sharetrader
Page 391 of 465 FirstFirst ... 291341381387388389390391392393394395401441 ... LastLast
Results 3,901 to 3,910 of 4649

Thread: Harmoney

  1. #3901
    yeah, nah
    Join Date
    Mar 2017
    Posts
    491

    Default

    Chart of Expected Defaults by grade (because I hate the Hazard Curve and often find annualised data misleading):

    This is calculated from historical data (unique.csv) with points on the line representing the percent of loans that are expected to default sometime in the future for a given loan age. (this is not annualised!) I've applied a window (description in title) to remove some outlier detail that plays havoc with the graph.

    To give an example of how to read the details:

    C Grade loans that are 10 months old have a ~3% chance of defaulting sometime in the future.

    Attachment 10084

    The first point on the line is the expected default rate for all loans in that grade (based on historical data) - NOT annualised.

    Some interesting crossovers and some rising and some falling grades...

    Fix: Had a date constraint wrong which influenced the right hand side of the chart. Note that the right hand side is still incomplete as number of loans still to low for meaningful detail at that age.
    Last edited by myles; 17-10-2018 at 07:17 PM. Reason: Fixed

  2. #3902
    yeah, nah
    Join Date
    Mar 2017
    Posts
    491

    Default

    Quote Originally Posted by beacon View Post
    A bivariate analysis is much more useful than segregated filtration in a multivariate environment. At the very least, it proves again that risk falls with age, until you hit 60.
    If you liked that then you might like these even more - same but with grade as well pdf runs to 67 pages though...

    l2_default.csv

    l2_default.pdf

    The csv is particularly easy to use in a spreadsheet with simple column filters - lets you 'dig' into the data without having to get bogged down with pivot tables initially.

  3. #3903
    Member
    Join Date
    Aug 2017
    Posts
    212

    Default

    Quote Originally Posted by beacon View Post
    Brilliant work leesal. Thanks for proving again, that loan picking ISN'T simply happening for love, and so, ISN'T a sheer waste of time.

    I'm assuming 1. is your default minimiser. That is an awful lot of filters, overridden by top level elimination/modification of filtered loans. Be interesting to know, what criteria you chose for your second default set.
    Thanks. Myles data stimulated a lot of thought and good ideas... and as a loan picker, its useful to have an overarching frame work - and one backed up with historic data all the better!

    I've progressed beyond where I was And now have a streamlined single "default minimiser" that can ideally put into an autofilter. The criteria I have is B,C,D,E grades; age band 20 -60 year; debt to income < 10%; time at residence 2+ years; all purposes except new vehicle, new boat, wedding expenses, tax, other. With those alone can boost return and half default from the average 7.0% to 3.5%. Then there are items which cannot autofilter on that have a big impact - enquiries 3 or fewer, and debt to income < 5% for E grade loans - which bring the default down to 2.2%. I have gone further then that without sacrifricing return, but it can get very complicated.

    Will start off running the default minimiser, and will be handpicking others I like from gut feel; and track how they perform.

    Quote Originally Posted by beacon View Post
    There is bound to be selection bias, since only 17 or so investors contributed to the data pool, versus the 8,800 odd investors Harmoney has on its books today. Hence the disclaimers Myles has put in the report. Still, we got input from loan pickers and index buyers to some extent, and I'm not holding my breath that Harmoney will (ever) publish default info by individual variables for its whole dataset. So, we've got the next best thing.

    What I find interesting also, is that Myles' efforts at cleaning and reporting data put Harmoney's output-to-date to shame. Four years, and they still have the type of data errors humvee recently reported. They also lack in investor education and communication, especially in the area of protect loans. But overwriting/hiding rewritten status/info of defaulting loans, I find absolutely appalling and unforgivably misleading, bordering on fraud. I hope they desist and rectify the wrong they have done - as it impacts on investment decision making.
    I wonder whether its possible that rewritten loan data can be fixed going forward. I recall a thread some time ago, someone mentioned that employment status changed part way through a loan, which means that the data we are viewing is taken from a live relational database - eg any of the "client data" links such as employment etc is subject to change at any point in time.

  4. #3904
    yeah, nah
    Join Date
    Mar 2017
    Posts
    491

    Default

    Hazard Curve for loans in unique.csv:

    Notes:

    • The x-axis is the month where the default occurred - the Harmoney graph appears to work on when they declare a loan has defaulted. (Which is more correct? I prefer to know when it actually occurred and payments stopped?)
    • Individual grade curves can be quite different to the overall average, but in very general terms, the peaks are similar (e.g. peaks at the 6 monthly points tend to be obvious across most grades)
    • The peak at 6 months would not be reported as a default until month 10 - 12 (or later) but the actual default occurred at 6 months.


    Attachment 10087
    Last edited by myles; 18-10-2018 at 09:06 AM.

  5. #3905
    Senior Member
    Join Date
    Apr 2002
    Location
    , , New Zealand.
    Posts
    726

    Default

    Quote Originally Posted by myles View Post
    Chart of Expected Defaults by grade ...

    Some interesting crossovers and some rising and some falling grades...
    Contrarian late life behaviors by C and D. Why I wonder? Maybe because they were the good E and Fs which rewrote and metamorphosed into C and D. This graph is worthy of a longitudinal analysis...

