Sunday, October 11, 2009
To sort this out, I have graphed the default rate of all prosper.com loans side by side with the U.S. unemployment rate. If bad economic conditions cause prosper borrowers to change their behavior, then you should see the prosper.com default rate go up when unemployment goes up, and vice versa.
These two curves don't look correlated at all!
In the last year, unemployment has roughly doubled, which represents a huge change in economic conditions. During that time the default rate on the prosper portfolio has bounced around, but it certainly hasn't gone up in tandem with the unemployment rate.
In a recent blog, Prosper's CFO Kirk Inglis was describing Prosper's new system for rating borrowers, and he said "...the historical performance that underlies the Prosper Rating System is derived from a poor economic environment. As a result the estimates of loss are biased higher than if the economic environment had been more benign."
In other words, he argues that because he did his calculations using loan data from the recent past, when economic conditions were bad, then surely future loans should do better than his model predicts. His unstated underlying assumption is a correlation between prosper loan performance and the economy. I see no evidence for his conjecture. Wishful thinking.
It may seem counter-intuitive, but we can't deny the data. Loan payment performance doesn't seem to track economic conditions.
Boring methodology footnotes: To compute the instantaneous default rate of the entire prosper portfolio, I looked at prosper's performance web page. I asked it to show me all loans (for all time) as observed on the first day of each month. I copied down the number of loans that had defaulted as of that date. Subtracting these numbers for two adjacent months tells me the number of loans that defaulted during that month. To convert this to a default rate, I divide by the total number of loans prosper had originated as of four months earlier.
I chose the four-months-ago total because loans take four months to default. Using a later total would have included loans in the denominator which could not possibly appear in the numerator. Finally, I converted this monthly default rate to an annualized default rate exponentially, ie da = 1-(1-dm)^12 .
This process produces an instantaneous "whole prosper portfolio" default rate. Beware that this number doesn't help us judge the performance of loans prosper originated at any point in time, or at any credit grade, becaue here they are all mixed together. It does allow us to observe trends (if any) in borrower payment behavior over time, so it seems appropriate for this inquiry into how borrowers behave during economic hard times.
You probably note a downward tilt in the default rate curve during the months that Prosper was shut down by the SEC. This is a side effect of the loan portfolio "aging" during that period. No new loans were being added, and the existing loans were all getting older. Therefore the age distribution of loans in the prosper portfolio was changing. This aging produces a lower default rate, because the default rate naturally falls somewhat as loans age. This happens because the portfolio is a heterogeneous mix of loan quality. The bad loans (think HR, E, D, ...) tend to fail early, leaving a more aged portfolio with a higher quality mix.
The unemployment rate shown is the whole-country U-1 series produced by the bureau of labor statistics, and which I obtained from www.economagic.com . This is the unemployment rate that you most often find quoted in the press.
PS: The best discussion among P2P lenders can be found at prospers.org .
Sunday, October 4, 2009
These charts show statistics for the performance of all prosper.com 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 or "charge off" as it is now called). The horizontal axis is the observation date. All data comes from Prosper.com's performance web page.
The worst month so far is October '06. Of the loans originated by Prosper.com in October'06, 43.5% have now gone bad.
I'll show that calculation as an example: In October '06, Prosper originated 743 loans. Three of these loans later disappeared from the stats Prosper provides. Prosper's stats now show having originated only 740 loans in October '06. The difference, 3 loans, presumably are loans that were repurchased by Prosper because of identity theft, but we don't know for sure. I count those 3 loans in the set that have gone "bad". Prosper reports that 307 of their October'06 loans have now defaulted. Finally, Prosper reports 13 of these loans are now in the "1 to 3 month late" category. I then calculate (3+307+13)/743 = 43.5% gone bad.
October'06 is not alone. Many other months have produced similarly bad performance.
More detail can be found in my earlier posts.
New tidbits this month: Two of these curves are "done". The May'06 and Jun'06 originated loans are now all either in the "paid off" or the "charged off" category. With no borrowers active, the numbers have reached their final resting place. I've decided to make the curves end when this happens, rather than continuing a long horizontal line into the infinite future. You can see this more plainly on the larger version of the chart.
I've started charting some of the loans originated after Prosper's SEC-sanctioned reopening. There weren't enough loans originated in July'09 to make useful statistics, so I've started with Aug'09, which makes its first appearance this month as a single dot. In July'09, Prosper raised the borrower credit score cutoff from 520 to 640. A huge change. Therefore the new loans in aggregate should go bad at a much lower rate (ie lower slopes on the new curves). We'll see.
Here's a chart of the same data in which each curve has been slid to the left to a common origin. The horizontal axis is now days since loan origination month.
Explanation of methodology can be found in my prior postings in this blog, and in forum discussions on the old prosper forum, now archived at www.prosperreport.com
Many of the very early posts in this blog are still on point, and provide background on prosper, from a lender's perspective. If you're new to this, please read old posts before sending questions.
From time to time I get a comment that I should separate out the credit grades in these charts. Some folks would like to see a separate curve for loans to AA credit grade borrowers, A borrowers, B borrowers, etc. Their thinking is that this would help lenders understand the risks of lending to these individual grades of borrowers. There's some merit to that idea, but I'm not doing it for two reasons.
First, there's another fellow doin' it already. Check rateladder's blog. Here's one of his recent charts. (Click on it for a larger copy).
From the chart you can see that about 5% of grade AA loans went bad during the first year of the loan, etc. That's useful information. In the early days of Prosper.com, prosper used to give lenders statistics from Experian that showed Experian's experience of a 0.20%/year default rate for grade AA borrowers. Unfortunately, that data came from credit card accounts, which we can now see don't behave anything like Prosper loans!
I recently asked rateladder whether he would be updating his charts now that prosper is running again. He says does not expect to do any more updates. That's a shame. Rateladder has published the SQL code he wrote to produce those charts. You can grab his code, download the database from Prosper, and run those calculations yourself. (Only works for database / SQL experts.)
The 2nd reason I'm not charting loan performance by credit grade is that credit grade is just one criteria. There are many variables you can use to choose borrowers. Most lenders use some combination of more than one variable. (Example: Credit grade A or better, no past delinquencies, and doesn't mention religion in the writeup.) I can't chart all combinations.
I do believe there's value in observing the behavior of Prosper.com loans in aggregate, even if that doesn't give lenders insight into exactly how they should pick loans.
In the early days of Prosper, it was common to hear a wide variety of speculations and myths about borrower behavior. These charts provide the data to understand whether some of those speculations were correct.
Example: "Once you get two or three payments made, the loans are good." I used to hear this all the time. Turns out it is not true. There's no bump in the default curves at the beginning. They're pretty smooth. Loans go bad over time.
A common myth you hear recently is that borrowers are defaulting a lot because of the recent bad economy. It ain't so. Look at the curves. (Look at the top chart above.) The horizontal axis is calendar date. If borrowers were behaving very differently now (during the recession and high unemployment) than they did earlier (say last year) you'd see it in this graph. You would expect ot see the curves bend upward as unemployment rose. Didn't happen.
PS: The best discussion among P2P and Prosper.com lenders are found on prospers.org. See you there!