Harmoney

[humvee]

E4 with partial pp, just up.

Previously would have mulled this over, and on a day with few loans would have taken it. Thanks to Myles's data will leave this one

Attachment 10072

I was looking at this one too but what caught my eye was actually the level of income on a benefit , By my calculations to get that level of after tax income would require a yearly before tax income of around $78,000. Are we really paying benefits that high? no wonder i pay so much tax (or is harmoney income data wrong again?)

That is equivalent of 1 person working full time for $37.50/hour or 2 people working full time at $18.75/hour or 91 hours of work/ week at minimum wage

[leesal]

I was looking at this one too but what caught my eye was actually the level of income on a benefit , By my calculations to get that level of after tax income would require a yearly before tax income of around $78,000. Are we really paying benefits that high? no wonder i pay so much tax (or is harmoney income data wrong again?)

That is equivalent of 1 person working full time for $37.50/hour or 2 people working full time at $18.75/hour or 91 hours of work/ week at minimum wage

that got me too for a moment, but the beneficiary seems to be earning only $30k per year. Still if thats the case, my wife can pack in her job and bunk up on the dole.

[myles]

How useful/useless is this mapping of 2nd level loan criteria?

What is that you ask...I've mapped what I think are the more practical criteria against all others (2nd level) and then sorted (descending) by default rate (and then population - higher population means more significant). Limited to a population of 10, just to get rid of the chaff.

It takes a little bit to generate, but I should be able to run it for individual grades and either report it or dump it out to csv. However, population numbers might be too low at the grade level, even with this data set (>20,000)...

I've not spent a lot of time going through it, but it contains some key info - I think?

Added: Just to explain the first row in case it's not obvious:

This is taking all loans with a 'Residential Status' of 'Fully Owned - No Mortgage' AND a 'Loan Purpose' of 'Debt Consolidation' and calculating the number of defaults, the total number of loans and the default rate. In this case there are 118 loans that fit these two critera, none of which have defaulted.

This is the top of the chart, which is a fairly obvious set of criteria to take that lead place...

[leesal]

How useful/useless is this mapping of 2nd level loan criteria?

What is that you ask...I've mapped what I think are the more practical criteria against all others (2nd level) and then sorted (descending) by default rate (and then population - higher population means more significant). Limited to a population of 10, just to get rid of the chaff.

It takes a little bit to generate, but I should be able to run it for individual grades and either report it or dump it out to csv. However, population numbers might be too low at the grade level, even with this data set (>20,000)...

I've not spent a lot of time going through it, but it contains some key info - I think?

Added: Just to explain the first row in case it's not obvious:

This is taking all loans with a 'Residential Status' of 'Fully Owned - No Mortgage' AND a 'Loan Purpose' of 'Debt Consolidation' and calculating the number of defaults, the total number of loans and the default rate. In this case there are 118 loans that fit these two critera, none of which have defaulted.

This is the top of the chart, which is a fairly obvious set of criteria to take that lead place...

If it shows the same types of categories occurring again and again, then useful as a data mining tool!. Being able to compare against some form of level of return however would be helpful, its difficult to contextualise the risk/reward benefit without.

[leesal]

Further to Myles efforts, couldn't resist doing a multivariate model.

comparing two loan sets:
1. Debt to income < 20%, enquiries 3 or less, time in job of 3 or more years, exclude new vehicle - new boat - tax - wedding expense - other loan purpose, exclude lendors aged over 60, exclude loans having 2 or more defaults, exclude partial PP, exclude B and C grade who aren't homeowners, exclude all A and all F grades.
2. loans that don't match the above criteria

Method:
* Included only loans dated prior to 1/1/2017.
* All $ values are common sized. That means all Interest, Charge-Off,Arrears, Investment, brought down to the size of one $25 note.

The results show that the model selection produced an average default of 3.6%, while the control had average default of 7.5%. Grade results, showed defaults were approx half in BCD, while E was only slightly improved compared to the control.

Need to catch up on sleep , time to give data-modelling a break..

