Right, I was talking about the Prosper cap (which used to be higher by the way).

pninen,

Excellent analysis and very good presentation. Personally, I have done exactly what you mentioned below (owning the prosper portfolio). My actual results fall into line with your predictions. I will lose money.

Do you have recommendations for Prosper on how to bring their portfolio going forward into line with what Experian would predict? I am interested in hearing your thoughts on that.

Regards,
AW

Guys,

I am junior (very junior) statistician at a cancer hospital (to remain unnamed). If you give me a dataset, I'll do some proper survival (reliability) analyses on the data.

I need the following variables for each loan: date of origination, date of failure, and the experian data including Current DQ, past DQ, and the works.

With that information I can properly stratify the data and give you true survival curves.

Feel free to email me a dataset: SamuelRTemple@aol.com.

I promise you I'll make some nice analyses worth your while.

Here's more fun: I've calculated annualized default rates for loans originating March-July where the loan was 0-$5000 and DTI<20%.

My rationale was that since we're comparing these default rates to credit cards with DTI <20%, we should look at statistics more typical of credit card line balances and DTI<20%. I get:

Still much much worse than Experian's rates, but the middle of the distribution is considerably improved. Sample size isn't huge, so the big jumps up and down between grades aren't unreasonable.

Very random question--what statistical software do you use?

I prefer to use SAS. I do use SPSS on occassion, and SPLUS very rarely. Generally the biostatisticians in the hospital perfer to use SAS. A lot of the doctors use STATA, I think.

Interested in Statistical software? Feel free to PM me with more questions.

Its time once again to update my analysis of Prosper's late rates.

I follow the rate at which loans become "late" in order to estimate the fraction of prosper loans that will default. Toward this end I follow the number of loans that have become "1 month late or worse". This number is the sum of two numbers that prosper reports "1+ months late" and "default".

Loans that are late a shorter time often convert back to current, but Prosper loans that have become 1 month late seem to stick in the late mode, and mostly head toward default.

Loans go late over time, therefore it is essential in the early life of a portfolio to track late statistics over time. In the early months, the late rates vs time will look nearly like a straight line. The slope of that line is the rate at which lates occur. If we know the probability that a late will become a default, we can multiply the late rate by this probability and get an estimate for the default rate of the portfolio. You know from reading my other thread http://forums.prosper.com/index.php?showtopic=6523 that only about 5% of loans that become 1 month late are being cured by the collection agencies. Therefore, something like 95% of these lates will become defaults. If the collection agencies were doing a better job, the math would be more complicated, but as it is, almost everything that goes 1 month late defaults, so complicated math isn't necessary.

Many people have been looking at late rates for the total prosper portfolio. This is usually misleading. The prosper portfolio evolves over time. New loans are always being added, and of course brand new loans aren't late yet. This distorts how lates are evolving over time. To solve this problem, I have divided the prosper portfolio into separate portfolios for each month of loan origination. The contents of these portfolios is fixed, so the evolution of the late rate over time tells us simply how that fixed set of loans is doing.

The late rates are ratios. The denominator is the total number of loans in the portfolio (which prosper calls "loans originated"). The numerator is the number of loans in the portfolio which are 1 month late or worse.

Here's the evolution of late rates vs time. Each curve represents one month of prosper originated loans. The dots are data points observed on the 1st and 15th of each month. The data starts at 60 days past the beginning of the month, because that's the first time you could possibly have any "1+ months late". You can see that while the curves are similar, some months are clearly worse than others. April was a bad month. May was a good month.

This chart is updated thru 12/01/06. (in other words two data points have been added since my 11/02/06 post)
(IMG:http://img.villagephotos.com/p/2006-6/1187065/prosperlate-2006-12-01.gif)

What we all would like to do is use the data we have to estimate how these loans will perform later in their life. Unfortunately, that is probably impossible. All the Prosper loans we are able to observe are still early in their life. I've tried fitting different models of delinquency vs time to the data. The problem is that the fits of different models all look alike toward the early part of the loan (because that's where the data is) and then different in the later part of the loan (because that's where the data isn't) and there's no way to judge one to be better than another. Meanwhile the curves all keep goin' up.

Another way of saying this can be found in "Credit Scoring for Risk Managers" by Elizabeth Mays. She says "It is important to have seasoned loans in your data set if you want to estimate lifetime default probabilities..." Right Liz. Unfortunately we don't have seasoned Prosper loans.

