One hallmark of a great idea is that it's "obvious" (ex post). Fantastic recent work by Bollerslev, Patton, and Quaedvlieg (BPQ) certainly passes that test.
BPQ build on the classic Barndorff-Nielsen and Shephard result that the precision with which realized variation and covariation are estimated is time-varying but can be estimated (let's just speak of "variation" for short, whether univariate or multivariate). Put differently, the measurement error in realized variation is heteroskedastic but can be estimated. Hence, for optimal variation prediction, one should presumably weight the recent past differently depending on the estimated size of the measurement error. BPQ do it and get large predictive gains. Check it out here. (This is the new and multivariate (covariance) paper, which cites the earlier univariate (variance) paper.)
Why didn't I think of that? I mean, really, the Barndorff-Nielsen and Shephard result is more than a decade old, and I know it well. Can I not put two and two together? Damn.
But seriously, congratulations to BPQ.