P2P Lending / NFT Lending Forum

zpbsfg,

What's the horizontal axis? Days since a certain time?

Bryce

I don't think we need to do any more research. We've established a major change in the data from week to week. It's time to call LC.

Here's a table by vintage quarter with the percentage of loans with a public record. Now, these are issued loans, so perhaps a TON of the ones on the platform won't issue, but 46% is so far out of line with the past, it makes me think something's wrong. That said, a public record isn't the end of the world.


. table iq, c(mean publicrecord)

--------------------------
       iq | mean(public~d)
----------+---------------
   2007-2 |              0
   2007-3 |       .0740741
   2007-4 |       .0653951
   2008-1 |       .0856553
   2008-2 |       .0979167
   2008-3 |       .0734266
   2008-4 |       .0698925
   2009-1 |       .0552359
   2009-2 |       .0541825
   2009-3 |       .0361538
   2009-4 |       .0465632
   2010-1 |       .0584795
   2010-2 |       .0680249
   2010-3 |       .0613372
   2010-4 |       .0560286
   2011-1 |       .0562956
   2011-2 |       .0614628
   2011-3 |       .0634016
   2011-4 |       .0386435
   2012-1 |       .0314555
   2012-2 |       .0304205
   2012-3 |       .0296941
   2012-4 |       .0096432
--------------------------

I agree on the egg-nog bit.

It's hard to have a clear discussion on the web, as the way one defines the outcomes, makes their data set, and conducts their method has much to do with the results. Often people don't lay that out clearly enough to have a fruitful discussion, or readers gloss over key words that have important meaning.

We should have an in-person conference or workshop! Then we could dig in and have more time to learn from each other.

If anyone has collected information about *which* notes have gone into grace period, late 16-30, and late 30-120, I would be very pleased to do an analysis of them to find out any predictive factors that I can. It's too bad LC doesn't provide this variable in their extracts.

I'm a full-time college administrator and adjunct professor in finance at a medium sized teaching university. In a former life, I was an analyst at the RAND corporation, where I earned my doctorate in policy analysis. I have a baccalaureate in mathematics (pure), and have enjoyed programming for 20 years.

Many of those questions will be answered on my website within a couple weeks. However, I don't really want to give statistics lessons over the net. A few stats courses and you'd have enough knowledge to do what I'm doing.

By the top 10% thing, I mean that I programmed numerous strategies, including filters, and then compared the returns against each other. I used inception to Q2 2008 to train my modeling and tested all strategies on Q3 2008 to Q2 2009 loans. Their outcomes are known at this time.

Public record is just one thing among many that affects the probability of a charge off. An attractive loan along many other dimensions may still be attractive with a public record. I buy them occasionally if the math works out.

Employer is not in any file LC provides. I think some of these websites have been writing screen scrapers to get some extra fields.

I've always used $50 per note. The way I decide on loans in which to invest doesn't suggest too much advantage in making some notes outsized. They are all very similar performers in expectation.

If I do another large round of funding (I invested 20k in July and August), I will probably bite the bullet and go with $75 just to shave a couple weeks off the investment time.

Write a script in your favorite statistics package that loads all the data and allows you to enter a parameter of a loan id, reads the characteristics of that loan, and finds all loans that are within certain tolerances of that loan on as many characteristics as you like. Then output whatever performance metrics you want.

My concern is that this type of information asymmetry probably won't last. You are getting an advantage in the secondary market. At some point the market maker will recognize that this is harmful to their business model long term and fix it so that anyone can look up anything about any loan, including pre-grace period, days "processing," etc.

I'll give you this one for free: there's no relationship.

Defined "double_or_more" to be just as you say: double or more the amount requested vs. the revolving balance, but only defined among credit card refinance or debt consolidation. By the way, it's not that common to request twice as much or more--around 14%.


. summarize requested_to_revolving if double_or_more~=. & loss~=., det

                   requested_to_revolving
-------------------------------------------------------------
      Percentiles      Smallest
 1%     .0618095       .0286648
 5%     .1461917       .0298174
10%     .2376426       .0305769       Obs                3467
25%     .4672504       .0310281       Sum of Wgt.        3467

50%     .8508224                      Mean           139.4503
                        Largest       Std. Dev.      1157.935
75%     1.316779          10000
90%      2.70636          10000       Variance        1340813
95%     5.772595          10000       Skewness       8.343586
99%        10000          10000       Kurtosis       70.75363



Analysis restricted to 36-month credit-refi or debt consolidation loans completed prior to November 2012 (N = 3467). Defined loss to be being repaid less than the amount loaned. Chi-2 revealed no difference in proportion of loans repaid.


+----------------+
| Key            |
|----------------|
|   frequency    |
| row percentage |
+----------------+

double_or_ |         loss
      more |         0          1 |     Total
-----------+----------------------+----------
         0 |     2,551        425 |     2,976
           |     85.72      14.28 |    100.00
-----------+----------------------+----------
         1 |       419         72 |       491
           |     85.34      14.66 |    100.00
-----------+----------------------+----------
     Total |     2,970        497 |     3,467
           |     85.66      14.34 |    100.00

          Pearson chi2(1) =   0.0504   Pr = 0.822

Filter strategies necessarily restrict loan volume too much. Each characteristic only has a probability of causing no repayment. I come up with a probability of a problem and consider the interest rate, then invest in the top ten percent of a ranked list. Takes two minutes to create a list using programs I developed and a stats package.

Stata for the win. Any stats package will be able to handle this.