Fighting ecommerce fraud can be a lonely game.
When everything goes right and chargebacks are few and far between, you’re just doing your job. If things go wrong and fraudulent orders increase, well then there is a problem. And it’s your problem.
So, it’s nice, now and then, for fraud fighters to gather together, trade best practices, maybe a horror story or two and listen to an acknowledged expert talk feedback loops, features, labels and AUROC/AUPRC.
Signifyd’s second Payment Fraud Meetup served the purpose as 50 or so fraud and payment professionals from companies like Uber, WePay and Vodafone came together at Signifyd’s Silicon Valley headquarters. They were there to hear Google data scientist Swami Vaithianathasamy run through the strengths and weaknesses of machine learning when it comes to detecting fraudulent online orders.
There were valuable takeaways and we’ll get to that. But there was also an opportunity for catharsis and commiseration — because if one theme came through loud and clear in Vaithianathasamy’s talk, it was that designing the ideal machine model for fighting fraud is difficult work.
“You cannot wish away fraud. It’s a problem that will keep growing. It’s not going to go away,” Vaithianathasamy told the gathering. “I always say, the time when we think we know everything about fraud is when we are going to lose.”
Ecommerce Fraud Protection Pitfalls and Considerations
Vaithianathasamy proceeded to run through some of the considerations and pitfalls in designing systems to detect and prevent fraud:
- Some systems become so complicated that they cease being useful.
- Others are so conservative they actually harm the business by declining as fraudulent orders that actually would be legitimate, revenue-producing sales.
- A good model is nothing without good data.
- A good model is good for only so long. Invest in analytic tools and be vigilant about monitoring your models.
- Be careful when creating goals. Reducing false positive sounds good, unless you ship everything and bankrupt the business. Conversely, you could eliminate fraud by declining every order and go broke that way.
- Anomalous behavior is not always bad behavior. Legitimate customers make anomalous orders. How often do you buy a $1600 4K HDTV?
- There is a tradeoff between freshness and comprehensiveness when it comes to fraud patterns. Machines detect behavior immediately. But detecting the result of that behavior can lag behind. Consumers, for instance, might not notice fraudulent activity on their accounts for weeks, creating a lag in learning of a fraudulent order.
Vaithianathasamy’s technical dive was illuminating, but one non-technical theme repeated itself throughout his talk: Getting fraud prevention right requires constant vigilance and nimbleness. Fraudsters are constantly reacting to new models and new models are constantly being designed to react to fraudsters’ evolving tactics.
“Where there is money to be made, there is opportunity,” Vaithianathasamy said of fraudsters’ motivation.
And make no mistake: Fraudsters are brazen and determined, as was made clear by the recent Equifax breach, a heist of 143 million consumers’ personally identifiable information.
Vaithianathasamy joked during his talk that he’d be charged with a felony if he didn’t mention the breach — one of the largest ever. But on a serious note, he added that the Equifax affair leads to one unavoidable conclusion: “There are way too many gaps in the system.”
It is the gaps, and fraudsters’ determination, ingenuity and brazenness that keeps Avani Sharma up at night. Sharma is a Silicon Valley data scientist who works in the payments field. She says that despite all the preparation and technology, it’s the prospect of a big fraud loss that weighs on her mind.
But on the flip side, it is preventing that loss that jazzes her. It’s why she changed careers from computer programming to data science. It’s why she does what she does.
“It’s very interesting,” she says of her fraud-prevention work. “It’s very tangible.”
Yes, tangible and about working to “catch the bad guys,” chimed in Sharma’s co-worker Srividya Sunderamurthy, who also works in payments.
Fraud protection field provides fulfillment
Sunderamurthy said there is great satisfaction in “learning about existing fraud patterns, and viewing patterns coming in, and thinking about how you can build systems” to prevent fraud.
“It’s different every day,” she said. “We have to dig into the different layers of what happened from a risk perspective.”
Yes, Sunderamurthy said, the specter of chargebacks is stressful. But that’s only half the equation. There is also the worry of creating a system that declines orders that actually were placed by legitimate customers — paying customers, or customers who would have paid, had their order been approved.
“You want to make sure you don’t decline everything,” Sunderamurthy said.
They are definitely common concerns in the fraud prevention field. And that’s the thing about an event like Signifyd’s meetup: Professionals like Sunderamurthy are reminded that they are not alone. There are many fellow travelers worrying about the same things.
In fact, Sunderamurthy said she got more than textbook learning out of the meetup. She also saw the gathering as a chance to meet fellow professional and to better understand what was going on in the field.
A field that no doubt felt a little less lonely that night.
Want to continue the discussion around online fraud and how humans and machines are teaming up to prevent it? Join the Fraud Squad LinkedIn community to network with your peers and to share your ideas and questions.
Photo of keyboard and magnifying glass by iStock. Photo of Swami Vaithianathasamy by Mike Cassidy.
Mike Cassidy is Signifyd’s lead storyteller. Contact him at firstname.lastname@example.org; follow him on Twitter at @mikecassidy.