Ever since the latest phase in Millburn’s strategy evolution began in 2013, its flagship diversified strategies have delivered a Sharpe ratio near one. These strategies have also significantly outperformed most traditional CTAs, including positive performance in each of the last five consecutive calendar years. Beyond this, Millburn’s niche, relative value commodity strategy (which is now soft closed) has produced some truly exceptional returns. Investors have clearly appreciated the performance — AuM has grown steadily over the period to more than USD 6.5 billion today.
As always, Millburn’s process aims to combine the firm’s substantial market knowledge — decades of experience systematically trading global futures and FX markets — with techniques to understand data. Today, though, the process is distinguished by the application of powerful machine learning technologies (sometimes referred to as statistical learning) — tools more commonly used for single corporate securities — to macro markets. And far from an overlay or sub-strategy, the approach is used holistically for signal generation in the form of both return and trade cost forecasting.
Some CTAs remain to this day substantially or entirely committed to traditional trend-following (which we define as using purely technical or price data and only momentum-based models). While Millburn was one of the original innovators in trend-following — its roots date back to 1971 —interestingly it was also among the earliest managed futures firms to move away from traditional trend-following in a significant manner. Millburn began reducing allocations in mid-2000, complementing pure trend-signals with signals from independent, hypothesis-driven, single-strategy models that were based on fundamental, behavioural and other non-price data.
These are incredibly powerful technologies, but we are humble in how we apply them.
Grant Smith, Chief Investment Officer and co-CEO, Millburn
The move in 2013 to the ensemble multi-factor, statistical learning-based framework, blending a variety of models, was therefore less a move away from something than a move towards what the firm saw as a better technology to deploy and harness the growing range of data that had already been feedstock for the models. “Over the last decade, it became clear to us that the explosion in data had the potential to have a major impact on understanding markets. This meant simple rules-based approaches to using momentum were perhaps not going to work as well going forward. And while some peers were moving in the direction of providing trend beta, we made the decision early-on to remain on the hunt for alpha —which meant investing heavily in a framework that could make better use of a range of data, both price and non-price. But we also wanted a process that had a built-in ability to adapt, to better match the increasing pace of change we were seeing in markets. We saw this as the source of our edge,” says Barry Goodman, Millburn’s co-CEO and Executive Director of Trading.
By mid-2013, the initial stage of implementation of the multi-factor framework was complete. Fast-forward six years to today and Millburn trades multiple billions of AuM using this multi-factor and multi-data statistical-learning framework. In the world of machine learning, this size and experience probably makes Millburn one of the leaders in the futures and FX trading arena.
While July 2013 was one milestone, the firm now sits on the edge of another. As of February, it has phased out traditional trend-following entirely in favour of its contextual statistical-learning framework, across its main long/short Diversified, Multi-Markets and Commodity programs, completing the firm’s transition from a rules-based to a pure data-driven approach.
The CTA universe has become more diverse over time and the question arises as to how Millburn should be categorised. Illustrating the steady move away from a hard-wired momentum approach, Millburn’s rolling beta to trend followers has been on a downwards journey, as shown in Fig.1.
In recognition of Millburn’s declining beta to the trend-following space, SocGen removed the firm from its Trend Index several years ago. While trading liquid and mainstream markets similar to a classic CTA, and while still delivering on the low- to near-zero correlation to traditional equity and other hedge fund investments that many investors seek, the firm may fit just as easily into a multi-factor quant or quant macro category as a managed futures category. “Strategy classification labels are in fact something that the firm isn’t too concerned with,” admits Goodman.
“We are looking to deliver a certain type of performance for our investors. In general, we are seeking alpha that is non-correlated to other investments in an investor’s portfolio, so the diversifying characteristic is key. But we are also trying to build a strategy that does not rely upon a particular market environment to have the chance to profit,” he continues.
This has been borne out historically, where the real differentiation in the firm’s return stream has come during non-trending or “choppy” markets. Here Millburn has been able to generate some very good results. And during reversals Millburn was able to minimise losses, as shown in Fig.2. This has been crucial because since 2013, markets have been in non-trending environments about 80% of the time, according to Millburn’s classification.
