EX 8 decrease of 20 points @ 36 months AoOA!

[BuckyB]

Sure, trying to understand it at a general level is great, but obsessing about every few points is a waste of your precious time, because the algorithm is not completely logical.  It has to hew to political and societal "palatability."

Understanding a TRUE machine-learning algorithm would be far more interesting.  But FICO doesn't use one (at least not publicly) because it would have "discriminatory" effects.  That is to say, even though the pure machine-learning algorithm is race-agnostic, its final score determinations mimic race-based classifications.  Which is quite interesting but completely politically untenable for Fair Isaac to put out into the public.  Thus it has to use a "watered down" version which uses scorecards for more "transparency."

I actually wrote an extended post on this recently but the thread was removed by moderators.

[Thomas_Thumb]

Garbage in garbage out. CRA data files lack the granularity to properly analyze some factors. One example being recent account activity and another being revolver/transactor behavior. Inability to consider R/T behavior affects predictive accuracy.

Of course Fico is a business and a thriving one at that. Stock price has tripled in the last 2 years. Fico's primary customers (lenders) influence what factors are used in scoring algorithms. Regulators are involved as well. Nothing new there.

What I find far more interesting (or concerning) is how companies such as Cambridge Analytics gain access to personal data. Then use the data at the direction of their cohorts to influence outcomes. In the case of CA they gained access to privileged data which was exploited to influence voter decisions. (See the Great Hack)