Optimal design of investment committees

Abstract

Investment committees are widespread across asset management firms, private and public institutional investors or family offices. Poorly designed boards can potentially destroy substantial value in the investment management industry, yet little research has been undertaken on their optimal design. From my 30-year experience as an investor, CIO for various firms and academic researcher, I believe that typical investment committees come with unaddressed challenges. Using qualitative group discussions to create a consensus view results in biases (group shift bias), incentive problems (free-rider) and aggregation problems. How can we ensure that all investment views enter the investment committee equally? In my opinion, we can learn from evidence gathered in social psychology how committees can make better investment decisions. I suggest creating an algorithmic consensus by averaging anonymous member portfolios instead of informal qualitative discussions towards the end of an investment committee meeting.

Introduction

Investment committees are widespread across asset management firms, private and public institutional investors, or family offices. Poorly designed boards can potentially destroy substantial value in the investment management industry, yet little research has been undertaken on their optimal design. From my 30-year experience as an investor, CIO for various firms and academic researcher, I believe that typical investment committees come with many challenges. This contribution aims to address these challenges and suggest a potential solution.

Industry typical investment committees use qualitative group discussions to create a consensus view. Ideally, the committee enriches its information by benefitting from different perspectives, skills, and experiences. In reality, I believe that investment committees suffer from biases (group shift bias), incentive (free-rider) and aggregation problems (does the final portfolio equal the group consensus).

Group polarization is a phenomenon that occurs when group members become more extreme in their assessments after interacting with each other. In investment teams, group polarization can lead to suboptimal investment decisions, as group members may become overly confident in their views and ignore alternative perspectives or information. Aggregation problems occur when we can not ensure that all investment views enter the committee equally. Incentive problems arise when the individual contribution is difficult to evaluate. Who was responsible for swinging the investment committee to take an equity overweight in March 2020?

In my opinion, we can learn from evidence gathered in social psychology how committees can make better investment decisions. I conjecture that these challenges will likely become smaller once an investment committee moves towards creating an algorithmic consensus by averaging anonymous member portfolios instead of informal qualitative discussions towards the end of an investment committee meeting.

Decisions by experts are noisy.¹ Investment professionals make no difference. Portfolio managers (PMs) within the same organization will arrive at different investment views and positions. The objective of the modern chief investment officer (CIO) is to ensure the organization makes better decisions by removing as much noise as possible and aggregating alpha information effectively between all decision-makers. I argue in “The case for perfectly working investment committees” section that asset management firms with investment skills will create investment committees (ICs) to aggregate views across portfolio managers (PMs). Pooling views lets ICs take less extreme risks. The benefits of diversifying individual forecast errors outweigh the disadvantages of collectively underperforming across all mandates if the IC makes the wrong decision. Strong firms will choose centralization. Instead, weak firms (defined as asset managers with little skill) are not incentivized to aggregate opinions but rather to let individual PMs invest according to their outlook. They place a larger value on diversifying their income stream than diversifying across investment views and are likely to choose individualization.

Besides collectively making better decisions (diversification of forecast errors), ICs have many advantages. They allow an organization to become independent from "star managers" that might expropriate large value from the firm when threatening to leave, as Kovaleski (2000) mentioned. Teams also perform better. Empirical research by Bliss et al. (2008) finds team-managed funds to attract more significant flows and take less extreme (factor-) risks. Bär et al. (2011) confirm these results.

ICs, in theory, create diversification of opinion. Much of this diversification will not materialize, as I explain in “The challenges to creating perfect investment committees” section. Decision-making by finding a consensus at the end of a committee meeting suffers from coordination costs, group shift bias, dilution of incentives and an increased disposition effect.² To mitigate these weaknesses, I suggest in “A better decision rule for investment committees” section that all IC members anonymously provide their own long/short portfolio (vector) of investment positions. All vectors must carry the same active risk to ensure equal influence. Individual portfolios are aggregated via averaging and scaled to a desired risk target. Anonymity will reduce group shift bias while constructing (compensation-relevant) individual portfolios guarantee diversification, incentivization and a reduced disposition effect. “Conclusion” section concludes.

The case for perfectly working investment committees

The centralization of investment decisions in investment committees results in better portfolios (diversification of individual forecast errors of team members) at the expense of performance dispersion across mandates. Centrally managed asset management mandates tend to under or outperform collectively. We illustrate this point and its implications with an example.

