LC FICO vs Loan Grade

[fliphusker]

I never did get my free toaster.  Just be sure to price those never late notes at -5% for me. 
I have not read this thread closely will do so later as it is quite interesting stuff.

Yeah, I think you're exactly right.
It's not often mentioned but odds versus FICO score is not a straight linear relationship.
Actually the log of odds (i.e. fully paid's / charge off's) is very close to linear FICO scores.
Straight line on a log graph and all that. Or so I've heard ...

BTW: I chose D grade loans to analyze simply because they make up 65% of my portfolio.
Maybe I should be doing a bit of Folio selling again.

[SLCPaladin]

Yes, exactly; you got it. Breakeven at 0% Mkup/Disc (including 1% service fee) appears to be a 25 point FICO drop. Who woulda thought ...
Now all we have to worry about is that I wake up tomorrow with another epiphany invalidating my results published to date. Caveat Emptor!

Don't necessarily think I screw up more than most (certainly not much less), but I freely admit when I do.
Think that's been an important contributor to whatever success I've had in my other endeavors.

However, I'm sleeping better and don't have anything rattling around in the back of the head except "rawraw's" time value of money.
That would argue the results are even somewhat pessimistic.

Keep in mind we are talking about D grade notes here! I'm sure the results for less risky loans (lower interest rate) would be quite different.

There is still something about your data that I can't wrap my head around. On the right-hand side of your graph where you have a -100 drop in FICO you show about a 1% return for notes sold at a 5% premium and a 0.5% return for notes sold at a 5% discount. I can't figure out how a swing of 10% would only make a .5% difference in returns. It seems like there should be a bigger delta between these two data points. The front part of the curve seems to make sense, but the data converges after about -50 and I can't see why this would be the case.

[Rob L]

As you proceed from left to right on the chart to which you refer above, the amount of FICO drop required to initiate a Folio sale increases. There are fewer and fewer loans that meet the sales criteria, hence more loans held. The buy and hold total receipts is always constant, $27,221,921.10. The graph shows simply (total receipts of loans held or sold at a given FICO change) / (total buy and hold receipts; a constant).

Here's a table including numbers of loans sold or held:



So, the dollar difference contribution of Mkup/Disc for the 286 loans sold at -100 point FICO change is small relative to the buy and hold total receipts constant of $27,221,921.10. Maybe this is a poor way of displaying the results, but that's what you're seeing. If you have a suggestion of a better way I could possibly look into it.

BTW: This also explains why the month over month FICO change graph converges to zero benefit or loss since there are almost no loans with a month over month change of -100 or more FICO points.

[rawraw]

I'm still amazed that the threshold is so low.  I always looked for 50-75

[Rob L]

I'm still amazed that the threshold is so low.  I always looked for 50-75

I always had Anil's numbers in the back of my head from his analysis a few years back and -40 was the number I remembered.

[rawraw]

I'm still amazed that the threshold is so low.  I always looked for 50-75

I always had Anil's numbers in the back of my head from his analysis a few years back and -40 was the number I remembered.

The interesting thing is fico drops of such a small amount sell way easier.  Perhaps I need to push my threshold back somewhat on high grade notes

Sent from my SAMSUNG-SM-G935A using Tapatalk

[Rob L]

I never did get my free toaster.  Just be sure to price those never late notes at -5% for me. 

I had deal where Wells Fargo was sending me a free toaster for opening an account I didn't know I opened.
When it arrives I'll send it straight off to you.
Just to be sure I don't get sued by WF, this was a joke.
(Although, I do have a checking account and an event happened that made me furious with them!)

[Fred93]

I have more or less validated your results.  I say more-or-less, because I never seem to do anything quite the way somebody else did it.  For example, I ran all completed 36 mo loans originated after 2011, whereas you ran just 2013.  I ran the "test" only in months where the loan was current, whereas I presume you ran it in months where the loan was current or late.  I did this because I thought it would be better to look for current loans to sell, as late loans sell at very poor prices.  There are probably more differences, as there are many details.  In any case, I get curves of the same shape as yours, although the numbers are a little different.  So I call that validated -- more or less.

But then...  This analysis depends mightily on what fraction of loans in the universe are detected.  In other words, the set you intend to sell.  I believe you used the word "sold" to describe them, as certainly the simulation presumes they are sold.

For grade D  and a 60 point FICO drop (presuming I am understanding your numbers), you see 27% detected.  633/(633+1704)  I see 20.5% .  Ok.  Given our differences in algorithm and data set, those are similar numbers.

But then... I ran this same test on my current portfolio.  It only detects 4.2% of my loans!  Yipes.  That's way different than either 27% or 20.5% !  Red flag.

