헤지펀드 수익률과 총요소생산성

Hedge Fund Returns and Total Factor Productivity

Article information

Korean J Financ Stud. 2024;53(2):309-331
Publication date (electronic) : 2024 April 30
doi : https://doi.org/10.26845/KJFS.2024.04.53.2.309
최수정,
Associate Professor, Soongsil University
(숭실대학교 부교수)
* 연락담당 저자, 주소: 서울시 동작구 상도로 369 숭실대학교 경영학부 07027, E-mail: sjchoi1@gmail.com. Tel: 02-820-0579.
* Corresponding Author. Address: School of Business Administration, Soongsil University, 369 Sangdo-Ro, Dongjak-Gu, Seoul, Korea 07027; Email: sjchoi1@gmail.com; Tel: +82-2-820-0579.
Received 2023 December 30; Revised 2024 February 18; Accepted 2024 February 25.

Abstract

본 연구는 구조적 벡터자기회귀 모형(SVAR)과 벡터오차수정 모형(VECM)을 이용하여 헤지펀드들의 투자 행태가 거시 경제적 총요소생산성의 변화를 예측할 수 있는지 살펴보았다. 즉, 주식 시장의 정보보유 거래자(informed traders)로써 헤지펀드는 다른 투자자들보다 시장 분석에 있어서 더 우월한 능력을 가지고 있다고 알려져 있으며, 이에 헤지펀드들이 거시 경제적 생산성의 향상과 관련된 뉴스 충격을 선제적으로 포착하고 대응할 수 있는지 확인하였다. John Fernald(2014)가 산출한 미국의 분기별 TFP 자료와 크레딧 스위스/트레몬트 데이터베이스에서 제공하는 헤지펀드 지수(HFI)의 수익률 자료를 활용하여 1994년 1분기부터 2023년 2분기까지의 표본 기간 동안 두 내생 변수 간의 시계열적 연관성을 검토하였다. 분석기간 두 변수의 동 분기 상관계수는 0.9791로 나타나 두 변수의 움직임에 유사성이 매우 높다는 것이 확인되었다. 그랜저 인과관계 검정 결과, “ΔLn(HFI)가 ΔTFP를 그랜저 인과하지 않는다”라는 가설이 기각되었으며 이는 헤지펀드들의 행태를 보여주는 헤지펀드 수익률 정보가 총요소생산성의 변화를 예측하는 데 유용하다는 것을 암시한다. 마지막으로, VECM 모델을 사용한 예측오차 분산분해에 따르면 ΔLn(HFI)에 대한 충격이 20분기 이후에도 ΔTFP변동의 65% 이상을 설명하는 것으로 나타났다.

Trans Abstract

This study explores whether hedge funds’ investment behavior can predict variations in productivity levels using a structural vector autoregressive model (SVAR) and a vector error correction model (VECM). As informed traders in the stock market with superior skills, hedge funds may quickly capture news shocks regarding future production growth in advance. With quarterly series of TFP provided by John Fernald (2014) and hedge fund index (HFI) returns obtained from the Credit Suisse/Tremont database, I find a contemporaneous correlation coefficient of 0.9791 between two endogenous variables over the sample period from 1Q:1994 to 2Q:2023, indicating a high degree of similarity in their movements. A Granger Causality Test rejects the hypothesis, “ΔLn(HFI)does not Granger Cause ΔTFP”, suggesting that the information inferred from the hedge fund index are valuable in predicting future economic productivity. Finally, the forecast error variance decompositions using the VECM model indicate that over 65% of the variation in ΔTFP even after 20 quarters can be attributed to a shock to the ΔLn(HFI).

Keywords:

1. Introduction

According to Fama (1990), stock market return variations are fairly correlated with the growth rate of future production since both may reflect information about shocks to expected cash flows. Beaudry and Portier (2006) argue that stock price movements incorporate the expectations of economic agents regarding future economic conditions. Given that changes in agents’ expectations are related to future economic conditions, I posit that news shocks inferred from current stock price movements may explain future business cycle fluctuations. Therefore, I expect that investors participating in the stock market will move first, and then one can predict future economic conditions based on the behavior of the stock market participants.