We start with a decentralized firm (or book of business within a multi-asset firm) with ten PMs and an information ratio of 0.2. Each PM manages a unique client portfolio. All PMs display skill (measured by nonnegative information ratios, IR) and receive independent rewards (bonuses) based on performance. There are several consequences of this setup. First, portfolio managers have little incentive to share their best idea. After all, they receive rewards for their performance relative to other PMs. Second, each client participates in a portfolio manager lottery, i.e. the received performance will depend on the portfolio manager allocated to the particular client and portfolio managers differ in skills, styles and luck. The likelihood of outperforming in any given year is a function of skill. An information ratio of 0.2 translates into an outperformance likelihood of 57.93% for every single mandate.³ It is almost impossible to end up with a year in which all mandates underperform $\left( {1 - 0.5793} \right)^{10} = 0.02\%$ in Table 1. Most of the time (in 84.65% of all cases), three to seven funds outperform their benchmarks.

Table 1 Centralized versus decentralized asset management—organization with skill

Let us instead assume the same firm above moves from decentralization to centralization of decisions. For this purpose, it builds a perfectly working investment committee. By this, I mean an investment committee that does not lose individual information in the aggregation process. All team members remain fully incentivized even though they now share a joint responsibility, and their contributions are challenging to evaluate (and their effort to arrive at investment views is not observed). Perfectly working ICs will leave clients better off, as sharing information on active portfolios must lead to a better (higher information ratio) portfolio.

Uniting ten uncorrelated PMs with an IR of 0.2 each in an IC results in an IR of 0.63 for the IC and a corresponding outperformance probability of 73.65% (in 3 out of 4 years, all funds outperform at the same time).⁴ Centralization leads to a much higher percentage of satisfied clients (7.36 out of 10 on average) than decentralization (5.79 out of 10). However, in about every fourth year (26.35%), all active portfolios will underperform collectively. The risk of this happening is minimal for the decentralized firm (0.02% or twice in 10,000 years).

Centralized investment processes have a higher probability of disaster years (years where all funds underperform simultaneously). While from a legal/ethical perspective, all clients within a given product should be treated equally, the trade-off between higher average performance and a higher probability of disaster years exists. Table 2 shows that centralization is disastrous for firms with no skills. The expected value of outperforming mandates is the same under decentralization, but the volatility around this average will be challenging to tolerate.

Table 2 Centralized versus decentralized asset management—organization without skill

Well-run firms recognize that diversity of opinion is desirable, but dispersion in portfolio outcomes is not. This insight rests on four observations. First, professional clients do not want to participate in a PM lottery but instead want to invest in the product with the highest likelihood of outperformance. Second, the normative theory on portfolio choice unequivocally recommends the aggregation of views. Third, the empirical literature shows that team-managed funds outperform individual PM-managed funds. Forth, portfolio manager lotteries are inconsistent with fiduciary asset management requiring all clients to be treated equally (access to the same alpha information).⁵

Bartolomeo (2018) summarizes the historical development of centralized portfolio management in a paper and a series of presentations. According to Bartolomeo, the idea started with Bill Sharpe about using reverse optimization to estimate security alphas from paper portfolios from independent managers and then form a consensus alpha to run a portfolio. He emphasizes that "since then, several firms, including Vanguard and Parametric Asset Management, have successfully used Centralized Portfolio Management with significant AUM".

Centralized portfolio management requires portfolio factories which Scherer (2022) defines as "… the automated implementation of an active model portfolio on a wide range of client portfolios that differ across benchmarks, regulatory constraints or client-specific constraints". Rather than using implied returns, Scherer shows that the same result can be achieved using well-known tracking error minimization. In the first step, the active model portfolio is added to a given benchmark portfolio to define a new target portfolio. Under client-specific constraints, the second step minimizes the tracking error between the target and the long-only portfolios. This procedure is equivalent to first backing out implied returns and then using these returns in active portfolio optimization.

The challenges to creating perfect investment committees

Perfectly working ICs are a Nirwhana solution. Table 3 compares the main differences between traditional as well as newly suggested investment committees. The reality is much more difficult. Here is why. In an IC, the CIO typically chairs a monthly discussion between senior portfolio managers and economists on how to invest, given the current investment opportunity set. At the end of each committee meeting, the CIO aggregates the investment opinions by finding a consensus among IC members. While the previous section outlined the benefits of "diversification of opinion", investment committees also have potential disadvantages and side effects.

