Anti-Money Laundering is a particularly challenging area of regulation for banks and even more so for large, geographically diverse institutions. What makes AML such a difficult problem to solve is that it involves complex data, detailed workflows, and significant human involvement.

In short it is perfect for AI.

This is what HSBC determined in their work with Ayasdi on a notoriously difficult problem - that of bank’s customer’s customers, also known as KYCC (know your customer’s customer). The vast majority (~95%) of AML investigations do not result in a suspicious activity report (SAR). This is called a ‘false positive’ - transactions that are flagged for investigation, but are not suspicious based on a deeper review.

Working with Ayasdi, HSBC was able to achieve a significant reduction in false positives (more than 20%) while keeping their suspicious activity reports at the same number.

The Modern AML Challenge

AML is a massive problem and is estimated to be around $1 to $2 trillion per year and growing. Given that the vast majority of those transactions end up funding bad people doing bad things, governments have stepped up their regulatory requirements, oversight and fines.

Banks have scrambled to respond to the increasing regulatory burden. The result is that the cost of compliance is increasing 50% year-over-year and quickly becoming a drag on earnings at a critical time for financial institutions. The complexity and costs are amplified for large, geographically diverse, financial institutions.

At the heart of the problem is the balance between signal and noise. Too much noise in the form of false positives increases costs for the bank as they need to investigate the suspicious activity reports. Too little signal means that the bank might miss criminals triggering a number of negative regulatory outcomes ranging from fines to physical observation.

Compared to its peer banks, HSBC’s AML infrastructure is the world’s best - so much so that it is often treated as a reference workflow. Current AML processes typically have hand-coded or, in the case of more sophisticated players like HSBC, machine-coded money laundering rule scenarios to evaluate each transaction for each geography or type of business. Subject matter experts encode established patterns such as high repetitive number of small transactions (structuring), money flows in and out of high-risk countries, etc.

Sophisticated institutions like HSBC go one step further and create segments of customers. The segments that emerge, however, are typically static and coarse since they only take into account a limited set of factors about each customer or entity. Moreover, they are done manually and separately for customer data and transaction data and have trouble effectively capturing the complex feature interactions.

This has a massive effect downstream as the rules are applied on a set of coarsely defined segments. When AML rules are triggered, the offending transactions must be investigated and adjudicated - creating the Hobbesian choice of regulatory exposure or throwing bodies at the problem.

HSBC saw past this false choice and transformed its AML process using enterprise grade AI.

Stop Changing the Rules, Change the Game

HSBC needed a principled approach to rule creation and segment creation to get their false positives under control. They also needed to be sure that any effort strengthened their ability to detect real criminals (false negatives). They also needed this approach to integrate with existing processes.

HSBC applied Ayasdi’s analytical approach to transform their process. We have written on that – so we will focus on the application of that process to AML in this case.

• First, they used unsupervised learning techniques to automatically create groups of customers and customers-of-customers (in correspondent banking situations).
• Second, for each group discovered earlier, using historical data and supervised/predictive approaches they ranked the likely launderers.
• Third, the solution created a justification for both the segments and the predictions.
• Finally, the system integrates with other applications (Act) and learns from new data (Learn), alerting the risk team to re-evaluate as the criminals evolve their methods.

Let’s start with the concept of unsupervised discovery. In most AML processes, subject matter experts make the determination of the rules or scenarios around what should trigger an investigation. HSBC put their subject matter experts onto this problem, however, when information was limited (as in the case of KYCC where only the transaction data was available), the thresholds for triggering the rules were creating far too many false positives. One major source of these problems was that they used K-Means clustering (with a Bayesian Information Score to determine the number of clusters). While K-Means clustering is a wonderful algorithm, it suffers from the following issues:

1. It requires O(n^2) distance computations per step. For a dataset with 10 million customers, this translates to 100 trillion distance computations and 800TB memory/disk. Note that the 10 million number is rather modest.
2. Since computing segments on the entire dataset is not feasible, the team down-sampled the data to a few thousand customers, to be able to fit the data in memory and process it in a reasonable amount of time.
3. Sampling combined with the inherent instability of K-Means meant that the segments were not representative of the dominant modes in the data.
4. Finally, since K-Means is an unstable algorithm, the general recommendation is to not use it with high dimensional data. This meant that the team had to select a small handful of features to use in the segmentation.

HSBC and Ayasdi used Topological Data Analysis (TDA) and machine learning (ML) to automatically assemble self-similar groups of customers and customers-of-customers. This exercise was done entirely unsupervised, with Ayasdi’s software making the selection of the appropriate algorithms, creating candidate groups and tuning the scenario thresholds within those groups until the optimal ones were identified. In this case, the platform automatically normalized the data columns and combined multi-dimensional scaling and single linkage clustering algorithms to create the topological model. This was then passed through an agglomerative hierarchical clustering algorithm which was optimized to produce balanced segments.