To identify the relationship between stock market movements and future economic conditions, this study examines the cross-impacts between hedge fund index returns indicating the performance of hedge funds in aggregate and macroeconomic productivity using a structural vector autoregressive (SVAR) model and a vector error correction model (VECM). I assume that the investment behavior of hedge funds, as informed traders in the stock market, captures news shocks about future production growth more quickly than the movement of the total stock market index, such as the S&P 500 index examined in Beaudry and Portier (2006).

The movement of the total stock market index reflects the behavior of all types of investors, whether informed or not, in response to material information. If the stock market exhibits some degree of informational inefficiency, then the behavior of uninformed investors (i.e., noise traders) may also contribute to delayed stock price responses. Therefore, the total stock market index, which reflects the expectations of all types of investors, is less appropriate than the hedge fund index for capturing news about future economic conditions. In other words, if hedge fund managers are well-informed and have superior predictive skills, then the hedge fund index (i.e., aggregating the behaviors of informed traders) can be a more effective proxy for capturing agents’ expectations regarding future production growth. Recent studies examining hedge fund managers’ performance demonstrate their superior ability to identify mispricing in the stock market and gain an informational advantage in trading using various sources (Agarwal et al., 2009; Ben-David et al., 2012; Gao and Huang, 2016; Huang and Jain, 2024).

Following Beaudry and Portier (2006), I test the bi-variate aspects between the aggregate hedge fund index returns and total factor productivity (TFP). TFP is the part of output that is unexplained by inputs for production, such as labor and capital. TFP represents technological efficiency and explains economic growth (Easterly and Levine, 2001). Many studies empirically estimate country- or industry-level TFPs, but only annual frequency measures have been proposed because of data limitations. However, Fernald (2014) provides historical quarterly series of TFP based on a growth-accounting database for the U.S. business sector. This study employs the Fernald’s quarterly TFP series to measure productivity.

Using quarterly series of TFP and hedge fund index returns, this study examines the macroeconomic dynamics between hedge funds’ behaviors and economic productivity. The results are as follows. First, the contemporaneous correlation coefficient between the two variables is 0.9791, meaning that their movements are highly similar over the sample period. Note that the correlation coefficient between TFP and a S&P500 Index is 0.6856.1) Second, a Granger Causality Test rejects the hypothesis, ΔLn(HFI)does not Granger Cause ΔTFP, implying that the information inferred from the behaviors of hedge funds seems to be useful to predict future economic productivity. Finally, the forecast error variance decompositions with the VECM model indicate that more than 65% of the variation in ΔTFP after 20 quarters can be attributed to a shock to the ΔLn(HFI).

2. Model Specification

2.1 Structural Vector Autoregressive Model (SVAR) and Impulse Responses

Sims (1980) introduced a new class of econometric models, SVAR, in which identification focuses on the errors of the system rather than identifying the autoregressive coefficients (Lutkepohl and Kraetzig, 2004). The innovations of the original VAR model are orthogonalized with a Cholesky decomposition of the covariance matrix; thus, a recursive structure is imposed on the instantaneous relationships between the variables. The drawback of applying the Cholesky decomposition to obtain impulse responses is that choosing the ordering of variables may produce different shocks. Thus, I need to check the robustness of the impulse responses by mixing the ordering of the main variables unless any theory supports the specific recursive structure.

SVAR models are similar to the VAR model but have more constraints to identify parameters. For example, the identification of shocks in an SVAR is based on economic theory, which suggests that the effects of some shocks are zero in the long- or short-run. In other words, I need to impose certain restrictions on the model parameters to identify the SVAR model. In the model of Sims (1980), the triangular (or recursive) identification scheme implies that an additional shock to the second variable does not affect the first (the most exogeneous) variable in the same period. The SVAR model is as follows:

(1)Ayt=A1* yt1+A2* yt2++Ap* ytp+C*xt+Bεt

where yt is a (K×1) vector of the endogenous variables, xt contains exogeneous variables, and the underlying structural shocks are εt ~ (0, IK). Matrix A specifies the instantaneous relations among the variables in yt, then the reduced-form disturbances are ut=A-1t. Using a Cholesky decomposition of the covariance matrix Σu and B to be a lower triangular matrix Σu = BB’’, I obtain the process yt = Φout1ut−1 +… = Ψoεt1εt−1 +…,, where Ψi = ΦiB. I employ a long-run restriction on the SVAR following Blanchard and Quah (1989) and Beaudry and Portier (2006). Blanchard and Quah (1989) identify supply (demand) shocks with persistent (transitory) effects on output. Then, the (1,2)-element of the long-run impact matrix Ψ=i=0Ψi is equal to zero.

