How Machine Learning Closes the Loop on Hidden Auto Fraud

Frank McKenna, chief fraud strategist, PointPredictive

An unscrupulous finance manager can manipulate a loan application to disguise a fraud scheme and fool even the most diligent lender. To address this, lenders are increasingly enlisting the help of machines to find hidden patterns of fraud. The technique is new in the fight against automotive lending fraud, but it has been used for 25 years to control credit card and mortgage fraud.

Machine learning software works by analyzing thousands of factors about a loan and pinpoints when something doesn’t look right. It might be issues with the income. It could be issues with the employer. It can even detect issues with a car and the way it is priced. It doesn’t matter. If any information looks fabricated, machines can spot it.

Machine learning software works because of how the models are trained to spot fraud. Machines are trained, just like underwriters are trained, but on a mass scale. Machines are shown millions of good loan examples, and tens and thousands of fraudulent loans. Overtime they learn subtle differences so when they see the pattern again, they identify it immediately.

Machine learning closes the loop on auto fraud by automatically learning every new pattern of fraud as it presents itself.

5 Hidden Patterns of Fraud that Machines Expose

So, what are the emerging patterns of fraud that machines are finding? What are machines telling us about fraud that we didn’t know before? Most importantly, what are best practices lenders can do?

1. Many Early Payment Defaults Have Patterns of Fraud in the Application

As many as 70% of loans that default within the first six months have evidence of material misrepresentation. Whether it is fabrication of identity, income, employment, or collateral, those misrepresentations might be missed because lenders may not be looking for those patterns.

Best Practice: Lenders should use a predictive models to screen applications for fraud risk during the approval process to identify loans that will likely default due to misrepresentation.

 2. A Small Percent of Dealers Create Most of the Risk

The models are showing us that as few as 3% of car dealers account for most of the fraud issues a lender might experience. But the models also show that those same dealers perpetrate their fraud scheme from one lender to the next. When one lender shuts them off, they move to the next.

Best Practice: Lenders should use predictive models to identify those few risky dealers that cause most of their risk and take preventative action.

 3. An Unscrupulous Dealer Will Defraud Consumers and Lenders Alike

When dealers engage in systematic power booking, inflate borrowers’ income, manipulate employment information or create phantom deposits, the risk of default is high and puts both the consumer and the lender in a bad situation.

When an unscrupulous dealer attempts to defraud a consumer, they often attempt to hide it from the lender as well. Our models analyzed lender fraud reports and consumer reviews and found them to be remarkably similar. When a lender had a bad experience with a dealer, the consumers always did as well.

Risky Dealer Profile Good Dealer Profile
Lender EPD Rate > 10% of Loans 0%
Lender Fraud Rate > 10% of Loans 0%
Consumer Rating 3 stars 5 stars
% of reviews that contain words “fraud”, “scam,” or “rip-off” 30% < 1%

4. Synthetic Identities Look Too Good to Be True

Completely fabricated identities, also called synthetic identities, appear to be ideal consumers. They have little debt, great credit scores, and sizeable stated income. These profiles are so good that they often pass through lenders’ automatic approval processes. Machine learning models have proven to expose subtle patterns in the credit, stated income and employment that appear to be misrepresentations and anomalies.

Best Practice: Screen applications using a predictive model to look for patterns of synthetic identity before passing the loans for automatic approval.

5. Sales Pressure May Lead to Increased Risk

Models have identified that certain times of the month represent higher risk to lenders and dealers. In some cases, reported cases of fraud and early payment default spike the last week of the month.

Best Practice: Don’t automatically approve more loans to pass volume through without proper underwriting checks and balances.

Closing the Loop on Fraud

Machine learning technology is emerging as a strategic tool for automotive lending. By leveraging machine learning and models, lenders can reduce their early payment default losses by up to 50%, catch systematic fraud up to six months sooner and better protect consumers who might be defrauded by finance managers or unscrupulous dealers.

 Frank McKenna is chief fraud strategist at PointPredictive Inc. McKenna is an advocate for fighting fraud and has worked with more than 150 banks, lenders, and companies throughout the world, designing strategies, solutions, and operational practices that help them reduce costs and increase efficiency. Connect with Frank on Twitter @frankonfraud or email

Leave a Reply

Your email address will not be published.