= Subscriber content; or subscribe now to access all American Banker content.

Future of Consumer Lending Belongs to PhDs, Risk-Takers

We all know that the FICO score is less predictive at the lower end of the credit spectrum. A thin credit file, or no file, makes it hard to grant loans on the basis of a FICO score alone in the subprime segments of the market, and those below.

But we now live in a world swimming in data. Advertisers can find out what a person buys at the supermarket, what magazines she subscribes to, what education she has, what websites she visits, even who her friends are on Facebook. They use all of this information to do a better job of targeting ads in real time. Would we get better results out of underwriting if we included all of these datasets, too? Of course we would.

In subprime, people have been doing this for years. Teams of smart guys develop a hypothesis about a new signal from the market, back-test it rigorously and build a new model to update their underwriting. But this relies on smart people and smart insights.

Today, with tens of thousands of data points available, it's hard to know what will be significant, and which hypotheses to test. Your team of smart guys might identify some predictors, but will they find them all? Probably not.

Advances in technology mean that now you don't have to develop your hypotheses in advance. You can simply make loans, wait for some to go bad, and look for what those bad loans had in common. Each time you update your model, it gets better. Rinse and repeat.

This isn't easy. In fact, it's really, really hard. It requires teams of math and computer science PhDs, with expertise in big data and machine learning (a branch of artificial intelligence).

Some companies have started to do it, and it's working. They typically focus on serving the underbanked via payday and installment lending. Four examples are Zestcash, Wonga, Progreso Financiero, and Global Analytics. These startups are getting big, fast. (Full disclosure: my venture capital firm is an investor in Zestcash, and I hold a board seat there.) They are making borrowing quick and easy, and passing the savings on to consumers in the form of lower rates. Wired magazine reported last year that Wonga made over a million loans in its first three years of lending, and it continues to grow at a similar rate. But all four of these companies are doubling revenues, or better, year on year.

Watch this space. More companies will apply big data and machine learning to different lending products (small businesses, automobiles, etc.) and in different geographies. Most will fail because of the technical complexity. But the winners will really upend the industries they compete in. They will win because they are willing to take more risk. They will make more loans, knowing that they will make more bad loans, in order to teach their algorithms how to distinguish good risks from bad ones. That is the tuition price for learning how to build and train new underwriting algorithms.

Does your bank have the courage, and the knowledge, to compete in this new world?

Jeremy Liew is a managing director at Lightspeed Venture Partners in Menlo Park, Calif. 


(6) Comments



Comments (6)
I think a couple of things are different that make it different from "the old days":

1. Many of these models are being applied to very short duration loans, which allows for both faster iteration and improvement of models, and faster reaction time to changes in underlying conditions. Underwriting 2 week payday loans or overdrafts is quite different to underwriting 30 year subprime mortgages because you can actually see if your model works pretty quickly
2. The compute power is dramatically different today vs the past, which allows machine learning approaches, vs the simple backward looking regression modeling of the past. This means your PhDs don't have to construct algorithms, they simply need to build a model that can identify patterns from past data, test it forward, and amplify the patterns that are predictive.

All that being said, judgement and instinct can all greatly accelerate the process of building an underwriting model - some things don't have to be learned by the machine if they are already known
Posted by jeremy liew | Monday, April 30 2012 at 6:20PM ET
The use of algorithms to sort thru an identify patterns is nothing new. IRS intiated a program in 2007 to apply such logic sequences to identify anomolies in fuel distribution system data which might suggest non-payment of excise taxes-------the govt use is to identify anomolies. Commercial use is to identify patterns so as to identify likely customers.

the credit analysis concept is more akin to govt anomoly analysis--including link analysis aka guilt by association.

In practice the PHd's are very poor in developing algorhithms. They are effective in constructing data retrieval and table compaisons--but lack practical understanding of the drivers of typical human decision-making. PHds are ivory tower types that do not have sufficient grasp of the intersets and concerns of real people to be effective. The real people analysis is best performed by elders--to whom human understanding is 2nd nature learned by decades of human interaction--not the ivory towr menatlity of people that have pent 85 % of their lives in classrooms. This was what the govt found--the suggestion that PHds are the solution to this new technology demonstrates the writers lack of comprehension of both algorhtms and humanity--what would i xpect from a bank-writer
Posted by OLDER&WISER | Monday, April 30 2012 at 6:06PM ET
Good banking in the end, like a good baseball manager, requires good information, good data, but it can never get away from good judgment. And in the end, it still seems that the ones with good judgment have the competitive edge.

Dodd-Frank and attendant regs would like to change all of that and drive judgment out of banking. The data demands assume that you can bank by numbers alone, and maybe even assume that you can regulate by numbers alone. I don't wish them luck on that.
Posted by WayneAbernathy | Monday, April 30 2012 at 4:44PM ET
In the old banking school, before models, we worked with customers. Yes, mistakes were made, but not on the scale that models have produced. I am not ready to have models make all my loan officers decisions.
Posted by phenry100 | Monday, April 30 2012 at 4:34PM ET
Perhaps, Household Finance had over 100 PhD quants. HSBC bought the thory that their models had the risks well defined. Ooops got that really wrong. Data is retrospective. The best lenders use data AND "gut instinct".
Posted by Old School Banker | Monday, April 30 2012 at 4:22PM ET
The sub-prime lending of the past decade applied PhD's and their models with a vengence. They were all wrong. The PhDs should have been paid in residual interests of sub-prime securitizations.
Posted by kvillani | Monday, April 30 2012 at 4:21PM ET
Add Your Comments:
Not Registered?
You must be registered to post a comment. Click here to register.
Already registered? Log in here
Please note you must now log in with your email address and password.