Beaudry and Portier (2006) also adopt the VECM with the long-run restriction suggested by Blanchard and Quah (1989) for identification with two differenced variables: stock prices and TFP. Following Beaudry and Portier (2006), the estimation strategies of this study are in two ways, SVAR and VECM. First, the cointegration test with the Johansen method (allowing a linear trend in data and exogeneous variables) to examine whether a cointegration relationship exists between HFI and TFP level variables indicates no cointegration relationship between the two variables. However, for the first differenced variables, the cointegration test shows a statistically significant cointegration equation between the differenced endogenous variables. Thus, I employ the SVAR with the differenced variables and impose a restriction on the long-run impact matrix to observe the impulse responses of the shocks. Second, I adopt the VECM with differenced endogenous variables without restrictions as a robustness test.

The bivariate system is expressed as yt=[ΔTFPtΔHFIt]. <Table 1> lists the definitions of TFP and HFI. As deterministic terms, a constant term and a linear time trend term are included; and the HP-detrended stock market index, SNP500, and the HP-detrended University of Michigan Consumer Sentiment Index, UM_SENT, are also included as exogenous variables in the SVAR to compare the results with Beaudry and Portier (2006).2) Additionally, this allows me to identify the news shock inferred through the behavior of informed investors who expect better future production growth than the one observed by the behavior of all informed and uninformed stock market participants. I chose an optimal lag order of 1 for the main endogenous variables through the VAR lag-length test and include one-lagged detrended exogeneous variables.

Definition of Variables

2.2 Vector Error Correction Models

When two or more variables are cointegrated, their time-series observations move together as a pair, and then, a linear combination of the two series exists to form a stationary process, although the individual time-series are non-stationary (e.g., the first-difference stationary I(1) process). Many macroeconomic variables are non-stationary in their levels, but their first differences are stationary, or I(1), and are often cointegrated with other variables. For instance, the GDP and the stock market index in the levels are individually non-stationary I(1) processes. However, these two time-series can be tied together, and the linear combination of the GDP and the stock market index becomes a stationary process.

A common way to analyze non-stationary I(1) data is to take the first-differences of the variables to make stationary processes. However, this approach may lose important information if the two variables are cointegrated. In other words, a Vector Autoregressive (VAR) model with first-differenced data does not capture the long-run relationship between two cointegrated variables. The Vector Error Correction Model (VECM) could be appropriate to analyze the cointegrated time-series by adding a vector of lagged error-correction terms in the VAR equations. In the case of the two variables, these error correction terms are lagged residuals from the cointegrating regression between the two variables in levels. The terms indicate the prior disequilibrium deviated from the long-run relationship, in which those residuals would be zero. The multivariate VECM specifications are as follows (Cottrell and Lucchetti, 2016):

Consider a VAR of order p with a deterministic part given by μt (typically, a polynomial in time) and a unmodeled stochastic or exogenous part, xt. One can write the n-variate process yt. as

(2)yt= μt+A1yt1+A2yt2+ +Apytp+Bxt+ϵt

The equation (2) can be expressed as

(3)Δyt=μt+Πyt1+t=1p1ΓiΔyti+Bxt+ϵt,

where Π=i=1pAiI and Γi=j=i+1pAj. This is the VECM representation in equation (2). If the rank of the matrix Π is between 0 and n, cointegration occurs and Π can be written as αβ’. Then, β is called a cointegrating vector and α is called an adjustment vector. In this case, yt is I(1), but the combination (i.e., an error-correction term) ect = β’yt is I(0). In fact, ect represents the deviation from the long-run equilibrium relationship among the variables and may not be zero. The equation (3) is expressed as follows:

(4)Δyt= μt+αβ'yt1+t=1p1ΓiΔyti+ Bxt+ϵt

If β is known, then ect-1 =β’yt-1 would be observable and all the remaining parameters could be estimated via OLS (ordinary least squares). For example, if there are two endogenous variables with one cointegration relation, then

(5)Πyt1=[α11α21] [β11β21] [y1,t1y2,t1] =[α11α21] (β11y1,t1+β21y2,t1)=[α11α21] ec1,t1

In this study, the VECM for the bivariate system is expressed as Δyt=Δ([ΔTFPtΔHFIt]). The definitions of TFP and HFI are described in <Table 1>. I include a constant and error correction terms as well as the HP-detrended stock market index, SNP500, and the HP-detrended University of Michigan Consumer Sentiment Index, UM_SENT, as exogenous variables in the VECM. I chose an optimal lag order of 1 for the main endogenous variables through the VAR lag-length test and include one-lagged detrended exogeneous variables.

