What a Credit Score Calculation could look like

nyermen

Following a thread about credit scores, I noted that I've been involved with some scoring models. So on request, I'm providing one as an example to give an idea of what a calculation can look like.

Disclaimer: this is a score specific to the credit institution, as well as the product and the sub product. I've trimmed it down to the parameters, as the full model has lots of maths. This is a new application, so almost entirely based on external history. Every application starts with a base number, which is then adjusted according to parameters.

Home Loan - Application scorecard Parameters

• Intercept (This is the initial number for everyone, that subsequent parameters adjust - 600 when the model was created)
• Type of Loan (Some types reduce score, otherwise no deduction)
• Geographical Area (named postcodes reduce score, otherwise no deduction)
• Minimum account age (age of the most recent account opened - older gives a larger deduction)
• Total Savings with the institution (over 5000 with them will increase score)
• Worst Areas in the Last 3 months (Deduction based on how many missed payments in last 3 months)
• Worst Arrears in the Last 12 months (Again, adjustment based on how many missed in the last 12 months. Smaller than 3 month adjustments)
• Worst Arrears last 3 years (Again deduction in line with above)
• LTV (Loan Amount divided by property value - More than 50% LTV has a varying deduction)
• PAYE or Self Employed (Increase or reduce score)
• Income Grading (Increase or reduce score)
• Permanent Employment Grading (Increase or reduce score)
• Employment Sector Code (Increase or reduce score)
• First Time Buyer Flag (Increase or reduce score)
• Income of First Applicant (Increase or reduce score)
• Income of Second Applicant (Increase score)
• Property Grading (Increase or reduce score)
• State of Construction (For houses being built or needing repair to be habitable)
• Term of Loan
• Loan to Income Ratio

So some observations - for this specific institution of course...
- It's a Building Society. I can't say the name, sorry, the models aren't confidential but I can't reveal who they are, and also I've also not given all the deduction numbers.
- The final number falls into one of 20 score "bands", ranging from the highest possible score to the lowest. The first 3 bands cover 85% of customers, the remaining 17 bands are the last 15%. Those first 3 bands cover the scores from 780 down to 680. (The bottom potential score is zero). This doesn't mean most applicants fail, rather that successful applicants normally fall in the top 15% of the potential scores
- The adjustments to the base score depend on the specific parameter, but for example, arrears in the last 3 months reduces the score by 300. Since band 3 is 680 and higher, with no other adjustments, you're now at 300, so you're unlikely to be accepted unless every other parameter is very positive. 12 months is a little better (reduces score by 200). Other parameters are much smaller (add or deduct between 0 and 50 in most cases).
- First 3 bands are normally an acceptance (all acceptances are reviewed by a human first)
- Otherwise the next few bands are referral, and the rest are a decline.
- That score plugs into a formula that then estimates the "Probability of default", aka the likelihood that the customer will default at some point. Default is defined as having any payment 90 days (or more) past due. Note 3 months is not 90 days - ie. you could be in default before missing the third payment. Default is their internal classification, it doesn't necessarily mean a marker is placed on your file at that point.
- They don't take into account credit history that is more than 3 years (They may well have changed this now, they didn't at the time)
- A different scorecard exists for a customer who has had credit with the institution before, in that case, additional input from a "behavioural scorecard" is applied.

Hope this is of some interest. Last comment - I'm erratic on when I can visit (Normally at least a couple of times a week unless I'm travelling on business), so if there are questions, I'll answer as soon as I can

Peter

(8 months until debt free)

fruitisbad

Interesting that they only consider info going back 3 years when everything stays on file for 6.

Thanks for posting

nyermen

For that specific customer, indeed they didn't. I should add I didn't "design" the model (those people get to retire at 50 on a private island), I just implemented it, hence I'm also picking holes in it I also note that they don't (or didn't) look at outstanding balances on the credit report either.

Perhaps because this is a "secured" model - so they were as interested in chance of default (Mortgages are often the last payment to be missed). There is another model (which isn't part of scoring, but is part of capital requirements) called "LGD" (loss given default). Basically, if the credit score relates to the "chance you might default", the "LGD" defines what could actually be lost in the event of a default. The LGD is focused on things like LTV, house price changes in the area, and "recovery costs" (repossession mainly).
I believe in the event of a referral, they would then seriously consider the LGD to determine the expected loss (since expected loss can be simply calculated as the Exposure x PD x LGD. There's also "unexpected loss", but that's another, much more enthusiastic formula)...

I'll see if I can find their model for "Buy to let" and note how it differs, won't be today though, as the one above took an hours searching of emails to find.

uncreative

Interesting that geographical area has an affect on the score. Is that linked to poor performing property prices potentially affecting the security address or more just a "postcode lottery" for the borrower??

nyermen

If I recall, it was to do with nominating certain areas of concern, specifically industry related (eg. Town May also have been to do with areas where house prices are more volatile (though that should matter more in the "potential loss" calculation than the "likelihood of default). I seem to recall they "upgraded" certain postcodes based on situation - eg. Guildford, where the University means house prices are more stable even during a downturn.

Of course, by postcode, we mean the first 2 / 3 letters.

David_White

Very interesting, thanks for posting.

I'd be interested to know what postcodes caused a reduction in score!

nyermen

Apologies, I think that is pushing the NDA... In reality, the initial lists we configured will have been extended. It was quite high level (eg. N1 in London, GU1 in Guildford, etc).

One thing worth adding - they also had a lending limit per postcode (regardless of postcode). If they had more than X mortgages in an area, it was an automatic rejection regardless of score. I think this was a liquidity thing - as a Building Society, they wanted a spread of mortgages.