The CTA industry is moving away from the “crisis alpha” narrative as this historical property of CTA returns is clearly dependent on the pattern and path of financial market returns. While February 2018’s equity market reversals saw Millburn’s positioning somewhat aligned with that of typical CTAs, in late 2018, Millburn actually posted some strong profits during the selloff; its Multi-Markets Program, for example, made 6.90% between September 20th and December 24th 2018, a period where the S&P500 Index lost more than 19%. Millburn’s performance was in marked contrast to traditional trend followers and other CTAs, which generally lost money over this period. So while there are clearly no guarantees, Millburn’s approach does retain the possibility of profiting during major risk-off episodes that coincide with trend reversals.
“We are certainly not leaving momentum behind, and we do expect to be able to pick up on trends when they happen. We still believe momentum can be instructive in terms of understanding forward-looking price movements in markets. So every model we construct incorporates price momentum. The same range of trend inputs used in most traditional trend-following CTAs are provided to our models, and more,” says Grant Smith, Millburn’s Chief Investment Officer and Goodman’s co-CEO counterpart.
“But our difference versus more traditional approaches is we leave it to the process to determine when these trend inputs are predictive, and when they are not, based on context. In some cases, momentum may not be useful at all. The key question is: how can you identify such periods?” he continues.
Machine learning is the answer, but it comes in many flavours. In line with the firm’s approach to diversification, the statistical learning techniques used vary across models, and in general no preference is given to a particular technique.
“There are practically unlimited ways to apply machine learning. Different techniques, different training sets, different data inputs, different algorithms within each technique, different retraining frequencies, etc. So there is very much an art of application. This is not ‘cut-and-paste’. We have been working on this for more than seven years and take a very risk-managed, diversified approach — I think that’s been the key to our success to date, versus others who have tried to use these statistical approaches in a deep way,” says Goodman.
Some managers may apply machine learning only to specific sleeves of portfolios or certain functions. In contrast to those who may use machine learning techniques to pick signals or weight standalone models, and in contrast to others who may use the technology only to execute more efficient trades, Millburn applies its multi-factor approach fully to return forecasting and execution cost forecasting; 100% of each forecast comes from these machine learning techniques.
“Research into new factors, or ‘features’ in machine learning parlance, is one of the most important dimensions of our process. We’ve been adding new factors at a steady rate since 2013, and today use roughly six times the number we used in July 2013,” says Smith.
In a typical model, price data might constitute 40% or more of the inputs, reflecting the firm’s continued belief in price as potentially good information for return forecasting. Non-price data, however, represents a large and growing proportion of inputs in a typical model. This could include more traditional fundamental or market structure data, for example, but increasingly the firm is evaluating the effectiveness of so-called “alternative” data sources.
However, in certain senses it is slightly artificial to split the data between price and non-price, as each is conditioned by the other – and both can be context- or regime-dependent.
This can result in signals that are, sometimes, very different from simple price-driven trend-following signals. The graph in Fig.4 compares trade signals from Millburn’s machine learning approach with those from a traditional trend-following approach, using Brent Crude trading as an example. For some multi-week and multi-month periods, Millburn’s positioning appeared to track a trend-follower more or less in lockstep, but for other periods, like in mid-July, Millburn held the opposite positioning to trend models, in an attempt to find profit during choppier periods. And in some cases, like during the strong reversal starting in October, Millburn’s multi-factor approach was able to move much more quickly, establishing a short position nearly a month sooner than the firm’s trend-following strategy.
Millburn also thinks that the distinction between traditional and alternative data will become blurred over time. The firm is approached regularly by vendors who have taken what was once “unstructured” data —such as satellite imagery — and turned it into something structured that can be evaluated by their framework. Millburn expects this trend to continue, providing more and more datasets to evaluate.
Given the exponential increase in the volume of data available globally, to have used six times as many inputs as five years ago may not be as dramatic an acceleration as it sounds. The expansion of data inputs at Millburn has taken place at what the firm would describe as a “measured pace” because this growing library of datasets has to be carefully curated. For instance, some datasets have been rejected as they may not have long enough histories, or even if long enough the histories may not be comparable and consistent over time.
“For us it is all about data, but less about volume than about quality. We don’t worry too much about whether a data source is labelled traditional or alternative. We’re concerned with things like: is the data clean; can it be verified; do we have enough history to let the models ‘learn’ on; but most importantly, is it valuable to the return forecasting process. These are all relatively straightforward criteria for us to test,” says Smith.
