Don't Be Too Smart About History
Why filtering for “similar” market regimes can make forecasts less reliable
Statistical models used to predict returns typically make a quiet assumption: That the past is uniformly informative about the future. A standard predictive regression assigns the same weight to turbulent periods, such as October 2008, as to calm periods. The data goes in, the estimate comes out, and the differences between market regimes are averaged away.
The natural fix is to condition on relevance, to identify which historical episodes most resemble today and place greater weight on those observations when forming forecasts. It is an intuitive idea with an elegant mathematical foundation, and it has attracted interest in both academic research and applied quantitative investing.
But there is a deeper question beneath the intuition:
Does conditioning on “relevant” history actually improve forecasts, or does it simply amplify noise precisely when markets become most unstable?
