The 05/01/07 update.
(IMG:http://img.villagephotos.com/p/2006-6/1187065/prosperlate-2007-05-01.gif)
I update this chart twice per month, using data thru the 1st and 15th of the month. The source of this data is the Prosper performance page. See earlier messages for details on methodology.

Here's a version of this chart where each month's loans are slid to the left to make the horizontal axis days since beginning of loan origination month instead of date. This allows you to visually compare the slopes of these curves more easily. Note that we now have the first data point at 425 days (ie 14 months). 14 months is the first data point which contains data from 12 months (ie 1 year) of borrower nonpayment events. (First payment is due after 1 month, and is recognized as "1+ late" after another mont has passed, so there is a 2 month lag.)
(IMG:http://img.villagephotos.com/p/2006-6/1187065/prosperlate-2007-05-01-slid.gif)

There it is! Thanks again pninen.

Thank you, pninen. As always.

Very, very helpful - and scary.

please stop showing this :) damnit.... this is like Groundhog Day

would this be difficult to recreate (the top graph) with HRs excluded?

Well as a start, I'd rather somebody found a way to recreate May '06. :D

The second graph is the one I have been waiting for.

Is it just me, or does it appear to show NO significant improvement from when we first started lending, blind as bats, in February/March of 2006?

This really cannot continue.

this would be a real bummer if we were buying a basket of all originated loans, sort of like a mutual fund or possibly Zopa?

But we're not. Some portfolios are not following these trends (I'm thinking MidasMulligan for instance). There are just too many variables in play to affect possible default outcomes in any given Lender's portfolio. Plot clean D's versus D's with a certain level of defaults or PR. Business related loans vs. debt consolidation loans, or student loans, or health related loans, or family related loans, etc...

Most of the lenders (myself included) think they can beat the averages and tailor their lending and search criteria accordingly. But this analysis is excellent and I hope continues to dig deeper in the future as this great Prosper experiment plays out.

yes. the picture is much different when you break out the loans by grade.

It would appear that the Feb. 11 credit expansion and cutoff of sub-520's did not help DQ's one whit either.

That's discouraging.

The only consolation is that the best 10% probably does a lot better than average...

It is too soon to observe the effects of the Feb 11 changes. They took effect for loans that began their listing on Feb 12. Feb originations are dominated by loans that began their listing in late Jan and early Feb. We need to wait for the March data, which will be almost all loans for which the new data was displayed. March loans may take until the June 1st data point to show up. Patience.

Awww, anything but that!!

Thanks pninen. Your analyses are always so sobering...

Thanks, Pninen, for the horror story!

The second chart makes it easier to see that it doesn't look like lenders are making use of the information that could help them make better bidding decisions. I think the answer would be to make educational material available to lenders and do so at a spot where they will find it. Considering that about a tenth of the lenders find the forums, maybe a link could be made to a tutorial on just what tools Prosper makes available (e.g., Performance page, the "additional credit data", and give them tools to see correlations themselves and how to use Performance page to gauge their expected ROI).

Until an effort is made to bring the non-forum lenders up to speed, I doubt the curves will improve. :(

Everyone keeps asking for loans broken out by credit grade. Personally, I'd like to see loans broken out by "stupid loans" vs "apparent good loans". There's a lot of "stupid money" coming into Prosper from newbie lenders who simply chase the highest rates. Occasionally you get someone like MuleShoes dropping ridiculous amounts of money on borrowers at low rates who should never have funded to begin with.

I have no idea how one could do this but a first cut might be to simply look at the performance data for all loans with 0 DQs. Alternatively, one could look only for non-AF loans.

My point is that there are loans on Prosper that should never have been made and wouldn't have been if some of these "large lenders" weren't closing them out. Alternatively, pump & dump groups and the lemmders who support those practices by underbidding are skewing the numbers as well. As long as these practices continue, and I don't see any way to stop them, the delinquency/default numbers are going to be higher than average.

The points that folks make about smart lenders picking better loans is certainly valid. There are always going to be some defaults, even from good credit grades but I believe the numbers on Prosper are adversely affected by uneducated lenders as much as bad borrowers.

Actually what this leads me to believe is that the problem isn't Prosper(the company), it is the constant stream of new Lenders to Prosper(The platform).

Prosper has given a lot of new data that theoretically allows us to make better decisions. If you choose suitably restrictive criteria, you can find lots of parameter combos that return reasonable loans, with low default rates.

So why do default rates seem so similar even though we have the ABILITY to pick better loans?

We(the collective Lenders) are still picking crappy loans and ignoring the data. I doubt there are too many experienced Lenders(6mo or more lending) who are still lending to HRs at 22%.

