This competition is quite interesting but there are some aspects that I don’t see here. I would like to know your opinions
Legal aspects
According to GDPR regulation in Europe I don’t think we are allowed to use gender features. Very surprised that the bonus round, the “age” but not “sex” features were removed.
Ethical aspects
Some solutions added some random extra cost to the price. Hmm, could we put that in production? Well I double that. Firstly, not sure if this strategy passes the fairness regulations. Secondly, if we put the pricing models in production, people would notice that very soon. The customers would visit the insurance company many times to get the best offers.
Obligations on pricing
I don’t think it is fine to put risky customers in a black list by selling super high price. There must be a celling limit.
Few more aspects that nobody mentioned such as price per travel KM.
And do we miss other important aspects of the game?
Explainability - your customers won’t be very happy if you put their price up from last year just because a black box algorithm said the risk went up.
Expenses - An insurer wouldn’t survive long with a 0.01% market share
Return on capital and investment income - regulators require insurers to hold capial and investors require a return on capital that reflects the risks
Then there are all the classical actuarial issues of trying to estimate future claims costs. Claims can take years to settle, and there can be changes in the external environment which change the future costs or the liklihood of claims. (Eg inflation, changes in the law, changes in road safety, pandemics meaning fewer people are driving to work etc…)
Insurers will generally have underwriting criteria where they can decline to provide quotes to customers outside their risk appetite, which is effectively the same as giving a high quote to get rid of a customer. Regulations vary bewteen countries about how this can done - with regulations in some countries stricter than others. Generally if it is not a form of discrimination it will be allowed. An insurer declining to quote commercial vehicles or certain high risk vehicle makes and models would be ok, but an insurer refusing to quote people of a particular ethnic background would not be ok.
I’ve heard rumors of European insurers charging different price depending on the day you were born (Monday, Thursday…) to allow them to calculate price elasticities. Would love to have it confirmed or denied though.
I’m not going to dig into the things that the competition may be missing in general (since there are many!) but some are rightly highlighted by @Calico. That’s a whole other, and very interesting, conversation. (Another one was cumulative profit leaderboards that was suggested by @simon_coulombe very early on)
On the legal and ethical aspects one thing that hasn’t been mentioned is that there was no regulation for model explainability (ignoring customer satisfaction). Otherwise we’d all have GLM after GLM most likely.
Some countries regulate pricing very heavily, with regulator approval required for prices and limitations placed on the rating factors that can be used (Australian Compulsory Third Party Motor and Isreali Motor Bodily Injury insurance for example), so they are limited in way they can set their rates and the explanations for changes will follow the prescribed rating structure. In Ireland, when i worked in motor pricing a number of years ago, i remember getting some phonecalls from underwriters asking for help to explain the pricing to some unsatisfied customers and brokers when their prices changed at renewal. We used a GLM, so it was possible to answer. I’m not sure how i would have responded if we used a neural network for pricing.
This is pretty interesting because of how they define “explaining the model”.
With a GLM you can define it deterministically very clearly but with a random forest you can only give a statistical answer on how the model is built, but you can very clearly deconstruct each internal tree and show the flow of the logic. Not sure that’s enough and not sure if the explosion of modeling heterogeneity will translate into regulations being adapted or at least updated in this regard.
@alfarzan: Don’t get me wrong. You are doing a great job (there are of course always something missing). @simon_coulombe: Your contribution to this game is amazing. I even submitted one time your baseline with a small modification on the pricing. It helped me to confirm that my model is not so bad and I am on the right track. I think I will finish good in the final round.
thanks mate! I’m a big believer in the value of “learning of public”. Someomes I look like a fool, but much more often I get some really cool insight from knowledgeable people I wouldnt have received otherwise. Overall, it’s totally worth it
One of my former managers told me that in the UK market they can create random control group samples within segments and modify their prices since it is unregulated and they do not need to file rate changes. They even have the freedom to change premiums daily if they want to. The purpose of such an experiment would be to measure the price elasticity of each segment directly. New business, of course. Maybe it would work for renewal business too? I don’t know.
Actually from one of the early interviews, this is what I heard as well, specifically in Colombia this one.
In this case they called it deep random-discounting to test where the market was pricing a specific segment of the portfolio. To me sounds pretty interesting (and a tad mad) but I guess the reasoning is that if you have the capital requirements in place already, the insurer can take this kind of risk if they wish.
… I hope 