3. Data Descriptions

3.1 Hedge Fund Industry

Since 1994, the hedge fund industry has grown rapidly and drawn much attention from academia and the financial industry. <Figure 1> plots the estimated total assets under management (AUM) of the hedge fund industry. Although there are no standardized central statistics estimating the overall size of the hedge fund industry, its total AUM have increased from approximately $118 billion in 1997 to $5.14 trillion at its peak as of 2nd quarter of 2023, according to the Barclay Hedge Fund database.3) The hedge fund industry fell significantly once in 2008, but the total AUM size has been continuously growing in recent days since the 2008 global financial crisis.

<Figure 1>

Estimated Total Assets under Management (AUM) for the Hedge Fund Industry

This figure presents the historical growth of the assets managed by the hedge fund industry. The AUMs are voluntarily reported by contributing hedge fund managers to the Barclay Hedge Fund database.1) Individual AUMs were aggregated and provided as annual and quarter-end numbers from the website. Historical information is available after 1997. In this figure, all observations are annual (1997~1999) and 4th quarter-end AUMs (2000~ 2022). Only 2023 represents 2nd quarter-end AUM. As of 2nd quarter of 2023, the total AUM for the hedge fund industry was $5,138.7 billion, and the total AUM for the funds of funds was $335.2 billion according to the BarclayHedge website.

In general, hedge fund managers are known for having exceptional skills to predict overall economic growth because of their outstanding performance. They are usually more informed about future economic news than individual or noise investors. Therefore, the economic predictions by hedge fund managers can be considered more precise than those of other types of investors in the stock market. In addition, they are less regulated than other institutional investors (e.g., mutual fund) because many offshore centers provide business-friendly regulations for the hedge funds. Thus, the hedge fund index is a better proxy for the news shocks to future production or economic conditions than the stock market index, which reflects the expectations of all market participants.

Among the various hedge fund indices, the Credit Suisse Hedge Fund (CSHF) Indexes has released the benchmark performance summary on both the aggregated composite index and ten style-based sector indexes since January 1994.4), As for the sectors of hedge funds, hedge funds generally notify their trading strategies or styles to investors in their prospectuses.5) The CSHF Indexes separate funds into ten primary sectors based on their declared investment styles. For example, the global macro funds attempt to generate excess returns by making leveraged bets on price movements in equity, currency, interest rate, and commodity markets. The macro section of the name explains that managers use macroeconomic principles to identify arbitrage opportunities in asset prices.

3.2 Sample Construction and Summary Statistics

To examine these hypotheses, I use the Credit Suisse Hedge Fund (CSHF) index provided by the Credit Suisse/Tremont database and a quarterly TFP series obtained from the Fernald’s website. The Credit Suisse/Tremont database, which tracks more than 9,000 funds, is used to determine the Index Universe. This selection universe is defined as only the funds with a minimum of $50 million under management, a minimum one-year track record, and current audited financial statements.6) The CSHF index is calculated with asset-weighting and rebalanced semi-annually. The index includes new funds using a rule-based construction method; therefore, it tracks at least the top 85% of the AUM in each of the 10 universe or strategy categories. The historical monthly indices from January 1994 to the present are available on the website.

Fernald (2014) provides a quarterly series on TFP for the U.S. business sector. Shocks to quarterly estimated TFP are regarded as an important factor driving business cycle fluctuations (Kydland and Prescott, 1982). Because the TFP series are only available at a quarterly frequency, I also use the CSHF index data from the end of each quarter. Therefore, the sample period for all the analyses is from the first quarter of 1994 to the second quarter of 2023 (118 observations).

