Terrorist activities are the central problem of our modern world. Over the past decade, the average number of annual deaths due to terrorism was 21,000. Terrorist organizations are the major actor of terrorism. According to the Council on Foreign Relations, front companies used for money laundering are among the three funding sources of terrorist organizations. In this case, money laundering is an essential tool for terrorist activities. So, what is money laundering? According to the Financial Action Task Force(F.A.T.F.), money laundering is the processing of criminal proceeds to disguise their illegal origin. It is of critical importance, as it enables the criminal to enjoy these profits without jeopardizing their source. This definition brings our minds another question; how is money laundered?
According to the FATF, firstly, the launderer enters the financial system with his illegal income. This can be accomplished by dividing large quantities of cash into smaller, less visible sums that are then deposited directly into the bank accounts, or by buying a collection of monetary instruments (cheques, money orders, etc.) that are then collected and deposited into accounts at a different location. Then, To separate the funds from their source, the launderer performs a series of conversions or movements. The money could be channeled into the purchase and selling of investment securities, or it could simply be wired through a collection of accounts at different banks around the world. Finally, the funds are reintroduced into the legal economy. The launderer may decide to invest the money. The launderer might choose to invest the funds into real estate, luxury assets, or business ventures. The intention to prevent these criminals from money laundering is called as Anti — Money laundering. As it can be seen, money laundering is an extremely complicated process to detect for humans, especially between the millions of clean daily transactions. Thus, machines are here to help to handle this difficult task. We name these machines Anti- Money Laundering Software.
What is an Anti- Money Laundering Software?
Anti-money laundering software (AML software) utilizes technology to help legal and financial institutions comply with financial regulators’ legal requirements to identify money laundering and combat financial crime. In short, AML Software utilizes technology to detect potential money–laundering transactions. Specifically, almost all of this software utilize Artificial Intelligence, for most cases its subbranch: Machine Learning, to detect possible money–loundering transactions. In this case, Artificial Intelligence(A.I.) and Machine Learning(M.L.) are the two essential concepts to know.
According to IBM, Artificial Intelligence (AI) refers to computers that can imitate human intelligence by performing tasks that can be described by a human being. IBM also defines Machine Learning as follows: Machine Learning is computer ability to keep learning without being reprogrammed. It uses algorithms that learn from data and create foresight based on this data.
How Do AML Softwares Utilize AI?
Financial Institutions like banks group transactions according to the different features. These features could be about the amount of money flow during the transactions, how does the transaction is made, or from where the transaction is made in the first place. After this process, banks apply some limits to these categories and flag those of transactions that overcome these limits. For example, in the U.S. and Canadian banks, transactions of more than $ 10,000 are automatically flagged by the system. This is very helpful for Anti — Money Laundering process because most of the money launderers are lazy enough not to divide their dirty money into small amounts. Banks are also flags to transactions from several countries. Transactions from these countries have a high potential to become money laundering transactions. Comparing the other transactions, the number of flagged transactions is drastically low; 99.9999% of transactions are unflagged, and 0.0001 % are flagged. Thus, they are accepted as exceptions. In the Artificial Intelligence word, these exceptions are named as anomalies. Machine Learning Algorithms are very helpful to detect these anomalies, they can detect these anomalies in a more time-saving way, and they make fewer errors than humans.
Which AI technologies are relevant to Anti- Money Laundering Software systems?
Machine Learning Algorithms detect anomalies by anomaly detection. Anomaly detection is the process of determing unexpected items or events in data sets, which differ from the norm. In our case, a particular branch of anomaly detection, clustering-based anomaly detection, is used to detect flagged transactions.
When new data is presented, the key concept behind using clustering for anomaly detection is to learn the usual model(s) in the existing data (train) and then use this knowledge to point out whether one point is anomalous or not (test).(test).
ML algorithms are trained & tested with data before they are faced with real-world problems. If this data is labeled, this training & testing process is called supervised learning. If unlabeled data is provided to the ML algorithm for the learning process, and the ML algorithm tries to make sense of it by extracting features and patterns independently, it is called Unsupervised Learning.
In our case, financial institutions group transactions as ‘k’ similar clusters of data instances. Data points that fall outside of those groups could potentially be marked by an algorithm as anomalies. This algorithm is called as K-Means Clustering Algorithm.
K- Means clustering algorithm is an unsupervised learning algorithm. However, financial
institutions label their data. This data is used for training ML algorithms. In other words, this data is suitable for supervised learning. ML algorithms use labeled data and implement clustering-based anomaly detection are called as Semi-Supervised Learning Algorithms.
