The High Costs of AML Transaction Monitoring: Can Machine Learning Help?

The Bank Secrecy Act (BSA) obligates financial institutions to follow a stringent transaction monitoring process to track/monitor, investigate, and report suspicious transactions to U.S. government agencies. According to OCC, these obligations comprise imperatives such as monitoring cash purchases, reporting daily aggregated cash transactions in access of $10000, and reporting activities that might indicate money laundering. As per anti-money laundering (AML) regulations, banking and financial institutions need to screen dubious transactions and notify the authorities through a document called Suspicious Activity Report (SAR).

Filing the SAR places a considerable cost burden on banking and financial institutions, which is incurred in detecting/screening, and investigating suspicious transactions. Additionally, legal penalties and reputational damage are looming threats that may arise due to potential regulatory violations or procedural lapses concerning AML mandates.

This blog outlines some critical challenges financial institutions face with transaction monitoring processes for detecting suspicious transactions. It also sheds light on the emergent machine learning techniques that can help financial institutions track, analyze, and report suspicious transactions effectively, and mitigate the risks of not complying with the regulatory norms.

Detecting and Investigating Suspicious Transactions: The 3 Key Challenges

1. The upfront cost of AML transaction monitoring

Monitoring millions of ongoing transactions is inherently cost-intensive due to the IT infrastructural requirements. It needs to infuse capital into acquiring anti-money laundering systems such as Automated Transaction Monitoring System (TMS) – software to detect and report suspicious activities. According to Reuters, banks have invested billions of dollars in these Anti Money Laundering (AML) systems, which is a sizable capital expense.

2. High cost of false positives

The capital cost spills over to operational expenses incurred in “manual” investigation. Reuter reports that 95% of the software-generated alerts are inaccurate or false alarms, and 98% of these do not result in a SAR. According to SAS Analytics, only about 0.5-7% of the transactions detected by a TMS are genuinely suspicious and warrant SAR submission.

The low efficacy of AML systems such as TMS compels financial institutions to trudge through manually investigating the “false positives” that can run in millions of records. As per SAS Analytics, these false alerts could be anywhere between 8-124 million per firm annually, snowballing the investigation costs.

As per the Reuters report cited earlier, this wasted investigation time and manual resource cost billions of dollars to the banking industry annually.

3. Cost of regulatory violations

Banks and other financial institutions face a potential risk of legal penalties and reputational loss in the event of violating the prevailing SAR filing norms. Failure to identify suspicious transactions or delayed reporting can attract regulatory fines and other legal action such as cease and desist orders, remediation costs, and more. A recent example is Commerzbank AG which was fined £37,805,400 in 2020 for failing to deploy adequate AML systems. According to BCG’s Global Risk Report, financial institutions have incurred more than $394 billion in fines since 2009 due to regulatory violations.

How Can Machine Learning Help Banks Detect Suspicious Transactions Effectively?

Machine learning systems leverage “evolutionary algorithms” that can empower AML transaction monitoring processes and AML systems to learn emergent behavioral patterns dynamically as they evolve within the specific circumstances or bounds of the financial service provider’s network. Here’s an outline of how ML can help:

1. ML models allow the AML system to mine, ingest, wrangle, and analyze large, unstructured datasets (big data) to draw meaningful and accurate insights. Some of these techniques include semantic analysis and statistical analysis to create a realistic risk profile of customers based on their KYC information and transaction patterns.
2. Unsupervised machine learning models can detect anomalous transactions using K-means clustering and One-Class Support Vector Machine (SVM) approaches. Combined with conventional rule-based logic, machine learning techniques can equip the TMS with a superior ability to detect suspicious transactions with higher accuracy.
3. Neural network-based supervised learning models are also being developed to detect suspicious transactions. The Radial Basis Function (RBS) neural network is an example that uses clustering and recursive algorithms to build transaction monitoring capability for anti-money laundering (AML) systems. The method is found to have the lowest false positive rate among other ML techniques based on support vector machines and outlier detection.

The primary benefit of ML in transaction monitoring is the remarkable drop in false-positive alerts. According to industry estimates, machine learning and artificial intelligence can detect more than 95% of false alerts.

Adopting Intelligent Solutions with Digital Knowledge Operations (DKO)™

The time is ripe for banks and financial institutions to leverage machine intelligence and build scalable AML systems to detect suspicious transactions effectively. However, considering the ongoing developments in AI, machine learning, and intelligent digital solutions, institutions need to observe due diligence and care while making the decisions when evaluating a machine learning-based AML system.

Digital Knowledge Operations™ is a solution framework that provides a consultative, data-driven, and tailored approach to implementing intelligent digital solutions across BFS industry verticals and functions. It facilitates seamless adoption of digital tools based on technologies like machine learning, data analytics, robotic process automation, etc.