Group Shift Bias. Kahnemann et al. (2021, chapters 8 and 15) describe a series of controlled experiments where committees amplified noise by making more extreme decisions than the average opinion across team members would have implied. By noise, they mean that similar groups would make different decisions, or the same group would make other decisions if only minor changes to how information is presented were undertaken. The authors argue that information sequencing (informational cascades) and group polarization can shift teams towards extreme decisions. Information cascades describe a situation where the sequence in which information enters the decision-making process matters. Suppose the CIO comments first and enthusiastically makes his investment case. In that case, this might swing other team members. Either out of respect for the CIO or because they see their impact on the committee outcome as minor. Instead, members might want to manage their career risk by sucking up to the CIO. This applies particularly to junior IC members and compresses the adequate diversification of opinions. Group polarization describes a situation where team members end up in more extreme positions than before they entered the team discussion. Suppose most IC members think that inflation is likely to rise further. In that case, the remaining holdouts believing in falling inflation might move to the rising inflation camp to safeguard their reputation within the group. While there is ample experimental evidence for a group shift bias by Davis and Hinsz (1982), Mulvey and Klein (1998), and Cooper and Kagel (2004), few empirical studies exist. To our knowledge, the only work in an asset management context is the paper by Bär et al. (2011). The authors reject the group shift bias as they find that funds managed by investment teams take less extreme decisions than funds managed by single managers. Teams display less extreme factor exposures and less extreme performance. Experimental and empirical evidence sometimes disagree. The author's experience with investment committees over the last 30 years confirms the noise from information sequencing and group polarization.

Disposition Effect. We know from Odean (1998) and related research in behavioural finance that individual investors are reluctant to realize their losses but, at the same time, sell winners too early. How does the disposition effect affect decisions by investment teams? Cici (2012) finds that team-managed US mutual funds display more significant disposition effects than funds managed by individual managers, as in Rau (2015). The authors confirm that teams show more significant disposition effects than individuals in an experimental setting. The most likely cause for this effect in an asset management setting is the prominent role and heightened attention (inside and outside the organization) towards the IC decisions. In my experience, a polarized team tries to manage its reputation. It thus tries to avoid admitting to mistakes by selling a loss-making position. ICs can become stale and wait until their decisions are proven right.

Incentives. While individual performance is relevant for bonus payments, IC performance is often not. Even if it was relevant, the impact of a team member on IC performance in a discussion is small, and the efforts in forming an educated investment opinion are unobserved (free rider problem). Not all team members are usually under the impression that their contribution enter a discussion with the same weight as the other team members.

Coordination Costs and Number of IC Members. Coordination costs increase nonlinearly with team size. For $n$ team members, there are $\tfrac{{n\left( {n - 1} \right)}}{2}$ mutual discussions to find a consensus portfolio. This is not realistic for larger committees. Rising complexity will also limit the size of ICs as the marginal costs from adding a new team member will rise exponentially.

In my experience, some of these issues are partially addressed with conventional means.

such as

• Encouraging diversity and selecting equally strong-minded team members (to ensure diversification and reduce group shift bias),

• The invitation of varying outside speakers (to unanchor group think),

• The creation of a team red and blue or a devil's advocate (to argue different investment cases),

• The analysis of past decisions (to uncover past biases and misjudgements),

• Standardized and comparable meeting material,

• The random assignment of speaker slots (to avoid the same person repeatedly framing the discussion).

However, none of these measures ensures that team members make independent and fully incentivized decisions or that all team members are equally heard and positions are unmistakably aggregated.

A better decision rule for investment committees

How can the CIO address the shortcomings of traditional investment committees outlined in the previous section? If investment committees do not work well, the case for centralization will become considerably weaker. The easiest way is to hire better IC members. However, this might still prove costly if we do not simultaneously remove conventional investment committees' deficiencies.

I suggest two modifications to identify the common consensus within an IC better. First, each team member anonymously provides a long/short vector of active positions. Rescale each vector to carry an identical tracking error to make each vector equally important. Second, replace consensus building using group discussions (votes) by averaging equally risky long/short vectors (portfolios scaled to carry the same tracking error). How do these changes impact investment committee decisions? Averaging comes with a mathematical guarantee to reduce noise in individual decisions⁶, while the provision of portfolios by each team member maintains the maximum incentive to perform well. Clearly defined decision rules lead to better performance (even for amateur teams), as outlined by Lovallo and Sibony (2010) or Tetlock and Gardener (2015). The anonymity in decision-making (only lifted to the CIO at year-end) ensures diversification between PMs (no alignment of positions). It also facilitates the absence of a group shift bias (no polarization and no reputational hedges) and removes the free rider problem. (Each team member has the maximum incentive to perform well.)