<Table 1> lists the detailed definitions of the main endogenous and exogenous variables. For the bi-variate SVAR and VECM, total factor productivity and the hedge fund index, (TFPt , HFIt), are employed as endogenous variables and the HP-detrended University of Michigan Consumer Sentiment Index and the HP-detrended stock market index, (UM _ SENTt , SNP500t), are included as exogenous variables. Barsky and Sims (2012) emphasize the importance of consumer confidence in forecasting future economic activity as confidence may reflect agents’ expectations of future economic productivity. Therefore, both the stock market index and consumer confidence survey results could also closely relate to future economic conditions.

<Figure 2> displays the time-series plots of the variables over the sample period from 1Q:1994 to 2Q:2023. For comparison with the productivity measures, <Figure 2> also reports the time-series plots of log-transformed utility-adjusted TFP series and business output series provided by Fernald (2014). Surprisingly, the hedge fund index (HFI) and log-transformed TFP series (TFP) show extremely similar patterns over the sample period. The correlation coefficient reported in <Table 2> between the two variables is 0.9791, implying that the behavior of informed investors, such as hedge funds, is closely related to the productivity in the U.S. business sector.

<Figure 2>

Time Series Plots over the Sample Period: From 1994Q1 to 2023Q2

Summary Statistics

<Table 2> presents summary statistics and the correlation coefficients among the variables in the SVAR and VECM system. HFI and S&P500 index are inflation-adjusted with the Gross Domestic Product: Implicit Price Deflator and transformed as per capita numbers. For estimation, exogenous variables such as UM Sentiment and S&P500 Index are detrended with the Hodrick–Prescott filter, but the correlation coefficients are calculated without detrending. The sample period is from the first quarter of 1994 to the second quarter of 2023 (118 observations).

The correlation coefficient between TFP and business output (Output) is 0.9807, indicating that the difference in the various productivity measures is minor from an empirical perspective. The correlation coefficient between HFI and TFP_Util is also very high, at 0.9502. However, the correlation coefficient between SNP500 and TFP is only 0.6598. Therefore, the stock market index seems to be much less related to the real economic productivity than the hedge fund index. If the stock market is somewhat informationally inefficient, then the stock prices reflect noisy information.

4. Empirical Results

In this section, I explain the results of the SVAR with a long-run restriction and the VECM without any restrictions as a robustness test. <Table 3> reports the estimation results of the VAR with a lag order of one, and <Table 4> summarizes the Granger Causality test between the two endogenous variables.

VAR System with a Lag Order of 1

<Table 3> reports the estimation results with the VAR system. A lag order of 1 is chosen based on LR, FPE, AIC, SC, and HQ tests. TFP and Ln(HFI) are included as endogenous variables and lagged exogeneous variables such as HP-detrended UM Sentiment and S&P500 Index. T-statistics are in the parentheses.

Granger Causality Test with a Lag Order of 1

<Table 4> presents the result of the VAR Granger Causality/ Block Exogeneity Wald Tests between the two endogenous variables. This test examines whether one variable contains useful information for predicting the other variable in the VAR system in <Table 3>.

In the first column of <Table 3> predicting ΔTFP, the t-statistics of one-quarter lagged ΔLn(HFI) is 3.9036, meaning that the differenced hedge fund index is statistically significant to predict the current changes in TFP productivity. Moreover, the hedge fund index is statistically stronger than the contemporaneous and lagged SNP500 incorporating behaviors of all informed and uninformed investors. On the other hand, in the second column of <Table 3> predicting ΔLn(HFI), the lagged ΔTFP is not statistically significant to predict the current return or performance of the hedge fund index. The Granger Causality tests in <Table 4> also present results similar to those in <Table 3>, where the first null hypothesis is rejected, but the second null hypothesis cannot be rejected. Therefore, I conclude that informed investors, such as hedge funds, would recognize news shocks on economic productivity in advance and move faster than other market participants.

<Table 5>, and <Figure 3> and <Figure 4> present the orthogonalized impulse responses with a long-run restriction, following Beaudry and Portier (2006), and the forecast error variance decompositions. A long-run restriction is imposed to identify the effects of a news shock on productivity. The left (right) panel of <Table 5> shows the results of the Structural (Cholesky) decompositions. I use Structural or Cholesky decomposition methods of the covariance matrix in error terms to make one underlying shock to be uncorrelated with the other shock.

