Western Union Clamps Down on Fraud with ML
Money transfers are a rich target for fraudsters, who have devised all sorts of creative ways to steal the funds. But thanks to a sophisticated machine learning system running atop a large Hadoop cluster, Western Union has taken a sizable bite out of the criminals’ take.
When one needs to send money, the name that often comes to mind is Western Union. The 167-year-old company offers ways to send money in person, over the phone, and across the Internet. In 2017 Western Union successfully wired more than $300 billion on behalf of clients to recipients in more than 200 countries and 130 currencies. On average, it completed 32 transactions every second.
Unfortunately, such a large financial services operation also attracts its share of fraud. As the notorious bank robber “Slick” Willie Sutton pointed out in the 1950s, criminals will naturally go “where the money is.” But thanks to the Internet, bad guys don’t even need to show up in person to rob banks anymore. Instead, they can use computers and telephones to steal other people’s hard-earned money through a variety of schemes – some sophisticated, some not quite so much.
Depending on the specific sector of the financial services industry, fraud losses can account for upwards of 5% of the value of transactions. In the sector that Western Union works in, the benchmark figure is about 1.2%. Those numbers might seem small, but they represent billions of dollars in losses that financial institutions are forced to absorb as a cost of doing business.
Because it impact profits so directly, financial services firms have a huge incentive to battle fraud. Most realize they can never completely eradicate fraud (the only way to do that is to stop lending or moving money), so instead they build teams who focus specifically on fighting fraud.
Digital Fraud Fighters
Like other firms, Western Union employs a multi-pronged attack to thwart fraud. When money is sent in person, there are ways to check the identities of the money transfer participants, which helps to keep fraud down.
But some of those authentication techniques can’t be used for validating transactions that occur digitally, including the money transfer methods the Denver, Colorado-based company starting making available in 2011, such as funding of credit card and debit cards, transfers to online accounts, mobile app pay services, and digital bill payments.
This new digital money transfer service posed a serious challenge, says Ernesto Boada, the vice president of digital engineering for Western Union.
“We need to make sure not only that the individual is the right individual from a legal and compliance” perspective, Boada says, “but we need to make sure that the one being paid is the right one, too.”
Fraudsters try to pull a variety of scams that Western Union needs to look out for, lest it be held liable for the lost money. “It’s very complex,” Boada tells Datanami. “There’s a bunch of schemes, unfortunately.”
One of those is account takeover, where somebody gets unauthorized access to a credit card account and starts wiring themselves money. Western Union also on the lookout for email scams, such as those where victims are tricked into sending money to get access to a large inheritance. There are also cases where children or grandchildren may avail themselves of relative funds without permission, Boada says.
“We’re very highly regulated in terms of compliance,” he says. “By the time we open digital, we just expanded the potential set of risk. Because by the time that any of these transactions happen — my credit card is stolen or somebody sends money using my credit card…we’re then liable for the money that was sent.”
Enter Data Science
To fight the fraud in its digital money transfer business, Western Union first looked to advanced analytics. Its first approach was based on a rules-based system, according to Tracy Li, the director of decision sciences at Western Union’s digital office in San Francisco, California
“We only had the rules to decide whether we will actually decline or approve the transaction,” says Li, who leads the office’s data science team.
Over time, Li’s team started using more supplicated tools, including scorecards, various SAS tools, open source R and Python tools, the Jupyter data science notebook, and H2O.ai. The diversity of tool usage was considered a strength, Li tells Datanami. “People choose whatever tool they feel comfortable with,” she says.
As Western Union’s data science program evolved, so too did its makeup. In addition to having classically trained data scientists with PhDs in statistics and mathematics, Western Union also brought folks with business backgrounds into the fold.
In early 2017, Li decided that her data science team needed a common platform to enable people with different backgrounds to collaborate, and that decision led to the selection of Cloudera Data Science Workbench, (CDSW) a software application designed to streamline the development and delivery of machine learning models.
According to Li, the combination of machine learning models developed in CDSW and a 900TB CDH cluster is helping to tamp down the fraud to record levels. “In our system we have more than 100 active machine learning models,” she explains. “We pull all the data from the Hadoop cluster and find the patterns, find the anomalies, and then we …. validate the model and put the model into our production system.”
When a money transfer order is initiated, Western Union routes the request to its Hadoop cluster, which returns its determination of whether the transaction is legitimate or not. “If we see this transaction is suspicious, we just decline the transaction,” Li says.
While the result is a simple “yes/no,” the collection of machine learning models – including logistic regressions, random forests, and some boosting algorithms — analyze a complex host of variables and make many calculations, including how reliable a potential user is and how much risk they pose to the company, Boada says.
“We have hundreds of variables that we calculate from understanding your behavior with us: How long you’ve been with us, where are you actually sending the money form compared to where you live, how often do you send money to someone,” he says. “So there’s a lot of information we gather to try to make a decision.”
This approach has brought dividends for Western Union in the form of dramatically reduced fraud rate. The company declined to state for the record exactly what its fraud rate is, except to say it’s significantly below the industry standard of 120 basis points, or 1.2%. “It’s much, much lower than before,” Li says.
If there’s one downside to Western Union’s approach, it’s an increase in false positives, where the machine learning models incorrectly flag a legitimate transaction as fraud. But that’s a trade-off that the company is willing to take.
“We prefer to take the risk of probably having a false positive,” Boada says. “But we want to make sure that we are not letting bad transactions through, and so far Tracy’s team has been very successful in preventing fraud.”
Machine learning is backwards looking, by definition. Models are trained to identify and react to patterns in the data that are highly correlated with bad things that have occurred in the past. But fraud is a highly dynamic enterprise, and so Western Union does its best to anticipate new fraudulent trends so it can stay in front of the bad guys.
“The fraudsters are really smart, unfortunately,” Boada says. “We actually monitor the dark Web as well and see what’s going on out there, but our success, or the success of Tracy’s team, is to be always forward-looking.”
Western Union is constantly looking to see where fraudsters are trying to game the models by constructing their schemes in such a way as to not trigger the thresholds that Western Union sets for certain variables, such as location of initiating transaction, destination of transaction, amounts, etc.
To get a leg up on these fraudsters, Western Union plans to keep evolving its data science capabilities, including possibly through the use of deep learning technology and neural networks.
“The fraud rate coming to our door is 10%, so then the maximum approval rate should be 90%. It’s very straightforward,” Li says. “But we are not there yet because of false positives. We do impact some good customers. So I definitely think there’s still room to improve, and definitely I think we will keep trying the new technology methods and keep pushing the boundaries.”