As with most ML/AI solutions, ours has an insatiable appetite for data. The more data sources that are available, the better the grouping. This includes, customer and transaction data. Unlike most ML/AI solutions Ayasdi is particularly adept at high dimensional data and where more of that data tends to break most ML/AI solutions, we eat it up.

What was important in the HSBC work was that Ayasdi was able to work with unlabeled data . What that means is that the SWIFT traffic activity are enough to produce meaningful segments. This increases the number of data sources considerably and brings into play a bank’s customers customer (KYCC).

These new groups were assembled in a completely different fashion than the previously constructed groups. One key feature was that the groups were more evenly distributed than the previously constructed groups.

As a general rule, the more powerful the model the more of a black box it is. This black-box problem is well known to the financial services community and is particularly problematic in the AML setting since the banks need regulatory approval for these models. HSBC’s savvy Internal Model Review team understood that a black box was an unacceptable option – no matter how powerful or explanatory. They required transparency.

Fortunately, for HSBC, Ayasdi was able to deliver both explanatory power and explainability. In this case, the explanation came in the form of a decision tree that could explain the new segmentation and rules to both the internal model review board and the external regulatory stakeholders. The general pattern of using simpler models such as decision trees to understand more complicated models is going to be critical as we build trust with heavily augmented or automated solutions.

The final step to creating a deployable solution was creating the appropriate thresholds by which to trigger investigations.

The bank has over 40 distinct money laundering scenarios, each of which have several parameters that must be tuned for each customer segment. Using the process that the bank had previously used on its existing segmentation, HSBC tuned each of these parameters for each of the scenarios tested based on the updated segmentation. As in the prior production approach, all parameters were tuned to ensure that every Suspicious Activity Report (SAR) previously issued would still have been flagged by the updated rules.

With this final step all of the elements of a transformative, operational AML solution exist.

It has unsupervised learning to create the segments, uses supervised learning to predict what groups new customers should go into, justifies those groups and changes to those groups, feeds the transaction management system and learns from new data - adapting to the changing behavior of criminals.

The ultimate goal of this new AML workflow (rule engine + transaction monitoring) is a highly principled, defensible, ranked list of transactions/entities to investigate.

HSBC by the Numbers

As noted, HSBC is a sophisticated player in the AML space and was already creating segments. Further, they were dedicated to continuous improvement and were willing to invest to achieve those ends.

The bank had an internal goal to improve operational efficiency in the elusive KYCC area by 3% with a stretch goal of 5%. They had a timeline to achieve this that was measured in years. They generated in excess of 800,000 alerts per month and had more than 5,000 members on their investigative team – a number that is expected to grow to 7,000 – reflecting their commitment to the AML challenge.

The initial work involved SWIFT data for over 70 million transactions. Ayasdi had access to the same features as the client, however, our technology was able to use more of them and do so more intelligently. Ultimately, Ayasdi discovered close to 1,000 features whereas the bank’s current technology was only able to use less than 10 features. While the project ultimately employed over 120 features, this 12X increase drove far superior performance. The increased feature count was a function of being able to incorporate and create features on transactional data (type, direction, value), customer data (geographical, chronological) and risk data.

By applying AI to the problem, the bank was able to achieve reductions in investigative volume of more than 20% while simultaneously lowering their regulatory exposure by discovering new risk segments that had gone unnoticed.

While the ROI on the project is confidential, the savings to the bank is into the tens of millions of dollars per year. Perhaps more importantly, the bank now has the most sophisticated system in the world and is able to represent that work to their external stakeholders.

The initial project took twelve weeks and involved a team of two data scientists and a project manager from Ayasdi, and a team of two domain experts from the bank. The twelve weeks included data access, transforms, etc. Once completed the initial “run” from a compute perspective took eight hours to go from raw data to the segments and predictive segment models. The plan is to re-run every month to ensure that the segmentation captures the modes of the data as it evolves. The run time of eight hours compares to six months of manual segmentation and review by the business unit leaders.

Summary

In the evolving world of regulation there is considerable pressure to achieve a target of “zero failure.” The approaches that banks have employed to date, both historically and in response to enhanced regulatory scrutiny is fundamentally driven by humans. The complexity of the AML challenge, however, is not well suited to the hand-coded approach that dominates most institutional thinking.

There is an opportunity to apply sophisticated techniques from AI to this mission-critical regulatory function that can dramatically improve efficiency while simultaneously reducing regulatory exposure. Furthermore, this new generation of technology provides exceptional transparency on what is driving the rule sets and segmentation.

To find out how to leverage AI for your AML challenges reach out to us at sales@ayasdi.com to arrange for a demonstration.