In my earlier post I discussed the simplest possible model, ie one where the default rate is constant vs time. Several people expressed displeasure with either that methodology or the result. So... use whatever methodology you like to extrapolate this data.

pninen, if it's convenient, would you please show us this same chart broken down by credit grade? Or, to keep it simple, perhaps just one chart for HR's only and another one for C's only? I'm thinking that the mix of loans has changed over time (i.e. HR's used to be preferred, but now C's are the "sweet spot") and I'm curious if we can see any patterns in their individual curves.

Update. I just added the 12/15/06 data points. I've also added the October loans. Nothing much has changed. The curves just keep moving up up up.
(IMG:http://img.villagephotos.com/p/2006-6/1187065/Prosperlate-2006-12-15.gif)

Thanks, Pninen!

I'm surprised there isn't an industry standard loan that is more similar to Prosper's so we could make comparisons.

That's what happens when your platform is the pioneer in peer-to-peer internet lending.

There are no 'industry standards'.

Thanks as always, pninen.

Thanks again Pninen! As usual, your visuals are always elucidating and certainly serve as a REALITY CHECK for the NEWBIE LENDERS who would be well-served by reading this BEFORE "investing" on the platform - especially since this data is not readily accessible (published) to prospective Prosper Lenders.

Having said that, for us more "veteran" Lenders, the data is always sobering - enough so that it's put many of us back into a helpless fetal position at times...

The more experienced Lenders (and regular posters) can all mourn for the Newbies we USED to be. I'd be hard pressed to find a Lender who doesn't reflect back on the earlier loans of their portfolio and moan while their head spins.

MESSAGE TO NEW LENDERS: Lurk...Hush (don't speak - no respectable/seasoned Lender here will listen to you anyway and any silly posturing will fall on deaf ears since it's all been said before). Learn from our mistakes so you don't have to repeat them all yourself...

That is correct, I have been ignored at times, and have been speaking too much

But it's just so exciting when you are new, and tempting to speak.

I shall start being quieter, i think

this thread, discussion forum and business concept are fascinating to me. i'm easing myself in very slowly and just experimenting.

prosper giving direct access of data to it's consumers for analysis is simply brilliant. i think the model is so immature that it is difficult to form any true conclusions about the long term success.

thank you for downloading, analyzing and posting this data/information.

Nice work. 2x's Experian default rates may be conservative for the entire portfolio. (IMO, this is a reasonable expectation for a portfolio of consumer debt re-fi loans heavily weighted to individuals that probably could not be funded elsewhere.) I just don't understand why we continue to bid these requests down to unprofitable rates.

pninen (or anyone really who has an opinion):

this is probably very elementary, but do you suppose the failure rate is simply because of a disproportionate number of HR borrowers actually getting funding vs. experian data?

hmm, that question isn't clear and i doubt my spelling.

regardless, what i'm trying to ask is, do we know, as compared to the experian data are more HR borrowers getting funded? it would make sense in a p2p model with inexperienced lenders.

thanks again for the processing of this data. simply fascinating.

If you read the first post, pninen broke the late %'s by credit grade.

All credit grades are performing horribly so far compared to the experian stats...

thanks loan chimp. there are no simple explanations that would magically make everything all ok.

84% success rate thus far (I realize we're still only a few months in on those original loans) is not that sobering considering the conduciveness of this platform for abuse.

I would scratch that high failure rate up to lender's lack of judgment as far as choosing borrowers.

Then again my opinion is worthless according to islandmele...so I'll stop chiming in now.

No, that's a good opinion.

You should just choose better borrowers than early adopters did. Let us know how you do.

I know it's been mentioned before, the A's underperforming on defaults, but this A & B data is shocking! Especially since the C's and D's are doing so much better (comparatively).

Do A & B's come here because they would have a problem at the bank that somehow we're not capturing in the data? Maybe their Experian report is higher than their TransUnion or the other one I can't recall, and that shows up when you pull all three?

Or does this market only make rational sense for C's and D's? Anyone else in it is acting for reasons we can't capture in the loan requests?

I'd be curious to see if the numbers change if you get rid of "social" and "family" loans-- you should be able to get rid of most of those by eliminating loans funded by fewer than (say) 10 lenders.

Good work, pninen.
As the defaults come in it is well that we reflect on your data.

I think that you make a good point.

Experian numbers appear closer for grades A-E and are explicity for DTI's less than 20%.

Just to check, I did a pull on DTI's under 20% by month and manually pulled out the HR's and NC's from both sides of the equation.

