@RobertEG wrote:Fair Isaac has clearly and publicly stated that it develops and licenses different algorithms to each CRA, customized to the requests of that customer.
They have stated that, with identical data,there can be as much as a 20-30 point variation between the resulting scores produced by each CRA,
The CRAs dont alter the algorithms. They are produced by Fair Isaac and licensed to each CRA.
AS a result of their trade-secret nature, the CRAs dont have access to the algorithms themselves, and the CRAs are in direct ompetition with Fair Isaac by their use and sale of their own Vantage scoring algorithm.
The CRAs perform the actual scoing at their site under license, running the licensed algorithms against their own data.
FICO does not possess the data, and thus does not run the scoring in-house.
^This
Also, to REG's point FICO blogs (at fico.com) had an article a few years ago discussing why the algorithms are tuned for each CRA. Basically, there is a history of differing geographical coverage as well as difference in creditor reporting when they don't report to all three. Because of these, relatively minor differences, FICO scores are "tuned" for the data each CRA has. The result is claimed, and I believe correctly, to be more predictive even though if exactly the same data was in each report the FICO scores of a specific generation would still slightly vary.
One of VantageScore's earliest claims is that it was developed with a common dataset amongst the CRAs and used the combined large sample for all three to develop VantageScore. VS presumably lost some predictiveness by tossing out those areas where reporting differences occur but possibly gained from the large combined samples.
I'd like to see some actual published statistical comparitive data from a third party, possibly a university with a good stats dept. (Princeton?). It would be nice if Hell froze over too.
@flan wrote:
@cashnocredit wrote:
@RobertEG wrote:Fair Isaac has clearly and publicly stated that it develops and licenses different algorithms to each CRA, customized to the requests of that customer.
They have stated that, with identical data,there can be as much as a 20-30 point variation between the resulting scores produced by each CRA,
The CRAs dont alter the algorithms. They are produced by Fair Isaac and licensed to each CRA.
AS a result of their trade-secret nature, the CRAs dont have access to the algorithms themselves, and the CRAs are in direct ompetition with Fair Isaac by their use and sale of their own Vantage scoring algorithm.
The CRAs perform the actual scoing at their site under license, running the licensed algorithms against their own data.
FICO does not possess the data, and thus does not run the scoring in-house.
^This
Also, to REG's point FICO blogs (at fico.com) had an article a few years ago discussing why the algorithms are tuned for each CRA. Basically, there is a history of differing geographical coverage as well as difference in creditor reporting when they don't report to all three. Because of these, relatively minor differences, FICO scores are "tuned" for the data each CRA has. The result is claimed, and I believe correctly, to be more predictive even though if exactly the same data was in each report the FICO scores of a specific generation would still slightly vary.
One of VantageScore's earliest claims is that it was developed with a common dataset amongst the CRAs and used the combined large sample for all three to develop VantageScore. VS presumably lost some predictiveness by tossing out those areas where reporting differences occur but possibly gained from the large combined samples.
I'd like to see some actual published statistical comparitive data from a third party, possibly a university with a good stats dept. (Princeton?). It would be nice if Hell froze over too.
There are many fewer creditors that don't report to all bureaus now. (There're still big differences in how they get and report public record data.) Historically, data was sent to the bureau on paper and then magnetic tape, so it cost non-significant amounts to report, and the value of reporting to all the bureaus, and not just the local one, was less. Reporting is now done electronically, and the cost of doing that is low, once the systems are in place. Creditors don't like that scores vary with the same data, as it makes it hard to sometimes use one and sometimes the others, or to do mid scoring, or averaged scoring. The more recent FICO scores are supposed to vary less between bureaus.
VantageScore 3 also takes advantage of the increased data that's reported now, compared to what was standard even 10 years ago. Actual payment data, mid cycle reports, better classification of account types, etc. Also, there is much better standardization of the data reported to each bureau. I no longer deal with software that generates reports, but I used to. The first place I did it, a long time ago, sent vastly different data to everywhere, because they wanted different things, and wanted it encoded their way. The current formats used are shared by all three bureaus, which means that the definitions of things like what a 30 day late is, what the responsible party for an account is, and the type of account are all the same. (I distinctly remember adding program logic to report the same payment data as different number of days late; someone wanted it calculated from the invoice data, everyone else wanted from the due date, but I have no idea who was in which camp.)