Lots of Lenders got a big reality-check when the first batch of Lates was sold off. The big red Delinquent got noticed, and Prosper started showing big negative numbers in their interface. It has been a long time since the last sale. Not enough reality-checks for people.

I agree that the arrival of fresh-faced new lenders contribute in some part to the poor statistics. I do, however, find it hard to believe that the drunken bidding patterns of March 2006 (where we had virtually NO information) should be replicated to this degree - months later, armed with a lot more data-points and the ever-stronger cries of caution.

I guess people who blindly jump on Prosper seeking 20% returns without doing any research are pretty much getting what they deserve.

I second this request.

I'd very much like to see how the lates look when you knock off the Es and HRs.

the results might be surprising to some people if E & HRs are lopped off.

Reason: there is an assumption being made that "smarter" lenders will not bid on E&HR, thus lopping off the E&HR loans will reflect picks of "smarter" lenders.

I haven't seen the data yet, but I would be very surprised if this was the case. I think that smarter lenders will bid on performing loans irrespective of credit grade. Also, the "smart" money may simply decide not to participate.

Maybe, zarp, but looking at this chart:
(IMG:http://img.villagephotos.com/p/2006-6/1187065/prosper-late-loan-stats-vs-time-2_apr06bygrade_2007-04-15.gif)
I'm not so sure.

If 4/2006 is an indication of a typical month, then the indication would seem to be that the linear increase in lates is predominately driven BY the Es and HRs. The others show a much lower and much less regular progression.

more importantly they seem to have leveled off (if I am reading this graph right?) Ds and better that is....

You might not be drawing the right conclusions from what you see. For example, when the "A" curve jumped up back in October, that jump was a single loan going bad. It has been level for some time since then, but that's what you would expect at a low bad rate consistent with the "A" credit grade. At some point another loan will go bad and it will take another jump up. Remember that when you restrict to one month and one credit grade, you've got a bucket with a small number of loans in it, so the resulting ratio will move in a jerky fashion. Stand back across the room and look at the chart.

I don't see any evidence here that the high grade loans have "levelled off". They look like they're continuing to ramp up.

I miss Muleshoes. I could use some $$ at 4%. ;)

289 loans in April 06
448 in May 06

Maybe push up the chart 1 month, as your sampling size almost doubles and any 1 loan won't 'adjust' the chart so much

and you only lose 1 month of 'data' i.e. its 1 month younger ...

I guess I am confused by the chart then

For example Cs, from April 06 seem to have leveled off between 10-11% default (steady) since August 06 til now...meaning no more sizable % of defaults in that range, no? Understanding its a smaller sample size...

ETA: I am making comments solely on your "April 06 originated loans" chart

**poof**

No I am not talking about the original pninen graph, the one that graphs each months loans

I am speaking the last few comments to the secondary graph that graphs the april 06 loans by grade

it appears for april 06, that the Ds and Cs and Bs etc are leveling off

since the sampling size is low, I'd hope to see if May 06 is showing the same thing

looking at only May 06, by grade... pninen indicates the sampling size for Apr 06 is small so 1 loan going bad really changes the graph... May 06 is almost double the size of April.

ETA: should I poof? :)

No do not poof. I just realized which graph you were talking about. Others might make the same mistake so the additional explanation is good.

Read pninen's response again. You think they are leveling off, but since the sample size is so small and defaults are supposed to be so small, it's inclined to do that - i.e. one goes late, it stays late for a few months, looks like it "levels off." Another goes late, the line jumps, stays there for a few months, looks like it levels off, another goes late, it jumps up, etc. Right now it looks level, but that's not because suddenly a bunch of loans are defaulting at a steady (not increasing) rate, but because the already lates are staying that way, and new ones haven't popped up yet.

In any case, the Bs are at almost a 15% default rate. They're supposed to be at, what, 1.8% according to Prosper's Experian numbers? And they're only getting worse (i.e. another loan will go late soon, making it "jump" again).

-arebelspy

I like this idea.

