In the last blog, we broke down types of fraud, who tends to commit them (1st party vs. 3rd party vs. 2nd party aka dealer) and why (misrepresent income, misrepresent home address or contact info, etc.). In this blog, we will explore how fraud happens. There are two primary ways information is misrepresented during the loan origination process – application data misrepresentation and document misrepresentation:
- Document Misrepresentation – documents submitted as part of the application process are tampered by either party
- Application Data Misrepresentation – application data is misrepresented by either applicant itself (synthetic ID, overstatement of income) or a dealer or a 3rd party (Identity theft)
Document Misrepresentation
Of the two primary types of misrepresentation, document misrepresentation is more obvious. One can see and then agree or disagree that the document is tampered with. In lending, the documents are primarily collected during the funding phase so this fraud is found down the funnel during loan funding. This is why certain lenders with certain customer segments might choose to ask for identity or income proof “up the funnel,” in Underwriting.
Document misrepresentation is identified in many ways but categorized into 4 key types:
- Content based identification – With document intelligence, one can extract data from documents and verify the content in the document. For example, in a bank statement, do the transactions add up to beginning and ending balances in the statement? In a passport, does the MRZ line add up to the information in the document?
- Consortium based identification – This is where the power of the consortium comes into play. If the consortium is strong, one can compare documents across the consortium to find similarities. For example, are the same transactions in a bank statement found in other statements? Is the employee number (within the same employer) on a paystub repeated across two paystubs for different employees?
- Visual elements based identification – When documents are tampered with or doctored, there is often a visual trace that can be identified. Are the colors and the fonts consistent with what is expected? Are the elements in the document properly aligned on the page? Are there any shadows or reflections that look suspicious? For example, do the name and address in the document have a different font or color scheme?
- Metadata based identification – When a digital document is modified, it often leaves a trace in the document metadata. This can be detected when documents are captured as PDFs, JPEGs, or other digital formats allowed in a document upload portal. However, in the case of auto lending where the dealer scans the documents provided by the borrower, the original document’s metadata is lost, so this methodology will not work.
With advancements in AI, it is very difficult for lenders to identify doctored documents because they look identical to real ones to the naked eye. The good news is a good guy using AI can catch a bad guy using AI! Fintechs are building AI solutions leveraging machine learning to automatically detect document fraud so humans don’t have to.
Data misrepresentation
Unlike document fraud, data fraud happens at the very top of the funnel during the application process when some information in the application is misrepresented by either party. This is either personal identifiable information (PII), income (most common) or contact information (residence address or phone number).
Many data providers and lenders have attempted to catch data misrepresentation in different ways over the years which can be categorized 3 ways:
- Source of truth – One can use sources like AAMVA to know if drivers’ license / identification info is accurate per the issuing authority. Some lenders use vendors that pull income/employment info from payroll providers (expensive). In cases of direct to consumer models, one can use open banking APIs to get bank statements to verify income, bank account information etc.
- Model – Based on historical data, one can establish a pattern of a person’s income when employed by a certain employer in a certain location with a certain title. These predictive models are as strong as the data used to train them. So the stronger (breadth and depth) the data, the better the predictive model.
- Consortium – Just like in document fraud, one can use compliant consortium data to identify if the same person is misrepresenting information across 2 different lenders.
If you are a lender seeing both kinds of fraud (document and data), it’s imperative to pick a fintech partner that can help you with both misrepresentations. Over the years, I have seen lenders pay attention to data misrepresentation but not document misrepresentation, which has traditionally been left to the funders to flag for further review. With the advancements in AI, make sure you are using AI to your advantage.