While the firm is tight-lipped about the exact nature and volume of data used, they make a point of avoiding the “big data” label. “Ours is not an approach whereby we give the machine hundreds of thousands or millions of esoteric data inputs and just let it run. We’re using data sources that in most cases, were I to show you the list, you would think seem quite reasonable in terms of having a potential influence on returns. Outside of the momentum inputs, which we have been using for decades, many of these data sources have been used at Millburn for ten years or more. So, we’re pretty confident they can add value. The difference in the framework we’ve been using since 2013 is in the contextual approach,” says Goodman.
The contextual approach, Millburn believes, is a better way to extract alpha out of information. “The whole idea of contextual understanding of markets is incredibly intuitive to us as researchers, and also, it turns out, to our investors. I don’t think anyone would argue that the release of a crude oil inventory number may have some potential impact on near-term price movements in that same market. But perhaps an inventory number released in July may mean something different than if that same number is observed in April, because of heating seasons and driving seasons. And what else might influence the price? Perhaps the price might react a certain way to an elevated inventory number released in April in the context of a surging equity market, as opposed to a falling equity market,” says Smith. Sentiment, seasonality and momentum data could also be combined in various ways.
100
The firm’s most diverse programs trade north of 100 global instruments, affording good opportunities for diversification but also, given each instrument can trade multiple times per day, many chances to make successful trades.
“So, while the interrelationships of data make intuitive sense, the question has been how to tease out these relationships, or influences, in a robust, scientific way. The framework we have built is engineered to do precisely this,” continues Smith.
But this is not easy, especially in the context of financial markets. “Solving for return forecasting in financial markets is quite different to predicting what movie you may like on Netflix, or to predicting what book you may like to buy on Amazon. Financial markets are extremely noisy, and extracting a signal from this noise can be more difficult compared with other industries. It requires real market knowledge, in our view, and experience across many dimensions of investment,” says Goodman. Parameters of algorithmic techniques need to be more bespoke, and are combined in a unique way to obtain a robust forecast.
Though machine learning is now 100% behind the models, this is not a black box approach. It is supervised by human discretion. “Humans need to define inputs and parameters including risk tolerance, and assess data quality, before you flip the switch and let the machine try to find relationships. But it would take too much trial and error for humans to identify complex and conditional multi-factor and multi-data inter-relationships that a computer can identify much more quickly and accurately,” says Goodman.
While the risk budgeting and setting of the model parameters are done by the researchers (with the help of quantitative tools), 100% of the active signals and active trading in a market are a direct result of the statistical learning process, in a true hands-off fashion. Within a framework of relatively stable risk budgets per market, the statistical learning processes take over, making near real-time determinations of how much of that risk budget to utilise. The process combines raw return forecasts (generated from an ensemble of learning models in each market) with execution cost forecasts, in order to determine ex ante whether to take a trade, hold off, or perhaps break the trade into pieces.
“Signals are only executed when the system thinks there is an opportunity for profit after transaction costs. This can mean deeper, more liquid markets may have a somewhat lower hurdle and trade more frequently, while less liquid markets may more often be governed by transaction costs,” says Smith.
While far from high-frequency, Millburn is sampling data and adjusting signals and return forecasts continuously, 24 hours a day, in light of the stream of new data. Straight-through Processing (STP) allows orders to be transmitted directly to executing brokers, and helps to handle larger volumes of trades as research has found benefit in increased signal frequency.
Of course, transaction cost predictions, which also rely on a full machine learning framework, can vary significantly — depending on time of day, depth of the order book, and other conditions considered by the models. While the costs themselves are important, also key is the ability of the systems to predict them accurately.
The result is a strategy that might be best described as “nimble”, focused on a “dimmer switch” approach to staying as close to optimally positioned as possible at any particular time. When investors examine movements in positions, while holding periods still tend to be in the 20-day range on average, it is not uncommon to see positions move from long-to-short or vice versa over the course of even a few days.
One argument for traditional trend following is that momentum was always a weak signal (based on Sharpe ratios) for each individual market. The strategy, it was thought, relied mainly on diversification to generate a decent portfolio level Sharpe, and now simply needs a much wider investment universe in order to replicate historical diversification benefits and maintain its historical performance. Some CTAs have therefore stuck with traditional models but varied their market allocations or started applying trend-following to more esoteric markets.
In contrast, Millburn’s innovation has so far been principally in models rather than markets. It has not dynamically varied its market weightings so much, nor, historically, added many new markets.
“These are incredibly powerful technologies, but we are humble in how we apply them. Our goal is not to over-engineer or over-optimise. Risk management is critical for us, and we still believe one of the most reliable ways to control risk is through diversification. So putting the portfolio together starts with what we think is a thoughtful but relatively simple risk budgeting framework, trying to roughly equalise risk allocations across the traded sectors for each program and then, within each sector, across the relevant tradable markets.