Anti-Money Laundering is a particularly challenging area of regulation for banks and even more so for large, geographically diverse institutions. What makes AML such a difficult problem to solve is that it involves complex data, detailed workflows, and significant human involvement.

In short it is perfect for AI.

This is what HSBC determined in their work with Ayasdi on a notoriously difficult problem - that of bank’s customer’s customers, also known as KYCC (know your customer’s customer). The vast majority (~95%) of AML investigations do not result in a suspicious activity report (SAR). This is called a ‘false positive’ - transactions that are flagged for investigation, but are not suspicious based on a deeper review.

Working with Ayasdi, HSBC was able to achieve a significant reduction in false positives (more than 20%) while keeping their suspicious activity reports at the same number.

The Modern AML Challenge

AML is a massive problem and is estimated to be around $1 to $2 trillion per year and growing. Given that the vast majority of those transactions end up funding bad people doing bad things, governments have stepped up their regulatory requirements, oversight and fines.

Banks have scrambled to respond to the increasing regulatory burden. The result is that the cost of compliance is increasing 50% year-over-year and quickly becoming a drag on earnings at a critical time for financial institutions. The complexity and costs are amplified for large, geographically diverse, financial institutions.

At the heart of the problem is the balance between signal and noise. Too much noise in the form of false positives increases costs for the bank as they need to investigate the suspicious activity reports. Too little signal means that the bank might miss criminals triggering a number of negative regulatory outcomes ranging from fines to physical observation.

Compared to its peer banks, HSBC’s AML infrastructure is the world’s best - so much so that it is often treated as a reference workflow. Current AML processes typically have hand-coded or, in the case of more sophisticated players like HSBC, machine-coded money laundering rule scenarios to evaluate each transaction for each geography or type of business. Subject matter experts encode established patterns such as high repetitive number of small transactions (structuring), money flows in and out of high-risk countries, etc.

Sophisticated institutions like HSBC go one step further and create segments of customers. The segments that emerge, however, are typically static and coarse since they only take into account a limited set of factors about each customer or entity. Moreover, they are done manually and separately for customer data and transaction data and have trouble effectively capturing the complex feature interactions.

This has a massive effect downstream as the rules are applied on a set of coarsely defined segments. When AML rules are triggered, the offending transactions must be investigated and adjudicated - creating the Hobbesian choice of regulatory exposure or throwing bodies at the problem.

HSBC saw past this false choice and transformed its AML process using enterprise grade AI.

Stop Changing the Rules, Change the Game

HSBC needed a principled approach to rule creation and segment creation to get their false positives under control. They also needed to be sure that any effort strengthened their ability to detect real criminals (false negatives). They also needed this approach to integrate with existing processes.

HSBC applied Ayasdi’s analytical approach to transform their process. We have written on that – so we will focus on the application of that process to AML in this case.

• First, they used unsupervised learning techniques to automatically create groups of customers and customers-of-customers (in correspondent banking situations).
• Second, for each group discovered earlier, using historical data and supervised/predictive approaches they ranked the likely launderers.
• Third, the solution created a justification for both the segments and the predictions.
• Finally, the system integrates with other applications (Act) and learns from new data (Learn), alerting the risk team to re-evaluate as the criminals evolve their methods.

Let’s start with the concept of unsupervised discovery. In most AML processes, subject matter experts make the determination of the rules or scenarios around what should trigger an investigation. HSBC put their subject matter experts onto this problem, however, when information was limited (as in the case of KYCC where only the transaction data was available), the thresholds for triggering the rules were creating far too many false positives. One major source of these problems was that they used K-Means clustering (with a Bayesian Information Score to determine the number of clusters). While K-Means clustering is a wonderful algorithm, it suffers from the following issues:

1. It requires O(n^2) distance computations per step. For a dataset with 10 million customers, this translates to 100 trillion distance computations and 800TB memory/disk. Note that the 10 million number is rather modest.
2. Since computing segments on the entire dataset is not feasible, the team down-sampled the data to a few thousand customers, to be able to fit the data in memory and process it in a reasonable amount of time.
3. Sampling combined with the inherent instability of K-Means meant that the segments were not representative of the dominant modes in the data.
4. Finally, since K-Means is an unstable algorithm, the general recommendation is to not use it with high dimensional data. This meant that the team had to select a small handful of features to use in the segmentation.

HSBC and Ayasdi used Topological Data Analysis (TDA) and machine learning (ML) to automatically assemble self-similar groups of customers and customers-of-customers. This exercise was done entirely unsupervised, with Ayasdi’s software making the selection of the appropriate algorithms, creating candidate groups and tuning the scenario thresholds within those groups until the optimal ones were identified. In this case, the platform automatically normalized the data columns and combined multi-dimensional scaling and single linkage clustering algorithms to create the topological model. This was then passed through an agglomerative hierarchical clustering algorithm which was optimized to produce balanced segments.