-arebelspy

Well this is why this second graph is so interesting to me

our assumption by looking at the pninen main graph is lates have constantly increased on a slope of 45degrees, which as a group sure seems pretty darn close to fact

but when I look at april 06 only, and break out grades I see flatlining happening after some initial bad loans VERY early in the process

so this would argue against a never ending slope that graph 1 seems to indicate is happening... i.e. the 45 degree angle continues on to month 12, month 18, month 20, month 26, forever

now the sample size is perhaps too small in apr 06, so pushing it up to May 06 (where my sampling size would almost double) I'd like to see if it shows a similar story and similar slopes on each grade to apr 06

the other thought is, if Bs and Cs (picking 2 groups that have flatlined in April 06 graph) pretty much stay as they 'are' with maybe only 1 or 2 more loans defaulting in next 24 months, BIG ifs I know, but IF this is the case this would tell us for apr 06 loans at least

yes the overall default rate is roughtly 15%, but
1) most of it happens EARLY (months 1-10?)
2) if you get past months 1-10 in the higher quality grades (i.e. B or C vs HR or E) you are 'reasonably' safe - barring a job loss or catastrophe to borrower

And if this 15% does end up flatlining for next 2 years, 15%/3 years = 5% annual default rate for Bs/Cs and we feel like Prosper might have a future....

Again these are big assumptions but just some thoughts

1) I'd like to take a bigger sampling month such as May
2) Until we see 24 months worth of data its all extrapolation anyhow...

I am just saying if we dont see a big jump up in lates in the ensuing 2 years of this data set, it would be the basis for some conclusions as I made above... but who knows at this point...

Of course May is going to look better. It is the main outlier month, despite the sampling size having doubled.

Ok then June 2006

I dont care which month, just a month with enough loans so that 1 loan in 1 grade doesnt move the line up 7%

June has slightly more loans that May 06

Each month we move forward we have to wait that much longer to get 'conclusions' at the back end, hence why I chose May.

I did a quick look at August. The first big month after April. Almost 3 times as much volume as April. It still is not enough data points. I excluded AF which lowered the numbers somewhat. For 'A' credit I got 2 out of 32 lates for 6.25% while for 'C' credit there are 3 out of 61 lates for 4.92%. That is as of today. As you can see a single change in late status can flip the numbers.

Perhaps a 2 month aggregated graph would be more useful. Instead of reporting on loans created in April, you could report on loans created in April-May. That would double the sample size.

Yes but even April and May combined is less than August. IMO we need triple August's number to get meaningful stats. Hey that probably is this Month's stats a year from now. <_<

I've made an attempt at posting monthly late data on the Prospers.org wiki. Here is a link to the May late data: http://prospers.org/wiki/2006MayData

Edit to add: I wasn't sure exactly how to post a link to that picture on this post so if someone else can send me a link, I'll edit this post with the actual graph.

thanks!

Sample size per grade would help that chart

Done.

how much leg work (time) does it take you to do one of these from scratch? I'd love to see one for every month starting when they are about 6 months old, i.e. in May 07, the November 06 would be the last one needed at this time

Not asking you to do that, just curious :)

and can you do a quarterly one

Apr-May-Jun 06, Q2 2006, should pull out some of the low sample count issues and should be a good base going forward for the next 12 months to follow on a monthly basis....

It takes awhile and I don't have time to do a lot more. I have done June 2006 here: http://prospers.org/wiki/2006JuneData

I will try to do more as time permits since I now have a spreadsheet for the ones I've done that I can copy from/to.

I did have a possible "ah-ha" moment when doing these. I wasn't around in May 2006 but didn't Prosper have a 23-24% max rate policy around that time? If so, that would explain why there were a lot fewer E/HR loans made during the month of May and why May's numbers have consistently been better than most of the rest of the months. Just a thought...

agree on aha to some degree - April should of shown the same good 'news' since the rate cap was in effect all of April

May 22, 2006
Interest rate cap: Prosper's maximum borrower interest rate changes from 24% to 29%.

March 30, 2006
Interest rate cap: Prosper's maximum borrower interest rate changes from 36% to 24%

****

Prosper saved us from ourselves

****

I'd say if you have limited time, instead of doing them individually for May and June

Just do a quarterly Apr - Jun 06, will be a better 'read' since the May effect will be washed out to some degree by the Apr and Jun "normal" data, and then 3 months worth of loans will be a bigger sampling of loans

Thanks for the work

wow, June 06s class of "D"s really suck! :o

Thanks for the charts guys they are very interesting

It is interesting how when you plot out the grades so you can see the internal structure, each month's portfolio has such different personality.

(IMG:http://i196.photobucket.com/albums/aa45/golfingluigi/2Q06Lates-1.jpg)
The number of loans in each CR is to the right of the CR in the legend.

I'm still more interested in the results since 2/12/07 added info. This year should beat the crap out of last year.

Ahh, so my theory concerning late/defaults in the first year and slowing in the 2nd year is starting to have some statistical representation.

There will be lates in every month of a loan's lifecycle, but I think as the lines continue to flatten out, we'll see a future for prosper and us.

I think Prosper should change their standards to <600 credit rating.