Is FICO log-linear?

[voter01]

So of course the suggestion that FICO scoring utilizes ANOVA got me wondering... And of course I am stalling on the conference paper I am supposed to be writing tonight... And then on page 39 of Google’s results for "fico score formula," I come across this (quote is from http://forums.yellowworld.org/archive/index.php/t-5069.html ):

 

snuffles  02-19-2003, 08:33 PM

okay. here's what i've learned after working for fair, isaac... (they're the guys who came up with the whole concept of a credit score in the '70s and are now milking that nice cash cow...) <Voter’s snip of irrelevant text>

 

- the FICO score is a mix of things - its kinda like a really huge tree of logic that tries to figure out what's the best way to assess a person's credit risk based on what their accounts look like. It's a point system, where there are different things they look at and give you points on. The points are rewarded so that your "score" gives companies an idea of your credit riskiness. The formula is log-linear, such that at a score of X, there a Y good accounts for every 1 bad account, where as X gets higher, Y gets higher. Things they look at are shtuff like: how many lines of credit/loans do you have, what's your utilization, how old are your accounts, have u been delinquent, how many times have you been delinquent, what was your maximum delinquency, how many tradelines have you been delinquent on (yes, delinquency is a pretty big matter, but there are a gazillion other things they look at too)

Heg, when you say with certainty that it’s not ANOVA, does that mean you have some particular info? (Note how I am desperately trying to be subtle here!)

 

Anyway, I’d be interested to hear anyone’s thoughts on the above quote.

V.

[hegemony]

Heg, when you say with certainty that it’s not ANOVA, does that mean you have some particular info?  (Note how I am desperately trying to be subtle here!)

 

<{POST_SNAPBACK}>

I will ask for permission to speak freely and get back to you. Expect a PM.

[hegemony]

logit

[Rezin]

Pshaw,

 

You guys really don't know thier scoring forumula? I learned that when I was 8 years old or so I think. I, of course didn't know what it was supposed to do, but I did completely understand the logic behind it all.

 

 

OK, this IS rather complex, so those who are not good at math, may want to grab a pencil.

 

 

 

 

Ok, the FICO forumula =

 

 

 

 

 

 

 

Grab the Magic 8 ball. Shake it, but not too harshly (must aviod bubbles, as they can cause a split file)

 

Look into the window, and wait....

 

Edited February 8, 2005 by Rezin

[hegemony]

Grab the Magic 8 ball.  Shake it, but not too harshly (must aviod bubblesw, as they can cause a split file)

 

Look into the window, and wait....

 

<{POST_SNAPBACK}>

that is funnier than you know...I have a friend who is "in the industry" and has a magic 8 ball in his/her office !!!! (as a joke...)

[hegemony]

and one of the problems with trying to "reverse engineer" is that many risk scores employ derived variables...not just "raw" data.

scoring_article.pdf

Edited February 8, 2005 by breeze

[hoapres]

Unfortunately, as an outsider having no access to the FICO model, I have by using my own CR attempted to reverse engineer the FICO model,

 

Obviously, having consistent access to only one data point coupled with the fact that I only have FICO scores starting from 2001, limits one's ability to make a definitive guess of the model :

 

Given that caveat :

 

I suspect, can't prove it!, but I believe that the FICO model is based on a weighted least square predictor methodology. Basically, in a nutshell, FICO takes your past credit history with a strong weighting towards the last 24 months to estimate your liklihood of defaulting on future accounts.

 

The public FICO statements clearly state the model uses a number of factors, most important being past credit history and utilization to determine the FICO score.

 

Also, there probably are other factors by placing you in a particular credit group. i.e. individuals with thin credit files are lumped in a different group than thick credit files.

 

I suspect the next-gen or Pinnacle 2.0 model will add additional variables and to handle outlying cases - i.e. individuals with thin files and minimal history, only 1 credit account for a long period of time, etc. Pinnacle 2.0 was supposed to standardize or at least reduce the balkanization of the numerous FICO models.

 

It is also conceivable, that the FICO modelers started off with a large sample size to use least square analysis to come up with a deterministic model. In that case, some sort of logarthmic-linear plotting scheme would be plausible.

 

Again, I have no access to the FICO model and/or access to FICO sources, so as an outsider looking inside, any inside information should be taken as much more believable than my guesses.

[hoapres]

and one of the problems with trying to "reverse engineer" is that many risk scores employ derived variables...not just "raw" data.

<{POST_SNAPBACK}>

 

ABSOLUTELY!!! With that said, my reverse-engineer attempt is based on a classic design of experiments by minimizing the number of variables. I have a thin credit file over a number of years and used that to make an educated guess.

 

Obviouslly, ideally - (short of having the FICO model) - blackbox - (don't know the model just the results ) - would be to take numerous scenarios and varying only one factor at a time. -i.e. start with the same file and only do ONE of the following at any particular point of time - change utilization, open more credit cards, and yikes! deliberately get delinquent entries. None of this is realistic, so I simply took my credit report as old derogatories faded away and worked with a thin credit file.

[hegemony]

The article posted above gives a good, relatively non-technical, introduction to the various risk score models, how they are used and why.

 

especially note what Hand says on p.152 regarding data quality "all too often the data sets from which the models must be constructed are poor: the samples may be distorted...there may be many missing values, values may have been incorrectly recorded..."

