I can't overstate how well received Mike's article was here (at FundingKnight). We're one of the smaller platforms in the industry - having recently passed the £10m loans issued hurdle, but with a way to go before we hit £100m.
As such, we take the view that we need to do better than the bigger players in order to appeal.
So, we're left pleased that we're already doing some of the things Mike's calling for (for example, personalised numbers, and publishing of our version of the VAR model performance - what we call "original estimated defaults" against "actual defaults" on our loan statistics.)
But we're even more pleased that he's come up with a bunch of things to add to our "to do" list.
There are some points I'd like to make in response to Mike's article that hopefully explain why we don't (yet?) do all the things he'd like us to.
The four points I want to talk about are:
· statistical confidence,
· service, and
· systemic risk.
Firstly, we're in a regulated industry. And that regulation is not symmetrical - turning down a borrower that would have made our lenders a profit isn't nearly as bad as accepting a borrower that defaults and makes our lenders a loss.
As Mike writes, both the lenders and borrowers are our customers. However, the regulatory playing field isn't level...
When we're evaluating loan applications, we either come to an agreement with the borrower that we'll start an auction, or we don't.
Sometimes we'll reject a loan application, either because we believe the applicant is too risky, or (in the case of applications from startups, of which we get a huge number) because there isn't a match between the kinds of things our lenders are looking for (and for which we've built an evaluation team) and the applicant. Because of some committed lenders on our platform, we've managed to sustain a 100% success rate at filling auctions (and meeting the agreed reserve price.)
And, based on our track record over the last few years, it could be that we've been a little too cautious. To date, our lenders have suffered lower losses than our estimates suggested they would.
If our mean losses are better than expected then, as far as the regulator's concerned, that's a good thing. It's entirely acceptable to the FCA to have us estimate losses of 2.57% and achieve losses of 1.78% (as we did in 2013 for loans in our 3 shield risk band).
As my 10 year old son piped up from the back of car the other week "the speed limit is a limit, not a target."
Based on our results to date it could be argued that by not taking enough "risky" loans, we have undershot our risk estimate. That means that our lenders have done better than we'd predicted.
Of course, this would also mean that we have drawn the line too conservatively and that we may have turned down some borrowers who we should have accepted. And on that basis, we're letting down those borrowers.
Purely selfishly, I can live with that. Like our Chief Executive, I've taken the view that I should invest with every single borrower on our platform. And while I'd like personally to have more loans to invest in, I still have a portfolio that's been diversified across over 100 business borrowers. I'm keen that our analysts, our Head of Lending, and our Credit Committee should keep saying "no."
One of the benefits of the way we do credit evaluation, though, is that we take a broader view than some lenders - more about that under "service."
There are many advantages to lenders of being in the P2B rather than P2P spaces (from a "quality of applicant analysis" perspective, as well as a tax perspective.)
However, we make much bigger loans. So, our £10m mark means that we've only issued loans to just over 100 borrowers.
We'd love to be able to present "expected returns after defaults" numbers in (almost) the way that Mike suggests.
We currently write (and I quote from my personal dashboard.)
“9.0% predicted return on your performing loans, after estimated annual defaults”
Mike suggested wording, from his article, was that we should phrase it like:
“Your ex post return is thus estimated to be between 4.5% and 5%”
We don't think that's going far enough, and would love to be able to write something more like:
“Your estimated return after defaults is estimated to be between 8.5-9.5% with a 97% confidence.”
The point being that, without that "with a 97% confidence" wording, we don't believe that showing a range is more meaningful than showing a single number.
And this is where statistical confidence cuts in. My own degree is Mathematics and Computation. That means that I'm the third most qualified statistician in FundingKnight (the winner, in case you're wondering, is our Head of Compliance) and will be pushed into fourth place when our newest team member starts on Monday.
With our loan history to date, we can't generate statistical analysis at the required level of confidence - the nature of the underlying maths is such that we need a much bigger data sample (by 1-2 orders of magnitude) before we can show that.
This also explains why we split out the estimated return for "performing loans" - we don't have anything like enough information to show a rate on the "non-performing loans" and "loans in recovery" (more about them later.)
Of course, in coming up with our credit evaluation model, we've drawn on significant academic consultancy, built a team of experienced managers (with a mix of former bankers, accountants and former underwriters), and back-tested against huge databases.
But, until we've run our own loans against 10,000 borrowers, and seen those loans finish, we don't, and can't, have what a mathematician would call "statistical confidence."
We don't charge fees to lenders for registering or investing in loans. (The only fee paid by lenders is if they choose to sell loan parts on our secondary market - The Loan Exchange.)
However, we do have a phone number that is manned by our team during "office hours", and have service levels for responding to the phone and email queries that are massively better than those required by the regulator.
We can only sustain this level of support if we can design and engineer our systems to be sufficiently straightforward that, in the average month, less than 1% of our lender base contacts us for help.
The more esoteric the numbers, the higher levels of queries we have. There are customers who want to know how we calculate certain things (and some things we can't explain without using the sort of mathematical formula that includes a Sigma.)
So, from a design perspective, we give more prominence to the simpler numbers ("what return am I getting, before and after expected defaults", "how much have I invested in loans", "how much free cash have I got sitting around".)
We also give prominence to personalised numbers. That default rate on my dashboard is based on an analysis of my personal portfolio, not some hypothetical portfolio based on 100 loans.
When it comes to borrowers, though, this is I think where we do much better from Mike's perspective. Our analysis isn't just based on a naive calculation of the security of the assets, but on our assessment of the people. This is why, in the few cases where we have had problems with loans, we've found, more often than not, that the guarantors of the businesses have continued to make payments to our lenders, even though their businesses have collapsed.
Our ethos of "service to lenders", or course, requires that we flag those up as "in recovery", which is why our "non-performing" , and "in recovery" numbers always look big - because we flag them up so much earlier rather than waiting for multiple missed payments.
"Systemic risk" is my gut reaction to over-reliance on the VAR models Mike mentions.
This is basically because the only two ways of doing VAR are Mark-To-Market and Mark-To-Model.
We have an active Loan Exchange - but we don't have a liquid Loan Exchange, at least not if you're a seller, though this is improving.
Because of this, I don't believe that we can (yet) use a Mark-To-Market methodology for analysing VAR.
Which leaves us with Mark-To-Model.
In order to get there, we'd need to solve the "statistical confidence" problems I mentioned earlier, but we'd still be left with the systemic one.
I'd like to write a lot, lot more here, maybe another day - not least because I'm fed up of mainstream journalists not reliably knowing that "systemic" is a completely different word from "systematic."
We need to come up with a better alternative to VAR models for our industry... so if you have some ideas, please get in touch with us....
... though if VAR does turn out to be the only game in town, then let's not stop where Mike suggests and just do it for our book - let's provide a custom VAR for the loan book of every single lender.
So, symmetry, statistical confidence, service and systemic risk... my four Ss of Risk.