Came across an intriguing post by Markku Kurtti: Employing volatility of volatility in long-term volatility forecasts (outcastbeta). The gist of it is that the volatility of volatility and average volatility is more predictable. So, predict them separately and then put them together to get a volatility forecast.
Our previous attempts at forecasting VIX have led us the conclude that nothing beats locf. Just extending the last value of VIX forward has beaten all of the models we’ve looked at.
- Some background on Zerodha Varsity: GARCH vs locf.
- Using Meta’s Prophet library: Prophet for VIX.
- And, more recently, Conditional Gaussian Mixture Models for VIX.
What if, we applied the process outlined in Mr. Kurtti’s blog for VIX?

Not looking so good if you forecast out for 20-days.

A good forecasting model should be able to estimate large spikes and reversion to mean from them. If not the former, then at least the latter. So, if a model is good at the mean-reverting bit, then you can combine both locf and the model to get better predictions.
For example, if you only consider the 1-day forecast, then obviously locf has the upper hand.

However, as you try to predict farther away, you expect the model to nail the mean-reversion bit.

That is not the case here.
Once again, locf wins.
Code and charts up on github.