What I found was this:

May: 266 active and billed, 18 1+ month late, 6.76% (11.59% annualized - flat)
June: 257 active and billed, 20 1+ month late, 7.78% (15.56% annualized - flat)
July: 269 active and billed, 15 1+ month late, 5.58% (13.39% annualized - flat)
August: 369 active and billed, 24 1+month late, 6.5% (19.5% annualized - flat)
September: 284 active and billed, 14 1+ month late, 4.93% (19.7% annualized - flat)

Having said this, I don't really believe that the annualized numbers will be flat, I believe that the default curve will become flatter. Not flat, but flatter.

Based on Experian numbers, you'd expect the 1445 <20% DTI loans for grades A-E that originated during the May-Sept time frame to be 108 defaults (annual). Thus far we've seen that 91 of them have already hit 1+ month late, with the average having made about 5 payments cycles. If you flat-line annualize this you'd get 218 defaults. Take away the 5% that we can expect to be cured and you get 207 defaults.

It looks like a default rate of about double the Experian rates for DTI's <20% (which is what the Experian default rates are based on) is what we can expect for grades A-E. We don't have the Experian default rates for higher DTI's, but if we were to guess that these would approach the high-risk default rates of under 20% DTI HR's, then we're looking at a Prosper default rate of around 21% for A-E's with >20% DTIs.

I don't think many of us are likely to have planned for 1 out of 5 of our A-E's to fail EACH YEAR. I think most folks were looking at a number between 1:10 to 1:20.

If the trend continues, it could get quite interesting (and not too posite) around here, as that would extrapolate to nearly 50% of the high-DTI A-E's defaulting at some point during their 3 year life cycle, again, assuming a straight line default "curve".

Someone please check my "math". I'm not feeling so well all of a sudden.

Its time again for an update on the late-rates chart. I've just added the 1/1/07 data point.

(IMG:http://img.villagephotos.com/p/2006-6/1187065/prosperlate-2007-01-01.gif)

The big picture is that the curves just keep going up, looking very much like straight lines. Also, with the exception of the purple curve (May) these lines appear to have nearly identical slopes, ie nearly identical late rates, which I believe will imply nearly identical default rates.

Details re how this chart is constructed may be read in my posts above.

This month there are new difficulties in calculating these late rates. Prosper has "bought back" several loans, and in the process has removed these loans from their database. This has polluted the late & default statistics. Loans that were earlier defaulted are now no longer counted as having defaulted. But I consider them defaulted! ... so I added them back. My purpose in doing so is to provide a consistent presentation. If I had not done this, the curves would have taken a jump down today, not because the loans got better, but just because prosper has removed some from view.

The details work like this. The above statistics are a ratio. The numerator is (was until recently) the sum of the "1+ months late" and "loans in default" lines on Prosper's performance page. The denominator was Prosper's "loans originated" line. Now that prosper has removed loans from the database, the "loans in default" line and "loans originated" lines have been decremented. For example, the database used to show that 294 loans originated in April, but now shows 289. Luckily I recorded these numbers before the bought-back loans were deleted, so i still use 294 (consistently over time) as the denominator in the April statistics. Similarly, as we know that 294-289=5 loans have been removed, I have added 5 to the numerator. The numerator is now "1+months late"+"defaulted"+"bought back or otherwise removed from the database". Unfortunately, you can not reproduce this data using Prosper's present database. You need to have access to numbers which are no longer displayed. I find this situation distasteful, but I believe it is necessary to make these adjustments to avoid a stupid result.

In a line from the movie, the character Gollum from The Lord of the Rings was somewhat prophetic regarding pninen's chart and the late loans it represents;

The tunnel... :o

Thanks for the update, pninen.

This is a sobering picture. Thanks for posting.

ugh. ugh. ugh.


Jeff "ugh" D.

It would be cool to see these charts broken down by Credit Rating.

Do AA, A and B loans show the same slope?

Given the high number of HR loans on Prosper, the compositie chart above isn't so surprising.

Yes, sobering.

And bearing out the subjective impression.

Thanks, pninen.

This is great work, but all this tells me is what I can already see just by looking at a selection of the loans that are 100% funded:

People are funding EXTREMELY risky loans, probably without understanding the level of risk.

Hi all,

I'm a current borrower with a question. I'm ready to make my first payment on my loan today(early I might add) and I'm wondering if prosper will also take the Auto payment, on the due date 1-17, in addition to the payment I'm making today. OR will have to make a second (planned) manual payment. I can't seem to get a response from Prosper in under 3 weeks. I intend to make several monthly pmnts and just want to keep my books straight if making an early payment cancels the auto pay for that month.
I need to get at least 1/3 of my loan paid b4 I start lending.

