Updated: Nov 20, 2020
By its very nature the VIX indicator should be able to provide some level of predictive power for SP500 stocks. VIX gives an indication of the implied volatility of the underlying stocks using options valuations. Wikipedia claims, grandly: "The VIX is a 30-day predictor of volatility given by a weighted portfolio of out-of-the-money European options on the S&P 500." If it were such an excellent predictor the behaviour and operational reality of the US stock market would be (and had been) very different. The predictive power of the VIX and the reality of the market do not reconcile well, this predictive power may not exist, or at least not exist as a simple, direct prediction of future volatility. We are going to assume that the index does indeed contain information that can be used to enter and trade stock market positions one way or another.
This paper from 2013 concludes: "We find that the regression R-squared is 0.6425 and
therefore suggest that VIX does contain some explanatory power in regards to the SPX
future 22-day volatility. However, within SPX high-volatile periods, we can easily observe
a great number of outliers." This other paper (revised 2018) concludes, in a similar line, that: "...without any diffusion assumption, the model-free formulation of the CBOE volatility index (VIX) is actually a linear combination of ex-ante return-moments, not the expected volatility."
As we are always looking for sources of future insights we put the conclusions of these papers (partially) to the test using new data from the very high volatility period starting February 2020 generated by the COVID19 crisis. We will try to predict the market looking at VIX and obtain profits in that difficult period.
Lets first go back in history a little bit to analyze the behaviour of the VIX and SPY indexes in the 2005-2015 period and attempt to extract conclusions for the 2015-2020 period. Beginning in 2005, the general shape of the daily closing prices and value of SPY index and VIX index are these:
The VIX index behaves in such a way that it remains stable during low volatility periods and spikes during high volatility periods, fine, it is supposed to act so. The behaviour includes certain directional bias, VIX is apparently not treating all volatilities equally, downward movements of the SPY are associated with more pronounced movements in the VIX. Options are generally used as hedge or protection against the underlying, in this case the SPY, so it may make sense that periods of exuberance in the stock market do not affect the VIX as much as the threat of large drawdowns. This bias by itself would not destroy the predictive power of the VIX (if there is any), it will only show a preference for a directional prediction. We can adventurously claim at this point that VIX will spike with the expectation of large drops and stay calm with upward movements. This is a very interesting and skewed method of measuring or predicting volatility, still useful for our oracular objective in any case. How negatively correlated are then SPY and VIX? For this period the correlation coefficient is -0.49 confirming the preference of the VIX for indicating more the fear of loss than volatility itself. From this period chart we can also conclude that the VIX index value time series is stationary. The Augmented Dickey–Fuller test yields a p-value of 0.028 indicating a stationary series in this time frame.
The following chart shows the value of the VIX against annualized rolling standard deviation of the SPY price for 22 market days, 22 market days into the future. We are trying to find how well VIX does its job of predicting future volatility, and using the housing bubble crisis looking for more intense movements:
The VIX indicator consistently lagged and underestimated volatility during this unstable period. This scatter plot shows what the VIX was trying to predict and the actual value:
With this coefficient of determination of 0.13 (r-squared) and all those outliers on the highest SPY volatilities show a very weak capacity from the VIX to predict volatility during a risk-critical period. Can we at least predict the return of SPY after 22 days looking at the VIX instead?
Then, being VIX apparently a lagging indicator that we could long or short, can we use the SPY index 22-day volatility to predict VIX in 22 days during the 2008 crisis?
Apparently not extremely well but better that trying to predict the SPY. With an r-square value of 0.36 we may be able to find the subset of SPY 22-day volatility that yields predictable VIX changes during this crisis. The next scatter plot shows the same SPY 22-day volatility against the 22 day future change of VIX index:
In general it appears that the 22-day volatility does not predict the change in the VIX. Although there might be an effect on VIX for extremely high SPY volatilities. Let's take a look at this effect at extreme volatilities for the 2005-2015 period and check that it persists, if it does we may enter long positions in leveraged, VIX tracking ETFs (for example: UVXY) when the volatility hits very high levels.
For very, very high SPY volatilities the effect persists in the 2005 to 2015 period. It is apparent that if the SPY volatility exceeds 80, the change in 22 day VIX value is positive and large. Before we try anything, as we are looking at these "large" SPY volatility values, we will check this same scatter plot for the 2015-2020 period including the large movements during the COVID19 crisis:
The volatility values increase to very high absolute values, there were 25% swings in a weekly basis for almost 60 days that resulted in extremely, never seen before, 22-day volatility values for SPY. Zooming in into the interesting region were volatility is above 80:
The prediction of positive VIX changes for large volatilities breaks down for extremely large volatility. The effect in this case could be that not many players in the market (retail investors, institutional investors, regulators, national authorities) understood the possible effects of the crisis, leading not only to drops in the value of the stocks, but also large swings showing a lack of clear direction, the worst possible market is not a market that falls, is a market that behaves erratically. If we limit ourselves to predictions in the 80-150 volatility historical ranges the scatter plot shows:
Quite inconclusive. There are in this case more points with negative returns for VIX after 22 days, but the movements upwards are larger and can (maybe) be captured. In any case we can anticipate that a trading model based in these SPY-VIX effects will have a very large deviation in the returns and large drawdowns.
It is time to try and model a trading strategy: we will buy and hold SPY and if the 22 day volatility exceeds 80 we will enter long positions in leveraged VIX-tracking ETFs until it drops below 80. If the volatility exceeds 150 we will close all positions and go to cash. No other risk controls or assets are to be used. The test starts November 2011, after the creation date of the UVXY ETF.
Implementing the model in Quantconnect leads to the following backtest results:
The strategy performs well up to 2018, at this point UVXY has been triggered 6 times providing moderate profits, it is impossible for the model to decide the best time to release the UVXY positions at the beginning of the 2018 market instability. Then in 2019 and 2020 nothing helps us even if we exit the market between February and March 2020 due to unseen volatility.
Here are the model code and the test results:
We can see, in any case, a potential for a volatility-predicted VIX strategy that we are going to develop in the future. Our first approach will be using deep learning for time series trying to discover a solid relationship between SPY volatility and VIX returns that we can consistently profit from. It will start with SPY information predicting VIX future as it is apparent that there is more power in this prediction direction than in the opposite.
Remember that the posts in our blog are not financial advice. We do not hold any positions on any of the mentioned instruments at the time of publication of this post. If you need further information, asset management support, automated trading strategy development or tactical strategy deployment you can contact us here.