Estimating the Long Term Default Rate

Great analysis, thank you.

I was going to ask where you found the data, but I think I may have cracked it. Are you taking the difference between column M of the loan book download (term in months) and column P (payments) to be the number of months paid before default?

Not sure I get your APRs, though. Why isn't the APR on an 8.5% loan 7.5%, after FC's 1% charge but before defaults?

Something wrong with the link. You may have meant to link to here?

forum.fundingcircle.com/showthread.php?9890-Weekly-Lending-Review-week-45.

I understand APRs, but I was unaware that FC's headline rates are not APRs. How embarrassing is that?!

It took me a while to find the one post in the thread which tells me how FC interest is calculated, viz. accrued nightly at R/365 on the balance outstanding (where R is the annual rate). This is not quite definitive, however, partly because it is not an official source and partly because it is not crystal clear from the post whether the interest compounds nightly, or is simple interest on the balance. I could not find the answer on FC's website. Do you know?

Thanks Chielamangus. I have only had a chance to scan this, but it looks impressive.

Interesting that you come up with the opposite conclusion to me: "2. The FC estimates of provision for defaults are conservative and appropriate to use. Actual repayment performance to date has been better than anticipated."

My view is that defaults are around 50% higher than FC suggested provision. We can not both be right. I will study your paper in more detail and see if I can work out whether there is something wrong in my view.

This is one case when I would not be too unhappy if you were right and I was wrong!

Maxmarengo & Transo - your assumptions about real default levels were the same as mine, so I was quite surprised at my conclusion!  I double checked, and I could see nothing obvious or major to change.  Yes, I could have used a weighted average for FC's estimated default rate (and probably should have) but looking at the data, though not calculating anything, I didn't think much would be changed.  So thanks for the comments.  I'll wait until they are all in (well, a week) and then I'll revise as appropriate. 

An observation:  it is not in FC's interest to understate the expected default level as they would lose a lot of credibility and ultimately investors.  I just wish they would be more explicit/transparent about how they get their estimates.  As a cynic, inter alia, I don't trust any data that comes out of thin air.  Everything needs to be verifiable - including my own conclusions.

I will have a look at this after Christmas. In the meantime, a couple of thoughts:
- I think it is worth looking at the mix of loans in your completed cohort. They will include almost no C's.
- I have a gut feel that on the whole small loans are more likely to default - that is where the loan size comes in. (NB: have not done the analysis on this).

On the suggestion that FC have no interest in understating the default rate, I have two thoughts as well:
- FC may not know any more than us when it comes to estimating the default rate.
- Right now FC's Directors best strategy is to grow the platform then sell out. So they may not be around when the chickens come home to roost.

So resisting my natural cynicism, I think FC have put together an estimate that they can justify and they may even believe it. But I doubt they lose any sleep over it.

I believe that to be true. Also, some sectors show significantly higher failure rates than others and some geographical areas show significantly higher failure rates than others.

But IMHO none of those factors should be "weighted" when trying to get long term default rates. Broken down by, yes, but not weighted.

I've been doing analysis on this for several months now, although mine is based on losses per pound-year (i.e. if you have £100 invested for a year, collecting interest, how much of it will evaporate in bad debts). I picked that way because, as has been mentioned, there are very few completed loans (even fewer ran the full distance) and plenty of losses on loans which have barely started, which really eat into your profit margins. The numbers so far are as follows (well 'were' the last time I ran it). columns (ha, I assume the forum will eat the tabs so you'll have to work it out) are total loan, loan months, late£, losses£ (after recoveries iirc), and then %PA loss, and the last number assumes that some %age of the currently late will also go mammaries skyward 'ere long. This is all subject to change every time FC updates the loan book data (and has not been audited, so use at own risk).

Risk Band: 0 - A+ 362 5457 £224,347.75 £221,851.86 1.2% 1.7%
Risk Band: 1 - A 1066 11776.5 £511,051.79 £648,730.43 1.7% 2.3%
Risk Band: 2 - B 1000 11898 £772,932.13 £1,056,921.97 3.0% 3.6%
Risk Band: 3 - C 926 8229 £453,356.46 £648,734.19 2.2% 2.8%
Risk Band: 4 - C- 152 403 £146,951.00 £0.00 0.0% 2.9%

There is a significant worsening of bad debts at month 9, and again at month 20. I have tossed all the loans into the same pot - 12,36, 60 month etc. It is also hard to figure out when early repaid loans were repaid, so there may be some loan months in there which never really happened (which would make the default %ages worse). There are few 'statistically significant' things you can use to pick out bad loans - credit score isn't bad, and the period when the loan was taken out (even after allowing for the extra age of earlier loans). Risk band is NOT (how scary) a statistically significant determinant of loans defaulting. Partly this is because the sample sizes of bad loans are still too small (thankfully) - when only 2% go bad, the deviation you can expect on a sample of a few hundred is rather large. At the current rate it won't be long before C-s are statistically significantly the most reliable loans. Ouch.

Personally I think your only defence is very wide de-worsification, and/or getting rid of squeaky wheels wherever you can. Alternative might be to invest in very few companies and track them religiously on a weekly basis, but even then getting useful data is likely to be impossible (cf Burden, iirc). I can't demonstrate the Fc numbers are wrong, but I can't see where they come from .. it's the leap from the measured defaults / loan book (which completely ignores duration and age or loans) to the 'expect 5% PA losses on C-' which I can't follow.