Thx, Jon
:unsure: :unsure: :unsure:

Thanks for the update. :o

Although the numbers are worrying, I do believe they need a grain or two of salt. I think that most people on this forum have looked at loans that have gone through and shaken our heads. That reaction comes from the fact that we have different bidding criteria.

I ran the statistics for my preferences during the months that I?ve been a lender and the number of ?1+ months late? dropped dramatically, from 5% to 1.5%. When I look through Pninen?s numbers, I draw a couple of conclusions:

1) Biddng across the board on all Prosper loans is a very bad idea.
2) Lenders are bidding on many loans that I would never consider for myself.
3) It is certainly possible to make money on Prosper and I have every intention of doing so

To use an analogy, it is certainly possible to sit down at a blackjack table, split 10s and double-down on 12s ? but you?ll lose money a lot faster than someone who plays more carefully. I think which loans you bid on is more important than the overall late rate.

It's good these are only 36 month loans... according to the chart they will be 100% late in 50 months! :D

pretty impressive!
I think every lender (especially NEW lender) should see this post BEFORE they do any loans, not just for the sake of money, it's for Prosper's own good

I basically quit doing any lending after 1/3 of my loans went to late (about 10), thought late rate was unacceptable and started to withdraw fund once they were paid... today (about 4 months later), I saw all lates become "3 month+" and so-called "professional" collectors did NOT cure ONE cent of them. OK, I lost 1/3 of my money but I still saved 2/3 of it (hopefully!). I'm actually released after seeing the statistics plot, no more worrying about my "lending skills or attitudes" whatsoever, it's just a FACT that almost every lender is going to face: stop lending and withdraw your funds (unless something fundmentally changed in Prosper), otherwise you will loose 50% (not even 1/3) of your money!

I strongly recommend this post to be PINNED (and updated monthly if possible)
OR
everyone start to put the link in sigs ;)

I have no doubt that some "elite" lenders can still make money, come on, even in the so-called "third world county" there are billionaires, right?(ok, billionaires won't even bother being here but you get the idea) However as a general business model, I don't see how Prosper can benefit both borrowers and lenders should this trend continue.

Does anyone else see a jag upward in the last few datapoints here?

Even if you assume the best case scenario, Experian default rates, maximum return is about 11%. Once you factor in the higher Prosper default rates, average return rates drop significantly.

(IMG:http://img.villagephotos.com/p/2005-11/1104071/prosper5.gif)

If prosper survives (no pun intended), good investors will begin demanding a higher interest rate to compensate for the higher default rates. And this will eventually drive the borrowers away because their rates won't be competitive.

I don't think it's as hard to make money as some think. Don't fund crappy loans. Those with 3+ Current DQ's are currently at 13.93% Late (190 loans) (of post May 1 loans)

There are currently 328 post May 1 loans late, those 3+ Current DQ's account for almost 58% of all the lates out there. Just by not funding those you'd make these numbers a lot better.

Is there any way to extend this straight line over the course of the 36 months?

Using this estimate, then we could answer these questions:

Is there ANY improvement between earlier loans and loans originating last month?

What is the ending defualt % at the 36 month mark?

There are two views. My opinion, from reading a little on the subject:
Classical models of debt assume that the defualt rate peaks at about a year. However, due to the 3 month selloff of Prosper loans, we will probably not see this.
However, subprime loans tend to exhibit different behavior- they go bad at random times.
I'm not sure what constitutes a sub-prime in a traditional lending market but I suspect it may be C or lower. Prosper C would be considered fairly high risk for a traditional lender.

Prosper sells off loans at 120 days, which is industry standard, AFAIK. Unless I'm misunderstanding what you're trying to say?

I should have clarified.You are correct.

QFE

A ruler, perhaps? ;)

While what you say is true of those with TOP credit grades who are WELL qualified for mortgages.... from my sources, currently about 80% of loans are NO DOC loans. They check nothing, except if you have a down payment of some sort, sometimes as little as 2%. The mortgage credit quality is much much worse than the banking system realizes.

A No DOC at 2% down ? Who?

P.

If you have a decent FICO, getting a 100% no doc is a breeze.

Try any broker. If they say they don't offer one, try another.

I'd be interested in seeing the total % of loans that default over the 36 month period at this rate...

Having a ruler would be nice...