I am still working on some spread (ie pair) and basket mean reversion models, and finding that mean reversion of the residuals is as often as not caused by the model adapting (too quickly?) to the spread dynamics, so while everything looks good up until simulation, in actual fact, the spreads as-entered don't or don't have time to mean-revert.
Has anyone dealt with this issue, or read papers/posts about it? I am looking for ideas, in addition to:
- slower dynamic beta/spread models
- recomputing the betas on a fixed schedule, like monthly
- 'fixing' the betas temporarily when a trade is entered