But payment protect is a prime example of adverse selection bias. Those that have good stable jobs and are confident of their ability to pay don't get PP, while those that know work has been a bit slow lately etc do get PP.
Printable View
I agree with you that that must be whats happening, but its interesting if you look at the figures for full payment protect in particular this group of people must be pre disposed to events or problems that cause defaults that are NOT Death, Terminal Illness, Disability or Involuntary Redundancy, because if the cause was any of these it would be a payment protect claim rather then a default
Had wanted to work out how to create a 'Fence' chart (finicky to get right) so thought I'd give it a go for this. Result below:
Attachment 10054
Not sure if there is enough detail to be meaningful? Would running it for individual grades (i.e. A1, A2) be useful? Let me know what you think.
Perhaps there is a better way to show what you want - or it might be best to use a Pivot Table once you have the data and slice and dice as you want?
Argh, lost two posts :( stupid mobile interface has far to big of a Delete button...
Anyway, 3rd time lucky.
I realised I got my wires crossed with what you guys are asking for - I thought you wanted 'Enquiries last 6 months' vs Grades rather than 6 monthly period borrower cohort vs Grades.
Unfortunately, I don't think we are going to be able to generate this. When Harmoney started selling off debt they used the 'Last Payment Date' field for some unknown reason. By doing this they overwrote the date when the loan actually defaulted (i.e. when the last payment was made).
Will have to have a look and see...but that last time-lapse showed a very large chunk of loans being sold off and effectively having their default date wiped into the future.
Myles,
No, I meant the date the loan originates. So the cohorts are loans originating in 1h2015, 2h2015, 1h2016, 2h2016 etc.
So that, for example, we can see out of all the loans originating in 1h2015, how many A, B,...F have been sold/charged off etc..
X axis 1h2015A, 1h2015B...etc
Y axis percentage of loans that are charged off.
No time lapse, just bar charts.
This is so that we can see the percentage charge off rates for older cohorts compared to younger cohorts.
Despite not requesting this data :), the result of grade vs enquiry is interesting.
Shows that grade is a far more important then # of enquiry in predicting a default. Hard to gleen much (as there may be confounders such the cohort mix), but it seems to suggest a 6+ defaults at A-C grade spikes default; D to E grades it doesn't seem to matter, and at F there appears to be an inverse relationship ie more enquiries results in less defaults. Fascinating!
Yep thats exactly what requested.
I'm guessing with your IT skills you don't need me to tell you. However I run cohort off a vlookup on column B "Date". Run a column on the combined export, and you should have all the data you need. Like the Fence post graph btw, very sweet presentation!
Attachment 10057
Early release of basic chart set for comment:
summary.pdf
I'll remove or break this link when the final data set and summary are done.
Looking for comments on the graphs etc. Will probably add to this over time, but this is just to get it out there for some initial comments so I can fix most of what needs fixing before the final data set comes together. Decided on pdf to keep it all together and for ease of viewing.
Notes:
- The charts are all based on the full set of data (except where stated), which means the default values are watered down a little due to the approx. 9 month gap between when a loan starts and when a default is flagged as having occurred (possibly ~7 months in the past). Do we just want to live with this or make some arbitrary cut off? (or run another set)
- Some of the earlier loans, I feel, don't reflect the more recent loans i.e. there was some 'dodgy', 'poorly' graded loans in the early days. These will likely be inflating the more 'current' default rates. Should we consider dropping some of the earlier loans out (some are still current)?
- Perhaps the above two cancel each other out to some extent?
- Any time based default details is broken due to 'incorrect' use of the 'Last Payment Date' field so some caution needs to be taken around this type of data.
- Dollar values of loans i.e. 'Outstanding Principal' etc., are of no value, since they are only for the particular portion of the unique loan record in the data set - so it needs to be understood that using these values would result in meaningless detail. (A couple of values like the 'Total Loan Value' can be used - ratios might be okay.)
- I've limited quite a few data sets to a population of at least 10, just to remove outliers so core detail in charts etc. aren't influenced.
This is great Myles- awesome work putting it all together.
Can I get you to double-check the Rewrite graph? (Am I reading it correctly that 0/6610 have defaulted?)
Included in the dataset should be this loan, which although it had 0 successful payments before being re-written, still got classified by Harmoney as a Re-write. So there should be at least 1 on that graph? (It should have been included in the orders.csv uploaded by 'Wilma@Bedrock.Flinstones', yesterday afternoon)
Attachment 10058
**edit**
I've gone and looked at the csv, and it looks as though Harmoney have adjusted the 'Previous Loan Pay-off (re-write) column to N/A, whereas in the Harmoney portal/dashboard under reports, it shows this loan was in fact a re-write. I get the feeling that the re-write info on charged off loans is unreliable, it looks to me as though the re-write info shown in the csv changes once the status goes from in-arrears -> charged off.
Attachment 10059
Many thanks Myles. This is great. First thoughts after a brief squizz: 1. Some cohort based data would be helpful or maybe just exclude loans below a certain age eg 12 months. This could be at letter grade level if that helps. 2. Some measure of net return would be of interest though of course with changes in rates and algorithms there will be comparability issues and the time factor would need to be accommodated. Maybe apply to just completed loans? 3. When you say full data set I assume you are not including duplicates.