A 2018 white paper “Ensemble Active Management,” boldly promised in its subtitle “The Next Evolution in Investment Management.” It was produced by the EAM Research Consortium, although there’s no information online about that group and the web address for it (found elsewhere) is not in service.
Most of what has been written about the concept since then has come from one of that paper’s co-authors, Alexey Panchekha, via the news and research page of Turing Technology (he is its president) and postings for CFA Institute’s Enterprising Investor.
Ensemble models
Booz Allen Hamilton’s The Field Guide to Data Science provides an accessible introduction to a wide range of analytical methods. Here are some excerpts from its summary of ensemble models:
An ensemble takes the predictions of many individual models and combines them to make a single prediction.
Ensembles overcome individual weaknesses to make predictions with more accuracy than their constituent models. If one model over fits the data, it is balanced by a different model that under fits the data. If one subset is skewed by outlier values, another subset is included without them. If one method is unstable to noisy inputs, it is bolstered by another method that is more robust.
Without delving deeper into the details, let’s shift back to the strategy at hand.
Ensemble active management
The premise of “ensemble active management” (EAM) stems from research showing that the highest conviction bets (relative to index weights) drive the alpha of an equity portfolio. The remainder of the holdings (often aggregating to a majority of the fund) serve as a “beta anchor.”
In EAM, a portfolio is formed from a group of underlying funds using these guidelines, set out in one of the Enterprising Investor articles:
All of the managers must share the same investment objective, such as beating a standard index like the S&P 500.
Most of the fund managers need to demonstrate better-than-random stock-selection skill for at least their highest conviction picks.
Ideally, there should be at least 10 underlying funds. [This sentence was hyperlinked to an academic paper by Eugene Pinksy.]
The investment processes must be independent. This is critical. Diversification at the predictive engine level is how Ensemble Methods solve the Bias–Variance Conflict.
From those overweight positions, EAM determines the “predictive engine” of each fund. That analysis is used to drive the algorithm by which a fifty-stock “best-idea-centric” portfolio is created based upon the preferences displayed by the underlying managers.
(To support the creation and ongoing adjustment of the portfolio, Turing replicates the underlying funds during periods between releases of portfolio data, using changes in the values of the funds and their component securities.)
Breaking it down
On a conceptual level, EAM fits with both the ever-broadening use of ensemble methods across a variety of disciplines, as well as qualitative work on the importance of diverse inputs to decision making processes.
To further examine the ideas behind it (or to potentially extend them into new applications, or to consider investing in an EAM portfolio) would require more information beyond that which is publicly available. What follows are some suggested angles of approach.
Another Turing document lists this as a first step in the process (emphasis from the original):
An institutional investor or investment manager selects 12‐15 proven investment managers, using their own insights into manager selection. The underlying strategies should reflect the same investment mandate (e.g., large core), but should be diversified regarding firms and approach.
There’s that word “proven,” tossed around throughout the ecosystem but rarely defined. What is its meaning here? As stated before, what’s important is “better-than-random stock-selection skill.” Over what period of time and in what ways is that determined? Given that performance analyses trip up more allocation choices than anything else, why would having “an institutional investor or investment manager . . . using their own insights into manager selection” be the best place to start?
The same problem besets the choice of “independent” investment processes, the importance of which is stressed throughout the various documents. Every manager will tell you that they are different, but many are superficially different at best. What are the characteristics that determine the degree of difference from one to another, and are those determined qualitatively or quantitatively?
Beyond this are the details of implementation from day to day, bringing up issues of how proprietary the methods are and how transparent everything will be to outside investors (which will probably vary from provider to provider). At a minimum, there should be more clarity regarding the points at which and the ways in which ensemble models are used.
Indications of performance
The initial paper was based upon research analyzing a small number of funds (37) to randomly create 30,000 clusters of ten funds each. The backtested performance numbers were quite remarkable, showing strong performance against both the benchmark (the S&P 500) and category (Large Cap Blend) of those funds — in terms of the frequency of outperformance over rolling periods and the amount of excess returns.
The first live portfolio started in late 2018 and Turing reports that currently there are 78 strategies in “the industry,” across a number of different kinds of vehicles, covering every part of the style box. The early numbers are compelling as well, although greater clarity is needed regarding the different kinds of portfolio and how the information is combined — and net-of-fees numbers should be reported in a straightforward manner.
But given that the first of these portfolios is reaching its three-year anniversary (and therefore qualifies for the artificial and ineffective age requirement imposed by many investors), you can bet that the marketing engines are revving up right now.
Considerations for asset managers
It is ironic that many of the materials regarding EAM dwell (accurately) on the inability of active managers to beat their benchmarks, while relying on those same managers for the raw material that fuels EAM. That highlights the portfolio construction practices that proponents of the approach think is at the heart of the active management underperformance problem.
Many managers offer high conviction portfolios and, for some, that is all they offer. If such portfolios come out of a unified investment process, are they likely to outperform or underperform something that taps a more diverse set of investment approaches, as EAM promises to do?