Staying within the bounds of these budgets helps,” says Smith.
Because of the diversification of models and signals, and because of the “dimmer switch” approach to trying to stay optimally positioned in every market rather than making binary shifts, generally the firm will have some non-zero position in each traded market at any time.
The firm’s most diverse programs trade north of 100 global instruments, affording good opportunities for diversification but also, given each instrument can trade multiple times per day, many chances to make successful trades. While the firm has to date avoided some of the more esoteric markets like electricity and others, research is ongoing.
Explicit relative value, or “spread” trades are currently only used for commodity markets. These spread strategies, which have been reserved for the firm’s Commodity Program, are reasonably capacity-constrained and were the reason for Millburn to soft-close its highly-sought-after Commodity Program early in 2018. The Commodity Program returned more than 23% last year, after fees, due in large part to those spread strategies, although outright trading in commodities was also highly profitable. In any event, RV trading uses the same signal and execution cost engine as is used in outright markets. This means improvements made in one component, or in one sector, are generally applicable to all others.
As the firm’s strategies have evolved, so too has its approach to communicating with investors. Whereas traditional trend-followers can easily explain that they have cut and reversed in response to recent price action, it can be harder for Millburn to explain exactly what has caused positioning to change given the multiple inputs of dynamically varying weights.
Craig Gilbert, who has a mathematics background and joined Millburn from Two Sigma in 2011, just as Millburn’s research turned towards statistical learning, notes that the team has worked hard to make these strategies understandable to its investors. “With quant strategies there will always be a certain part of the process that will be opaque. But much of the process — all the elements that surround the opaque bit — can be explained. Understanding these elements — things like risk controls, operational procedures, data sourcing and cleaning, portfolio construction, signal combination — is, we think, just as important.”
To help, the firm once again leans heavily on data, producing detailed visualisations and animations to demonstrate the mechanics, with the goal of demystifying the approach. Investors can also have sight of signals in each market, and can track how these have evolved over time through the current date.
And while explaining the precise genesis of particular signals in markets can be more difficult — the “opaque” piece — the firm does have tools at its disposal to help do this on a case-by-case basis. But the magic lies partly in the mystery.
“By definition, statistical learning approaches seek to uncover patterns or tendencies in historical data more efficiently than could be done by humans. In some cases, these techniques are uncovering real, robust, statistically-significant patterns that would have been entirely impossible for a human being to uncover. So, our natural desire to explain the ‘why’ somewhat misses the point,” says Gilbert.
In addition, even with a certain amount of opacity baked-in, most investors take comfort in the process being grounded in some degree of intuition, while understanding that at some point the machine will take over and do the heavy lifting. “Our researchers tend to worry less about why we are positioned a certain way, because of the confidence they have in the process. For investors, because the inputs going into the model are curated and generally quite common sense, we think the trust in the output should also be higher compared with some other quant approaches,” says Goodman.
Financial markets are extremely noisy, and extracting a signal from this noise can be more difficult compared to other industries. It requires real market knowledge, in our view, and experience across many dimensions of investment.
Barry Goodman, co-CEO and Executive Director of Trading, Millburn
Millburn’s asset growth and strong performance clearly make it attractive as an employer and recruitment has also risen to meet the new bars. The approach to talent at Millburn cuts across all areas of the firm. “We seek smart people, of course, but also place a great deal of emphasis on the right cultural fit. From a research point-of-view, what is common today, and what differs from even ten years ago, is that every hire comes in to the modelling team with strong skills in mathematics and statistical analysis, and the ability to leverage machine learning techniques,” says Gregg Buckbinder, Millburn’s President and Chief Operating Officer.
While Millburn’s talent team is often recruiting from some of the same universities as those toured by the likes of Amazon and Google, or even some of the firm’s quant brethren, researchers at Millburn can potentially be much bigger fish in a smaller pool. “We think we offer an environment where a researcher can have somewhat more immediate and direct impact…this is not an environment where a new researcher will find themselves working on projects to eke out marginal basis points,” says Smith.
But precisely because each team member can have a major impact, hires are very carefully chosen. As Smith elaborates, “For hires in Research and Development, while we are always looking, we hire very selectively. Our preference is for adding talent with skillsets that may be orthogonal to those we already have, and/or that can be leveraged in multiple ways by the existing team. Because we operate in a no-silo environment, we think this approach makes sense.”
