How Patent-Pending HAWK Technology Improves the Detection of Money Laundering
Money laundering is abnormal behavior, which should make it easy to spot. However, money launderers change their methods to avoid prosecution, which makes it difficult to detect financial crime. Emerging AI technology provides anti-money laundering teams with the anomaly detection power they need to catch and fight financial crime. That said, the technology is a single tool in the AFC belt. Success depends on how companies understand and use that tool. That’s why it’s important for anyone involved in this process to understand how AI works, how frontline operators can trust its results, and how they can use it to help stop financial crime.
Supervised and Unsupervised Learning for AML and Financial Crime
There are two basic approaches to AI data science: supervised and unsupervised learning. Both methods effectively detect anomalies in large datasets. With unsupervised learning, you only need input data. The algorithm “learns” from the data you feed it and outputs recognized patterns. This is an effective tool for AML anomaly detection. Companies can feed the model massive amounts of transaction data, and it will identify anomalies on its own.
On the other hand, supervised learning models require labeled data for their pattern recognition to work. With that labelled data, the algorithm knows what to “look” for, and can categorize additional data accordingly. For example, you would feed the model cases, and the labels would tell the model whether the case is money laundering or not, and the risk factors, such as multiple transactions with round amounts or abnormally large transactions, that explain why. Supervised learning can be effective from an AML standpoint, since you train the model on exactly what to look for. However, there are significant drawbacks to using this approach. Labelling data is labor intensive. For a small team of people, categorizing enough transactions to train the model would be impossible. In addition to that, if you base all pattern recognition on historical behavior, you could potentially bias the model to miss anomalous cases that don’t fit past patterns. For market-informed anomaly detection, unsupervised learning is generally the best solution.
In addition to supervised and unsupervised learning, there is a blended approach, known as semi-supervised learning. Hawk’s technology often uses semi-supervised learning models to take advantage of the strengths of both. Semi-supervised algorithms work with partially labelled data. In an AML context, you would only have a subset of the data labelled as “money laundering” or “not money laundering,” while the bulk of the transactions would have no label. Since no company can label millions or billions of transactions, this is the situation for most. Semi-supervised models thrive with partially labelled datasets, as they can leverage the unlabeled data to perform the supervised task of predicting labels.
Anomaly Detection for AML and Financial Crime
How does an algorithm go about detecting anomalies with only a set of data? One of the technologies Hawk AI uses is the Isolation Forest method. Isolation Forest measures a data point’s isolation, or how far it lies from the rest of the data. Based on how isolated the point is, the algorithm can generate an anomaly score. This score will tell a user just how anomalous that point is. Dr. Felix Berkhahn, Chief Data Scientist at Hawk AI, said, “Isolation Forest achieves high accuracies throughout a wide range of use cases. It is a very robust method that allows for quick model building, and it is transparent. Practitioners widely adopt and accept this method for these reasons.”
Money laundering and other financial crime happens online now, and criminals can move money in the blink of an eye. Companies need to detect and respond to threats the moment they present themselves, while still providing a smooth and seamless customer experience. The Isolation Forest method can detect anomalies as they occur for real-time response.
Spotting financial crime is a nuanced task. Every case has multiple risk factors, such as cross-border transactions and round payment amounts. On top of that, each company has a different range of risk profiles among its clientele. Two or three risk factors could converge in a particular case, but it could still reflect normal behavior. For example, a customer could be from a monitored jurisdiction and have a large transaction, significant risk factors, but the transaction could be an average home sale. Certain risk factors, such as cross-border payments, may even be integral to a company’s business model and occur regularly in the transactions they process. The challenge for every company is determining what constitutes normal behavior and what constitutes suspicious activity.
What Makes Hawk AI’s AML Anomaly Detection Different
Many existing anomaly detection methods give insights into the factors that led to an anomaly. However, they don’t put out what would have constituted a normal value. Hawk AI’s patent-pending explainable AI technology tells users the expected range of normal behavior. It explains the score for each risk factor of a case with natural human language. That context is necessary for evaluating whether a given case qualifies as suspicious activity or not.
The model puts out multiple pieces of essential contextual information:
- The Range of Expected Normal Behavior by Risk Factor
- The Anomalous Behavior Score
- Real-Time Cluster Plots and Analysis
Machine learning models can also improve themselves over time. For instance, cases flagged by the unsupervised model as suspicious can be either closed, escalated, or labeled as appropriate. The model can use the aggregated data from each of these case decisions to optimize anomaly detection. As more cases go through the workflow, the model learns what to look for. With training data generated by normal operations, your team can take part in making an efficient and effective financial crime detection machine.
AML Anomaly Detection in Action
What does this look like in practice? Consider an example. A company reporting itself to be a restaurant records a lower-than-expected count of nighttime transactions. The model identifies this pattern as anomalous, as all other data from restaurants show a contrasting behavior pattern. In fact, the model shows that the transaction data aligns more with the expected behavior of an automotive store. To detect this anomaly, the model has learned that there is a positive correlation between revenue and nighttime transactions for restaurants. In this example, not only is there no positive correlation, there is a negative correlation between nighttime transactions and revenue. The HAWK:AI model can detect and visualize this information for the end user. In theory, it is possible that a legitimate restaurant behaves in this way. However, the anomaly detection model shows that it is not normal behavior. A compliance analyst can use this contextual information to investigate the case, decide whether it’s suspicious, and file an SAR if necessary. Without AI-powered anomaly detection, this case likely would not have surfaced at all.
With unsupervised and semi-supervised learning, as well as efficient and effective Isolation Forest methodology, AI anomaly detection is a powerful anti-money laundering tool. Companies can use this technology to learn what financial crime looks like from the market itself. The technology can also take direct feedback from frontline operators and incorporate it into the model, making it even more effective at surfacing the cases that analysts need to see. This creates the best kind of feedback “loop,” where the model identifies new insights, and users validate those insights.
A growing business will naturally see an increase in transactions, making it easier for financial criminals to hide in plain sight. Without state-of-the-art data science technology, detecting financial crime will become even more difficult. Now that you understand the basics of how AI anomaly detection works, you can use it to better detect and report money laundering. Patent-pending anomaly detection is just one of the ways Hawk AI uses AI and Data Science to fight financial crime.