Now you might think, well the fraction detected just affects the level of profitability, so I would have less profit from a selling program.  Yea, but if the number is this far off, then the other stats, like the fraction of detected who go on to default may be way off too.  This leads to possibly bogus selection of threshold, etc.  I feel like the robot from Lost in Space is waving his arms and yelling "Danger danger". 

Here's what I think is going on.  When I use all loans in the history file, that's a really broad set of loans.  I broke them down by grade, but nothing else.  My portfolio quite different than that, because when I bought those loans, I used a bunch of filter criteria beyond grade, such as: inquiries, DTI, years of credit history, income, and so forth. 

What I'm saying is that it seems likely that historical info from the full payments file is not representative of my portfolio, so statistics gathered on this broad unfiltered set do not apply to my portfolio. 

So maybe the problem is even harder, and we must filter the history, selecting only loans that pass some filter criteria similar to the filters we used to select the loans in our portfolios, and only then will we have statistics that we can use to guide us in how we should expect our portfolio loans to behave.

Doing one grade at a time was a good idea.  The statistics are different for different grades.  Maybe it just wasn't nearly enough differentiation.

[nonattender]

Yes, if you're really choosy up-front, you won't have so many that drop and need to be sold.  (I heard you say, in nine paragraphs.)

Another way to say that might be:  "For every up-front selection criterion, there is an uniquely optimal back-end sell-signal trigger."

(And you can go more granular from there, until it stops being worth it.)

[Fred93]

Yes, if you're really choosy up-front, you won't have so many that drop and need to be sold.  (I heard you say, in nine paragraphs.)

Dude!  I may be wordy ... but YOU count paragraphs!

And sadly, the return improvements from these two activities (upfront selectivity and later fico-drop-watching) aren't additive. 

Another way to say that might be:  "For every up-front selection criterion, there is an uniquely optimal back-end sell-signal trigger."

Yea.

(And you can go more granular from there, until it stops being worth it.)

May have occurred last Tuesday.

[Rob L]

I have more or less validated your results.  I say more-or-less, because I never seem to do anything quite the way somebody else did it.  For example, I ran all completed 36 mo loans originated after 2011, whereas you ran just 2013.  I ran the "test" only in months where the loan was current, whereas I presume you ran it in months where the loan was current or late.  I did this because I thought it would be better to look for current loans to sell, as late loans sell at very poor prices.  There are probably more differences, as there are many details.  In any case, I get curves of the same shape as yours, although the numbers are a little different.  So I call that validated -- more or less.

I'm really glad to hear that!! I always wonder if I've forgotten something and/or made some major screw up.

My test was to potentially sell current AND neverlate loans only. Loans not neverlate were never sold no matter whet the FICO drop. I feel like selling loans that are damaged goods (not neverlate) is a completely different process. My attempt was to model a process that monitors only never late loans one month at a time and to immediately sell the ones where the FICO drop threshold is tripped in the month the drop was detected.

But then...  This analysis depends mightily on what fraction of loans in the universe are detected.  In other words, the set you intend to sell.  I believe you used the word "sold" to describe them, as certainly the simulation presumes they are sold.

For grade D  and a 60 point FICO drop (presuming I am understanding your numbers), you see 27% detected.  633/(633+1704)  I see 20.5% .  Ok.  Given our differences in algorithm and data set, those are similar numbers.

But then... I ran this same test on my current portfolio.  It only detects 4.2% of my loans!  Yipes.  That's way different than either 27% or 20.5% !  Red flag.

Now you might think, well the fraction detected just affects the level of profitability, so I would have less profit from a selling program.  Yea, but if the number is this far off, then the other stats, like the fraction of detected who go on to default may be way off too.  This leads to possibly bogus selection of threshold, etc.  I feel like the robot from Lost in Space is waving his arms and yelling "Danger danger". 

Here's what I think is going on.  When I use all loans in the history file, that's a really broad set of loans.  I broke them down by grade, but nothing else.  My portfolio quite different than that, because when I bought those loans, I used a bunch of filter criteria beyond grade, such as: inquiries, DTI, years of credit history, income, and so forth. 

What I'm saying is that it seems likely that historical info from the full payments file is not representative of my portfolio, so statistics gathered on this broad unfiltered set do not apply to my portfolio.

Probably the big difference is that I chose a set of loans that were all fully matured (3 year loans, more than 3 years elapsed, final MOB loan status either Fully Paid or Charged Off). Testing a portfolio in progress will naturally yield many fewer sales occurrences since loans still have lots of time in front of them to trip the sell button. Of course there will be a difference due to loan selection criteria and a big difference across loan grades (intererst rates) in the portfolio.

