Using Machine Learning to Optimize Chargeback Re-presentments

Ken Bui
Ken Bui

Latest posts by Ken Bui (see all)

Every year, enterprises lose billions of dollars to ecommerce chargeback costs. New fraud techniques based on “social engineering” are contributing to the problem.

Dr. Selena Zhong, Senior Data Scientist Lead at Microsoft, recently spoke on these issues at the Real Machine Summit, hosted by Feedzai and Money 20/20 in Las Vegas .

To see Dr. Zhong’s full talk, check out the video below.

Social engineering is an umbrella term that refers to psychological manipulation; in this case, taking advantage of humans as the weak link in data security systems. Because social engineering in fraud is spiking worldwide, it’s worth looking at what companies can do about it.

Dr. Zhong discussed social engineering-based fraud at Microsoft. Here’s an example of how it works. A fraudster calls up the support desk with just enough information to start a conversation:

Support: May I have your phone number? Fraudster: Is the one you have 1234567? Support: No, it’s 123456-8
Fraudster: Thank you.

The fraudster hangs up and calls back. This time when the support agent asks for the caller’s phone number, the fraudster says “123456-8.” The conversation continues:

Support: May I have the last four digits of your card? Fraudster: Is it Visa ending 1234?
Support: No, it’s Mastercard ending 2345. Fraudster: Thank you.

The fraudster hangs up. Now he has the phone number and last four digits of the payment card. The next time he calls, he asks to change the password. The account has now been compromised.

Dr. Zhong explained that the fraud vectors for different businesses vary greatly, and not all involve social engineering. Breaches and older means of account compromise (phishing, malware, brute force) are still regularly used. What they all have in common is that fraudulent e-commerce charges too often end up back at the merchants in the form of payment card chargebacks.

For Microsoft, these charges add up to many billions of dollars a year. Dr. Zhong described how the representment phase of the chargeback process offers opportunities for companies to minimize chargeback revenue loss.

Chargebacks happen when a card-issuing bank has approved a customer dispute on a fraudulent charge, and passed the charge on to the merchant. The merchant now has a choice to accept the charge and lose revenue, or refuse it and represent it to the bank if they think there are compelling reasons to do so. If the bank again denies the charge and returns it to the merchant, the merchant can let the transaction go, or take it to arbitration. Either way, representment is costly.
In this representment ping-pong game, businesses need an effective way to help decide which chargebacks are most likely to “win.” This is where machine learning can help.

Microsoft is using a machine-learning (ML) based system to predict the probability that certain chargebacks will result in a win. It’s a powerful alternative to an older, manual system that forced managers to make decisions based on arbitrary rules.

The outcomes have been very positive. With the rule-based, manual system, Microsoft won only about 10% of their total chargebacks. With the ML-based solution, they now win closer to 25%. This is a lift of 150%, and it means a difference of many millions of dollars for the company.

The Microsoft story is a just one, big-enterprise example of how ML systems can help businesses optimize chargeback representments.

Learn more about machine learning for fraud detection here.