How machine learning could prevent money laundering

Machine learning is being put to use in all sorts of areas today. From smart cars and homes and beyond, the use of artificial intelligence (AI) and machine learning (ML) are becoming a larger part of how many companies conduct business. As more and more businesses are hit with cyber crime rather than physical crimes, there has been a needed shift from commercial surveillance systems towards cyber security systems to protect confidential data. More recently, we’ve seen ML sink its teeth into anti money laundering (AML) with big potential impacts there.

Most current AML systems are founded on an extensive list of rules. Banks and institutions are required to comply with the Bank Secrecy act and implement certain AML rules. These regulations are meant to help detect and report any suspicious activity that could indicate money laundering or terrorist financing. As these regulations have become more demanding, the traditional rules-based systems has become more and more complex with hundreds of rules driving know your customer (KYC) activity and Suspicious Activity Report (SAR) filing. As financial institutions monitor billions of transactions a day, the data mined from each creates a silo of information and data that any person would find overwhelming to sift through. More and more cases are being flagged for investigation but more and more false positives pop up.

Along with in an increase in the false positive rates, another challenge found in AML is the fact that it hardly ever signifies as the activity of just one transaction, account, business or person. Because of this, detection cannot focus on singles instances but rather requires analysis of behavioural patterns of transactions occurring over time. Therefore it is nearly impossible for personnel to investigate all cases in a timely manner.

Additionally, there are very little historical data surrounding ML making it difficult to pinpoint exact tactics and methods for ML. From trusts to black market currency exchanges or loan-back schemes, there is, unfortunately, no typical ML case. This also lends to indefinable data labels and sets, which have required manual analysis in the past. Despite the growing difficulties to monitor AML, more financial institutions are turning to the various technological tools such as AI, ML and big data analytics to detect ML cases.

When combined, these systems can merge across massive spectrums of data sources and dig through boundless mountains of data. When ML or AI has been implemented, however, these hybrid systems are able to translate those unlabelled points of data into signals to detect behavioural anomalies and intent. These ML systems can establish “normal behaviour” patterns and then identify anomalous behaviour next to it, thereby weeding out the false positives from true ML cases. In this instance, there is massive reduction of false negatives and false positives.

Moreover, FICO has created an AML Threat Score that helps to prioritize investigation queues for SARs, utilizing behavioural analytics from Falcon Fraud Manager. FICO utilizes transaction profiling tech, self-calibrated models and customer behaviour lists all which can adapt to the constant changing dynamics within a financial institution. Other ML systems that have seen progress has been the “unsupervised learning” — a form of machine learning that uses algorithms to draw inferences from data sets that lack labelled responses. Due to the large gap in historical data on ML cases, there is a weighty need for ML technology to be able to analyse and gain insight from data without prior knowledge of what to look for. Unsupervised machine learning learns from that unlabelled data and results in the ability to differentiate between the relevant and irrelevant data and can then divide this unlabelled data into usable clusters. This is achieved through link analysis, temporal clustering, associative learning and other techniques that allow financial institutions to track entity interactions, behavioural changes and transaction volatility.

The benefits of using machine learning to prevent ML can be seen when labelled and unlabelled data from this slew of sources can be ingested into a system that is flexible enough to accept a multitude of data points across a myriad of sources while also analysing its potential for a ML case. As the AML regulations that are required of banks become more intense and fines for complying grow, you can be sure to see the implementation of machine learning built on top of the traditional AML systems already in place today.