Rotella Qdeck Machine Learning and Volatility Strategies

Smart portfolio hedging and FinTech customisation

Hamlin Lovell
Originally published in the March | April 2021 issue

Rotella Qdeck ML Long-Only Volatility Program (MLVP) received The Hedge Fund Journal’s 2021 CTA Award for Best Performing AI/Machine Learning Long Volatility strategy in 2020, when the strategy made 27.52% including a gain of 10.92% in March 2020 and limited monthly losses to single digits. 

This is currently Rotella’s only strategy that is 100% based on machine learning models, but the manager has since 2018 applied some machine learning to the signal generation process in several programs, including trend and short-term trading, which also generated strong returns in 2020. “The trend portfolios mix tactical models with machine learning models calibrated to a trend mandate. Other long volatility programs, such as the Equity Tail Risk Diversifier and the Tactical Long-Biased Volatility Program also blend machine learning with Rotella’s classical tactical models,” says lead volatility portfolio manager, Kemp Nicklin.

Short volatility without tight risk controls can blow up in the manner of picking up pennies in front of a steamroller.

Kemp Nicklin, lead volatility portfolio manager, Rotella

Performance objectives and role in portfolios

MLVP has a regime-conditional performance optimisation objective: it aims to profit from most equity market corrections while mitigating time decay costs and generating positive returns over time. It met all three criteria in 2020 and 2019. In the first half of 2021 time decay from maintaining long volatility exposure outweighed profits on long equity exposure. 

It is always long of implied volatility indices and usually also long of equity indices, but actively varies position sizes of both. Opportunistically monetizing volatility spikes, and profiting from uptrends in equities, are both designed to mitigate the costs of being long volatility. 

MVLP complements trend followers in a portfolio. “Trend followers will succeed in a long-term equity drop, like the 2008 crash, by riding the trend all the way, but they can miss sudden shocks such as the one in February 2018 or get whipsawed in shorter term corrections. The March 2020 crash was difficult to profitably trade since the equity market rebound after the initial drop in February could have stopped out some simple trend following approaches. The volatility program can even be seen as whipsaw insurance for left tail regimes when trend following CTAs do not work, and it has low downside correlation with momentum. It can provide some crash protection, but does not necessarily need a big adverse event,” says Nicklin. 

MVLP has learned from recent patterns of market behaviour. “The strategy complements traditional trend following CTAs because it is a good hedge for higher frequency, instantaneous bear markets and shocks that happened more often in recent years. Investors need a spectrum of hedges to help reduce the tails and smooth out the compounding of returns,” says Jagdeesh Prakasam, Rotella’s CEO since 2019, who joined in 2003. 

Liquid, scalable and responsive instruments 

The choice of instruments is simple, liquid and scalable to match the performance objectives: the front three months of the US VIX future and the front month of the European VSTOXX future. “Another reason for staying near the front end is that shorter-dated contracts generally show more responsiveness and convexity to exploit surprises. Longer term VIX contracts have their own interesting dynamics but not so much utility for this purpose,” says Nicklin. Though increases in realized equity volatility do not always filter through into a higher VIX, which measures implied volatility, the hit rate with short term contracts is pretty high. 

Though the European VSTOXX is highly correlated to the VIX, it can sometimes perform differently and provide a diversification benefit: “During the European sovereign debt crisis in 2011, the VIX and VSTOXX both rose early in the year, but the VSTOXX hung onto its gains throughout the year,” says Nicklin.

Why pair long equities and long volatility?

Various approaches to volatility trading bring their own challenges. “Short volatility without tight risk controls can blow up in the manner of picking up pennies in front of a steamroller. A simpler long volatility put replication is likely to lose money over time, because it would struggle to hang onto profits in a rebound. Staying only long of volatility spreads the losses over time and truncates left tail events,” explains Nicklin.

27.52%

The ML Long-Only Volatility Program received The Hedge Fund Journal’s 2021 CTA Award for Best Performing AI/Machine Learning Long Volatility strategy in 2020, when the strategy made 27.52%

One rationale for pairing longs in implied volatility and equities is that they usually move in the opposite direction, hence the equity position is described as a “hedge” for the long volatility, not the reverse. A breakdown in this correlation pattern can be either benign or malign for the strategy. A dual rally in equities and volatility can be profitable, as seen in 2020 during some brief periods though this only occurred 4% of the time on daily data, such as in parts of August 2020. A double bear market in equities and volatility can be adverse but is also rare. “It happens most often after a spike in volatility, which also means that it would tend to occur after a profitable period for the strategy. The most recent example was late 2018, which was somewhat unique for the size of the equity drop that occurred without a VIX increase. After the February 2018 “volmageddon”, equity markets moved to a higher volatility regime, which squeezed short volatility strategies. This meant that volatility unusually had a muted response to the larger equity market drop in late 2018. Our model was then adjusted to be more robust and perform better when volatility is not as responsive. Early 2016 also saw a VIX somewhat below its 2015 levels while equities corrected,” says Nicklin. The program posted single digit losses in between inception in September 2018 and December 2018.

Optimisation and monetisation

A passive approach to staying long of both equities and volatility in constant sizes and ratios could have lagged the Rotella strategy by about 15% in 2020, according to Nicklin. Active optimization and monetization are needed to balance the goals of portfolio insurance, cost mitigation and long term returns over multiple periods and scenarios. “There are many false alarms that quickly calm down and there are many different scenarios. Our signal inputs for optimisation include volatility and volatility of volatility. A spike in both together could be one stronger signal for taking profits on the VIX,” says Nicklin. The term structure of the VIX also feeds into the models since it is more expensive to roll a VIX market in contango than one in backwardation. Monetisation or portfolio rebalancing can involve reducing the VIX position, adding to the long equity position, or both, and can be done daily or intraday. Optimisation is the primary driver of monetisation, but occasionally risk management constraints, such as the volatility target of 12% or even margin to equity constraints, could force exposure reduction. “We automatically de-leverage and downscale if volatility, or volatility of volatility, become very high,” says Nicklin. “Optimising the multiple inputs would be hard to accomplish with only human rules, but machine learning can extract nuances that show if the cost/benefit analysis stacks up.”

