Having your credit card declined by a cashier can be a bit embarrassing and a hassle, especially if it’s not your fault or credit card fraud is suspected.
If a sale was blocked by your credit card company because it suspected fraud when you tried to charge a higher amount than usual, the initial decline can send you into a panic that you were the victim of credit card fraud. Calling your bank to resolve a sale that was flagged as suspicious can take time — time that can seem wasted if a legitimate transaction was incorrectly flagged and then blocked.
Thank your bank’s fraud-detecting technology for what data researchers call a false positive. Recently, some researchers at the Massachusetts Institute of Technology studied such transactions and found a way to more accurately flag suspicious activity with new machine-learning techniques.
False positives cost $109 billion more than fraud
False positive rates are as high as 10-15 percent, and only one in five transactions declared as fraud are truly fraud, the researchers reported. The high number of false positives may be costing merchants more than the fraud itself.
A simple solution at the checkout counter when your credit card is denied is to pull out another one and make your purchase. But even if you’re successful with that, the whole debacle may be enough to get you to stop using the card that first declined your purchase.
To your bank, that decision can be costly, even much more than the price of fraud.
When a false positive occurs, the customer may not buy the item they were trying to buy and may stop using that credit card. A 2015 report from Javelin Strategy and Research estimates that only one in five fraud predictions is correct and that errors cost banks $118 billion in lost revenue as customers refrain from using that credit card. The cost of real card fraud is $9 billion, according to the Javelin report.
The same study also reports that 26 percent of shoppers whose cards were declined reduced their shopping at that merchant following the decline, and 32 percent stopped entirely.
There are other costs for merchants when a customer is falsely declined, including the cost of acquiring a customer, such as bringing users to a website and converting them into customers. There’s also potential lost revenue from a customer over a lifetime if they stop shopping there.
Getting better help from computers
Machine learning has been used since the early 1990s to detect financial fraud. Behavior patterns from past transactions, called “features,” are put into models to signal fraud. Swiping your credit card starts the model and if fraud behavior is matched, the sale is blocked.
Sometimes the casher will be able to tell you why, but often they won’t know. It’s then up to you to find another way to pay, call your credit provider for assistance, or leave without the purchase.
Amount and location are the two features most commonly used. For example, if you charge more than $2,000 on one purchase or make numerous purchases on the same day, red flags may go up that your card is being fraudulently used. The same goes if you’re using it in a foreign country for the first time on a trip.
The MIT data scientists expanded on those features with an automated approach that extracts more 237 detailed features for each individual transaction, calling it a Deep Feature Synthesis. These include a user being present during purchases, the average amount spent on certain days at certain vendors, the hour of the day a transaction was placed, and on what days.
For example, if you regularly spend $15 on Friday mornings at Starbucks, the model will consider that if you spend a similar amount at another coffee shop on a Friday morning and will approve the transaction.
Time between transactions can also be weighed. If you buy something with a credit card at 9 a.m. at one place, then 30 minutes later use the same card to buy something 200 miles away, it’s unlikely you’ve traveled that distance in half an hour, and the second purchase is declined as suspected fraud. However, if the second purchase is made online on your phone, then it is likely legitimate.
Pitted against a traditional model used by a bank, the DFS model generated 54 percent fewer false positives. That, along with a smaller number of false negatives detected — actual fraud that wasn’t detected — could save the bank about $218,000.