Attachment 10077

Attachment 10078

[BJ1]

HM 161018.JPG

This loan tells us that when the borrower moved into the new home (presumably with a new mortgage) s/he had a large Harmoney loan in place already committing near to 20% of after tax income. I have difficulty accepting that the bank knew about it as the total expense to income ratio would have been unacceptable. Why didn't s/he use the bank to refinance Harmoney. The loan will fill reasonably quickly even though it doesn't make sense to rate this B1.

While I don't deny there is value in understanding more about the greater harmony portfolio, keeping loans like this out of one's own is the first step to reducing defaults.

[RMJH]

How useful/useless is this mapping of 2nd level loan criteria?

Interesting but too much detail for such a thin market to be of practical benefit? And isn't that what Harmoney's AI would be designed to capture? I think all of this stuff should give us confidence that HM's grading system has been good in the prevailing economic conditions. It certainly has challenged some of my filter prejudices. If you cherry pick too much you could end up with low diversification and too much cash. Thanks again for all your hard work and for engaging our minds and shining a light into some dark corners.

[myles]

Interesting but too much detail for such a thin market to be of practical benefit?

At grade level it can pull out some anomalies worth considering. Some examples (bold is default rate):

Potential Avoids:

B

Age Band

Above 60

Co-Borrower

Yes

3

42

7.14

C

Age Band

20-29

Residential Status

Living with Parents

5

36

13.89

C

Residential Status

Other

Loan Purpose

Other

5

24

20.83

D

Marital Status

Married

Residential Status

Boarding

10

63

15.87

Potential Picks:

D

Income Type

Employment or Self Employed

Loan Purpose

Education Expenses

0

86

0.00

D

Loan Purpose

Education Expenses

Co-Borrower

No

0

78

0.00

D

Residential Status

Renting

Loan Purpose

Education Expenses

0

67

0.00

and some significant repeating of particular criteria at the high default rate end e.g. Purchase New Vehicle, helps spot some combinations to avoid...

Some of these a clearly obvious, but they can help support assumptions, or not.

I'll keep playing with it and see if anything comes of it.

[RMJH]

Yes, keep up the good work Myles. One thing I would really like to see (from Harmoney) is employment vs self employment. Intuitively it seems bizarre not to distinguish between these two but maybe there is data to support consolidation on a portfolio basis. I don't like to invest in business loans but wonder whether many are hidden in other categories by the self employed.

[leesal]

At grade level it can pull out some anomalies worth considering. Some examples (bold is default rate):

Potential Avoids:

B

Age Band

Above 60

Co-Borrower

Yes

3

42

7.14

C

Age Band

20-29

Residential Status

Living with Parents

5

36

13.89

C

Residential Status

Other

Loan Purpose

Other

5

24

20.83

D

Marital Status

Married

Residential Status

Boarding

10

63

15.87

Potential Picks:

D

Income Type

Employment or Self Employed

Loan Purpose

Education Expenses

0

86

0.00

D

Loan Purpose

Education Expenses

Co-Borrower

No

0

78

0.00

D

Residential Status

Renting

Loan Purpose

Education Expenses

0

67

0.00

and some significant repeating of particular criteria at the high default rate end e.g. Purchase New Vehicle, helps spot some combinations to avoid...

Some of these a clearly obvious, but they can help support assumptions, or not.

I'll keep playing with it and see if anything comes of it.

Definately very useful. Over 60 with a coborrower; married and boarding! Both categories that jump out as being awkward. The data is pretty thin, but can be insightful. Your mining for raw diamonds.

For example Purchase new vehicle coming up bad, makes sense but didn't consider. On the other hand Business Cash Flow comes up pretty good. I wonder if the 25,000 loans the data set is missing is causing some selection bias.

Am now applying two default sets of criteria for loan selection - A default minimiser and the another that takes stable loans less likely to early repay. Both generate a return 3% higher then the data.

Funnily enough this E grade I normally wouldn't touch, met the conditions in my "default minimisation" set.

Capture.JPG