Analysis Featured

Using AI to Improve Underwriting

There has been an increasing trend of late of FinTech companies using artificial intelligence and machine learning technologies for a variety of purposes. Some financial technologies rely on AI or machine learning for core functionality. Others use automation as an add-on to an already powerful solution. Either way, machine learning is quickly becoming the wave of the financial future. Soon, every financial technology company will incorporate it in some way into the technology that is revolutionizing the way people relate to money.

One of the more fascinating presentation at LendIt USA 2017 last week was Marc Stein’s presentation titled “Using AI to improve underwriting.” Stein is CEO and founder of

“The practical application of using AI to improve consumer credit underwriting is profound,” Stein said. “It’s also applicable to any form of underwriting.”

Stein’s presentation centered around a case study of one of his client’s using the technology to improve consumer credit underwriting. The firm was unnamed, but the lessons that can be learned from Stein’s presentation don’t need a name. They are legion.

The Unchanging Nature of Consumer Lending

“Consumer lending has changed since the late 1950s,” Stein said at the beginning of his presentation. “The current model for evaluating credit is FICO.”

Since FICO is based on linear regression to build a credit scoring model, Stein said it isn’t accurate enough, nor does it go deep enough into consumer data to reach the more granular levels of a person’s financial situation. In that regard, then, the credit scoring system is due for an overhaul.

“Credit score cards are based on four key attributes,” Stein said. These are payment history, or revolving utilization, credit history length, types of credit, and inquiry count. “Three of the four roll into the fourth. Therefore, we make credit determinations based on a single score.”

That works as long as you limit your lending to the prime market. By screening out the population most likely to to default, lenders attempt to focus on good risk instead of bad.

“It’s not optimal, but it’s good enough if the cost of deposits is low enough,” he said. “But what about the 51% of borrowers below 680? How about the 35% below 620? The FICO model is very poor in those populations.”

As a result of this credit scoring model, consumer lending is restricted to people with a history of consumer borrowing. Prime lending is subsidized by overcharging subprime borrowers because it divides the general population into three classes–people with access to cheap credit, people with access to expensive credit, and people with access to no credit.

“Consumer Lending can’t be fixed with the current underwriting methodology,” Stein said. Why? Because consumer lending is an inherently non-linear problem. Therefore, lenders need to stop using a linear solution to solve this non-linear problem an migrate to a non-linear solution. That’s where artificial intelligence comes in.

How AI Can Solve the Consumer Lending Problem

One question Stein often gets is, “Why not add more attributes?” But the problem, he said, is the basic math involved.

“If we limit FICO to 550-850 and the debt-to-income ratio to 40%, limit trades to no more than five in each of the three time ranges and inquiries to 10 within six months, there are still 15,750,000 possible scenarios,” Stein said. “Adding another score with 700 values increases the range of possibilities to 11,025,000,000.”

That’s why Stein set out to build an online system where lenders can take online applications and apply a machine learning algorithm to them in real time to return a probability of a binary outcome. He said the key metric is the first payment default.

“It’s a good proxy for overall payment default,” he said. If a borrower defaults on the first payment, chances that they will default on the loan are substantially higher.’s first client was a deep subprime unsecured installment loan lender who had a first payment default rate of 32.8%. Stein’s company built a non-linear model set using ensembles and reduced FPD to 22% in the first month. In the second month, it fell to 18%, then to 15% in the third month, and finally to 8.6% in the sixth month.

“Their max yield sweep spot is 9% per month, we we tweaked the algorithm and have been running at 9% for two years,” Stein said. “An overall default rate under 14% is unheard of in that space.”

Optimizing Consumer Lending

underwrite.aiStein was careful to point out that using artificial intelligence in consumer lending doesn’t exempt lenders from the Fair Credit Reporting Act. It’s still critical in the U.S. to offer an explanation for who is declined. It’s called an adverse action report. You have to say who’s being declined a loan and the reasons why.

Artificial intelligence allows lenders to make loans to a larger segment of the population and at higher rates. It does this by getting enough detail in the marketplace to identify the danger zone more narrowly and specifically than the linear FICO credit scoring model. It’s all about the distribution of defaults.

“It’s very difficult to calculate from the default loan matrix the cause of decline,” Stein said. “You can use it for fraud identification, but not creditworthiness.”

At 32.8% FPD distribution, applied an initial algorithm to cluster all the defaults at the bottom of the above chart (the area shaded all red). That represents the volume that can be discarded without a cost to the lender’s business.

“The goal is not to solve for the larger area,” Stein said. “The goal is to identify the highest risk population in which there is not bleed over in positive value.” Stein discussed the reasons people default on loans. A and B category defaults do so for different reasons than D and E. “A defaults because somebody lost their job or there was a major life change. These things can’t be predicted at the individual level. As you move higher up, you deal with a higher form of statistical modeling.”

By identifying the 17% of loans at the bottom of the scale, can eliminate those loans with no negligible loss to the business.

“Bad loans will harm you more than good loans can help you,” Stein said. is a solution that learns on its own. No one on Stein’s team adjusts rules manually for any lender. The company’s programmers wrote a model that determines what algorithm out of 64 will present the best efficacy in the intersection of a lender’s product with their customer base. Each time a customer makes a payment, or fails to make a payment, Underwrite’s robot asks the lender to send a unique identifier to the solution and the model re-trains. By using this non-linear model, Stein’s team can expand the range of input variables and understand the quantum effect in consumer lending with the collapse of the outcome of uncertainty. The potential is for lenders to make more money in the long run by identifying the highest risk populations more succinctly.


Allen Taylor
Allen Taylor


default image

"Your daily letter is great!" , Ram , Founder and CEO, PeerIQ

default image

"Hi George - just want to tell you that you are doing a great work with Lending Times;-) Brgds, Kasper" , Kasper, Partner and Co-founder at Dansk Faktura Børs A/S

default image

"I've been following your newsletter for some time now and have been very impressed with the content." Charlie,Co-Founder | Bolstr

default image

"Hey George, I must say I really enjoy your site. It has inspired me to do some changes at our platform and we are the biggest consumer lender in Sweden." , Ludwig, CEO @ Savelend Sweden AB

default image

"Your daily email is very useful. It gives quick update on what's going in the market. Thank you very much for all that info." Yann Murciano, Head of Base Metals Trading at Morgan Stanley

Our daily p2p news digest

Daily News Summary Digest Sent Daily To Your Inbox