Fighting financial crime without excluding the underbanked

Register now

One prominent industry executive sees a way to use technology to help struggling immigrant communities in the U.S. who may be inadvertently prevented from accessing the financial system.

Eugene Ludwig, founder and chief executive officer of IBM's Promontory Financial Group, said artificial intelligence — already employed to help identify potential anti-money-laundering activity — is getting smarter, and can now be used to identify vulnerable groups of people who have been incorrectly labeled as high risk.

In an interview with American Banker, Ludwig detailed how Watson Financial Services plans to use its "dynamic customer segmentation tool" to improve the credit underwriting process for middle- to low-income people, helping it to make more nuanced decisions and allowing banks to offer financial products cheaper for these populations.

Following is an edited transcript of the interview:

You recently said that artificial intelligence could be used for financial inclusion in addition to reducing false positives in anti-money laundering monitoring. What did you mean by that?

EUGENE LUDWIG: This entire area of AML is challenging for financial institutions because small mistakes, particularly if there are a pattern of them, can give rise to huge fines. One of the things that institutions have done is to de-risk. De-risking has meant not doing business with certain financial institutions or not doing business in certain parts of the world or not doing business with certain populations. Often it’s not the financial institution itself pushing towards this direction, but it’s a regulatory agency that’s worried about the ability of the institutions to really identify bad guys. This all starts not because anyone wants to exclude, but because of practical reality.

Yet it creates a real problem for immigrant communities. Should we be excluding people who are coming here as refugees for safety? Nearly all of us are immigrants and all of our grandparents would have had more burden in terms of getting started in the new world then they would have had if they lived under these de-risking concerns.

What we have worked to develop and applied at a couple of institutions successfully is what some call dynamic segmentation, and I like to call multifactor behavioral analysis. It is a much richer analysis of the individual within groups, and we are more able to, on a cost-effective basis, differentiate between high-risk and low-risk customers so that the low-risk customers can be included in our financial system without putting the institution or the country at risk.

Can you explain your own preferred term, “multifactor behavior analysis”?

One way to imagine multifactor behavioral analysis is to think of it as a modern graph analysis. A graph database is designed and well suited for working with highly interconnected data; it allows us to capture and visualize complex transactions, relationships and networks.

If you use a two-dimensional analysis because you only have a certain number of axes, you put on one axis, “Number of years in the U.S.” and the other axis, “Countries from which you come.” You could graph higher- and lower-risk individuals and geographies on that graph, and a bank could say, 'Hey, they came from England but were here for four generations, that’s not very risky.' But if they came from Somalia and are here nine months, that customer might be placed into a much higher risk category.

The problem is that it’s only two dimensions. If you now imagine your graph as multidimensional (at least three), you can incorporate into that analysis other factors that could be determinative. We could include a stable checking or payments history from an employer where you know where the money is deposited from. What goods are people buying and from where? Is their living situation stable? What parts of the ZIP code are they in — are they in an area known for a money-laundering activity or is it a tamer one?

Focusing on the individual and other factors can give you a much richer picture of the individual, allowing that institution to bank that individual more comprehensively than it would otherwise do and improving financial inclusion pretty considerably in these populations. Then you can target the higher-risk populations and have richer information for suspicious activity reports. This means you can run your business in a less-risky fashion as well as in a more profitable fashion because you’re able to bank more people.

"We are able to differentiate between high-risk and low-risk customers, so that the low-risk customers can be included in our financial system without putting the institution or the country at risk."

More importantly, people who need the financial help and can grow small businesses and develop a stable economic environment are able to prosper. And that’s not trivial because to the extent that people prosper and there’s business opportunity and they get bonded to a community, they are less likely to get involved in illicit activity.

This can go even beyond AML. This kind of dynamic scoring and bringing other factors into the analysis can help, and is beginning to help, in credit scoring and credit analysis. Low- and moderate-income consumers do transactions in small denominations. Small loans cost institutions a disproportionately large amount of money because they are so hard to do economically. That in part is what accounts for higher cost products to low- and moderate-income consumers. So if one is able to target, from a credit perspective, better and lower transaction costs at the same time, one can provide financial products to these populations and dramatically decrease costs to them and therefore really bring more people into the financial mainstream of the country.

When did you realize that the dynamic segmentation tool had these other uses?

We recognized about 18 months ago that you could do the dynamic segmentation in what were characterized as higher-risk immigrant populations in the AML context. Our Promontory people went to work with IBM data scientists to develop tools for doing that. They’ve been implemented over the last six months. It’s not productized yet, but we’ve seen other, smaller entities doing similar things.

In a sense it’s obvious. If you’re able to segment people in terms of lower risk for AML, segmenting them in terms of lower risk for credit is just one more dimension. Because all of the things that I’ve mentioned about AML, stable payments, how do they use their money to buy goods, what goods did they buy, are they making ends meet — all of the things you would use in an AML analysis in terms of risk can also be used in determining whether or not they are creditworthy.

At American Banker's RegTech 2018 conference, there were discussions about finding ways to serve underbanked customers without accidentally discriminating against them. How do you mitigate for this?

You raise a very good question. We are absolutely focused on the negatives that might come from the use of these tools. One has to be looking at the outcomes so that one is not inadvertently discriminating.

I remember years ago, when we were not giving people these modern tools, the case of an institution that had a second-look program, which was an excellent program designed to effectively improve economic outcomes for less credit worthy customers but had the adverse consequence of arguably having age discrimination qualities. The notion there was we’ve already sliced and diced the population from the standpoint of people we business with — what if we took a second look and try to cull out more people to give credit to, which would be a good thing to do.

And they did it! But they hadn’t thought through that the mechanism that they were using could cause age discrimination. It made me quite wary of any tool or any analysis. We analyze ours with some rigor to make sure we’re not unfairly discriminating.

What were the calculations that resulted in the age discrimination in that particular example?

That institution had recognized that some people didn’t have credit scores because they never had credit. Anybody who didn’t have credit was being lumped into bad credit scores.

The problem was that people who didn’t have credit tended to be disproportionately young because they were just getting into the workforce and hadn’t borrowed.

It taught me that you have to look at these things from multi dimensions. And the dimensions that you use, you have to think through how they might be used by others as discriminatory and then try to correct for that issue.

Are there other technologies that you are excited about that you think would work well with the work you’re doing at IBM?

There are others. Blockchain for sure. Some of the older software technologies that are being constantly upgraded like [operational risk software] OpenPages is something we’re very actively involved in.

One thing I’ve learned is how important domain expertise is. Because the technologies are like looking at a canvas before you paint on it. What the canvas looks like at the end is largely dependent on how good the paints are, how many colors you can get from the paint and how good the canvas is. It’s very dependent to a degree on who paints on that canvas.

What kinds of results one gets and how useful it is is partly dependent on the sophistication of the tech, but it more so depends on who is causing that technology to be effective, and accordingly Promontory has become an integral part to the technology build.

For reprint and licensing requests for this article, click here.
Financial inclusion AML Regtech Credit scores Artificial intelligence IBM Promontory Interfinancial Network