Pretty much all countries have the no claim bonus for driving. (It's in the earliest actuarial exams.)

You can clock up several years of no claims bonuses, and the discount increases with each year.

Interestingly, the no claims bonus can be insured. Now, generally only risk averse people will insure their no claims bonus - and these people also tend to drive more risk aversely and hence have less accidents (or perhaps less likely to make claims - who the hell knows).

So, the very fact that somebody wanted to insure their no claims bonus should in theory give them a discount on their premiums. But you cannot offer a discount for an extra feature! This means you have to dive into your generalized linear model and fiddle with a couple of parameters to make sure there is some sort of nominal charge for the insurance on the no claims bonus.

My understanding is you have access to total monthly credit usage vs. credit availability, not what I'm spending money on. I would think there's information to be gleaned from whether the distribution of spending across categories looks like a deliberate household budget, whether purchases include luxury items, etc. that would let you spot people likely to bury themselves in credit card debt before they actually do so. Just as people's social networks might tell you about their attitudes towards money when influenced by friends.

I have not worked for the US market the last five years, but as far as I remember there is a law in the US that a bank cannot take adverse action on a existing customer holding the bank's product based on credit history from another bank - so for example, my bank cannot decrease the limit of a card because the customer defaulted in a personal loan held with another bank. Of course, at the time of origination of the credit card, we can use this information.

The whole point being that just because data is available does not mean I can use it even though it may be statistically significant for me to use it.

Again, I am not sure of insurance regulations so there may be a possibility around using more data, if available, in insurance models.

"Your demographic" is just a feature of your model though. Just because your methods are not very sophisticated, doesn't mean that others can't be upset that you arbitrarily group them by whatever rules you decided to come up with.