The Covid crisis showed how the rule can work well: “The system exits the equity position about 15% of the time and we completely exited the long equity position during the Covid crash. We also stayed in the long VIX position for enough time to accumulate some profits before monetizing part of it,” says Nicklin.

The research process 

These algorithms were developed using Rotella’s expertise in machine learning, which uses special techniques to model the asymmetry and fat tails of financial markets, all informed by Rotella’s decades of experience in building trading systems. “Genetic algorithms train models to optimize certain metrics, subject to constraints, such as the performance objectives for the strategy. Traditional regression analysis can predict small moves successfully, but long term the left tail of the distribution, which is harder to estimate, is what we are focused on and assign greater weights to. We specify a complex non-linear function, and optimize it to allow for back-testing, which includes execution and transaction costs and must be fast to evaluate the objective many times. We used VIX futures data back to the start of the contract in 2004. This captures nuances that do not map onto a simple classification and prediction approach, which can lead to brittle outcomes in the case of extremely non-Gaussian return distributions and asymmetric payoffs,” says Nicklin.

This is an art as well as a science because philosophical decisions can influence future results.

Jagdeesh Prakasam, CEO, Rotella

Parsimonous prioritising of signals

The algorithms are run thousands of times and rewarded for success and penalized for failure in the manner of reinforcement learning. Common indicators and networks are used for the volatility and equity indices and for the relationships between the two. The neural network has thousands of inputs but needs to prioritise them. “Too many parameters slow down the optimization and require more evaluations which also constrains the neural network. We have shrunk it without sacrificing the unique and informative parts of the data,” says Nicklin. “The dimensionality reduction process is conceptually similar to principal component analysis, with some twists. It identifies around 15 unique, orthogonal, mutually uncorrelated components, that map back to the original dataset to map the feature set. Rather than deep learning, this is really a parsimonious approach without too many parameters to bog down the estimation,” says Nicklin.

Rotella’s 15-year machine learning journey

Rotella has been working with machine learning and neural network frameworks for over 15 years, managing challenges such as classification, overfitting, complexity and back-testing. “The research budget was expanded to explore machine and statistical learning approaches and develop objectives that can be clearly explained even to those with no grasp of AI. This is an art as well as a science because philosophical decisions can influence future results,” says Prakasam.

One core belief is balancing efficacy against complexity. “We started off by trying to adjust leverage, applied to the same signals, to improve the Sharpe ratio by 10-20%. We required a good improvement in Sharpe relative to the number of parameters – we would not hyper-fit a model by adding 30 parameters for a 10% Sharpe improvement. We also need to allocate live proprietary capital to determine whether or not a model was overfitted,” says Prakasam.

The framework has also spawned cross market signals derived from within the investment universe traded. The volatility program has a two-way traffic of signals between equities and volatility. The multi-asset class trend programs, trading equities, bonds, currencies and commodities, might for instance use evolving signals from oil or Treasury bonds to inform trading of equities indices.

Collaborative research team

Rotella employs a broad spectrum of quantitative individuals including several physics PhDs, AI/ML experts and more traditional financial engineers. “We pair theoreticians and practitioners, so that obvious mistakes are not made. Either group alone could make huge mistakes or be in constant research mode for years and not develop a tradable signal,” says Prakasam. There are three teams – data science, quant engineering and investment – that work openly and collaboratively with open-source code rather than in siloes. Some people straddle more than one team or move between the teams over time. “The research group have ideation brainstorming sessions three times a week to share knowledge and understand why models work. They are incentivized to collaborate and would rather colleagues, than the market, call out their errors,” says Prakasam.

Customised strategies, model testing and portfolio construction 

Rotella’s other existing long volatility programs have some differences with MLVP: “The Tactical Long-Biased Volatility program also mitigates costs through calendar spreads. It had a strong 2020 though the MLVP was more responsive during smaller market shocks. The Equity Tail Diversifier is tactically long volatility and can be long or short of equity futures,” says Nicklin.

These are other possible variations on the long volatility theme: “We could adjust the amount of equity exposure or might even have no long exposure for clients who are purely seeking downside protection. There is also interest in portfolio insurance strategies with a zero or negative expected return, such as put replication. In Asia, which lacks liquid futures, Rotella could contemplate an options-based strategy,” says Nicklin.

More broadly, customisation requests across the strategy suite typically involve different sector exposures, markets traded or combinations of trading programs.

Customisation can also spawn new strategies. Rotella manages proprietary and external capital and has now also developed a FinTech offering – Qdeck – that can help other allocators to back-test and prototype models and signals and build portfolios, alone or with Rotella. “I do not envy funds of funds because they do not get enough visibility into what signals are doing. They may see phenomenal back-tests in 20 black boxes but need to tailor strategies to very different mandates with aims for absolute return, tail risk, and other objectives, but have no control over how managers evolve strategies. We can offer firms radical transparency over our intraday signals, which can be compared with other managers, in a cloud-based platform. There are also natural language programming modules. The vision is that allocators, advisers and endowments can then compete with robo-advisors who are building their own programs,” says Prakasam.