PPP Fraud Protection with Machine Learning

The United States Governmental Accountability Office Stated, “We estimate that through February 2006, FEMA made about 16 percent or $1 billion in improper and potentially fraudulent payments to registrants who used invalid information to apply for disaster assistance”[1]

With businesses across the nation requesting assistance through the Paycheck Protection Program (PPP), financial institutions could be facing similar percentages of fraud on their books. While, the Small Business Administration stated that lenders wouldn’t be held accountable for borrowers failing to comply with the terms of the program; there still remains a great deal of uncertainty on if lenders will be completely protected, or if a portion of the liability still will remain with the lending institution.

While financial institutions know their existing customers deeply – the risks increase during this time of crisis with new customers requesting to apply for federal aid. Fraud will take many different shades as a result of this round of federal aid. Some clients may slightly overstate figures related to payroll, number of employees, etc.; while fraudsters may use more illusive techniques like synthetic identify fraud or using a combination of real and falsified information to create a realistic identify.

Given the heavy reliance on taxpayer ID, in combination of other fraud techniques (e.g. synthetic identifies) – there’s an increased risk that different taxpayer IDs could be used to receive multiple loans for the same small business (loan stacking). Even with the SBA E-Trans system capturing most forms of loan stacking, there still exists risk beyond this first risk control: with small business often having multiple tax IDs and various degrees of ownership – the risk is still present.

With forgiveness documentation timelines looming, Financial Institutions will be facing immense pressure in helping clients navigate the terms associate with loan forgiveness. Many FinTech’s have emerged to help gather the associated data to service the loan, but if organizations haven’t put in place technology solutions – their client portfolios may be at risk of attriting to another organization if the client has a poor PPP experience.

To further complicate the issue, while institutions look to service loans as quickly as they can, some bankers may conduct a less than robust due diligence – potentially opening the institution to increased liability down the road.

One way to provide audit protection and conduct a layer of automated due diligence, is through leveraging advanced analytics to compare business based on their statistical similarities. By leveraging machine learning techniques to group (segment) clients based on their business attributes, outliers in forgiveness data can be flagged for additional review.

For example, your organization has 200 restaurants in your portfolio that are requesting PPP forgiveness. Of those restaurants, 130 have deposit relationships with your organization. By mining history operating account behavior, profiles can be generated on each restaurant (estimates on monthly payroll, rent, utilities, among demographics such as location, industry type, etc.). Using this profile data in combination of machine learning, the clients can be grouped together based on their similarities. These grouped clients will, on average, have similar needs and in the case of PPP, have similar loan forgiveness documentation. By applying the grouping logic to new clients your organization doesn’t have a deposit relationship with (or limited history), you can estimate the likelihood of the forgiveness documentation of being anomalous.

While these types of analytics will provide additional due diligence to the PPP loan process, they can also be reused to understand clients needs on a deeper level (e.g., “Next Best Product Article”) . For more information on how to apply these analytics to your PPP process today, please contact us at partners@dataproductsgroup.com .

[1] https://www.gao.gov/assets/120/114055.pdf