Machine Learning for Realized Volatility Forecasting

Check out this interesting paper.  Lots to think about.

Kim Christensen & Mathias Siggaard & Bezirgen Veliyev, 2021. "A machine learning approach to volatility forecasting," CREATES Research Papers 2021-03, Department of Economics and Business Economics, Aarhus University.
https://ideas.repec.org/p/aah/create/2021-03.html

The paper shows that various ML methods outperform HAR for daily forecasting of realized asset-return volatility.  Of course there's nothing particularly interesting about HAR.  What's interesting is long memory, the overwhelmingly dominant feature of asset return realized vol dynamics, and HAR is just an approximate way to capture the long memory while staying in a comfortable linear regression framework.  Anyway, the key unanswered questions raised by the paper is how are the ML methods approximating long memory, and why do they deliver better approximations to long memory than HAR?  It is well known that long-memory and (infrequent) regime-switching are closely linked.  Perhaps the ML methods are picking up infrequent regime switching?  Do they also deliver better approximations than exact long-memory models?