2011.Q2
In this question, what is the purpose of calculating the SSE of the holdout sample vs 1.00 i.e. Sum[(1-.6)^2+(1-.8)^2).... I don't get what that tells us - just that having real relativities for these quintile groups (rather than 1.00) makes sense to do? However it is a different comparison.
In my solution I just calculated the SSE for Raw - HO and Cred - Raw. This treats the HO sample relativities as the "true values" and thus allows us to compare which method got closer to that.
I feel like its things like this where I lose a few partial points - what am I missing here?
Comments
When it comes to making predictions in the Couret & Venter reading we have three options.
1.) Use the overall hazard group relativity.
2.) Use the raw quintile relativity from the training data set.
3.) Apply the multi-dimensional credibility procedure.
We use the sum of squared errors to compare which is best. To have a valid comparison - i.e. one which isn't distorted by scale, we must normalize the predictions so that the hazard group relativity is 1.000 overall.
So if we're using the overall hazard group to predict the relativity, it will always be 1.000. The purpose measuring the SSE here is to see if it's better to predict using the experience of the entire hazard group or whether it's better to use the raw class experience in each quintile, or if we should use the multi-dimensional credibility based prediction.