Table 3 Decision rules for investment committees

In my experience as CIO at previous firms, the portfolios submitted anonymously by IC members show a much more significant position variation than the communicated opinion difference in IC meetings. The average pairwise correlation between team members is typically low, and it was always difficult for me to imagine that I, as a CIO, could have unveiled this difference in opinion via a group discussion during or towards the end of an IC meeting. Since each IC member manages their long/short portfolio, a group disposition effect is much smaller.

Conclusion

Skilful investment firms will prefer perfectly working investment committees over individualized portfolio management. However, traditional investment committees are riddled with challenges. This results in biases (group shift bias), incentive problems (free rider), and aggregation problems (how to ensure that all member views enter the IC portfolio equally). I argue that these challenges will likely become considerably smaller once an investment committee moves towards creating an algorithmic consensus by averaging anonymous member portfolios instead of relying on qualitative group discussions. While investment committees based on these principles always performed well in my previous CIO positions, communication is one weakness in this design choice. Finding a coherent ex-post narrative that builds on a consistent top-down view is problematic because consistency across positions is neither enforced nor desired.

Appendix

Probability to outperform. Assume a decentralized investment firm (or book of business) with $i = 1,...,n$ portfolio managers; described by their skills (as measured by their respective information ratios: $IR_{i} > 0$). Clients of decentralized firms receive an information ratio $IR_{i}$ from their portfolio managers. Alternatively, clients of centralized firms get access to the joint information ratio of all portfolio managers. An investment committee combines and centralizes all manager information in a model portfolio. Assuming that the performance of all portfolio managers is uncorrelated, the centralized firm can produce a higher information ratio than the best manager can achieve on its own:⁷

$$IR^{c} = \left( {\sum\limits_{i = 1}^{n} {IR_{i}^{2} } } \right)^{{\tfrac{1}{2}}} \ge \max \left( {IR_{i} } \right)$$

However, the expected higher performance comes at the expense of much more significant P&L volatility. A fully centralized firm will offer identical active portfolios to all clients. Consequently, all client portfolios will underperform in a lousy year and outperform in a good year. Not so for the decentralized firm. As all mandates are run separately by uncorrelated managers, some mandates will exceed while others might not. Instead, the likelihood that all portfolio managers underperform for a decentralized firm in a given year is small.

For a given information ratio, we can compute the hit ratio $h$ (the probability that a single manager will outperform) directly as⁸

$$h = prob\left( {\alpha \ge 0} \right) = \tfrac{1}{{\sqrt {2\pi } }}\int\limits_{ - IR}^{\infty } {e^{{ - \tfrac{1}{2}x^{2} }} dx} .$$

For example, an information ratio of 0.2 translates into a hit ratio of⁹

$$h = \tfrac{1}{{\sqrt {2\pi } }}\int\limits_{ - 0.2}^{\infty } {e^{{ - \tfrac{1}{2}x^{2} }} dx} = 57.93\% .$$

Once we know the hit ratio, we can compute the probability of $m$ (out of $n$ funds) outperforming their benchmarks from

$$\left( {\begin{array}{*{20}c} n \\ m \\ \end{array} } \right)h^{m} \left( {1 - h} \right)^{n - m}$$

For example, for $n = 10$ independent portfolio managers with an information ratio of 0.2, the likelihood that all ten portfolio managers underperform in a given year amounts to

$$\left( {\begin{array}{*{20}c} {10} \\ 0 \\ \end{array} } \right)0.5793^{0} \left( {1 - 0.5793} \right)^{10} = \left( {1 - 0.5793} \right)^{10} = 0.02\% .$$

Scaling. Assume we have $n$ position vectors with an average correlation of $\rho$ each. What volatility $\sigma$ do we need to scale each vector $w_{i}$ scale up to achieve a target volatility of $\overline{\sigma }$ on average? Recall that the volatility of an equally weighted (same skill, i.e. same $IR$) is given by $\frac{{\sigma^{2} }}{n} + \left( {1 - \tfrac{1}{n}} \right)\rho \sigma^{2}$. The required volatility thus amounts to

$$\sigma = \frac{{\overline{\sigma }}}{{\left( {\tfrac{1}{n} + \left( {1 - \tfrac{1}{n}} \right)\rho } \right)^{{\tfrac{1}{2}}} }}.$$

For example, for $n = 10$ uncorrelated IC team members targeting a volatility of 1% for the collective IC portfolio, they need to target a volatility of

$$\sigma = \frac{0.01}{{\left( {\tfrac{1}{10}} \right)^{{\tfrac{1}{2}}} }} = 3.16\% .$$

To put it differently: If each IC member provides a vector with 1% target volatility, the aggregation process would need to leverage this portfolio by multiplying each position with 3.16.