Harmoney

[humvee]

Hi Myles

How many loans exist in the combined data set with the each of the following problems

Negative principal outstanding

Paid Off/Cancelled - but still owe money

Active/Arrears but $0.00 principal owing

As I assume others have the same problems as I have had with data quality - Or maybe even more concerning - if this data is correct

[myles]

How many loans exist in the combined data set with the each of the following problems

Negative principal outstanding: 49 (-$152.13)
Paid Off/Cancelled - but still owe money: 248 ($1350.66)
Active/Arrears but $0.00 principal owing: 8

Some of this could be due to timing, small rounding type errors, and there will be duplicates? Doesn't look excessive for the size of the data set?

[myles]

Thoughts on this style:

Attachment 10063

[leesal]

Thoughts on this style:

Attachment 10063

Really good. Thanks Myles.

Appears that HM had trouble grading C and D grades early on in '14. Otherwise looks to be running consistent. Wonder whether there's early signs of improvement in DEF grade defaults out of platform 1.5

[RMJH]

Thoughts on this style:

Attachment 10063

That one I really like. Big picture and time based. And as time moves on it should give increasing information on default variability/predictability. If you could show which risk model was being used for each cohort so much the better.

[beacon]

Thoughts on this style:

Attachment 10063

Great graph, Myles. Illuminating to see defaults reaching a quarter of the loans already, in the earlier cohorts for DEF.

[leesal]

Fantastic effort so far Myles. And a very good pickup by alundracloud on the rewrite data glitch.

Will however comment that many of the graph's lack predictive validity, as you haven't controlled for confounders I'm guessing there's significantly more AB grades in home improvement and 50-59 age group loans; and more EF grades in 20-29yo and household items (for example).

The question begs whether a "C" grade 20-29yo is more likely to default compared to a "C" grade 50-59, and similar with Home Improvements vs HH Items or Used Cars. If not those categories may already be fully explained by the grading category assigned by HM.

Commend you on your efforts so far though, top work!

Capture.JPG

[myles]

Will however comment that many of the graph's lack predictive validity, as you haven't controlled for confounders

It was never my intent to run numbers for all combinations, nor to validate the grading system...

By putting together some high level detail, individuals can drill down to 'discover' what they think are key predictors and perhaps share what they find if it is of value. As I said before, a pivot table can be used to do much of that type of work. There is plenty of research available that has covered much of this ground before - whether it applies is the question.

[beacon]

... there's significantly more AB grades in home improvement and 50-59 age group loans; and more EF grades in 20-29yo and household items (for example).

The question begs whether a "C" grade 20-29yo is more likely to default compared to a "C" grade 50-59, and similar with Home Improvements vs HH Items or Used Cars. If not those categories may already be fully explained by the grading category assigned by HM.

Commend you on your efforts so far though, top work!

Capture.JPG

I had observed in my data too that riskier grades were predominantly youngsters and safer grades elders, so there is an element of safety with age already built into HM grading, I think. But interesting insights on household items vs home improvements, leesal.

[myles]

I've updated the summary document with a few more charts: summary.pdf

Would appreciate thoughts on the last chart - Estimated vs Actual Default Rate - which I've tidied up and made more informative (I hope). Harmoney, in their definition of Estimated Default Rate, make no mention of limiting the 'window' of loans by ignoring the first period where loans don't default/aren't flagged as defaults. This is the approach I've taken to be consistent with their figures (i.e. no window).

However, other platforms, e.g. Lending Club, limit the window to loans older than 120 days:

"Annualized Charge Off Rate is calculated by dividing the total amount of loans in charge off by the total amount of loans issued for more than 120 days, divided by the number of months loans in charge off have been outstanding and multiplied by twelve. The loans issued for less than 120 days are excluded from the calculation because loans are unlikely to charge off during the first 120 days."
Source: Lending Club - How We Measure Annualized Charge Off Rate

If anyone has thoughts on this, I'd appreciate them.