Harmoney

[myles]

A couple of charts that are an attempt to find loans that are more likely to default than others. These are based on my loan set so may differ from others and some of my criteria may limit some details showing up.

The number in the brackets after the description is the sample size for that group.

Attachment 9976

Nice data-set that highlights that some borrower cohorts appear to be more likely to default than others. I find it interesting that a De Facto relationship is lowest in that group. I haven't taken a loan from a 'boarder' for a long time, but some of my earlier loans were. Most other details are kind of what you would expect I think. The 'Home' income type stands out from the rest (still not 100% sure what it is defined as).

Attachment 9977

There was some discussion a long way back suggesting borrowers location might influence defaults - this chart certainly suggests that might be the case. Note I split out the smaller sample size regions just for clarity. Other regions, for me, have not had a default.

Hope there are some details here that are useful to others for 'picking' loans.

Just as an example, so it's clear what these are showing - I have had (or still have) 14 loans from borrowers residing in Timaru, just under 15% of those loans have defaulted (so that would be 2 out of the 14). [The 14 is in the brackets after 'Timaru'.]

[IntheRearWithTheGear]

Their model takes into account all the easy stuff such as location, wage, sex, working status etc into account - so i dont think there a holy grail for better selection from our lender point of view. Their model would most likly take into account that in auckland if a loan is submitted at 23:44 at night by sombody earning 14000 per month it should be marked as a loan default etc.

At the end of the day they see more data about the client POPULATION to classify the loan than we do - so it simply comes down to a choice. All we have is the single loan application.


One thing i thought would be interesting is picking defaults from spelling mistakes in the free text field.

I suspect the harmony model dosnt evaluate this "idea", in its classification. I suspect they dont even use this field in their models and as such could be a further advantage. (however we still dont have enough data to it) (yes, i seen the study in regards to religious words).

[Bjauck]

Their model takes into account all the easy stuff such as location, wage, sex, working status etc into account - so i dont think there a holy grail for better selection from our lender point of view. Their model would most likly take into account that in auckland if a loan is submitted at 23:44 at night by sombody earning 14000 per month it should be marked as a loan default etc.

At the end of the day they see more data about the client POPULATION to classify the loan than we do - so it simply comes down to a choice. All we have is the single loan application.


One thing i thought would be interesting is picking defaults from spelling mistakes in the free text field.

I suspect the harmony model dosnt evaluate this "idea", in its classification. I suspect they dont even use this field in their models and as such could be a further advantage. (however we still dont have enough data to it) (yes, i seen the study in regards to religious words).

Interesting - is a simple mistake or lack of proof-reading before clicking on submit just a slip or symptomatic of stress hich could precede a meltdown and default?

Errors based on lack of education could well be already reflected in the grade given to the loan application.

[IntheRearWithTheGear]

Somtimes you can guess gender from the words used in the description field.

example "I brought the house off my x hubby" - hence female.

Do females statistically pay back better than males ?

Obvously harmony knows gender in the loan application - us as a lender do not.

So i do think there is value in the field.

[myles]

A couple more:

Attachment 9978

Some interesting detail in the above chart worth considering. Looks like skipping 'Tax Bills, New Cars and Funeral Expenses' might be a good option. 'Loan to Family Members' surprises me, but relatively low numbers.

Attachment 9979

Take with a grain of salt, but I thought it of interest. This relates only to my loan set and selection criteria. A single default in the A5 grade with only 10 loans makes it look bad. However, I will likely dial down my E4 and E5 intake (though I need to look a bit further into it as there may be some influence due to earlier, much higher interest rate loans).

defoth.jpg

'Previous Defaults' are not as bad as I would expect (but low numbers - by choice). 'Enquiries in last 6 months' looks okay up to 3, perhaps even 5 considering my overall average default rate.

[alundracloud]

I've not seen this before- These two loans hit the marketplace at 0% filled (as if they hadn't gone through the auto-lend) and with only 11 days time remaining, usually 14 days.

11days.jpg

[joker]

A couple more:

Attachment 9978

Some interesting detail in the above chart worth considering. Looks like skipping 'Tax Bills, New Cars and Funeral Expenses' might be a good option. 'Loan to Family Members' surprises me, but relatively low numbers.

Attachment 9979

Take with a grain of salt, but I thought it of interest. This relates only to my loan set and selection criteria. A single default in the A5 grade with only 10 loans makes it look bad. However, I will likely dial down my E4 and E5 intake (though I need to look a bit further into it as there may be some influence due to earlier, much higher interest rate loans).

defoth.jpg

'Previous Defaults' are not as bad as I would expect (but low numbers - by choice). 'Enquiries in last 6 months' looks okay up to 3, perhaps even 5 considering my overall average default rate.

Thanks for the data Myles. It's fantastic. How are you capturing it all? Are you entering it all manually in Excel, do you work it via Harmoney's reports or do you have a more sophisticated system? I'm interested in collating my data similarly. Cheers

[myles]

How are you capturing it all?

Up until this last lot of graphs I was just using a spreadsheet, but now that I feel that I have some useful data to mine I import it into a database. Just using the Harmoney Report Loans Export data (csv).

What I'm currently using (I've got half a lifetime of IT/Programming/DBA experience):

sqlite3:

• great little public domain (free) database, perfect for this sort of thing (helps if you know SQL and it's power)
• simple import of the csv file and then a bit of a tidy up of the data to get it in the correct datatypes etc.

gnuplot:

• one of the best plotting tools available IMO and open source
• It's been a long time since I've used gnuplot so having to re-learn a few things - hoping to try out some 3D graphs down the track if I can find the data to make it useful.

make:

• just a unix build automation tool which 'glues' it all together i.e. import csv file, data tidy up (with a bit of SQL), and then generate all the graphs.

Probably sounds more complicated then it is, but I'm finding it much better than a spreadsheet and it only takes seconds to regenerate all the graphs etc. and it's easy to add to over time.

[leesal]

Great data Myles! Very useful

Some of the data is too thin to draw meaningful conclusions though - eg "tax bills" as a lending reason having one default in 12 loans, or enquiries in the higher categories.

Items I found interesting, I'm only considering your data with a certain level of significance

Renting/Boarding vs Home ownership - suprised to see renting performing so badly and at similar levels of default to boarding

In a relationship vs alone

urban vs rural

business cash flow and household items categories stand out as performing badly.

Good work, keep the analysis rolling in

[RMJH]

Yes, good work. Would be good to use the whole data set! Or even pool our data if that could be done anonymously and securely.