Wednesday, November 12, 2008

Lendingclub - Late Loan Stats

Lendingclub has been around for more than a year now, so I figured maybe there was finally enough data to allow one to understand the default rate behavior of Lendingclub loans. I have prepared charts of the behavior of Lendingclub loans that are similar to the ones I've been making for Prosper.

These charts show statistics for the performance of all loans. Each curve represents the set of loans that were created in one calendar month. The vertical axis is the fraction of those loans that have "gone bad", in other words are 1 month late or worse (up to and including default). The horizontal axis is the observation date. All data comes from Lendingclub's performance web page.

The curves are "noisy" (ie they jump up and down a lot) and are not as orderly as the curves on the prosper chart. That's because the volume of loans at Lendingclub is still too low to get good stats. We can see that Oct'07 was a really horrible month. Those loans are now 1 year old, and 15% of them have gone bad! That's about as bad as prosper!

After Oct'07 they must have fixed something, because the later curves all seem to have a lower slope. We'll never know. Lets slide these curves over to a common origin, so we can visualize how common their shapes are...

Just look at where these curves crossed the 390 day line, (ie 30 days after the 360 day line, because it takes 30 days for a loan to become 1 month late) or visualize where they might cross the 390 day line as they extend, and that tells you what fraction of loans went bad in the first year. This is then an estimate of the annual default rate for Lendingclub loans.

There is a little problem. Although the data is still quite noisy becase there aren't enough loans, we can see that various groups of loans fall between 5% and 15% bad after 1 year.

Lendingclub has given us estimates of the annual default rate for their loans. This used to be on the web site, but is now found in the prospectus. For their best grade, "A1" they predict a 0.16%/year default rate. (Yes, that's really zero point one six percent!) For their worst grade, "G5" they predict a 5.25%/year default rate.

Ideally we would gather statistics for each credit grade and compare them to these default rates that Lengingclub assumed. Unfortunately, if we split the monthly pools of loans into all the different credit grades, there would be very few loans in each group, and the stats would be horribly noisy.

What we can do however is look at all the Lendingclub loans together. If Lendingclub's numbers were right, the default rate would be something between 0.16%/year and 5.25%/year, so it might come out something like say... 2.5%. However, looking at the picture, you can see they're way off. No month is anywhere near 2.5% at the 390 day line. The curves show us that the average might be more like 8%/year. Hard to be precise with such a small amount of data.

If Lendingclub has used faulty estimates of default rate, then the interest rates on Lendingclub loans are all too low. Seems likely.

In the beginning, gave their lenders default rate estimates that came from Experian (their data supplier), showing their experience with a large number of credit card accounts. Many lenders initially used those estimates to determine the rates they bid. Unfortunately, actual Prosper loans turned out to behave very differently than the Experian credit card statistics. Anonymous internet P2P loans apparently don't behave like historical credit card accounts. Sorry. Later, Prosper deleted the hilarously low Experian numbers from their web site, and showed lenders only their own historical performance data. This was a landmark improvement.

Likewise, Lendingclub obtained estimates from historical account data at Transunion (their credit data supplier) and naively used these to calculate the interest rates they offer on their loans. The assumptions aren't holding, so the rates are too low.

Luckily, Lendingclub added a term in the interest rate formula that they call "adjustment for risk and volatility". They set this term equal to 2x the expected default rate, and added it into the interest rate. (There's a good theoretical basis for doing something like this, which I won't bore you with.) The effect is that interest rates end up adjusted for 3x the assumed (bogus) default rate. That provides a little wiggle room, especially for the lower credit grades. For the high credit grades (ie A) however, I suspect they're just way off.

Lendingclub has originated between 100 and 400 loans in each of the months shown above. Unfortunately, the recent months have very low numbers of loans originated, because Lendingclub was not taking in lender money while they fought some regulatory battles. They're only now ramping back up. We really need them to get over 500 loans/month and then accumulate data for several months (a year?) before we can understand Lengingclub loan performance in a more seriously quantitative way.

Best discussion among P2P lenders occurs at

No comments:

Post a Comment