As ML is a subbranch of AI, ML algorithms and techniques could be count as AI technologies. In this case,
· (Clustering Based) Anomaly Detection
· Supervised Learning
· Unsupervised Learning
· K — Means Clustering Algorithm
are AI technologies that are relevant to Anti- Money Laundering Software systems.
As relevant AI technologies for AML Softwares are mentioned, let’s talk about how to evaluate them.
How to evaluate modern Anti- Money Laundering Software systems in line with advances in AI?
With advances in AI, Modern Anti–Money Laundering Software Systems can be evaluated with their AI algorithms’ success rate. There is no single answer for measuring the success rate of Anti — Money Loundering Software’s AI algorithms because each algorithm is designed for a different purpose. AI algorithms of Anti — Money Loundering Software’s performance is evaluated by performance evaluation metrics. Each performance evaluation metric measures algorithm’s success by its own standard. In our case, three of the most common performance evaluation metrics are Precision, Recall, and F1 Score. To understand these three evaluation metrics better, true positive, false positive, true negative, and false negative need to be clarified.
In our case,
True positive: Account is flagged as a money-laundering account by algorithm, and in fact, it is a money-laundering account. The outcome of true positive is a success.
False positive: The account is flagged as a money-laundering account by algorithm, and in fact, it is not a money-laundering account. The outcome of false positive is a waste of time (if time = money, then waste of money)
True Negative: Account is not flagged as a money-laundering account by algorithm, and in fact, it is not a money-laundering account. The outcome of true negative is a success.
False Negative: Account is not flagged as a money-laundering account by algorithm, and in fact, it is a money-laundering account. The outcome of true negative is imprisonment and monetary severe penalties and jail by authorities/regulators.
Precision is the fraction of relative instances among the retrieved instances. It is based on relevance. In other words, precision shows how precise/accurate a how many of those expected positive are actually positive, according to a model. When the costs of False Positive are high, precision is a good metric to use.. In our case, as false positives are drastically high(we’ll see its details in the case study part), precision is a major evaluation metric for Anti- Money Loundering Softwares’ Algorithms. Precision’s formula can be found below:
Recall is the faction of relevant instances that were retrieved. Recall is also based on relevance, in other words. Recall estimates how many of the True Positives in a model capture through labeling it as True Positive. Applying the same understanding, it can be known that Recall could be the model metric that is used to select the best model when there is a high cost associated with False Negative. In our case, as false negatives are our agenda not only for their frequency but also their consequences, Recall is other major evaluation metric for Anti- Money Loundering Software’ Algorithms. Recall’s formula can be found below:
F1 Score , also called the F score is a measure of a test’s accuracy. The F1 score is defined as the weighted harmonic mean of the test’s precision and Recall. F1 Score measures accuracy for the cases that if there is an uneven class distribution(Large Number of True Negatives and a very comparably small amount of False Negatives and False Negatives). In our case, as 99 % of transactions are not actually money — loundering transactions, F1 Score is an accurate choice of evaluation metric. F1 Score’s formula can be found below:
As evaluation metrics are mentioned, let’s talk about some weak points of AI; challenges of using AI technologies in AML Softwares.
What are the challenges of using Artificial Intelligence technologies in Anti- Money Laundering Software?
There are two major challenges of using AI technologies in AML Softwares. The first one is the explainability issue of AI solutions; the other one occurs when combining multiple AI point solutions.
Explainability Issue of AI Solutions
After the AI solution is implemented, AML Software, regulators, and staff of the AML Softwares may not be able to understand outcomes because each of the algorithms might have a different outcome, and it can be confused to keep them in mind. Educating AML Sofware staff for each algorithm is not only expensive but also time-consuming. Hence, the staff of AML Softwares may not be able to evaluate outcomes accurately. Thus, they approach these algorithms as black boxes. To overcome this challenge, AI Solution providers should prepare basic but inclusive cheatsheets for AML Software staff and provide 24 / 7 support for the algorithm issues.
Combine Multiple AI Point Solutions
Many vendors provide AI-powered point solutions for very specific aspects of AML Software challenges including entity resolution, connection analysis, adverse media screening, and sanctions screening. Because of its siloed structure, financial institutions are wary of this strategy. And if you implement two or three separate solutions, you’ll still have a big product. Implementing two or three independent solutions still leaves large productivity gaps in processes, while the unexpected clashes between point solutions will result in completely new and unexpected problems. (*). The challenges that come with “making your own” solutions often apply to this approach. AML Softwares should have a flexible integration policy to overcome this challenge, and they need to purchase adjustable solutions.
As the potential challenges and their simple but effective solutions are mentioned, it is time to ask the “why” question; why are Anti- Money Laundering Software companies integrating AI in their solutions?