Qdeck Rotella ML Long-Only Volatility Program (“MLVP”) is a fully-automated trading strategy designed to complement portfolios consisting of stocks, bonds and momentum.
Adding long VIX futures attains similar exposure to rolled put option exposure. This results in the potential benefit of profit during extremes, but the cost of time decay under typical market conditions. Time decay loss from holding such a position correlates strongly with short equity exposure. This is exactly analogous to the time decay and negative delta from being long put options. Thus, simultaneously holding equity futures acts as a time decay and delta hedge. An edge can be obtained by adjusting the volatility and hedge positions over time in such a way that time decay is reduced and tail benefits remain.
To do this we may take advantage of the following stylized facts:
- Volatility prices increases as equity prices decrease
- Volatility of volatility level increases as equity prices decrease and equity volatility increases
- Volatility term structure tends to backwardate as equity prices decrease. (Buying and holding volatility becomes a profitable strategy during sufficiently large equity drops, such as the 2020 COVID-19 financial crisis)
These general concepts map to statistically measurable features. It is possible to capture these effects using a rules-based tactical approach (for this see Rotella Tactical Long-Biased Volatility Program). Here we use a large set of statistical features and a proprietary machine learning algorithm, described in the following sections.
The portfolio of MLVP always consists of long positions in VIX futures (nearest 3 contracts) and VSTOXX futures (front month only). These positions are each hedged with a positive (or zero) position in E-Mini S&P 500 futures and Euro STOXX 50 futures respectively.
Machine Learning Model
MLVP uses a neural network with custom architecture to reach its investment objectives. Its feature set consists of over 4,000 technical variables that serve as different measures of relative price level changes, volatility level changes, and volatility of volatility level changes. These features are mapped to a lower dimensional space for computational feasibility, and then combined on a cross-asset basis. The latter combination step results in a purely statistical way of measuring relative term structure changes. The model is then trained in a non-standard way: no attention to predictive accuracy is made. Instead, the network is trained to maximize the worst-case benefit on tail events, conditioned on overall performance being positive. Finally, model outputs are normalized to constrain risk characteristics of the final position.
The upshot of this seemingly complicated model is simply a long-only volatility position that changes very slowly, gradually increasing during crisis periods. The volatility position is then assisted by a much faster changing equity hedge position, which takes advantage of the much greater liquidity of the equity futures. The equity hedge exits quickly when volatility term structure begins to flatten and invert, or when volatility of volatility increases. Figure 1 below demonstrates a sample period of the net volatility position (blue) vs. the net equity position (red) for the S&P 500 component.
Figure 1: Net positions
During the recent financial crisis, long volatility strategies have performed extremely well. It is difficult to balance gains from such a period with their time decay losses in calmer times, however. MLVP seeks to achieve the best of both worlds, delivering weakly positive returns during volatility contango periods and sharply positive returns this month. We have provided performance charts to illustrate its performance compared to an equity benchmark (S&P 500 Total Return Index) and a momentum benchmark (SG CTA Index).
Figure 2: MLVP experiences compellingly positive returns in the recent crash.
Figure 3 (HYPOTHETICAL): Adding MLVP to an equity portfolio has the potential to significantly reduce losses during stock market downturn such as COVID-19 crisis in March 2020.
Figure 4: Momentum as measured by SG CTA Index has significantly underperformed compared to MLVP in the recent drop.
Figure 5 (HYPOTHETICAL): Adding MLVP to a momentum portfolio has the potential to enhance its downside performance without adding time decay drag.
Qdeck Rotella Tactical Long-Biased Volatility Program
Qdeck Rotella Tactical Long-Biased Volatility Program (“TLVP”) is another volatility program with similar investment objectives to MLVP. It is a portfolio of several subsystems:
- An opportunistic long-short system trading VIX futures contract hedged with E-Mini S&P 500 futures
- Opportunistic long-short system trading VSTOXX futures hedged with Euro STOXX 50 futures
- Long only VIX and S&P E-Mini futures
The first two systems explicitly exploit the relationship between volatility term structure and its future returns as a form of directional carry. This long-short exposure serves as a time decay hedge, while opportunistically capturing tail events. The final system attempts to measure volatility breakout conditions, and piles on long volatility exposure in such environments while maintaining an equity hedge.
TLVP has benefited from the recent large volatility shock as well as the proceeding backwardation in volatility futures. Below are the relative and combined performance charts for the program:
Figure 6: TLVP has experienced a sequence of large magnitude, positive returns during the recent crash.
Figure 7 (HYPOTHETICAL): adding TLVP to an equity portfolio has the potential to result in net positive performance through the crash period
Figure 8: TLVP seeks to make up for momentum losses in the recent crash
Figure 9 (HYPOTHETICAL): TLVP seeks to add significant value to a momentum portfolio while introducing minimal time decay drag.
Avellaneda, M. et. al. “PCA for Implied Volatility Surfaces.” January 31, 2020. [https://arxiv.org/pdf/2002.00085.pdf]
Badran, Alexander & Beniamin Goldys. “A Market Model for VIX Futures.” April 3, 2015. [https://arxiv.org/pdf/1504.00428.pdf]
Chakrabati, Arnab & Rituparna Sen. “Some Statistical Problems with High Dimensional Financial Data.” August 8, 2018. [https://arxiv.org/pdf/1808.02953.pdf]
Kamalov, Firuz. “Forecasting Significant Stock Price Changes Using Neural Networks.” November 21, 2019. [https://arxiv.org/pdf/1912.08791.pdf]
Miot, Alexandre & Gilles Drigout. “An empirical study of neural networks for trend detection in time series.” March 2, 2020. [https://arxiv.org/pdf/1912.04009.pdf]
Prado, Marcos Lopez. “Advances in Financial Machine Learning.” February 21, 2018.
Vazquez, Samuel. “Options Pricing, Historical Volatility and Tail Risk.” February 6, 2014. [https://arxiv.org/abs/1402.1255]