  6. #3906
    Senior Member
    Join Date
    Apr 2002
    Location
    , , New Zealand.
    Posts
    726

    Default

    Quote Originally Posted by myles View Post
    If you liked that then you might like these even more - same but with grade as well pdf runs to 67 pages though...

    l2_default.csv

    l2_default.pdf

    The csv is particularly easy to use in a spreadsheet with simple column filters - lets you 'dig' into the data without having to get bogged down with pivot tables initially.
    Just had a quick look Myles. This is marvellous. Looks like I may be exercising my grey matter much more this weekend. Hope I get the time

  7. #3907
    Senior Member
    Join Date
    Apr 2002
    Location
    , , New Zealand.
    Posts
    726

    Default

    Quote Originally Posted by leesal View Post
    Will start off running the default minimiser, and will be handpicking others I like from gut feel; and track how they perform.
    You won't be able to attribute RAR improvement fully to your minimser then, but RAR improvement will be welcome nonetheless. How did you discern a loan was unlikely to early repay though?

    Quote Originally Posted by leesal View Post
    I wonder whether its possible that rewritten loan data can be fixed going forward.
    Yes, it is possible to debug going forward as well as rectify retrospectively.

    Quote Originally Posted by leesal View Post
    I recall a thread some time ago, someone mentioned that employment status changed part way through a loan, which means that the data we are viewing is taken from a live relational database - eg any of the "client data" links such as employment etc is subject to change at any point in time.
    Maybe, but changing whether a loan was initially rewritten after a loan has defaulted? Some may call it covering the tracks, and it is unnecessary to do really. I remember Myles producing a graph for his own data for rewrites a few posts earlier (after rechecking his 50 defaults to date). It showed his rewritten loans were a slightly better risk than the new loans. Most likely, this will reflect in the composite set too, but overwriting it makes the aftermath impossible to analyse/troubleshoot

  8. #3908
    Senior Member
    Join Date
    Apr 2002
    Location
    , , New Zealand.
    Posts
    726

    Default

    Quote Originally Posted by myles View Post
    Hazard Curve for loans in unique.csv:

    Notes:

    • The x-axis is the month where the default occurred - the Harmoney graph appears to work on when they declare a loan has defaulted. (Which is more correct? I prefer to know when it actually occurred and payments stopped?)
    • Individual grade curves can be quite different to the overall average, but in very general terms, the peaks are similar (e.g. peaks at the 6 monthly points tend to be obvious across most grades)
    • The peak at 6 months would not be reported as a default until month 10 - 12 (or later) but the actual default occurred at 6 months.


    Many thanks for this, Mylesy. So, if it survives the first 6 months, chances of its survival improve with age. In this light also, it is interesting to see the C and D behaving as contrarians in later life. By the way, like you, I'd prefer to know when a default actually occurred and payments stopped, if it doesn't inconvenience Harmoney to change the date they use to plot defaults.

  9. #3909
    yeah, nah
    Join Date
    Mar 2017
    Posts
    491

    Default

    Quote Originally Posted by beacon View Post
    So, if it survives the first 6 months, chances of its survival improve with age.
    No (I think), exactly why I produced the Expected Defaults graph. The Hazard Curve (i.e. Harmoney Hazard Curve), is plotting defaults in a month vs total loans (all loans from all time). It's not wrong - it is a Hazard Curve as per:

    Failure rate is the frequency with which an engineered system or component fails, expressed in failures per unit of time.

    But it is not the survival rate:

    The survival function is a function that gives the probability that a patient, device, or other object of interest will survive beyond any given specified time.

    Perhaps I'm wrong with the way statistical survival works, but shouldn't it be the chance of survival of those that are still alive? i.e. if we have 1000 loans and 10 default in month 10 (10/1000), the Hazard Curve plots month = 10 vs 1%. But aren't we more interested that, if 600 of the loans have already defaulted or been paid off by month 10, the 'none' survival (or chance of defaulting) is (number of defaults that have yet to occur)/400?

    This is what I tried to plot with the Expected Default chart, which shows, in lower risk grades, a fairly flat default rate across time, in D grades a rise before a fall and for E and F a fall.

    I've clearly misunderstanding what the Hazard Curve was showing, up until now, and suspect I may not be the only one?

    Clearly 'statistics', and it's terminology aren't my strong point , happy to take suggestions on the best way this should be represented. I'm just not sure it is the Hazard Curve?
    Last edited by myles; 18-10-2018 at 05:11 PM. Reason: Changed numbers for more clarity

  10. #3910
    Member
    Join Date
    Aug 2017
    Posts
    212

    Default

    Exactly right. Going to bump my rather verbose post from a month back....

    Actual annual survival rates in some grades are 4 times greater then HM quoted default rates. Lower grades are still worth it, you just have to factor in early repayment.

    EG in grade E3 - after early repayments of 50% pa, your probably only going to rake in 30-35% interest over the term of the loan. After defaults of 4.5% x 5 = 22.5% you generate a return of approx 10%. In complete contrast HM data which suggests 27.99 - 4.5% = 22.5%!!!!



    Quote Originally Posted by leesal View Post

    HM annual average default is misleading. Rather I prefer to look at cohort default across the full term. Taking HM forecasted stats for Grade "E", their default forecasts are approx 4.5% per annum, or 22.5% across a 5 year term. To validate this, taking the 2014 E grade performance off the "historical annual default rate tool " https://www.harmoney.co.nz/investors/default-rates - shows that the cumulative default of E grade at 22.7% (and running to a similar place on 2015 and 2016 cohorts). Critically the definition of cumulative default is based on the number of loans originally funded, not the loans outstanding.

    How is the cumulative default rate calculated?


    The cumulative default rate is calculated by dividing the total number of defaults by the total number of loans funded. For example in 2015, for grade C3, 447 loans were funded and 17 loans defaulted to the end of 2017 creating a cumulative default rate of 3.8%.

    How that reads to me, is early repayment is not factored in. If HM were to publish annual default based on time in lent, the number would be significantly different. ie If 22% of your E grade loans are going to default, and 78% remain good - how are your stats going to look if 40% of the good ones repay early in the first 12 months!

Tags for this Thread

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •