A Dive Into The Lending Club Data

Glad that my findings match yours !

I think it does beat a C.D. from a risk/return perspective. It's probably higher risk than a C.D. but returns largely compensate for it I believe.

To me this is possible because the Lending Club is desintermediating a business that was traditionnally 'high margin'.

> In finance terms, I will be surprised if they have a higher RAROC [1] as compared to large banks. If they really do, then congratulations (you will put banks out of business in a few years)??

Or, more likely, just drive down bank profit margins.

I don't know much about the banking industry but savings accounts and CDs have always felt scammy to me. They're marketed to the rubes that have no idea what they're doing so they can get away with not being competitive with other financial instruments.

Your background sounds interesting! I'm working on bringing a credit card to the subprime market. Would be great to connect with you.

You are correct, a comparison with CD/SA is an incorrect comparison that doesn't take into consideration the default premium and liquidity premium included in Lenders' return. CD/SA is as close to risk free return (with FDIC insurance) as you can get while Lending Club loans are unsecured loans (highest risk fixed-income investment).

Banks haven't proven too skilled in estimating risk (see 2007). The average Lending Club lender is surely worse than the banks, but the best ones are surely better.

Net it all out and hopefully we get a Darwinian marketplace for lenders, lenders who don't need acres of employees preparing powerpoints for each other in class A office space.

Of course, the true measure of estimating risk is hidden until a crisis occurs. Interesting to see where it all ends up.

Yes, exactly for all the criteria it is a matter of "the rates compared to marginal increase in risk".

Look at the 'A' grade. They're nice and safe intuitively, but in my opinion they're not a really good investment.

Also, it's pretty instructive to look at the Lending Club grading algorithm in details.

They made it public at some point. Now they are a little less transparent about it.

But some details can be found in their SEC prospectus.

I can link that up as well if you guys want.

I'm no expert here but current job duration may be counter-intuitive. We usually think that the longer the better but it may also correlate with inability to find another job, so in fact during a recession, these kinds of people may actually be more at risk than people used to job-hop.

It probably needs to be controlled for social class and job type but I think this could explain the phenomenon.

> if the rates are much higher compared to a marginal increase in risk

Exactly. Same thing with public records -- having a public record could very well mean you have a higher default rate but what is important is LC punishes it more than they should so its a value investment.

We invest via a model (not filters) and the whole idea behind the model isn't to find what criteria makes someone less likely to default in absolute terms, but what makes one D2 loan less likely to default than another D2 loan.

> but this says that <2 years of experience is better than longer

When you say better, do you mean risk or return? First of all, you are looking at the variable in a "univariate" sense,i.e, the relationship of the default rate or return by the categories of this variable. But their internal model is multivariate - there may be other factors influencing risk or return which is not obvious in the univariate dimension. It also depends on the power of this variable in predicting risk. And finally, lower risk may mean lower return - you just need to find the efficient frontier :) - http://www.investopedia.com/terms/e/efficientfrontier.asp.

For students going to a particular university, math SAT scores are inversely correlated with verbal. If students had a higher math SAT and a higher verbal SAT, they'd be at a better school (and worse math + worse verbal = worse school).

For debtors inside a certain grade, it looks like employment history and other creditworthiness metrics are inversely correlated. So I suspect that it's less employment length being an anti-signal, but rather within the grade people with short employment length have compensatory advantages to stay in that grade.

If you only focus on return, you will draw erroneous conclusions. The upper limit to return is restricted by the interest rate. An A grade loan that carries interest rate of 5% will never give you return of 10% while a D grade loan carrying 18% interest rate has some chance of giving you 10% return, assuming you hold loans to maturity. If you only consider returns, your findings will always be biased toward high interest (supposedly high risk) loans. Similarly, the return comparison across vintage will generate erroneous conclusions.

You need to take into consideration Interest Rate and Default/Loss Rate while considering the validity of the relationship.

The Employment Length on its own is a poor indicator (statistically insignificant). You need to at least combine employment length with credit age (when the first credit line was opened) and Income to improve predictability.

There is no vintage of Lending Club 5-year term loans that have fully matured. The first 5-year term loan was issued in early 2010 (IIRC, May) so the first vintage is just coming up to full maturity.

  Interest Rate (Rate of Return)
  = Real risk-free interest rate 
  + Inflation premium 
  + Default risk premium 
  + Liquidity premium
  + Maturity premium

You need to determine whether you are being compensated on 5-year term loan for potential change in inflation, higher default risk, and longer maturity over 3-year term loan.

The code to process the lending club data is done in python. I got the full history of all payments made on the platform and pre-process it to have something that's light enough to be explored with a good user experience.

For the viz', yes I used DC.js, I can open source the .js if you guys want it.

There are many pearls to find. It depends on your 'set of preferences'. You want return? you want low-risk ? you care to deploy a lot of money, or not that much?

But I'd say that for starter, anything in the 8%-9% range is a very good deal these days in this environment

Look for the Pearl! you can definitely get 8% interest over the long run with a good liquidity on your cash deployed. To me, it's totally worth it for money that you don't need in the really short term

For my own personal interest, can you speak at all on your experience using LendingRobot? (successes/misses, use duration, gains, etc)

A bit more courtesy would have been welcome. You sound very condescending. And I hope you talk to your clients in a different way!

Let me address your methodology comments nonetheless, which are for the most part unfounded.

* I don't have any finding about annual income. I don't think it is mentionned anywhere in my conclusions.

* "delinquencies and public records has a big impact on returns as newer loans are not aged enough": because I average across vintage, and because I don't average based on volume on the platform, I account for the aging biais.

* "Time is not risk [..] kurtosis etc.": I don't say that time is risk. I suggest the reader to look at the return series through time. Essentially to look at the volatility of the returns ( without pronouncing the word volatility to keep the content accessible to a novice reader). I essentially encourage the reader to visually assess his Sharpe ratio. Which is a good universal risk measure.

* "reconsider average across vintage": averaging across vintage is a first approximation. I acknowledge the fact that a better methodology would be to take a weighted average that matches the amortization profile of a loan.

* I maintain that any statistics you compute in 2006, 2007 or 2008 is less reliable (statistically). Yes it is an important period to have because of the crisis. And this is why I put on the chart. However, you can't compute very reliable returns when you have a dozen of loans to average across.

Anyway, I happy to exchange with you in PM on methodology if you would like to continue the discussion

You can fully automate investment 'filters' i.e. filter the notes you're willing to invest or not.