SVAR Impulse Responses and Variance Decompositions

<Table 5> reports the results of impulse responses and forecast error variance decompositions over the next 5 quarters in the SVAR system with a lag order of 1. A long-run restriction is imposed to identify the effect of a news shock on productivity. The left (right) panel presents a Structural (Cholesky) decomposition for the factorization matrix of the covariance matrix.

<Figure 3>

SVAR Impulse Responses over the Next 5 Quarters(Structural Decomposition)

<Figure 3> presents the orthogonalized impulse responses over the next 5 quarters in the SVAR system with a lag order of 1. To make one underlying shock to be uncorrelated with the other shock, Structural decomposition methods of the covariance matrix in error terms were used, and thus the shocks are orthogonalized and instantaneously uncorrelated each other. Δ TFP is ordered first.

<Figure 4>

SVAR Impulse Responses over the Next 5 Quarters(Cholesky Decomposition)

<Figure 4> presents the orthogonalized impulse responses over the next 5 quarters in the SVAR system with a lag order of 1. To make one underlying shock to be uncorrelated with the other shock, Cholesky decomposition methods of the covariance matrix in error terms were used, and thus the shocks are orthogonalized and instantaneously uncorrelated each other. ΔTFP is ordered first.

First, the upper-right panels of <Figure 3> and <Figure 4> show the impulse responses of ΔTFP to a shock of ΔLn(HFI) where the responses converge to zero regardless of imposing the long-run restriction in a year. That is, the effects of changes in hedge fund indexes on ΔTFP are significant7) in the following quarter but almost disappear in 2 quarters after the shock of ΔLn(HFI). Second, the lower-left panels present the effects of ΔTFP shock on the returns in hedge fund indexes, ΔLn(HFI). They appear insignificant even in the first quarter after the shock and gradually disappear within a year.

The variance decompositions in <Table 5> also show similar results between two variables. In the first period, approximately 10.906% of the variation in ΔTFP is from shocks to the ΔLn(HFI), and the contribution of ΔLn(HFI)to the variation in ΔTFP converges to 18.380% when a Structural decomposition method is employed. In the case of ΔLn(HFI), approximately 13.341% of the variation in ΔLn(HFI) is from shocks to ΔTFP in the first period, and it converges to 13.904% and becomes stable after 3 quarters. Therefore, I conclude that hedge funds, as informed investors, would be able to know the news on future productivity in advance, so they move faster than any other agents in the economy.

<Table 6> presents the estimation results using the VECM with a lag order of one.8), Beaudry and Portier (2006) identify news shocks inherent in the stock market index with the VECM to explain a factor causing future business cycle fluctuations. They argue that changes in technological opportunities are the main driver resulting in the business cycle fluctuations, and then stock market participants may notice these changes in advance. This study also employs the VECM as a robustness test because the result from a cointegration test with the endogenous variables in level is not statistically significant but the 1st differenced variables are cointegrated from the test. Following the literature, current and lagged UM_SENT and SNP500 are also included as exogenous variables.

VECM system with a Lag Order of 1

<Table 6> reports the estimation results with the VECM system. Since the Trace and Maximum Eigenvalue tests indicate a statistically significant cointegrating relation between differenced TFP and Ln (HFI), I estimate the VECM system with a lag order of 1 as a robustness test. T-statistics are in the parentheses.

In <Table 6>, the cointegrating vector (β) and the adjustment vector (α) are statistically significant, and represent the long-run equilibrium and short-run adjustment relationship between Δ(TFP) and ΔLn(HFI), respectively. In the first equation with a dependent variable Δ(ΔTFP), the current and lagged stock market index, SNP500 and SNP500(-1) are statistically significant although the lagged 2nd differenced hedge fund index, Δ(ΔLn(HFI)), is not statistically significant, implying that stock market participants may capture well the future economic conditions in advance, so information from the stock market seems to be useful to expect future productivity. In the second equation with a dependent variable Δ(ΔLn(HFI)), the current and lagged stock market index, SNP500 and SNP500(-1) are also statistically significant, implying that the changes in performance of hedge funds are related to the stock market performance.