“Platform” firms, where separate pods operate independently, fit one part of the EAM formula; would an allocation algorithm work better in terms of allocating money among the pods than qualitative decisions do?
How about those picking subadvisors for multi-manager vehicles? Would they be better off adopting the EAM methodology? It seems made to order for them, especially since many subadvised vehicles have struggled to add value.
Or imagine a situation where a firm decides to seed a number of in-house strategies, not with the goal of marketing them or aggregating them into a fund, but to use as the information engines of the algorithm. (You’d have to get the incentives and the sizing of those pools right, since you’d want the individuals doing the investing to be appropriately aligned, even though they are ultimately providing the signals for a different purpose.) Maybe this kind of approach will spawn new structures; you can see shades of it in some current ones.
Another example: An EAM piece suggested that an asset owner could pay managers “for their list of holdings and weights, apply ensemble methods to the combined list of securities from [a group of] active managers, and trade the resulting EAM portfolio themselves.” Many asset managers sell model portfolios already; to what extent are they being used in this way? What are the implications going forward? Will we see more asset-light firms? Will we still call them “managers”? What happens if you remove the asset gathering function from the business that produces the ideas?
For large, multiproduct organizations, other possibilities exist. The components of multiple in-house investment processes could provide a more granular level of signals to feed an ensemble approach. The kind of component reporting that would be required is atypical at most firms but provides possibilities for the future.
Considerations for others
For those analyzing, hiring, and firing asset managers, the principles regarding alpha engines and beta anchors — foundational for EAM — offer a template to review your investment beliefs and potentially redesign the evaluation of managers and the execution of your strategy.
Someday, you will likely be considering whether to invest in an EAM-like strategy, but the tenets that underlie it also can be addressed by restructuring selection and portfolio construction practices that result in beta (really, beta minus fees) becoming dominant when active management is the stated goal.
In general, the materials released regarding EAM have been promotional in nature, par for the course in the business and not unexpected given the apparent performance in backtests and early live results. The challenge will be to dig into the details, poke holes in that narrative, and create a map of likely risks across a broader sweep of time.
Research and development
Some of the above might be interpreted as saying that the emergence of EAM as a force is a fait accompli. Not in the least.
But it is a good means by which to get at some larger points.
Every investment organization, large or small, ought to have an R&D function dedicated to continuous improvement, focused on new methods internally and the emerging environment externally. Choices have to be made, especially when resources are tight, but not evaluating ideas outside your current state is untenable as a long-term strategy.
In this situation, there are general and specific questions to ask, no matter if you work for an asset manager or any other kind of organization. Where are you at in your thinking about the use of machine learning and other techniques to change your processes in the future? What new internal structures for investment organizations (and external competition among them) might result? You may be very far down that road or not even started; how you address the idea of EAM (or any other possibility) will depend on that backdrop.
Then, take the ensemble idea itself (not EAM) and consider how it can be used within your organization and the strategies that you employ. It is, on its own, a very powerful concept to apply within the investment realm.
Of course, there’s the notion of EAM itself. Bob Tull, a member of Turing’s advisory board asked some questions:
What if Ensemble Active Management is shown to be a superior means of delivering active management?
What if EAM portfolios are strong enough to re‐open the active vs. passive debate?
What if EAM‐powered active ETFs are readily available to the investing public?
If EAM gets traction in the industry and eventually becomes hard to ignore (a progression we’ve seen many times before), what are the implications for your organization?
The industry picture
Intense exploration is occurring in the industry regarding ensemble methods, machine learning, and other data science techniques. A number of strategies utilizing them are already available, but it’s safe to predict that many more offerings are on their way. We will see applications across asset classes and types of organizations, addressing a wide variety of problems.
Some examples can be observed in working papers and in published articles. For instance, see “Machine Learning for Stock Selection,” from a 2019 edition of the Financial Analysts Journal. In addition, there are books on the topics, such as Marcos López de Prado’s Machine Learning for Asset Managers (Cambridge University Press, 2020).
Usually, when alpha is to be found via such methods, the practitioners who discover it are interested in keeping it to themselves. Therefore, the state of practice across the industry is a mosaic at best and an evolving one at that. For example, while you don’t hear ensemble methods referenced as part of the narratives of traditional money managers now, that will be coming. Below the radar, though, those techniques are already in production at firms that don’t broadcast their tactics.
So, there may be other EAM-like concepts of note that haven’t been publicized. Interestingly, the approach comports with the theory behind TOPS, the “Trade-Optimised Portfolio System” of Marshall Wace, which harvests high-conviction ideas in a different way. TOPS has been used by the firm for two decades and was referenced in an earlier posting on The Investment Ecosystem.
As you can see, there are many potential layers of inquiry and analysis related to EAM. It may or may not become the next big thing, but examining it now and watching its progress can inform your approach to your current practices — and your plan for investigating the broader spectrum of possibilities to come.
Published: November 22, 2021
To comment, please send an email to editor@investmentecosystem.com. Comments are for the editor and are not viewable by readers.