Algorithmic AML: Making Visible What is Frequently Invisible. Tackling the 1% Problem
3–5% of global GDP is laundered through the global financial system. That’s a range of $800 billion to more than $2 trillion a year. According to the United Nations Office on Drugs and Crime, despite the nearly $2.4 trillion in illicit funds laundered each year, less than 1% of that money is detected.
Financial institutions are required by regulators to combat money laundering. They have invested billions of dollars to do so. Despite this, these institutions still face significant penalties for non-compliance. Penalties over the past five years have totaled more than $300 billion dollars and is expected to increase to well over $400 billion over the next few years.
Less than 1% of the $2.4 trillion dollars in laundered money per year is detected.
In September 2020, the International Consortium of Investigative Journalists (ICIJ) released more than 2,100 suspicious activity reports — known as SARS reports — filed by banks and other financial firms with the U.S. Department of Treasury’s Financial Crimes Enforcement Network, or FinCEN. SARs reflect specific concerns of regulatory agencies delivered to banks and other financial institutions based on fraud patterns identified or they intend to monitor. They are not necessarily evidence of criminal conduct or other wrongdoing. They are, however, triggers or “Flags” of concern. And consequently, serve as an “early warning signal” of what banks and others should pay attention, but these are the “leading indicators” of what regulators are and will be paying attention to.
80% of SARS are false positives. And every SAR reported requires investigations. No matter an imperative is to reduce false positives.
Banks have spent billions on transaction monitoring systems that scrub their accounts for possible money laundering schemes. Detection rules are action-based and target suspicious transaction behaviors, such as excessive cash deposits, structured transactions intended to avoid government record-keeping thresholds, and rapid money movement through one bank to another.
Customers who violate the detection rules trigger a system-generated alert, which is reviewed by an internal investigator. Despite decades and billions of dollars in industry investment, over 95 percent of system-generated alerts are closed as “false positives” in the first phase of review, with approximately 98 percent of alerts never culminating in a suspicious activity report (SAR).
False positives cost the financial industry billions of dollars in wasted investigation time each year but more importantly, expose banks to steep fines and reputational damage for failing to identify bad actors involved in organized crime, sanctions evasion, or terrorism. Banks can reduce risk by reassessing their detection strategies, which presently lack the focus or sophistication to identify illicit source behavior.
Whether or not banks are motivated to actually counter illicit funds flowing through their accounts is a topic for others to wrestle through. The recently released FinCEN files suggest perhaps not. It suggests that there is lots of “box-checking but little practical progress” with many focused on “technical compliance” rather than whether systems “are really making a difference.” (source: https://www.icij.org/investigations/fincen-files/global-banks-defy-u-s-crackdowns-by-serving-oligarchs-criminals-and-terrorists/)
People will dispute on whether or not banks should be more aggressive to counter money laundering and/or whether or not they will be. What is not in dispute, however, is that significant monies are spent by banks as a result of anti-money laundering (AML) activities and requirements. And where monies are spent, new capabilities will be developed. More than 335 start-ups globally have entered the AML space within the past 3 years. The draw? The growth of spend anticipated to counter illicit fund movement and fraud.
And here it gets interesting. Because, as the old saying goes, the more things change, the more things stay the same. Unless it doesn’t. And of all the energies and monies, start-ups and SARS being created, the domain of AML reflects both sides of this saying: there is much staying the same, yet some incredibly interesting new capabilities being developed that could (finally) start to tip how AML is conducted, fraud patterns identified and both penalties and investigatory costs reduced.
Let’s explore how.
Making it Pragmatic
There are two fundamental problems to overcome to strengthen one’s AML / Fraud process.
1. The False Negative Problem — which focuses on the financial exposure of fraud. Penalties imposed on a bank reflects this exposure rate. This problem is also known as the “upstream” — or initial part — of the “Know your Customer” (KYC) process. New customers need to be vetted, scored according to risk profiles aligned with the institutional criteria of who makes up an “acceptable” customer.
2. The False Positives Problem — which requires being current with regulators and the notices or “flags” they provide based on insights they’ve seen and concerns they have. Alignment with such SARS reports involves significant monies spent by banks as they investigate potential discrepancies. Due to the sophistication of contra-partiers, 95% of system generated alerts are false positives. False positives cost billions of dollars in wasted investigation time each year and expose banks to steep fines and reputational damage for failing to identify bad actors.
Both the False Negative and False Positive problems are costly. In the former case, it stems from penalties, in the latter case, from the costs of both investigations and reputational damage.
Bad guys are clever guys. They adapt to new capabilities and respond to SARS reports. The former stems from the never-ending introduction of new technical capabilities (e.g., witness the hundreds of new startups alone in the space). The latter indicate where they should focus since such reports are how the regulators tell the banks what is important to them and consequently where banks should allocate their resources to align with regulatory focus.
The implication of both? An ongoing game of cat-and-mouse juggling new capabilities to use and the “white space” of regulatory focus leaving plenty of “dark space” to continue to commit fraud.
Making Visible what is far too often invisible:
From Rules-Based Methods to Algorithmic Insight
There are two different approaches to tackle fraud; one has been the mainstay of the industry to date, the second reflects new capabilities based on pragmatic lessons from machine learning, algorithms and math.
