October 9th, 2019, 11:22 pm
Just a thought to get started. You could first do a simple (zero-intercept) regression using the regular session (log)-returns, say at whatever frequency is convenient:
[$] R^{VIX}_t = \beta R^{SPX}_t + \epsilon_t[$].
Then, the predicted overnight returns, at the same frequency, could be
[$] R^{VIX}_t = \alpha + \beta R^{SPX}_t[$],
where [$]\alpha[$] was changed each night to make each full overnight VIX return come out correctly. Log returns would be easiest for this idea.
You could test the method by also using it to predict some known regular session VIX returns in the same way and see how it does. In other words, just pretend some of your regular session VIX returns were missing and try the same prediction method on those.
