According to a recent Europol report, combatting global money laundering continues to be a challenge. “The banks are spending $20bn a year to run the compliance regime … and we are seizing one percent of criminal assets every year in Europe”, Rob Wainwright, director of Europol told Politico recently.
At an institutional level, the volume, variety, availability and location of data, along with changing regulations and fraud schemes keeps many anti-money laundering (AML) compliance officers up at night.
With this in mind it is not surprising that many in the field are looking to new solutions that include artificial intelligence (AI) and, more specifically, machine learning (ML) to help propel the effectiveness and performance of their AML programs.
However, while there is much enthusiasm for the promise of what new technology can offer, there is also much hesitation since many don’t understand how AI and ML work and what it can and will do for them.
Before getting into what the technology can do for AML programs it is worth noting that the one thing AI and ML will not do is completely replace the need for people. Human or natural intelligence has to be coupled with machine intelligence in an effective AML program. However, what machine intelligence can do is look at both structured and unstructured data to detect patterns or clusters of activity that may be out of the ordinary and potentially fraudulent.
It should also be noted that solution providers like do not “program” the AI system to detect threats – they help financial institutions to train it to do so. This requires people who are subject matter experts and leaders within the AML compliance and fraud space to work in conjunction with the AI system to determine whether abnormal behaviours or clusters of transactions are suspicious or fraudulent.
Starting with anomaly detection
The first model that should be implemented by every organization is anomaly detection. It is an advanced technique that uses an organizations’ data and data science to detect behaviour that doesn’t fit within the expected normal behaviour profile. There are different types of anomalies that can be detected, including:
- Point anomalies are single instances (or outliers) that fall outside of the normal behaviour. These include sudden large deposits or credits to an account.
- Contextual anomalies are data points or transactions that appear anomalous without the appropriate context. For example, increased spending during the month of December would appear anomalous in itself but in the context of the holiday season, would be considered normal. A similar situation may appear if a client is starting a new business
- A collective anomaly is when a number of data points are considered anomalous but the values of the individual data points are not themselves anomalous. An example of a collective anomaly includes a case where deposits of $4,500 are made into a chequing account. It may be unusual to have frequent deposits of such amount into a personal account but the value of the deposit is not flagged as suspicious.
It should be noted that to have an effective anomaly detection system, organizations need to implement a solution that detects all the different kinds of anomalies – not just point anomalies.
It should also be noted that not all anomalies are fraudulent – they are simply outliers that require further investigation. To help compliance staff manage the number of anomalies that need investigation, the technology should be configured to rank alerts by risk level. More information about ranking can be found in the next section.
Transaction monitoring using machine learning
Most organizations are familiar with rules-based transaction monitoring where instances like when transactions exceeding a certain value or fund transfers from certain high-risk accounts are flagged.
While these are still valid and need to be part of any transaction monitoring system, technology has allowed financial institutions to augment the effectiveness of these systems with machine learning.
Instead of focusing on monitoring and flagging transactions that did not pass a specific threshold, the system is designed to look for suspicious or fraudulent behaviour. Like in anomaly detection, the system using data science and machine learning to take in various inputs, including historical data that led to a filing of a suspicious activity report (SAR), to determine whether a transaction is likely to be fraudulent. This is not a binary response of suspicious/not suspicious but rather the likelihood of being fraudulent.
Effective AI technology solutions leverage a layered analytics approach which allows financial institutions to reduce the high levels of false positives generated by rules based systems. Less false positives allows compliance and fraud prevention staff to focus on investigating true compliance risks and fraudulent cases.
The compliance department sets the threshold for which they hold transactions for further investigation. Again, these can be tailored to the institution’s risk tolerance. Some may decide that they investigate transactions with a threshold of “50 or more” while more risk-tolerant organizations may increase the threshold.
Aside from the ability to catch more fraudulent transactions, another positive outcome of evolving from rules-based transaction monitoring systems is that implementing these models requires the organization to revisit their risk assessments. The information captured can then be used to show auditors and financial regulators that reasonable measures were in place to detect illicit activities.
Robotic Process Automation (RPA)
Analysts spend a significant amount of time reviewing a large number of alerts and cases each month. The process is exhaustive and prone to human error – which can be costly for an organization. Robotic process automation not only increases the speed and efficiency at which decisions are made during the review process, it also takes away the need for compliance staff to get involved with routine tasks. Organizations that incorporate automation into their programs enjoy many benefits including time and cost savings, increased compliance to policies and procedures, fewer errors and the ability to quickly adapt to new rules and regulations.
Tasks that can be successfully automated include those that are:
- Rules-based, rather than those that require human intervention
- Are triggered and supported by digital data
- High volume and repetitive.
When starting on their automation journeys, many financial institutions look at their tasks and processes that are very time consuming and routine. Examples of where automation and workflows can be used by financial institutions for AML compliance include
- Validating client information against internal and external sources during the customer due diligence phase
- Screening customers against sanctions, politically exposed persons (PEPs) and negative news lists
- Acknowledging and resolving alerts created by the transaction monitoring system that should be treated the same way
- Verifying account activity of high-risk activities
What I have provided in this article is a high-level view about how anomaly detection, transaction monitoring using machine learning and automation (RPA) can impact and enhance AML programs. In my next article, I will discuss the implementation of these models from the perspective of a data scientist.