Machine Learning Expert Features: How AI Can Help in the Fight Against Money Laundering

Money laundering is a consistent problem plaguing society. It funds the worst elements of humanity, facilitating terrorist and criminal activities around the globe.

By its very nature, money laundering is a complex problem to quantify. However, the United Nations Office on Drugs and Crime estimates the annual figure at 2 to 5 percent of global GDP. That’s $800 billion to $2 trillion achieved by illicit means becoming “legitimate” in the global financial system every year.

Given the scale of the problem, it makes sense that there is a sizable anti-money laundering (AML) industry. The AML software market alone is expected to reach $1.77 billion by 2023.

However, given that…

• 90 percent of laundered money remains undetected

• AML activities recover only 0.1 percent of illegally gained funds

• Over 95 percent of system generated money laundering alerts are false positives

• The cost of AML compliance for both the USA and Canada reached $31.5 billion in 2019.

…AML as an industry needs to do better. Improving software solutions to detect suspicious behavior accurately and find/recover laundered money would make a massive difference to financial institutions worldwide.

The truth is that a combination of a complex and fast-moving financial world and sophisticated bad actors means differentiating between legitimate and illegitimate money is a daunting challenge. Authorities and criminals find themselves in a technological arms race to develop new and improved strategies to catch and evade one another.

However, with the advent of AI and the rapid improvement we’ve seen in machine learning algorithms, the good side might have a new weapon to help them in this fight: Expert Features.

What is Artificial Intelligence?

Expert Features take AI to the next level, and are what makes AI truly intelligent. While the phrase Artificial Intelligence is thrown around a lot at the moment, what actually is AI?

AI is a broad term to describe technology that performs human-like tasks. AI machines can adjust based on new inputs and learn to improve their performance. A critical concept within AI is machine learning.

When discussing AI technology for fighting financial crime, what we are really talking about is machine learning. Machine learning algorithms can find complex patterns in vast amounts of data in ways that humans could not.

Compared to conventional algorithms, machine learning’s performance is constantly self-improving. It can also make predictions or decisions from the dataset without a human explicitly programming it to do so.

You can provide machine learning algorithms with data containing financial crime, fraud, for example. With enough examples of fraud, the algorithm builds a model of what fraud looks like. Then by feeding the algorithm new datasets, it can find new instances of fraud.

This is what makes machine learning ideal for catching financial crime.

Expert Features: The Building Blocks of AI

Expert features were first developed for fraud detection. They show a path from traditional analysis to a more intelligent machine learning approach.

Fraud detection contains four steps:

• DATA – received and analyzed

• RULES – applied to the data

• ACTION – e.g., should this transaction be allowed to pass?

• ALERT – is this suspicious and requires investigation?

Statistical metrics or features define the rules applied to the data. For example, they could include the number of payees, average transaction volume, and value. These features form a baseline that defines normal/expected activity. Deviation from the baseline produces scores. Enough anomalous scores raise an alert for suspicious activity.

Financial organizations have had some success detecting fraud using this approach. However, criminals can game the system by exploring the natural boundaries of the rules. How much can I deviate from the norm without triggering an alert?

To stay ahead of criminals requires improved analytics and the implementation of savvy, more intelligent rules. This is precisely what expert features are.

A good metaphor when thinking about expert features is crossing a traffic intersection. You look left and right; how many cars do you see? In this analogy, the number of vehicles is a basic feature or statistical metric for evaluating whether it is safe to cross.

Consider a traffic light instead; this is an example of an expert feature. It has intelligence built into it, understanding how the intersection works and how to optimize traffic flow. When you arrive at the intersection, the traffic light provides meaningful information to influence your decision whether to cross or not.

The difference between a traffic light and expert features is that you are not privy to how a traffic light determines its output. With expert features, you get a clear explanation of the calculations that go into its recommendation.

Introducing machine learning models to the rules applied to financial data allows for a malleable approach. Rather than a rigid traditional approach, you gain agility with user feedback and criminal behavior, improving how the model creates alerts. This produces much more accurate results and reduces the high level of false positives seen in traditional analysis.

The Benefits of a Feature-Centric Approach for AML

While there is some overlap between fraud and AML, fraud has to move quickly and requires detection in real-time. AML requires a greater level of scrutability. There needs to be lineage and evidence to ensure regulatory transparency.

Money laundering is typically not a single event. Instead, it involves multiple transactions to mask the source of the funds. How does this money move through a complex system, and can we find indicators for potential criminal activity?

A better approach for AML is not whether a pattern exists in the data, but rather – how risky is this pattern? Also, with the successive nature of money laundering, is this pattern occurring in conjunction with another pattern that compounds the level of suspicion?

AML investigators use statistical metrics that include:

• Average daily activity

• Total cash in for a given period

• Max cash in for a given period

• The volume of cash in for a given period

Typologies in AML produce expert features that, combined with machine learning, can provide a more nuanced indication of the overall risk present. Examples could be:

• Structuring for a given period

• Rapid in-out

• Rapid in-out recurring

• High-risk anomaly

These all have intelligence inherently baked into them that can fuel a model, improving AML alert accuracy significantly. By considering the context of a customer’s behavior and activity, we can substantially improve the identification of suspicious behavior. Thus, reducing false positives while still stopping laundered money from slipping through the cracks.

At NICE Actimize, we work with financial organizations to take statistical and expert features and build a model that is as sophisticated or as basic as you’re comfortable with. For example, this could include producing a scorecard model that takes these expert features as inputs or a more advanced machine learning model that can consider the customer segment for assessing risk.

Key benefits of an AML feature-centric approach include:

• Expert features are measurable attributes that are clear in their calculation, reducing concerns around lineage, evidence and justifying actions to regulators.

• Expert features provide better effectiveness across multiple models. They can inform future models, so you don’t have to start from scratch for a new typology or risk assessment.

• The model can constantly add new features to supplement your AML approach.

• The value of features is shared across different models to vary how risk is interpreted within the system.

• It provides support for model feedback and performance tracking.

• Governance at the feature level means the model is traceable and explainable.

• Because this approach is compatible with regulatory compliance, you can follow the data within the expert feature and delete it if legally required.

Conclusions

Using a machine learning feature-centric approach in AML provides a more nuanced and accurate assessment of the risks present within a financial system. With expert features, you can use intelligent money laundering traffic lights. But instead of just green, amber and red, you get an entire spectrum of colors, each clearly explained to inform your decisions better.

Expert features provide financial organizations the help they need to detect suspicious behavior, ensuring significantly less illicit funds make it into the hands of criminals.

To read more about NICE Actimize’s AML solutions click here.