So maybe the problem is even harder, and we must filter the history, selecting only loans that pass some filter criteria similar to the filters we used to select the loans in our portfolios, and only then will we have statistics that we can use to guide us in how we should expect our portfolio loans to behave.

Doing one grade at a time was a good idea.  The statistics are different for different grades.  Maybe it just wasn't nearly enough differentiation.

[Fred93]

Testing a portfolio in progress will naturally yield many fewer sales occurrences since loans still have lots of time in front of them to trip the sell button.

Oh, of course that's right.  Each borrower has many opportunities over time to let his FICO score drop, and in the history file all of 'em are evident.

But... head spins... seems like that effect should be on the average something like a factor of 2, or a bit more.  Loans in my portfolio probably on the average have lived about half their life vs the full loans we extracted from the history.  Then of course some of the loans in my portfolio that have dropped FICO have already gone bad, so don't show up on the snapshot, which is why I say a bit more than 2.  ...but that's not right either, because I'm ignoring history on the loans in my account.  I'm just looking at present FICO.  But it's not a factor of 36 either, because your FICO score isn't independent from month to month, although there is an independent (noise) component.  So I've got an armwaving argument that the factor should be between 2 and 36.   Hmm.  My measured factor of 4.9 is right in the middle.

I could reprocess the history file looking only at one random month in each loan, rather than every month.  That should be equivalent to sampling the current month in my portfolio.

But its worse than that.  When I took a snapshot of my account, it was the first time I'd ever done that.  Its not like I had been running this test and selling notes every month.  So I have accumulated dropped FICO notes, which will boost the number.  The true one month numbers for my account will be lower.  Yipes.  All these different ways of sampling should produce different numbers. 

I don't have any easy way to run a test on my current portfolio that just considers loans that have freshly triggered this month, because I don't have easy access to month-to-month FICO for those loans.  This info is in the payment history file, but I'd have to look up all my loans there one by one.  Seems like too much work.

This brings me back to a subject you've discussed a bit... the time element.  Need to get time into your calculations.  You've calculated how many loans trigger the test and/or how many $, but we need to know (I think) how many per year.  If you save me $1 but that's $1 per 3-year loan term, that's different than saving me $/year.   To get into the units of return that lenders think about, we need an annual number.  Should I just divide by 3, or do we need to do something more to figure this out.  Average length of a performing 3-year loan is something like 2 years because of prepayments.  So maybe a factor of 2?  Or...

While scanning the history file and running the test, we could compute the # triggers per month divided by the # loans seen that month, to compute a triggers per loan-month ratio.  x12 gets fraction of loans that trigger per year.  Perhaps this should replace the fraction of loans that trigger in their entire lifetime in your calculation. 

This is just stream of consciousness.  Anyone looking for a conclusion here will be confused.

[rawraw]

This is just stream of consciousness.  Anyone looking for a conclusion here will be confused.

LOL!  While I think that should have been at the beginning, it being the last thing I read made me literally laugh out loud

[Rob L]

I don't have any easy way to run a test on my current portfolio that just considers loans that have freshly triggered this month, because I don't have easy access to month-to-month FICO for those loans.  This info is in the payment history file, but I'd have to look up all my loans there one by one.  Seems like too much work.

The pmthistall file is only updated every three months. Not much value here. List all your notes on Folio for a ridiculous guaranteed profitable markup. Not that you want to sell, but that you want the latest FICO update ASAP.  You "should get the updated FICO within your sample time interval of downloading the smallnotes file. As a seller there's additional data in your mynotes file that can provide an additional bit of informational edge. The most recent FICO may actually be found there first; I dunno. There may be issues with staleness (where does LC publish the most recent FICO update first). I point out that possibility but don't really know the answer. If I were seriously into this I would know. I'm sure there are folks that do.

Simplifying things by ignoring time value of money is troublesome, but leaving it out is a worst case for selling the FICO drops. Returns should be better than those that ignore it.

Sometimes I think that if I'm not confused I'm not paying enough attention. Stream of consciousness wins hands down over the stream of unconsciousness I'm given to on occasion.
Those earlier charts I posted were embarrassing  LOL

[Fred93]

Another bit of (imperfect) validation...

Instead of running the test on current notes in my notes.csv file, with the sampling problem we discussed above, I realized that this file has all my Charged Off & Fully Paid notes as well, so I ran the test on them instead.  Now this isn't exactly the same as running the test on the history file, where I have updated FICO every month.  In this case I only have FICO at time of that final event.  Better than nothing, and I think final FICO and lowest FICO along the way probably aren't that far apart most of the time. 

The number  I got for FICO drop>60 detection was 22.4% (all fully paid & charged off in my portfolio, ie mix of grades), which very closely matches what I got on the history file.  So maybe my selection of notes doesn't change things so much after all.