Criminals launder trillions of dollars, directly impacting the US economy in a major way. A Machine Learning Scientist is enlisted to combat this. Sound like a high-tech spy film, doesn’t it? In actuality, it’s the very real-life situation of Keyu Chen, the genius who spearheaded the creation of what is known as the AML Model which is designed to detect patterns indicative of money laundering. According to the United States Department of State, criminals launder an estimated two to four trillion dollars annually, posing a substantial risk to global financial integrity. As a leading financial technology platform, PayPal has also encountered substantial losses attributed to money laundering activities. Ms. Chen’s work with PayPal makes a legitimate counter-offense by establishing a new method of hyper-defense.
Detecting fraudulent or misleading activity is nothing new or unusual for Keyu Chen in her line of work but the sheer amount of data needed to detect activity in this specific area was the most daunting aspect for this undertaking. What she needed to create was something which could learn quickly on a massive scale across multiple languages and geographical locations. Impressively, her AML Model has established a successful process by effectively analyzing and scoring all red flag indicators with more than 95% accuracy. Keyu used feature engineering to extract anomaly features from diverse datasets comprising historical transactional data, user behavior patterns, and known instances of money laundering. The patterns identified include transaction frequencies, amounts, geographic locations, circular payments, structured payments aimed at reducing transaction amounts below regulatory thresholds, and sudden changes in transaction information, etc. The AML model operates by iteratively optimizing decision trees to predict the likelihood of a transaction being associated with money laundering based on these extracted features. The use of a tree-based model not only enhances explainability but also ensures strict adherence to compliance regulatory guidelines. Moreover, it automates the filing of Suspicious Activity Reports (SARs) to the federal Financial Crimes Enforcement Network (FinCEN) for violation accounts. The automation of SAR filing results in a significant reduction in operational time by fifty percent.
Prior to the creation of Keyu’s AML Model, PayPal used rule-based systems which were written using SQL (a programming language for databases) and were based on the experience of its business and compliance teams. While this was the most effective system at the time, its accuracy was only about sixty-five percent. The AML Model revolutionizes accuracy, increasing it to a remarkable ninety-five percent. Key to this is Keyu’s utilization of iteratively optimizing decision trees in the AML Model that allow it to learn and anticipate false positives that were actually legitimate transactions. Unlike the static nature of rule-based models, this approach dynamically adapts to new data and evolving patterns of money laundering, such as transaction structuring and circular payments.
The potential for mishaps and misperception of data and trends was astronomical for this project. Keyu concedes that the human-meets-legal concerns were among the most difficult to navigate in her work. She remarks, “PayPal, like other financial institutions, must comply with rigorous anti-money laundering regulations, and any misstep could have legal ramifications. It required the model to be both accurate and explainable, which is often a tradeoff in machine learning. The AML Model provided some level of interpretability through its use of decision trees, but balancing the need for transparency with high predictive accuracy was a constant challenge. This required meticulous validation, ensuring that every flagged transaction could be justified in terms of regulatory compliance.” There will likely be more upgrades and new approaches needed in the future. Ms. Chen agrees, “Money laundering is inherently evasive and dynamic. Criminals are constantly developing new strategies to avoid detection, making it difficult to create a set of features that would generalize well across different schemes. While some patterns, like transaction amounts or geographic anomalies, are easier to detect, others—like circular payments or structured transactions aimed at regulatory evasion—are more subtle and can vary greatly depending on the laundering method used.” It’s remarkable to perceive how the methods of combatting international fraud are changing, redirected into a more high-tech system in which those like Keyu Chen will reinforce the protective decisions of the international community. The AML Model indicates that circumnavigating international legal boundaries will be much more difficult for those who wish to do so.
Writer : Basil Thomson
Punit