Dear all,
I am writing to ask your opinion relating to a very
hot topic in the financial sector: what is the minimum
number of events (bads) required to build a robust
logistic regression scorecard?
This question is very important in the context of the
Basel II accord where many financial firms have not
got high volumes of defaults.
The rules that the industry seems to follow is that
you need ~1500 bads(event=1) if you need to build a
full logistic model, while you can build a weight of
evidence model with as little as 150 bads.
-Do you know the statistical origin of these figures?
I believe the 1500 figure has purely a historic
origin, going back to when American statisticians had
to collect the data manually.
-Do you agree with them? I find quite hard to believe
them, especially given the fact that medical
statisticians often run perfectly fine models with
very low volumes of cases.
I would be very grateful if you could share with me
your opinions or/and direct me to the appropriate
references.
Regards
Piero Bassu
__________________________________
Do you Yahoo!?
The all-new My Yahoo! - Get yours free!
http://my.yahoo.com
|