Up until now, I've been focusing on special cases in my sampler. I started with iid, simply sampled in blocks rather than individual rows. That went to the hit rate being correlated in blocks and then the hit rate being correlated in partitions in blocks. It's time to finally take on the most general case: each partition gets its own distribution and hit rate. No caveats.
On the one hand, it's easier. Since we are refusing to make assumptions, we're pretty much left with empirical methods. On the other hand, the point of this paper is to rigorously state things that most people just wave their hands at. (This emphasis is not merely academic; when I take these ideas to the actuaries at work, they are willing to listen, but they absolutely want proof. This company has 6 trillion - yes, 12 zeros - dollars at risk; they aren't going to stake that on something that appears to work).
No comments:
Post a Comment