<Table 7> and <Figure 5> show the results of the impulse responses and forecast error variance decompositions over the next 10 or 20 quarters in the VECM with a lag order of one. The upper-right panel of <Figure 5> presents the impulse response of ΔTFP to a shock of ΔLn(HFI) where the responses converge to 1.1634 in a year. The magnitude of the impulse responses from 1st quarter to 2nd quarter is fairly large, at 1.5236, compared to the magnitudes of 0.8256 and 0.7932 from the SVAR model in <Table 5>. The lower-left panel of <Figure 5> shows the impulse response of ΔLn(HFI) to a shock of ΔTFP where the responses converge to 0.1034 in a year. The forecast error variance decompositions in <Table 7> also indicate that approximately 65% of the variation in ΔTFP is from a shock to ΔLn(HFI) in the horizon of 20 quarters, whereas approximately 18% of the variation in ΔLn(HFI) is explained by a shock to the ΔTFP.

VECM Impulse Responses and Variance Decompositions

<Table 7> reports the results of impulse responses and forecast error variance decompositions in the VECM system with a lag order of 1. I impose a Cholesky decomposition for the factorization matrix of the covariance matrix. ΔTFP is ordered first.

<Figure 5>

VECM Impulse Responses over the Next 5 Quarters (Cholesky Decomposition)

<Figure 5> presents the orthogonalized impulse responses over the next 5 quarters in the VECM system with a lag order of 1. To make one underlying shock to be uncorrelated with the other shock, Cholesky decomposition methods of the covariance matrix in error terms were used, and thus the shocks are orthogonalized and instantaneously uncorrelated each other. ΔTFP is ordered first.

In summary, the findings from both SVAR and VECM models indicate that the trading behaviors of hedge funds are more likely to forecast the changes in TFP. In other words, information inferred from informed traders is more valuable for predicting future fundamentals or productivity than that inferred from the behaviors of all informed and uninformed traders. However, its predictive strength in the opposite direction appears to be limited.

5. Conclusions

This study examines the dynamic relationship between economic productivity, TFP, and hedge funds’ behaviors, HFI. The investment behaviors of hedge funds as informed traders quickly capture the news shock about future production growth; therefore, the information inferred from the hedge fund index returns is useful for predicting future economic productivity or economic conditions. Using a quarterly series of TFP and hedge fund index returns, this study finds that a predictive power exists in the behavior of hedge funds on real economic productivity. The findings from the SVAR and VECM models are robust regardless of whether a long-run restriction is imposed.

References

1. Agarwal V., Daniel N. D., Naik N. Y.. 2009;Role of Managerial Incentives and Discretion in Hedge Fund Performance. Journal of Finance 64(5):2221–2256.
2. Barsky R. B., Sims E. R.. 2012;Information, Animal Spirits, and the Meaning of Innovations in Consumer Confidence. American Economic Review 102(4):1343–1377.
3. Beaudry P., Portier F.. 2006;Stock Prices, News, and Economic Fluctuations. American Economic Review 96(4):1293–1307.
4. Ben-David I., Franzoni F., Moussawi R.. 2012;Hedge Fund Stock Trading in the Financial Crisis of 2007-2009. Review of Financial Studies 25(1):1–54.
5. Easterly W., Levine R.. 2001;It's Not Factor Accumulation:Stylized Facts and Growth Models. World Bank Economic Review 15(2):177–219.
6. Fama E.. 1990;Stock Returns, Expected Returns, and Real Activity. Journal of Finance 45(4):1089–1108.
7. Fernald J.. 2014;A Quarterly, Utilization-Adjusted Series on Total Factor Productivity. FRBSF Working Paper 2012-19 (updated April 2014)
8. Gao M., Huang J.. 2016;Capitalizing on Capitol Hill:Informed Trading by Hedge Fund Managers. Journal of Financial Economics 121(3):521–545.
9. Granger C. W. J.. 1969;Investigating Causal Relations by Econometric Models and Cross-spectral Methods. Econometrica 37(3):424–438.
10. Huang Q., Jain P. K.. 2024;Informed Trading by Hedge Funds. Journal of Financial Reserach DOI:10.1111/jfir.12386.
11. Johansen S.. 1991;Estimation and Hypothesis Testing of Cointegration Vectors in Gaussian Vector Autoregressive Models. Econometrica 59(6):1551–1580.
12. Kydland F., Prescott E.. 1982;Time to Build and Aggregate Fluctuations. Econometrica 50(60):1345–1370.
13. Lutkepohl H., Kraetzig M.. 2004. Applied Time Series Econometrics Cambridge: Cambridge University Press.
14. Sims C. A.. 1980;Comparison of Interwar and Postwar Business Cycles. American Economic Review 70(2):250–257.
15. Internet Resources Credit Suisse Hedge Fund Index (Dow Jones Credit Suisse Hedge Fund Index) Available from: https://secure.hedgeindex.com/hedgeindex.
16. Federal Reserve Economic Data Available from: https://fred.stlouisfed.org.
17. JMulTi Manual, ver. 4.24 Available from: http://www.jmulti.de/ or https://jmultir.r-forge.r-project.org/.
18. OECD Labour Force Statistics;Short-Term Labour Market Statistics Available from: http://stats.oecd.org.
19. Quarterly-TFP series produced by John Fernald, September 2023 Downloaded Available from: http://www.frbsf.org/economic-research/indicators-data/total-factor-productivity-tfp/.

