Cutting Throught the Volume of False Positives
For financial institutions, uncovering anti-money laundering (AML) crimes can be an exhaustive process, and all the more so when 95 percent, on average, are false positives. At Valley Bank, predictive analytics helped comb through millions of transactions across three-quarters of a million customers, but it took weeks to create models manually.
“Our team has to work through a lot of noise to find productive information, which results in staffing challenges and a challenge in keeping staff engaged,” explained Jennifer Yager, Director of Financial Crimes Compliance. “When so many are false positives, you run the risk of missing the items that provide actual value.”
Building 100+ Models Automatically
When looking for ways to automate its fraud detection, Valley Bank found most solutions primarily offered segmentation and were difficult for non-data scientists to use. In DataRobot AI Platform, the bank discovered a unified environment for optimization across the entire AI lifecycle. They found that the AI platform could help cut through the volume of false positives in a way that was realistic for the bank to manage.
Partnering with the DataRobot team, the bank designed a solution to integrate with existing systems before proceeding with a proof of value (POV). From the start, the bank’s model risk management team was involved to help validate the results.
The bank fed predictions into its AML case management system. Alongside DataRobot data scientists, Valley Bank built and validated more than 100 models with a robust backtesting strategy and generated 175 features via DataRobot automated Feature Discovery.
In the POV, alerts that the bank identified as having a low likelihood of being escalated to a case were tested against actual alerts worked to test the predictions against actual outcomes. Following the success of the POV, the bank moved the model into production and began hibernating alerts that fell below the threshold for likelihood of escalation.
As part of an ongoing quality assurance check, the bank chose not to hibernate all transactions below the hibernation threshold; rather, it put 20 percent back in as a form of ongoing below-the-line (BTL) testing.
“We’ve been doing this for a year now with the AI platform, so we’re not seeing any weaknesses in the data.”
During the trial, the bank reduced false positives by more than 30 percent. For validation, the bank’s model risk management team recreated those models successfully in the platform.
Documentation, available with one click, helped gain the confidence of both the model risk management team and the Office of the Comptroller of the Currency (OCC), plus supported audits.
Automated End-to-End Process
During the POV, DataRobot trained and onboarded four bank users within just two weeks. And when the bank’s only data scientist left the role, the team could still operate productively – a testament to the platform’s ease of use.
“DataRobot AI Platform has hundreds of prebuilt models. We just feed in the data and it suggests models automatically,” explained Chris Mendoza, Director of Financial Crimes Technology. “The beauty of the platform is, once the model is deployed, it’s there.”
Previously, they spent weeks coming up with new models or retraining models. Now, DataRobot AI Platform does that in a day.
Mendoza also appreciates built-in safeguards that help spot data quality issues immediately.
Leading Innovation within the Bank
Since deploying the platform, Valley Bank saves approximately 22 percent in total alert volume every month. Additionally, the number of alerts escalating to case increased by three percentage points.
The efficiency gained with the DataRobot AI platform enables the bank to achieve more with a lean team.
“It’s allowed us to create more scale,” Yager said. “I think it’s created a lot of pride in the department in the fact that we have implemented and are leveraging this machine learning model successfully. And we can feel good that we’re leading innovation within the bank in this sense.”
Along the way, the support of the DataRobot team made all the difference.
“In the POV, DataRobot Customer Facing Data Scientists did most of the work for us,” Mendoza said. “And now, every time we have questions or challenges, they’re always there.”