I'm trying to do some work on bet size (empirical Kelly, etc), and investigating a rolling position sizing algorithm to grow or shrink bet size based on strategy performance. The goal is to mitigate draw-downs during unfavorable regimes without having to go and classify the regimes explicitly or learn them. However, it seems that this approach would only add value if there was in fact auto-correlation in the returns, no? Anyone have any research handy on this sort of thing?
The next step would be to make this part of a holistic position sizing strategy of strategies that did some sort of mean-variance analysis on the rolling strategies' results.