Programming the statistical procedures from SAS

score mappings

Reply
Super Contributor
Posts: 395

score mappings

Hi, has anyone mapped a score to another score ?  If so, using which porcedure ? and more importantly is it a "simple" enough task?

Explanation:  I have an internal score that we use and we want to map it to the FICO score.  Using odds, score distributions, etc ?

Thanks

Trusted Advisor
Posts: 1,195

Re: score mappings

Hi,

Simplest way is to check correlation between two score variables in addition to check distributions of both to make sure they are identical with regard to distributions. If score variables are following same distribution and highly correlated then it may be concluded that they are mapped.

Regards,

Grand Advisor
Posts: 16,901

Re: score mappings

1. Yes

2. Data Step/Proc SQL/PROC FREQ/PROC MEANS for rules to generate mappings.

3. Yes and No, depends on background

New Contributor
Posts: 3

Re: score mappings

As stat@sas said, correlation. Let me add just a bit more....

1. Windowed Correlation. A simple correlation coefficient will express the overall average closeness of the mapping across the entire range of values. Often we find the fit is good over some ranges and less good - even poor - over others. My friend and colleague Steve Raimi likes to call this the "flaw of averages". It's especially important for something like a FICO score than has a different meaning over different ranges. So, you might want to break down the FICO score into ranges, such as subprime, prime and excellent; and calculate the correlation separately for each region of the financial risk spectrum. I expect you will find your score works better in some areas than others. One model might be especially good at very low risk prospects while another is good at distinguishing between subprime candidates.

2. Predictive Ability. In the end, your score are going to be used to make a prediction. In addition to finding the correlation coefficient over different ranges, you will want to compare the accuracy of your predictions compared to those using the FICO score. Does you model predict the same outcomes? What percentage of a population get the same recommendation (accept of reject) as FICO? Does your score get the right answer - accept for loans that pay off and reject for loans that untimately default - than does the FICO score?

FICO is a very widely used standard for many reasons but it must be understood that it is generic. A model designed for a very specific population or developed using additional information such as past purchase and payment history to a particular company should be able to do better than the generic FICO score. So, while your score should resemble FICO, you will not want it map exactly because then there is no need for your score. Address the differences in the mapping between your score and FICO to demonstrate the areas where your score is better.

Ask a Question
Discussion stats
  • 3 replies
  • 298 views
  • 0 likes
  • 4 in conversation