Millburn also invests heavily in technology. Chief Technology Officer and co-Director of Trading Manu Kambhampati, who joined the firm more than 11 years ago, seeks to keep the firm on the leading edge across all aspects of the business. Says Kambhampati, “in everything we do, we focus on building a highly-resilient, scalable, zero-downtime environment. That’s what our process demands today. Sometimes this means utilising ’off-the-shelf’ products, but other times we will build customised packages to suit our needs. Like our signalling process, everything is context-specific.”
This translates to a mix of proprietary and open source packages, many of which are highly and specifically modified.
And it has worked. “This philosophy on structuring the R&D team, combined with an efficient research and technology framework, has enabled us to support our growth while avoiding the management challenges that can come with hiring armies of researchers,” says Buckbinder.
Staff have always eaten their own cooking at Millburn. More than USD 400 million, or about 7% of the firm’s AuM, is that of current and former employees, their family members and related entities. These assets are, wherever possible, invested right alongside Millburn’s external investors, without preferential terms with the exception of lower fees, ensuring a good degree of alignment.
These co-investments are spread from the most senior staff right down to the most junior, affording each employee the opportunity to invest (company-augmented) retirement pension savings in the firm’s Diversified Program. It’s a perk that is heavily utilised, underscoring the belief that the employees have in their firm.
With steady asset growth has come the question of capacity. While the firm demonstrated with the soft closure of its Commodity Program that it is disciplined in protecting its existing investors, in the near term at least it sees no such capacity issues with its main Diversified and Multi-Markets programs. And while Millburn’s CIO Smith is relatively non-committal, the firm’s current simulations show solid results maintained at levels beyond USD 10bn, even without any allowance for future research to improve scalability. So while not unlimited, there remains significant room to support asset growth – from performance, or net inflows, or both.
The firm’s UCITS strategy, launched on the MLIS (Merrill Lynch Investment Solutions) platform that has now been acquired by Generali Investments, trades virtually pari passu with its Diversified Program, providing exposure to each of the four sectors of equities, currencies, fixed income and commodities. Given the liquid and diversified nature of the strategy, among other things, only minor modifications were needed to be made to accommodate UCITS regulations, meaning tracking error to the flagship is practically nil.
While the firm has always served a wide array of investors, institutions have started to invest more heavily, including some of the largest and most sophisticated pension funds and insurance companies. “These investors have been attracted to our returns, of course, but also to our stability and longevity, and to the firm’s ability to adapt to stay ahead of the competition,” says Goodman.
And while Millburn’s US investor base continues to account for the lion’s share of its AuM, interest (and assets) have also been growing from Europe, Asia and Australia as the firm’s brand has become more visible, aided by the tailwind of strong performance. “On a global scale, we are seeing investors beginning to shift their views, looking to either diversify or, increasingly, entirely re-think their approach to the CTA portion of their portfolio. We think we provide a unique option,” says Gilbert.
There’s a certain electricity at Millburn today, as the team moves into bright new offices in Midtown Manhattan, at 55 West 46th Street, just off Avenue of the Americas and four blocks north of Bryant Park.
From a research perspective, the framework continues to provide a multitude of opportunities and research is progressing along the three dimensions of additional factors, additional statistical learning techniques, and improved execution.
From a new markets point of view, the firm is among a growing number of CTAs to have entered into joint ventures with local Chinese quant managers to help expand their futures offering for mainland RMB investors. This strategy started with the simple application of Millburn’s international models to the local Chinese futures markets. Performance has been strong and regulatory changes mean this program soon may be accessible by non-Chinese investors.
Moreover, some of the local China futures markets themselves are being made available to international investors—for example Iron Ore (traded on the Dalian Commodity Exchange) and Crude Oil futures (traded on the Shanghai International Energy Exchange) were both opened to foreigners in 2018. Millburn expects this trend to continue and its experience trading these markets locally will, they believe, give them a head-start on trading of these instruments in their international programs.
And on the horizon is the application of these models to tactical equity security trading, in the form of a family of tactical sector-based approaches and, ultimately, single-name equities — for the moment just a glimmer in the eye but a promising one. Proof-of-concept has been established and discussions with strategic investors are already underway. Millburn’s unique heritage, culture and process could provide fresh and differentiated perspectives for the quantitative trading of equity and credit.
“The firm has come a long way, but we feel like we are only in the early innings,” says Goodman.