Tuesday, January 18, 2022

Machine Learning for Stock Return Volatility

This new Filipović-Khalilzadeh paper is quite nice.


They do both tree methods and neural nets for realized volatility forecasting, using not only the RV history but also various "standard" observed predictors.  Fine.  But the really interesting thing is their implementation of the "long short-term memory model," which wins all their races: 


Hard to tell if it's really capturing long memory in the statistical/econometric sense (a crucial finding of Andersen et al (2003)), and they don't discuss or even mention statistical/econometric long memory.  Perhaps the workings of the "long short-term memory model" are close to those of the "Corsi approximation" to long memory used in Andersen et al. (2007).

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.