FICO 9 also takes advantage of the fine grain (payments in addition to balances) but VantageScore was the first to do so. FICO 9 also drops use of paid collections but VS3 led doing this as well. FICO's blog had a post last year pointing out how significant a predictor the payment data was. In a sample people that made 3x the minimum payment had half the risk of those making minimum payments.
I agree that the data is far more consistent between CRAs. Not totally* but very close. IMO, FICO fell back on the "tayloring" argument because VantageScore was getting a lot of play out using the same algoritm with all CRAs. Unlike VS3 which was built by a consortium of the 3 CRAs, FICO doesn't have the luxury of using data outside of the CRA they license.
More to the point, I suspect VS3 is quite good. They use an absolutely huge, collective database and they have publicly published more statistical data than FICO without an NDA.
As for differences between CRAs FICO algorithms I personally have seen one rather odd anomoly. EQ's FICO 8 includes charge cards in utilization based on using the high balance as a proxy for CL. The other CRAs do no. Also, FICO 4 from all 3 CRAs excluded charge card balances. I was rather surprised to see this to the point where I ran a specific test and had a 40k balance report on my Amex but only about 1k on another card. EQ FICO score dropped to around 760.
As for differences between CRAs, they remain to a degree. VS3 actually publishes data showing statistical differences in scores for the 3 difference CRAs on the same consumer. IIRC, the median difference was around 20 points so there are still variations. Whether these are significant enough that FICO can eke significant predictibility from customizing each CRA is an interesting question; It's probably a factor but a diminishing one.
@MarineVietVet wrote:
@cashnocredit wrote:FICO 9 also takes advantage of the fine grain (payments in addition to balances) but VantageScore was the first to do so. FICO 9 also drops use of paid collections but VS3 led doing this as well. FICO's blog had a post last year pointing out how significant a predictor the payment data was. In a sample people that made 3x the minimum payment had half the risk of those making minimum payments.
I agree that the data is far more consistent between CRAs. Not totally* but very close. IMO, FICO fell back on the "tayloring" argument because VantageScore was getting a lot of play out using the same algoritm with all CRAs. Unlike VS3 which was built by a consortium of the 3 CRAs, FICO doesn't have the luxury of using data outside of the CRA they license.
More to the point, I suspect VS3 is quite good. They use an absolutely huge, collective database and they have publicly published more statistical data than FICO without an NDA.
As for differences between CRAs FICO algorithms I personally have seen one rather odd anomoly. EQ's FICO 8 includes charge cards in utilization based on using the high balance as a proxy for CL. The other CRAs do no. Also, FICO 4 from all 3 CRAs excluded charge card balances. I was rather surprised to see this to the point where I ran a specific test and had a 40k balance report on my Amex but only about 1k on another card. EQ FICO score dropped to around 760.
As for differences between CRAs, they remain to a degree. VS3 actually publishes data showing statistical differences in scores for the 3 difference CRAs on the same consumer. IIRC, the median difference was around 20 points so there are still variations. Whether these are significant enough that FICO can eke significant predictibility from customizing each CRA is an interesting question; It's probably a factor but a diminishing one.
I asked a former administrator of this site about this and here was the reply:
This is news to me. If so, it would mean that at EQ charge cards are either reported as revolving or they're reported as "open" type credit, with this type now being included in utilization. If the latter, I doubt FICO would only design it to apply to EQ, so it would also be in place at the other two bureaus. I would definitely like to see the source of this comment.
So you're sure about your comment? Just curious.
I'm sure of it. I personally experienced it. I've tracked my EQ FICO score for years and Amex never impacted my score. I put way more spend and more month to month variation on Amex than my other cards and was shocked when an Amex triggered alert also resulted in FICO score changes after the conversion to FICO 8. As a consequence I ran a test by charging close to 40k while dropping my other card balances. EX and TU FICO 8 went up from the other card balance decreases while the EQ FICO 8 dropped like a rock more than recovering the next month when the balance reported dropped to normal levels. Interestingly, after I made a particularly large purchase that bumped my high limit to 60k Amex no longer changes my FICO scores. Looks like there is a CL/High Balance limit that takes the CC out of the calculation. This has been reported for other CCs as well.
I also suspected some sort of reporting difference might be at play and checked the reporting detail from a direct CRA pull with one of the mods here with an Amex charge. It matched.
I've had confirmation "elsewhere" from another person who saw this as well.
Here's the thread I posted Amex changes on.