I am looking at the daily returns of the VIX index (dVIX ) and the daily returns of the S&P 500 (dS).
I am running a linear regression (using 0 intercept) and get a regression slope of -1.4, i.e.
dVIX = -1.4 * dS/S + epsilon
R^2 is around 70%.
Now I am looking at dVIX + 1.4 * dS/S. The way I think about this is that I hedge out the spot movements from the VIX. As expected, the correlation between this construct and dS/S is very low, around -2%.
Naively I had first expected that this would give me something that mean reverts, but actually I have some strong drifts in there:
Also, even though the correlation to spot is almost 0 it still "looks" similar to the plot of the S&P:
I first thought that maybe using a static regression slope over the full time period is not accurate, so I used a moving regression using exponentially weighted moving averages, but the result is very close to what I got above. Also the regression slope is very similar for different parts of the full time horizon, so that cannot be the reason.
Can anybody explain what is happening here?
I am aware that if I actually wanted to trade this, I would have to consider the decay of the vol curve, i.e. my VIX future would move down the vol term structure (if the vol term structure is downward sloping as it would be on average). I have a way of calculating the vol slope from market data (explaining all this would make this post too complicated and long), but even after considering this I still have quite a significant drift.
For background information (not necessary to read to understand my question), the reason why I was looking into this the following: For S&P 500 options the ATM volatility is strongly correlated to spot moves. So I was hoping to modify my delta hedge, so that it does not only remove the delta move, but also the spot exposure in the vega move, so that the option price would mainly be moved by theta and the higher order terms, i.e. gamma, volga, vanna etc.