When credit card giant American Express began offering bank accounts for the first time last year, it had a foundation of fraud detection to bring to an entirely new product arena. That meant in some cases, the company could port over AI and machine-learning models used to spot phony identities or dodgy transactions for its credit card products to its consumer and business checking accounts. But it’s been a process, and now, AmEx plans to invest in bringing additional AI techniques used to protect against credit card fraud to its banking products.
“We have models which run to detect whether it's you or whether somebody else is logging into your account. Very straightforwardly, we transferred it to the banking product,” said Abhinav Jain, vice president for Global Fraud Decision Science at AmEx, who is responsible for the company’s fraud detection models. “We had at least the technical side of the model ready to prevent this kind of fraud happening for a customer.”
The fraud models are designed to recognize odd behavior or suspicious patterns of activity that are not typical of a particular customer. After kicking off a business checking account product in October followed by a consumer-aimed checking account in February, AmEx’s models already are picking up on an emerging trend.
“Fraudsters show up very quickly once you launch a new financial product. And one of the patterns that we have seen starting to emerge very quickly is fraudsters attempting to take over a customer's profile and apply to open a checking account in their name – so, identity theft,” said Ana Palaghita, vice president and head of Banking Fraud and Deposit Risk at AmEx.
“This has been a typical pattern that I've seen just in the month and a half to two months that we've been in-market,” said Palaghita, who leads the company’s fraud strategy team and works closely with Jain’s data science and ML modeling group.
Business checking accounts are attracting their own distinct fraud patterns, Palaghita said. “On the business side, we've seen different patterns. We've seen more of the synthetic identity where fraudsters are attempting to open accounts with businesses that are either not real or they're just a front in an attempt to deposit fake checks and extract money that way.”
Some companies are using machine- and deep-learning models to detect fraudulent behavior by entities sanctioned by the U.S. Treasury Department in relation to the war in Ukraine. AmEx sources said they would not address the war in Ukraine or sanctioned entities in more general terms.
False positives and speedy data
The AmEx fraud detection models react and optimize automatically by adjusting how they weigh certain data points in the decision-making process, for instance.
“In the places we’re seeing more fraud, they will be more aggressive,” said Jain, explaining that the models self-calibrate by attributing higher probability of fraud to certain data elements that are reflective of other recent fraud. They might weigh more heavily particular geographic regions, currencies or types of products associated with attempted transactions, Jain said.
Fraudsters show up very quickly once you launch a new financial product.
Of course, without proper tuning, automated fraud detection systems can be overly sensitive, halting legitimate transactions and annoying customers in the process. However, if a transaction sets off a fraud alert, AmEx doesn’t necessarily stop the transaction right away. Sometimes the company puts transactions on hold, then sends a text or email alert to the customer asking whether they’ve made the purchase or taken the action in question. That information gets fed back into the fraud model to optimize it.
“As soon as we identify this was a false positive, that information also feeds back in real time to the model,” Jain said.
Indeed, financial services companies increasingly are reliant on real-time data and data processing to run fraud detection models. Not only do they need sophisticated machine learning to keep up with evolving fraud approaches, but they need speedy data processing on the back end to ingest fresh data inputs into fraud models and ensure those models recognize and react to quickly-morphing fraud patterns.
“Within milliseconds, we should be able to link that IP address, that email address, to the fraud database. If a second attempt comes from similar entities, we should be able to stop it,” Jain said.
A new crop of database startups is emerging to help financial institutions and other enterprises grab real-time data and make use of it immediately to update machine-learning models for things like ecommerce recommendations and dynamic pricing.
Because AmEx has fewer “silos and fences” separating the data and technology systems behind its product lines than other financial services companies, it has been able to more readily transfer machine-learning models for use across products than other companies have, said Palaghita, who worked in various roles at Capital One since 2007 before joining AmEx in 2021.
That capability “is not necessarily something that every financial institution can do so quickly and seamlessly,” she said.
AmEx is still in the process of porting fraud models used on its credit card side over to its banking side, Jain said. Up next: incorporating time series data for neural networks used to detect identity fraud or online account takeovers. “Getting the time series view helps the model further drill down into exactly what is fraud and what is not,” Jain said. “There are neural network algorithms that help us do that.”
In data used for more standard modeling, each transactional data point used to train a model is assessed individually as good or bad. Models that incorporate time series data consider history and context by viewing not only the current transaction, but a customer’s previous 10 or 20 transactions. For example, if a customer’s last few recent transactions happened at a mall, but soon after a transaction is attempted in a faraway location, a model looking at time series data would be able to detect it as fraudulent.
“We are investing heavily into those types of algorithms, and we've had really good success in the initial launches that we’ve had, and we want to expand that more and more to different types of fraud — and even start leveraging it much more,” Jain said