CXO Bytes

Reshaping anti-money laundering in the machine learning era


Anti-Money Laundering (AML) compliance becomes knottier for financial institutions as perpetrators adopt more tortuous implements to dodge regulations.

From old-school smurfing, currency exchanges, and “mules”, to P2P transfers, online auctions, and cryptos, money laundering remains a shapeshifting nemesis that guzzles in trillions. UNODC had estimated ~US$1.6 trillion as the globally laundered sum as early as 2009.

In the FinTech era, digital products like e-wallets, cryptocurrencies, etc., afford criminals access to wider and subtler vulnerabilities, and their exploits of the labyrinthine money networks continue to mount.

Costs of AML efforts  

AML regulations like BSA, U.S. PATRIOT ACT, EU-Anti- Money Laundering Directives (AMLD), etc., obligate banks to detect financial frauds, screen customers, track transactions, and report anomalies (SAR/STR).

While banks continue to upgrade their domain expertise and AML technology to curtail money laundering, their annual upfront cost of financial crime compliance snowballs to exorbitant scales, reportedly upwards of $200 billion.

Traditional AML software drives up this cost due to excessive “false alerts” that go up to 90%; meaning 9 out of 10 alerts can be falsely tagged as “suspicious” resulting in undue investigative efforts, time, and costs.

Additionally, hefty legal penalties, revoked licenses, and reputational dents are looming risks on account of inaccurate/late filings, undetected transactions, sanction violations, KYC lapses, etc.

For instance, screening blocked entities, SDNs, and PEPs amid a sea of 24,000+ estimated global sanctions (as of 2022) can stress even the most equipped banks.

Tapping algorithmic intelligence with machine learning

Traditional “rule-based” AML systems use Boolean logic to label transactions and patterns based on a threshold value or binary True/False. Also, they cannot adapt to address real-world complexities.

Technologies like machine learning offer a significant advantage by tapping into algorithms to generate AML models that emulate anthropomorphic (autonomous) learning behavior. These machine-learning models can learn to detect, analyze, and label vast financial transactions.

Machine learning techniques perform algorithmic data analysis over datasets – transactions, behavioral patterns, KYC information, etc. – to train AML models that can detect patterns and predict outcomes with higher accuracy.

Trained with more data volume and variety, the AML model eventually gains higher precision that can rival human proficiency while performing faster.

Unlike logic/rule-based AML systems, machine- learning-based solutions can detect latent patterns to curb the broadening gambit of money laundering.

Pinpointing transactions with data-driven learning

Machine learning does an incredible backend heavy lifting when all might appear done and dusted on the front end.

And it sums up to autonomous parsing of “structured and unstructured” information to train the AML model through continuous learning and refinements.

For banks, machine learning capabilities apply to effective transaction monitoring, watchlist screening, KYC (CDD/EDD), and the like.

Supervised learning and unsupervised learning form the keystone of the training approaches.

Supervised learning (SL) involves training AML models that learn patterns from “labeled” datasets with input values mapped to the output. For instance, multiple values and attributes of a transaction can indicate layering, which can serve as the input for the model to comprehend intricate patterns.

Through adequate training, the model learns to predict the outcome of even unforeseen circumstances. AML tools built using the SL approach offer the ability to exploit transactions pre-tagged for anomalies.

With data analytics, the supervised learning approach can help banks improve transaction monitoring by exploiting structured information.

Further, AML systems with unsupervised learning capability can learn patterns from unlabeled data using approaches like clustering, anomaly detection, and latent variable modeling.

Ability to encapsulate patterns as probabilities is a standout feature of unsupervised learning, and is key to modeling financial transactions, customer risk profiles, historical records, etc.

Banking operations generate a glut of unstructured or raw data, and the real application of unsupervised learning unfolds here.

Due diligence, surveillance, and watchlist screening

Customer due diligence and sanctions surveillance are concerned with analyzing unstructured information in the external world.

For instance, global watchlists, government circulars, social media, and public databases are critical sources of topical customer information.

Real-time sifting of these vast and distributed data sources is key to effective due diligence for compliant onboarding and risk mitigation.

Natural language processing (NLP) provides seminal tools to augment the machine’s ability to parse unstructured data repositories.

NLP techniques such as information extraction, tokenization, pattern matching, pattern recognition, and syntactic analysis have fundamental applications in KYC and watchlist screening.

For instance, information extraction techniques leverage machine learning and NLP to rapidly extract structured information from unstructured documents. For banks, this ability translates into a robust due diligence system that can surveil and extract insights from unlimited information repositories on the web.

Likewise, AML monitoring systems with pattern matching and pattern recognition capabilities can help banks quantify obscure relationships between a prospect/customer and risk parameters like sanctions, political exposure, or precedence.

Pragmatism is the mainstay of driving the shift

Machine learning holds a tangible promise to scale up transaction monitoring, KYC, and surveillance in the banking industry.

While financial institutions look forward to embracing the colossal change, they need to evaluate the choices and their strategic impact. A realistic, domain-centric approach is the linchpin to successful adoption.

(The author is Mr. Anuj Khurana, Co-Founder and Chief Executive Officer at Anaptyss and the views expressed in this article are his own)



Leave a Response