TRANSACTION RISK BEHAVIOR BASED SEGMENTATION AND THRESHOLD FINE-TUNING
Sutra improved the productivity rate of alerts for one of the largest bank in Portugal and further reduced false positives by ~ 17% while capturing 100% of SAR customers.
Alert productivity was improved by 17% without compromising on the capture rate of actual SAR customers
Client: A leading Bank in Portugal
The client is one of the largest banks in Portugal. The bank had instituted an anti-money laundering solution based on a legacy system which has defined customer segments based on marketing/ risk levels. The client was looking to enhance the productivity of the investigation team. The current AML system was generating a higher number of false positive alerts. This was increasing the backlog of alerts for the investigation team resulting in non-identification of the true positives. Together with the surge in number of alerts, this was leading to a higher operational load on the investigation team
Sutra FCC Solution:
New customer segments were introduced – based on customer transaction behavior and risk level. An automated process for threshold fine-tuning was recommended. This process was applied at a scenario and segment level to reduce the false positive alerts. And it was found to be the best-fit approach to manage legacy scenarios.
Transaction risk behavior based segmentation
A model was created using last few months data and applying heuristics segmentation to identify transaction behavior based segments using 250+ factors. Executing k-means and stepwise clustering to extract meaningful segments based on key themes and patterns. Business rules were developed for the newly generated segments to classify customers at any point in time.
Threshold fine-tuning of scenarios
The automated process leveraged the underlying transaction distribution across different scenario parameters normalized by the customer segments. Insights were generated for user input and automated thresholds to analyze the false positives reduction and productivity rate at the scenario and segment levels.
Cross-scenario optimization
There’s a major problem in threshold fine-tuning. The scenario-based STR cases and overall customer level catch rate have a very high overlap. It means that, for example, more than one scenario captures the same customer with different degree of abnormal behavior.
So, to overcome this scenario combination, their thresholds are passed through an optimization engine to identify the most optimal set of thresholds. To understand the span of alerts and overlap, the team used brute force to test 190,000+ combinations of the threshold for scenarios to identify best performing threshold combination. This minimizes overlap of the same customers across scenarios while maximizing the capture rate (maximizing SAR customers).
Business Results – Significant improvements in alert productivity & profiling accuracy
The process generated 14 transaction risk behavior segments which were then sliced with the customer risk levels and demographic segments resulting in 90 odd segments. This further helped in understanding the granularity of the transaction distribution during the threshold fine-tuning process.
The engagement strived to improve the overall productive rate of alerts thereby improving effectiveness of the investigation staff. Despite the high threshold values for the scenario parameters in the current AML system, the solution was able to reduce the number of false positive alerts by ~ 17% while ensuring no SARs were being compromised.
About Us
Project Headed by: Abhishek Gupta, Sutra Management Consultancies
Presented by: Jay Modi, Sutra Management Consultancies
Lovish Kanther, Sutra Management Consultancies