I have a question which needs help from a Traverse algorithm expert, can u help me ? ]]>

What are you referring to by “latent factor”?

]]>Interesting… I am particularly surprised by the fact the AUC increases along with the dispersion, thus with the heterogeneity in your portfolio!

So, in the limit case where the portfolio would be perfectly homogeneous, with the exact same probability of claiming a loss for each policyholder, you could only obtain a very poor classifier (in term of AUC), even with a perfect model? That’s almost counterintuitive…

I would also be keen to understand the effect of a real-life model on the AUC. I expect it would be more challenging to capture the heterogeneity of a portfolio through a limited set of covariates, when that heterogeneity is high (= high unexplained variability and low R-squared, assuming some linear model). So, in the opposite case of a rather homogeneous portfolio and the same set of covariates, would you be more likely to get a better classifier (higher AUC) since the R-squared would be higher? This would somehow counterbalance the effect of portfolio heterogeneity on the AUC when assuming a perfect model.

By the way, how would you define heterogeneity? Is it to be understood the same way as variability? Reading your post, I feel heterogeneity is related to the variance of the covariates x_i, whereas dispersion would be related to the variance of the probabilities generated by the model. I may be wrong though, but I feel these two notions are not completely equivalent.

Thanks.

]]>Overall this seems intuitive: we’d expect that adding a useless predictor should make the model crappier. Or at least that’s the way I see it.

It would be interesting to know if this were always (on average) true for all models and all coefficients.

]]>Can you share details about the Computational Science seminar? Which city & which dates?

Thanks

]]>Many thanks!

]]>great book!

Can I access to the data sets in survival chapter? Is the dataset ‘data\\DataMortality.csv’ (10.2.1 part of the book). I can’t find it also in the ‘survival’ package.

Thanks!

Best ]]>

ic=maxf$objective-0.5*qchisq(.95,1)

with this revision:

Lower: #[1] 0.2203471

Upper: #[1] 0.6696117

Means, half of Chi-Sq value will produce similar results. ]]>

I totally agree with the conclusion on necessity of pooling risks together.

I think it also depends on what you are trying to achieve with the analysis.

If we are talking about pricing, then we just want our result to be predictive of the future. That means causation is not what we are looking for, since we just need to find some correlations we expect to be stable with time. So there is even less need to go for hyper-individualization, because it’s easier to estimate the correlations on high volumes.

Now, if we are talking about prevention, then causation is definitely what we are looking for.

We nowadays see a lot of programs in motor insurance, often based on telematics, aiming at reducing insurance risk by changing client behaviour.

But that suppose to be able to identify what generates claims. And to achieve that, the natural inclination is to go into hyper-individualization, thinking that we are closer to causation by doing this. But, as you mentioned, they are a lot of bias and risks to come to incorrect conclusions.

That’s why I am generally sceptical about such programs.

And I didn’t even mentioned the very low frequency/high cost claims where pooling is the only answer…

]]>Any chance you share the code for the reproduction or the Lasso/Ridge RSS lines plot?

]]>