An exciting time to be modeler. As I mentioned there are a number of parallels to the challenges of modeling this virus and the modeling of risk. Obviously the "good/bad" definition, but also the indefinite definition. In risk the indefinite definition is for those people who are not good (rarely late on a loan) nor bad (defaulted on loans). The idea in having this indefinite roup is to make the distinction between good and bad more explicit.
For the virus indefinite might be those that got sick but never sick enough for ICU.
Another aspect of risk modeling is "Reject Inference," the concept is to try to determine how those people that were never given a loan would have performed had they been accepted. In other words, infering the performance of those that were originally rejected. The reason for doing this is the original sample of known performers may be biased and some of those original rejects might have been good customers due to flaws in the original accept/reject strategy.
For the virus this would apply to groups of people whose illness status was unknown (not tested) and were never given any treatment.
I know there have been amazing advances in modeling tools since I retired; however, one of the flaws of the more recent ones was lack of ability to identify "reason codes" which in credit risk modeling had been required. The reason codes identified specific factors resulting in some negative action by the lender (rejection, reduction in credit limit, for example).
It seems to me that reason codes would be very important for virus modeling, especially when modeling what treatments are effective in recovery or what strategies reduce incidence in geographic groups (cities, counties, states, countries).
Again, I am retired and unfamiliar with the current state of research, but it seems like every company the has sophisticated tools and access to data should be working on this for the good of their states, countries, and the world.
AN EXCITING TIME TO BE A MODELER AND STATS GUY!!!
------------------------------
Michael Mout
MIKS
------------------------------
Original Message:
Sent: 04-21-2020 15:28
From: Michael Mout
Subject: Good and Bad
When I worked in credit and insurance risk (FICO, Capital One, and State Farm) we built models based on various definitions of good and bad risk and used logistic regression and other tools, random forests to find ways to separate the good and bad groups analytically.
Seems to me that this is a natural approach to finding ways to define what techniques work best for treating people with C19 or ways to minimize the risk of infection for cities, towns, states, countries, etc.
I'm retired and don't have access to the tools or data or new techniques, but it sure seems like this is something that someone with the tools and skills should be pursuing.
------------------------------
Michael Mout
MIKS
------------------------------