As with most ML/AI solutions, ours has an insatiable appetite for data. The more data sources that are available, the better the grouping. This includes, customer and transaction data. Unlike most ML/AI solutions Ayasdi is particularly adept at high dimensional data and where more of that data tends to break most ML/AI solutions, we eat it up.

What was important in the HSBC work was that Ayasdi was able to work with unlabeled data . What that means is that the SWIFT traffic activity are enough to produce meaningful segments. This increases the number of data sources considerably and brings into play a bank’s customers customer (KYCC).

These new groups were assembled in a completely different fashion than the previously constructed groups. One key feature was that the groups were more evenly distributed than the previously constructed groups.

As a general rule, the more powerful the model the more of a black box it is. This black-box problem is well known to the financial services community and is particularly problematic in the AML setting since the banks need regulatory approval for these models. HSBC’s savvy Internal Model Review team understood that a black box was an unacceptable option – no matter how powerful or explanatory. They required transparency.

Fortunately, for HSBC, Ayasdi was able to deliver both explanatory power and explainability. In this case, the explanation came in the form of a decision tree that could explain the new segmentation and rules to both the internal model review board and the external regulatory stakeholders. The general pattern of using simpler models such as decision trees to understand more complicated models is going to be critical as we build trust with heavily augmented or automated solutions.

The final step to creating a deployable solution was creating the appropriate thresholds by which to trigger investigations.

The bank has over 40 distinct money laundering scenarios, each of which have several parameters that must be tuned for each customer segment. Using the process that the bank had previously used on its existing segmentation, HSBC tuned each of these parameters for each of the scenarios tested based on the updated segmentation. As in the prior production approach, all parameters were tuned to ensure that every Suspicious Activity Report (SAR) previously issued would still have been flagged by the updated rules.

With this final step all of the elements of a transformative, operational AML solution exist.

It has unsupervised learning to create the segments, uses supervised learning to predict what groups new customers should go into, justifies those groups and changes to those groups, feeds the transaction management system and learns from new data - adapting to the changing behavior of criminals.

The ultimate goal of this new AML workflow (rule engine + transaction monitoring) is a highly principled, defensible, ranked list of transactions/entities to investigate.

HSBC by the Numbers

As noted, HSBC is a sophisticated player in the AML space and was already creating segments. Further, they were dedicated to continuous improvement and were willing to invest to achieve those ends.

The bank had an internal goal to improve operational efficiency in the elusive KYCC area by 3% with a stretch goal of 5%. They had a timeline to achieve this that was measured in years. They generated in excess of 800,000 alerts per month and had more than 5,000 members on their investigative team – a number that is expected to grow to 7,000 – reflecting their commitment to the AML challenge.

The initial work involved SWIFT data for over 70 million transactions. Ayasdi had access to the same features as the client, however, our technology was able to use more of them and do so more intelligently. Ultimately, Ayasdi discovered close to 1,000 features whereas the bank’s current technology was only able to use less than 10 features. While the project ultimately employed over 120 features, this 12X increase drove far superior performance. The increased feature count was a function of being able to incorporate and create features on transactional data (type, direction, value), customer data (geographical, chronological) and risk data.

By applying AI to the problem, the bank was able to achieve reductions in investigative volume of more than 20% while simultaneously lowering their regulatory exposure by discovering new risk segments that had gone unnoticed.

While the ROI on the project is confidential, the savings to the bank is into the tens of millions of dollars per year. Perhaps more importantly, the bank now has the most sophisticated system in the world and is able to represent that work to their external stakeholders.

The initial project took twelve weeks and involved a team of two data scientists and a project manager from Ayasdi, and a team of two domain experts from the bank. The twelve weeks included data access, transforms, etc. Once completed the initial “run” from a compute perspective took eight hours to go from raw data to the segments and predictive segment models. The plan is to re-run every month to ensure that the segmentation captures the modes of the data as it evolves. The run time of eight hours compares to six months of manual segmentation and review by the business unit leaders.

Summary

In the evolving world of regulation there is considerable pressure to achieve a target of “zero failure.” The approaches that banks have employed to date, both historically and in response to enhanced regulatory scrutiny is fundamentally driven by humans. The complexity of the AML challenge, however, is not well suited to the hand-coded approach that dominates most institutional thinking.

There is an opportunity to apply sophisticated techniques from AI to this mission-critical regulatory function that can dramatically improve efficiency while simultaneously reducing regulatory exposure. Furthermore, this new generation of technology provides exceptional transparency on what is driving the rule sets and segmentation.

To find out how to leverage AI for your AML challenges reach out to us at sales@ayasdi.com to arrange for a demonstration.