted from surveillance feeds, and undocumented threshold hikes engineered to keep alert volumes inside analyst headcount budgets. Governed by SR 11-7 (Guidance on Model Risk Management), the 2021 Interagency Statement, and EU DORA ICT & Algorithmic Governance, surveillance teams at firms such as E*TRADE, Coinbase, and CMC Markets enforce three-lines-of-defense model governance. Consult our FATF AML/KYC Compliance Playbook 2026 inside the FATF AML/KYC hub.

Four-Stage SR 11-7 Lifecycle for Rule & Machine-Learning AML Scenarios#

Every scenario—whether a deterministic peer-percentile velocity rule or an XGBoost/Graph-Neural-Network typology scorer—must maintain an auditable Model Inventory Dossier verified by an independent Second-Line Model Risk Management (MRM) unit:

Quantitative Distribution · Fatal Supervisory Red Flag
Fatal Supervisory Red FlagExamination Pass Criterion
012. Peer-Group SegmentationGrouping $5k re…023. ATL & BTL Threshold Tuni…Raising $25k th…
MRM Lifecycle PillarCore Engineering & Statistical DeliverableFatal Supervisory Red FlagExamination Pass Criterion
1. Typology Coverage & Data LineageEnterprise Risk Assessment (EWRA) -> Scenario map + GL-to-TM row/dollar reconNew instant-payment or crypto product launched with unmapped txn_codeAutomated daily 100.00% source-to-TM ledger hash match
2. Peer-Group SegmentationCluster homogeneous peers (Retail vs SME vs Prop Desk vs Institutional CASP)Grouping $5k retail savers inside same percentile distribution as $50M hedge fundsKolmogorov-Smirnov intra-cluster homogeneity + documented features
3. ATL & BTL Threshold TuningSensitivity grid + blinded BTL stratified sampling ([0.75*tau, 0.99*tau])Raising $25k threshold to $100k solely to clear a 12,000-alert backlogUpper 95% Clopper-Pearson bound on BTL productive SARs <= 1.5%
4. Ongoing Drift & SHAP AI AuditMonthly PSI (<0.20), champion-challenger shadow runs & top-5 SHAP driversBlack-box ML probability score (0.87) with zero investigator reason codesHuman-interpretable feature attributions archived per case
SR 11-7 AML transaction monitoring model validation workflow covering data lineage, peer segmentation, ATL/BTL sampling, and PSI monitoring
SR 11-7 AML transaction monitoring model validation workflow covering data lineage, peer segmentation, ATL/BTL sampling, and PSI monitoringVerified Exhibit

Preventing Silent "Unmonitored Product" Disasters With Automated Schema Gates#

Why did multiple Tier-1 banks suffer $100M+ fines after core-banking or brokerage upgrades? Upstream engineering teams added a new transaction type code (e.g., TX_INSTANT_SEPA_V2 or TX_STAKING_UNBOND) without updating the ETL filter feeding the AML engine—leaving billions in volume completely unmonitored for years! Enforcing a daily General-Ledger-to-TM Completeness Control (blocking any unknown txn_type_code into an immediate P1 compliance incident) eliminates unmonitored-rail risk.

Below-the-Line (BTL) Upper-Bound & Population Stability (PSI) Equations#

Independent model validators evaluate proposed threshold hikes using exact binomial upper bounds on blinded BTL samples alongside monthly Population Stability Index (PSI) drift tracking:

Formula or model from the articleNot independently verified
PSI = sum_(b=1..10) ( Actual_Share_b - Baseline_Share_b ) * ln( Actual_Share_b / Baseline_Share_b )
  • SARRateBTLhat = ksarbtl / nsampled_btl
  • EstimatedMissedSARs = Npopulationbtl * Upper95CIClopperPearson(ksarbtl, nsampledbtl)
  • Approval Gate for Threshold Increase (tauold -> taunew):
  • Require ksarbtl / nsampledbtl <= 0.005 AND Upper_95CI <= 0.015 AND 100% Prior Known SARs Retained or Captured by Companion Rule
  • Worked Validation (Proposed Velocity Hike Creating N_btl = 4,000 Suppressed Alerts; n = 250 Blinded Sample Finds k = 0 Missed SARs):
  • Rule-of-3 Upper 95% CI = 3 / 250 = 0.012 (1.2% <= 1.5% ceiling); PSI = 0.07 (< 0.20 stable) -> Approved by MRM Committee
Comparison of alert-to-SAR conversion rates and false-negative capture across legacy rules, tuned peer percentiles, and explainable ML overlays
Comparison of alert-to-SAR conversion rates and false-negative capture across legacy rules, tuned peer percentiles, and explainable ML overlaysVerified Exhibit

NYDFS 23 NYCRR Part 504 Annual Board / Senior Officer Certification#

For New York-chartered or licensed institutions—including global correspondent banks, broker affiliates, and BitLicense virtual-currency firms—23 NYCRR Part 504 elevates transaction-monitoring and sanctions-filtering governance to personal executive accountability: requiring an annual Board of Directors resolution or Senior Officer Compliance Finding backed by documented data-lineage, pre-/post-implementation testing, and threshold-validation workpapers.

Connecting Model Risk Management to Typologies, AI Governance & SAR Quality#

Link SR 11-7 model validation directly to structuring and smurfing bunching models, brokerage mirror-trade and money-pass rules, and FinCEN SAR / goAML XML conversion feedback loops so investigator dispositions continuously refine scenario precision.

Frequently asked questions

Does every deterministic AML rule have to be treated as a formal "Model" under SR 11-7?

The April 2021 Interagency Statement (Fed, FDIC, OCC, FinCEN, NCUA) clarified that simple deterministic single-criterion rules (such as an exact $10,000.01 CTR trigger) need not automatically be classified as statistical "models," though they still require rigorous change-control, data-completeness testing, and periodic threshold validation.

Why did regulators fine banks hundreds of millions for "capacity-driven" threshold tuning?

In multiple landmark OCC, FinCEN, and NYDFS consent orders, institutions raised transaction-monitoring dollar cutoffs specifically to cap monthly alerts at the number their existing analysts could review—suppressing thousands of genuinely suspicious SARs without conducting Below-the-Line impact studies.

How is Below-the-Line (BTL) testing executed statistically when true false negatives are unknown?

Independent validation teams temporarily run scenarios in shadow mode across a 10%–25% lower parameter band (e.g., lowering a $50,000 velocity trigger to $37,500), draw a statistically sized stratified random sample (95% confidence level), and have senior QA investigators review those near-miss cases blind to measure missed-SAR rates.

Can machine learning replace AML rules entirely, or how do supervisors expect AI to be governed?

FinCEN/Interagency AI Innovation statements, UK FCA, and EU DORA/AI Act frameworks permit hybrid ML anomaly scoring or alert-prioritization models provided firms maintain champion-challenger benchmarks, Population Stability Index (PSI) drift monitors, zero protected-class bias, and human-readable SHAP explanations.

Publication record

Published
14 November 2026
Updated
14 November 2026
Author
Victoria Ashworth, LL.M, CAMS
Review
Review not recorded
Automated fact check
Passed in the stored pipeline record
Source URLs
3 recorded

AI assistance recorded: anthropic/claude-opus-4.6. This field does not establish human review or source verification.

Academic & Institutional Attribution

Cite This Research & Embed Live Data Card

Open Attribution License (CC BY 4.0)

Standard academic & institutional research paper citation format.

Victoria Ashworth, LL.M, CAMS. (2026, November 14). AML Transaction Monitoring Model Governance: SR 11-7 Validation, ATL/BTL Tuning & AI Explainability. RegTech Codex. https://regtechcodex.org/briefings/aml-transaction-monitoring-model-validation-sr-11-7-backtesting
Explore Fatf AML KYC Transaction Monitoring Hub (20) →

RegTech Codex Data, Benchmarks & Contributor Records

AML Model Governance: SR 11-7 & ATL/BTL Tuning (2026) | RegTech Codex