How to Quantify Lender Risk in P2P

This article is a child of the parent article “Does UK P2P Communicate Lender Risk Well?”

Hints For The Scholarship Candidates

To throw a couple of bones out there for the scholarship candidates… stick your hand up as soon as you have got the right answer…

First bone.

Risk as I mentioned (laboured?) in earlier articles is the uncertainty of outcome.

Second bone (bearing in mind the parent article critique of the P2PFA loan data format not just being “not ideal” (cough) but fundamentally being producer-centric).

21stC Fintechs thrive by being customer-centric [and both the borrowers and lenders are your customers].

Anyone got it?

Let’s take a quick sketch at what this means for the four basic types of P2P lending/borrowing models.

I/IV Quantifying Lender Risk in the Simplest P2P Model

Let’s say I go to a P2P running the simplest P2P model – no bells or whistles.

Let’s say I “do the right thing” and diversify into a portfolio of a hundred loans.

Telling me my expected return [yield minus fees minus “expected losses”] is (say) 5.5% is utterly useless in terms of conveying any risk information to me.

It doesn’t. 

Even worse than that platforms seem to have their own philosophy of giving more like median loan loss expectations or pessimistic loan loss expectations [and others no doubt may give wildly optimistic “estimates”].

Anyway for avoidance of doubt the expected losses are NOT a risk measure – it’s a return parameter dudes!!! 

Let’s say you have a life assurance business  – the fact that everyone will die is not a risk it’s an expectation.  The risk is that your calculation/estimate is wrong and by how much. The average lifespan is 81.5yrs in the UK.  The risk in your business is that your folks might die on average at say 75 or 85.

Let’s say you insure a whole bunch of cars against theft. It’s not a risk that some will get nicked – it’s a certainty.  The risk is that say you assume 2% will be but it ends up being 6%.

Risk is uncertainty. 

Uncertainty means a range of outcomes.

[If you want to be geeky you can substitute “probability distribution” for range of outcomes but let’s (a) K.I.S.S. and (b) as I covered before excess mathematical thinking and one soon forgets the difference between sums and the real world.]

So back to my example – I have a portfolio yielding (say) 6%.

What I want from you is my risk (not your many years later information on how many loans defaulted in a whole calendar year a long time ago!).

So – and this is not beyond the wit of man – especially a man with a computer and the Blue Peter Guide to Statistics – what the investor wants from you is something like:

“Portfolio current yield 6%; based on our modelling, in normal circumstances, by the end of its lifetime you should expect total credit losses of between 0.5% to 1.5%; your ex post return is thus estimated to be between 4.5% to 5.5%”

Simple eh?

This also means that we can do away with this magical “one-hundred” number … for any investment (even a single loan, or a thousand loans) you can give me some modelling of the uncertainty of my outcome.

Which is the risk.

You can also tell me not just the risk in today’s purchases but how that affects my overall portfolio of loans with you.

Isn’t that cool? (so I can increase or decrease my risk profile and I can know by how much J).

II/IV Quantifying Lender Risk in a Credit Grading P2P Model

Let’s extend the simplest P2P model to include credit ratings.

Of course we retain the same quantification statement letting our investor know what the range of outcomes might be.

A further challenge/deficiency at present is that I don’t think I have seen any platform that demonstrates the risk in the rating categories :-O

All (?) credit rating based platforms do show expected losses per grade (“expected performance”) but don’t show the vital spread of potential losses (“risk”)! 

It’s the same groupthink/blind spot.

Higher risk assets are both expected to lose more and be more uncertain in outcome.

So eg a portfolio of 100 A grade loans might have expected losses of 0.5% +-0.2%; 100 Z grades 50%+-25% [or we can talk about probability of default but am trying to KISS].

It’s this wider spread of outcomes, it’s this greater uncertainty of uncertainty as one increases credit risk, that is what stops one from just working out which grade has the best apparent return of yield minus expected losses.

Sadly though as a lender right now you haven’t got a Scooby.

[Google translate says lol “Scooby” = “clue”]

Furthermore the vital function of a credit-rating-type-P2P, that one must be able to assess, is how well they assign loans to credit ratings.

If they do that well they are a great “marketplace platform”.

If they do it badly it’s a disaster – in extremis it almost ends up almost falling under the trade descriptions act [cf “I bought this hifi which said it was 100W and it wasn’t” with “I bought these B grade loans and they all defaulted”].  Interesting (preventative) work for lawyers here? J

Right now I can’t see any information that would allow one to say “yeh these guys correctly assign ratings 20%/50%/80% of the time”…

Definitely “scope for improvement” chaps.

III/IV Quantifying Lender Risk in Collateralised Lending P2P Model

Some platforms are do over-collateralised lending and haven’t lost a penny to date.

Obviously we can’t quantify any risk here – it requires a more qualitative assessment.

Picky points would be to look into how long it has taken to go through the legal process of taking the security and selling it (as “no loss” can mean “got it back eventually” or in the strictest sense “every cashflow was on time”). 

Also in terms of LTV ratios an incorporation of risk information into those would do away with overly-mathematical things like 67.4% (as you can’t be that sure of the value) and replace it with LTV 65-70% [rosy scenario, pessimistic scenario].

IV/IV Quantifying Lender Risk in Provisioning P2P Model

Provision funds add a further level of complexity to the issue of calculating risk.  Not least of which there are a variety of models (not always I feel well understood by non-FS folks?).

A simple provision model will be something like borrower pays interest rates + fees + contribution to provision fund.

On the other side lender receives this cashflow plus (potentially) payments from the provision fund.

Of course where it gets complicated is the whole opaque tangle around not just now around “expected losses” (which is subjective and where folks, it would appear, interpret the meaning differently in the first place) but also re provision fund contribution rates, (apparent) surplus etc.

Furthermore to simplify there is a spectrum between provision funds which “pay per loan” and those which are more designed around treating the entire loan book as a portfolio to which all lenders are exposed.  Here if the provision fund lacks funds to pay an individual everyone’s individual asset is effectively converted into a share of the whole loan book.

However in principle we are at the same place as in prior models.

It’s the platforms obligation to provide risk data to the lender in a simple, clear, useful way.

It needs to do some modelling (albeit rather more complex) to give a range of expectations for the lender – eg 5.1% p.a. over 5 years with expected credit losses of (say) 0%-1%.

Measuring and Comparing Platforms’ Credit Performance

With a bit more sleight of hand, stats, and an industry trying to help lenders (tease, tease lol) we could even compare platforms in the same way banks’ VAR models are compared with backtesting.

For those who don’t know how this works, simply put you estimate/model every day what your likely trading performance is to a certain probability.  Let’s say you say +-$100m today.  And you get a tick every time it’s within your range and a cross every time it’s outside.  It’s then easy to see how well your model is working [eg if your model was expected to deliver 90% ticks and 10% crosses and it got 88%/12% it’s fine but it’s not if it say got 12%/88% lol].

Similarly for lenders one would like to see that eg Platform A – 90% of the time investor returns fell with the range quoted upfront; Platform B – 10% of the time.

And for credit grades we could measure whether the expected return envelope for say Bs was correct or not.