zcommodore

"Loans go late over time."  So says Fred93 and others who have studied delinquency rates of loans on Prosper.  The question in my mind has always been, at what rate? and does that rate change?  Well, I've compiled a lot of data from Prosper's performance page and now have some interesting graphs to present.

First, I want to tell you that I've gone through and collected information from eric's on repurchased loans as I mentioned in a previous blog post. I have subsequently added all repurchased loans and their "late history" back in to all my graphs.  I will be updating some wiki pages as I get time with my updated graphs.  In the mean time, I wanted to share my findings of this study.

First, here's a graph showing what the percent of unpaid loans go late relative to the previous month. (Note that you can click on each graph to bring up a larger version.)

This graph shows all loans since March 2006 through March 2008.  The X-axis is the age of the loans in months.  The graph starts at 1 month and includes all loans 1-2 months old.  The red line is an average of all new lates in a given month as a percent of all current loans in the month.  As time goes on, some loans are paid in full or go late are subtracted out.  Also, lates tend to accumulate so each prior month late totals are subtracted from the current month's late totals resulting in the number of "new lates".

You will notice there are a small handful of months where the datapoint is actually below zero.  These are months where the total number of lates actually went down from one month to the next.  It doesn't seem to happen often but it does happen once in awhile.

One thing I was curious about was whether changes to the Prosper platform, particularly changes related to lender information, made any difference in the rate at which loans went late.  So I broke down the total data set into three different segments: Early loans (March 2006-February 2007), New Extended Data (March 2007-October 2007) and Lender Guidance (November 2007-March 2008).  Here are those graphs:

Keep in mind that the last graph, Lender Guidance, is really too new to draw many conclusions from since there are only ~5 months of data available.  Really, this is true for all the last 6-8 months of data from any of the graphs.  The last data point on any of the graphs is from only one month's worth of data (the oldest month in the data set) so as those loans age and I have time to update the information, the average may change.

Ok, so those graphs are nice but how do they compare with one another?  Well, here's my final graph that shows each of the "average" lines superimposed on one another with the same scale for all of them.

So, what can we learn from all this?  Well, looking at the oldest set of loans, it appears that over time, even considering the fact that the number of current loans is going down, the rate at which they go delinquent seems to go down as well.  The peak rate of going late seems to happen in the first 6 months of the loans age.  Some loans still go late over time but not at the same rates seen in the first 6-12 months.

The earliest loans seemed to go bad at a much higher rate out of the gate but as more information became available to lenders the rate went down initially.  However after a certain period of time, it appears that all loans start to behave in a similar fashion.  Two different things could be at work here: Prosper could be doing a better job of weeding out ID theft and other fraudsters and/or lenders may have enough new information in order to weed out the borrowers who will default right out of the gate.

I intend to track this over time and may post about it again sometime.  I don't anticipate things changing a whole lot but I'm especially interested in how the "Lender Guidance" loans behave so I'll be watching those for sure.  Unfortunately, it will be at least a year before we have enough data to start to make many conclusions about those.

Feel free to comment on this blog or, if possible, I'm hoping to get a thread started in the lobby of the .org forums about this as well.

ETA: After a little bit of discussion on the .org forums, I found a mistake in some of my calculations.  I've made corrections and have updated the graphs above.  It appears that the conclusions are still the same though despite the error.  I've also modified the text of this post to reflect the changes.