The Mainstay: Rules-based approach
A rules-based approach — whereby the system flags cash transactions over a certain currency amount, blocks transactions to certain countries, uses customer data to select accounts for additional monitoring, and categorizes merchant accounts based on prior transactions.
AML involves an ongoing game of cat-and-mouse juggling new capabilities to use and the “white space” of regulatory focus leaving plenty of “dark space” to continue to commit fraud… Requiring algorithmic rather than rule-based approaches to “make visible” what often remains invisible.
It is the foundation of many (the majority?) of current products and approaches used. As new flags are identified, rules are constructed to monitor similar transactions. Rules get updated into an ever-increasing set of rules to monitor.
Rules-based approaches have an inherent limitation. Bad guys change behaviors frequently requiring catch-up and the addition of new rules. The volume, velocity and variety of new bad behaviors overwhelms rule-based approaches. Either there is a constant “you can’t see what you don’t catch” problem and so only create rules of bad behaviors that are found or the rule set becomes brittle because each one needs to be coordinated with other rules creating a “log jam” of code which makes fraud monitoring heavy (e.g., expensive) and hard to maintain. Combine with this the comment earlier that only 1% of bad behaviors tends to be found, and you get a quick sense of the imperative of new capabilities.
Algorithms with machine-learning capabilities have three characteristics which significantly helps to reduce both false positives and false negatives. (Note: a recent engagement completed with PwC on one of their banking clients attested that our Minerva-AML approach was 3x more effective in finding fraud patterns than rules-based approaches. We’ll further explain how later.). Rather than mapping transactional data to known flags (e.g., bad behaviors), algorithmic approaches seek to “stitch together” patterns of behaviors, some seen before and some inferred from anomalous behaviors.
Our algorithmic approach differs from rules-based approaches in three ways.
First, it is based on rich fraud taxonomies. This solves the “10,000 acre forest problem.” Here’s the problem. Finding a needle in a large forest (e.g., fraud in a set of millions of financial transactions) is difficult. You need to cover lots to ground to do so. Common sense tells you that there is no need to walk every inch of the forest. The needle, and fraud, isn’t everywhere, but in particular parts of the forest, or in a subset of the tens of millions of transactions. Narrowing the ground to cover, from 10,000 to, say, 1,000 or fewer acres will a) increase the odds of finding what you want and b) accelerating doing so.
Rules-based approaches cover all 10,000 acres, the more the rules, the more the ground they cover. Algorithmic approaches do it differently. Narrowing this search space rests on the taxonomies and learning modules run across them. This involves two steps:
Step 1, prepare the data to be explored — the 1,000 acre search space. Often, significant data is missing from transactional records such as beneficiary flags, country office, region, perhaps the instrument name, and so on. The taxonomy helps to fill in blanks, inferentially again based on the underlying patterns that make up the taxonomies.
Step 2, and again based on the taxonomy, it’s important to “force a segmentation” of the data. This segmentation becomes the “common sense” search space — the 1,000 rather than the 10,000 acre forest to explore.
The following depicts the nature of transactional data often received that needs to be prepared, powered by the fraud taxonomies.
Second, the learning engine is strengthened through insight into the current set of SARs flags, their stability over time and over efficacy at targeting known frauds. This step involves building classification models identifying combinations of alerts which translate into “true,” known or knowable SARs. It is insight into the combination of, and the sensitivity of relevance among them, rather than a listing of them that matters. How? By isolating “faint signals” of which types of SARs influence others and thereby serve as “pointers” to derivative / new fraud patterns.
Third, network insights. One reason identifying fraud is that no single financial transaction may be problematic and hence not identified. It is the combination of many of them over periods of them that gets interesting. There are two blunt implications of this clever way to “remain under the radar.” One, statistical analysis across transactions are limited — since, again, any particular transaction may be legal. Two, relationship analysis needs to be performed to understand a) if and how any particular transactions may be connected with each other and b) the number of hops between two accounts (the origin and disbursement accounts) and (here it gets particularly interesting) the number of intermediaries between the accounts.
The only way “into” this problem is to use link, or network, analysis, transposing the data into a graph database to derive such insights.
The figures below visually depict how intermediaries can be linked together resulting in patterns previously invisible to become visible.
It’s important to note that the rules-based and algorithmic-approaches complement each other. Rules-based systems generate a set of flags to monitor. These become inputs into the learning modules for algorithmic insights. And the new fraud patterns identified by the algorithmic approach can be fed into the rules-engines to monitor fast-moving fraud patterns.
For the organization, this is key. How so? Because it means that the sunk costs of existing rules-based systems remain relevant. Algorithmic approaches to be “light-weight” in terms of investment; they can be thought of as “services on top of” or complementary to existing solutions. The second advantage of this is speed. As services, they can be deployed quickly — in days rather than weeks, thereby strengthening one’s insights while reducing both false positives and false negatives and the penalties and investigation costs associated with them.
What’s not to love?
About the Author:
Ralph Welborn, Ph.D. — Managing Partner and Co-Founder, ClearPrism
Previous roles included Head of Strategy & Transformation (IBM — Middle East / Africa), Senior Vice President Global Strategy, KPMG Consulting, CEO of Advanced Analytics Company. Author of three books on Business & Technology Transformation.
Ralph can be reached at firstname.lastname@example.org