Edited February 8, 2005 by hegemony

[hoapres]

The article posted above gives a good, relatively non-technical, introduction to the various risk score models, how they are used and why.

<{POST_SNAPBACK}>

 

The article does give a good quick overview of risk models. However, a basic knowledge of statistics probably is needed to understand the techical details.

 

Often lost in the search for the holy grail of the Fico model, outside of the intellectual pursuit of coming up with the answer!, is that one should concentrate on the factors needed to obtain the credit one desires. Contrary to popular belief, the FICO score is not the sole determining factor in granting credit. Perhaps an extreme example, you could have an 850 Fico score but if you have an Amex chargeoff then most likely you won't get another Amex card until you pay off the chargedoff account.

[hegemony]

The article does give a good quick overview of risk models.  However, a basic knowledge of statistics probably is needed to understand the techical details.

 

<{POST_SNAPBACK}>

well...it is the least technical paper I could find

 

at least I did not select one on the Kolmogorov-Smirnov statistics...

[hegemony]

on the garbage-in garbage-out vein...an interesting read is found in:

 

Avery, R.B., Bostic, R.W., Calem, P.S. and Canner, G.B. (2003) "An overview of consumer data and credit reporting," Federal Reserve Bulletin 89, pp. 47–73.

 

here's a link to the paper:

 

www.federalreserve.gov/pubs/bulletin/2003/0203lead.pdf

 

30% of the people in their RANDOM sample had a collections account!!! That seems very high to me...

 

12% had PRs

 

53% in the sample PIF

Edited February 8, 2005 by hegemony

[hoapres]

The article does give a good quick overview of risk models.  However, a basic knowledge of statistics probably is needed to understand the techical details.

 

<{POST_SNAPBACK}>

well...it is the least technical paper I could find

 

at least I did not select one on the Kolmogorov-Smirnov statistics...

<{POST_SNAPBACK}>

 

Touche! The paper, admittely at a very quick glance!, does appear readable to the layperson. However, unless you find intellectual pleasure in pursuing the Fico model, it is probably best to concentrate on the obvious aspects of improving the credit report and let the FICO score improve as a matter of course instead of putting the cart in front of the horse putting efforts in figuring out FICO at the expense of not concentrating on credit repair.

[hegemony]

bump for newbies

[Kona]

This undoubtedly has been the most fascinating reading I have done this weekend. Infinitely more exciting than the trashy romance novel where the heroine is seduced by the stranger with dark piercing eyes, hynoptic voice, and hands that left no territory unexplored!

 

(I haven't read one of those for a long time, but they are all the same. Just change the characters names, locations, etc.)

 

Time might be better spent on further improving credit without worrying about the mathematical insider info. One could get so caught up in statistics that he or she might forget to make a payment.

Edited March 13, 2005 by Kona

[JerseyBaby]

Often lost in the search for the holy grail of the Fico model, outside of the intellectual pursuit of coming up with the answer!, is that one should concentrate on the factors needed to obtain the credit one desires.  Contrary to popular belief, the FICO score is not the sole determining factor in granting credit. .

<{POST_SNAPBACK}>

Not lost at all. There are those of us here that point that very thing out at every available opportunity.

[Kona]

 

Often lost in the search for the holy grail of the Fico model, outside of the intellectual pursuit of coming up with the answer!, is that one should concentrate on the factors needed to obtain the credit one desires.  Contrary to popular belief, the FICO score is not the sole determining factor in granting credit. .

<{POST_SNAPBACK}>

Not lost at all. There are those of us here that point that very thing out at every available opportunity.

<{POST_SNAPBACK}>

 

Yep! One of my favorite phrases is that the score may open the door but the total credit picture determines if one will be allowed to enter.

[GEORGE]

and one of the problems with trying to "reverse engineer" is that many risk scores employ derived variables...not just "raw" data.

<{POST_SNAPBACK}>

YOU CAN'T "reverse engineer" F.I.C.O. SINCE A + B NEVER = C

 

1 + 2 DOESN'T EVEN = 3 WITH F.I.C.O.

[GEORGE]

on the garbage-in garbage-out vein...an interesting read is found in:

 

Avery, R.B., Bostic, R.W., Calem, P.S. and Canner, G.B. (2003) "An overview of consumer data and credit reporting," Federal Reserve Bulletin 89, pp. 47–73.

 

here's a link to the paper:

 

www.federalreserve.gov/pubs/bulletin/2003/0203lead.pdf

 

30% of the people in their RANDOM sample had a collections account!!! That seems very high to me...

 

12% had PRs

 

53% in the sample PIF

<{POST_SNAPBACK}>

"garbage-in garbage-out"

 

YOU HAVE FOUND THE SECRET TO F.I.C.O.

 

shhhhh KEEP IT QUIET!!!!

[hegemony]

bump for Rob

[hegemony]

cough

[GEORGE]

SNEEZE

COUGH

COUGH

SNEEZE

[richandpoor]

Is it conceivable to think in the not too distant future (10 - 20 years) that other factors may be added to the equation.....

 

1) Lifestyle

2) Genetic heritage

3) Location

 

Certainly, this is the marriage that Insurance corporations are drooling to come up with....