How To Reduce False Positives In Machine Learning

Compliance software enables financial institutions to detect a wide range of criminal threats, including increasingly sophisticated money laundering and terrorism financing methodologies. However, in broadening the scope and sensitivity of AML/CFT programs, automated software solutions also run the risk of increasing the amount of false positive alerts that they generate, misidentifying customers and incorrectly flagging transactions as money laundering or terrorism financing risks. 

False positive alerts represent a significant time and cost drain in an AML/CFT context since firms must remediate each case in order to meet their compliance obligations under jurisdictional regulations such as the US’ Bank Secrecy Act or the UK’s Proceeds of Crime Act. Similarly, the administrative friction associated with false positive alerts creates negative experiences for customers. One of the most effective ways to address this challenge is to reduce false positives using machine learning and automated monitoring within a compliance solution. Machine learning tools enable firms to interpret AML data with a greater degree of clarity and efficiency, reducing false positive rates and enhancing customer experiences. With that in mind, it is important to understand how machine learning fits into an AML solution and how machine learning can reduce false positives.

What Are False Positives?

A false positive occurs when innocent transactional behavior appears suspicious under the parameters of a firm’s compliance solution – and consequently triggers an AML alert. Many legitimate financial behaviors can exhibit the characteristics of money laundering and terrorism financing methodologies, while some customer names bear similarities to, and may be confused with, those that appear on international sanctions and watchlists. 

When a false positive alert occurs, firms must establish whether the threat is genuine and then, if necessary, submit a suspicious activity report (SAR) to the authorities. The remediation process requires firms to collect and analyze a range of data in order to establish the legitimacy of the alert or, alternatively, confirm its false positive status and allow the transaction to proceed. 

Research suggests that false positives account for over 95% of AML alerts and make up around 42% of companies’ compliance spend per year – amounting to losses of over $3 billion.

How Does Machine Learning Work In AI Compliance?

The data collection requirements of reducing false positives make machine learning a particularly useful AML tool. 

Broadly, machine learning builds on the speed and efficiency of AI compliance software while informing the remediation process with a depth of additional information. In more detail, machine learning systems utilize historic data to make observations about customers and their behavior over time, and then use that data to make intuitive decisions about emergent alerts – and even predictions about future outcomes. Importantly, machine learning systems are programmed to adjust their outputs automatically, generating new data points for the reducing false positive process without the need for any further input or direction from compliance employees.

How to reduce false positives in machine learning

Applied as part of a larger compliance solution, machine learning tools expedite the remediation process for AML alerts, identifying false positives faster than other types of analysis, and escalating true positives where necessary. The potential for machine learning tools to reduce the false positive rate is significant: research suggests that compliance solutions using machine learning may reduce false positive alerts by around 55%. 

Machine learning systems help to reduce false positive rates in the following ways: 

Structuring data: False positive remediation involves the analysis of vast amounts of unstructured data, drawn from external sources such as media outlets, social networks, and other public and private records. Machine learning systems can help firms better structure that data, learning to prioritize and categorize information based on its relevance to particular types of alert. 

Semantic and statistical analysis: Many false positive alerts are generated by redundant data, often involving outdated information or improperly matched names. Machine learning systems can be trained to recognize redundant data by semantic context in order to streamline alert remediation. Similarly, machine learning systems can be programmed to perform statistical analysis on historic and emergent transaction data to help establish the likelihood of a false positive alert classification. 

Intuitive screening: False positives often occur during politically exposed persons (PEP) and adverse media checks or checks of international sanctions lists as a result of misidentification of names or misinterpretation of data. In this context, machine learning systems may be used to enrich customer risk profiles by intuitively providing additional identifying information or clarification over naming conventions to help compliance teams distinguish false positives. 

How to reduce False Positives in Machine learning with Adverse Media Insight

Adverse media is a crucial component of risk-based AML compliance: negative news stories often indicate that a client is involved in criminal activity and presents an increased AML compliance risk. However, negative news screening often generates false positives as a result of misidentification.

With the benefit of machine learning, firms can train their adverse media screening solutions to identify and collate relevant news articles, and structure them more effectively for the alert remediation process. Historic adverse media stories can be referenced quickly and efficiently to inform decisions about incoming alerts, while new stories can be used to enrich client risk profiles in real time. Machine learning is particularly advantageous to adverse media screening measures because systems can be trained to improve their accuracy over time, taking into account duplicate names, similar spellings, the use of nicknames and aliases, and the use non-Western characters and naming-conventions.