Felix Salmon quotes Mark Gimein's story about Prosper.com's default rates. For people familiar with peer-to-peer lending, this is old news. I have more than a passing interest in this - I was at Zopa, which runs a peer-to-peer lending operation in the UK and ran a different business model in the US, through 2007-08.
Mark makes this insightful comment about Prosper:
To look at the results of Prosper's loan marketplace, though, is to see not a solution to the credit crisis, but a microcosm of it.
He follows it up in an email exchange with Felix:
Because ultimately a paradox of lending is that the people who are more likely to repay are those who *don’t need the money*. And Prosper attracts those who do need it.
I agree with all of this, but I would like to make a broader point about lending in general.
Lending is in a select club of business activities (along with insurance) where you have to turn away a large fraction of customers that you have spent money selling to. If you make too few loans, you will not get an adequate return on your marketing dollars. If you lend too much, you will lose your shirt with defaults. Lenders make 2-4% of the loan principal. But, when a loan defaults, they lose 100% of the principal. So, one bad loan wipes out the returns from 25 to 50 good loans.
Effective underwriting is so difficult because you have to walk this tight-rope. Moreover, you have to make individual decisions that estimates the credit-worthiness of the borrower over the course of the loan, a time measured in years (or decades in the case of mortgages).
This is also the reason why the allure of automated underwriting is so powerful. Underwriting that relies completely on scorecards or models provide an illusion of predictability that is misplaced. Most lending models rely on information present on credit reports to make a lending decision.
No doubt, credit reports serve a useful purpose in credit underwriting, but they have severe limitations. First, they are backward looking, and the abiding mantra in everything financial is that past performance is not an indicator of future performance. Second, there is usually a lag of a few weeks to a few months before adverse information makes its way to credit reports, a fact that is exploited by borrowers intent on not repaying their loans. Third, credit reports present only the liability side of the financial picture of an individual and has no information about income - models that have automated decisions rely on income that is stated by the borrower.Good credit underwriting requires synthesizing information from a lot of sources - credit history, current income and future income potential, potential for fraud - that requires a degree of skill that humans currently do much better at than computers. Technology has not yet advanced to a point where this kind of fuzzy decision making can be fully automated.So, does this mean automation and models are useless in making loan decisions, and we have to go back to the dark ages where everyone shuffles paper? Definitely not.Manual underwriting is expensive and you have to turn away a lot of borrowers. Using models and automation to turn away unqualified borrowers earlier in the process is an efficient use of your skilled underwriters. It is also a non-confrontational way for borrowers to be turned down for a loan (better for a borrower's self-esteem than being told by a person that they do not qualify for a loan). Models are also very useful in setting loan interest rates, so you have consistency in pricing.The goal of automated credit models must not be to eliminate skilled underwriters but to allow them to spend more time evaluating the loans that will get made.Disclosure & Shameless plug: oFlows provides tools that makes the lives of underwriters easier.