Notes

1)

Examples of investment strategies by hedge funds include 1) global macro, 2) convertible arbitrage, 3) event-driven, 4) long/short equity, 5) emerging markets, 6) equity market neutral, 7) fixed income arbitrage, 8) multi-strategy, and so on (https://lab.credit-suisse.com/#/en/index/SECT/SECT/faq). For example, hedge funds may take equity market neutral positions in anticipation of a bear market, or actively bet on directional market moves. Therefore, hedge funds do not necessarily target certain types of firms that are highly affected by macro-level productivity compared to the firms included in the S&P500 index.

2)

I include the real quarterly S&P500 composite stock market index as an exogenous variable to measure the marginal effect of investment behaviors by hedge funds, acting as informed traders, on TFP or productivity. This approach intentionally excludes the influence of mixed investment behaviors by all informed and uninformed traders. Of course, the S&P500 index could be included as an endogenous variable in the SVAR and VECM. However, this approach would make the system of equations more complex, leading to less clear interpretations. Additionally, to obtain comparable results, I adhere to the methodology and variable constructions suggested by Beaudry and Portier (2006), who use only two endogenous variables: the S&P500 index and TFP.

3)

Source: The Barclay Hedge Fund database (https://www.barclayhedge.com/solutions/assets-under-management/); The reported aggregated AUM size does not include the AUM of funds of funds, which is approximately $335.2 billion as of 2nd quarter 2023.

4)

The CS reports an aggregated composite index (i.e., The Credit Suisse AllHedge Index) and ten style-based sub-sector indexes since January 1994. In this study, I use the aggregated composite index which is an asset-weighted index set to 100 as of January 1994. The CSHF index is a rules-based measure of an investable portfolio, with its constituents rebalanced semi-annually according to the sector weights of the index.

5)

However, no standardized definitions for the strategies are used.

6)

The Credit Suisse Hedge Fund Indexes FAQ(https://lab.credit-suisse.com/#/en/index/HEDG/HEDG/faq).

7)

The mean (median) of ΔTFP is 0.8121 (0.8915), and the standard deviation is 3.0643.

8)

JMulTi software is used for VECM estimation (http://www.jmulti.de/).

Article information Continued

<Table 1>

Definition of Variables

Variable Definition
TFP Log-transformed quarterly time-series calculated with dtfp from the Fernald’s website.1) According to the explanation on the Fernald’s data, dtfp is the business sector TFP meaning output growth less the contribution of capital and labor. As dtfp is a percent change at an annual rate (= 400 change in natural log), the log-transformed TFP is re-calculated as TFP = TFP(-1) + dtfp/400.

Output Log-transformed quarterly time-series calculated with dY from the Fernald’s website. dY is the equally-weighted average business output with expenditure (product) side and income side. As dY is a percent change at an annual rate (= 400 change in natural log), the log-transformed business output is calculated as OUTPUT = OUTPUT(-1) + dY/400.

TFP_Util Log-transformed quarterly time-series calculated with dtfp_util from the Fernald’s website. dtfp_util is the utilization-adjusted TFP: dtfp_util = dtfp - dutil. Adjustments for variations in factor utilization are made on labor effort and the workweek of capital. Since dtfp_util is a percent change at an annual rate (= 400 change in natural log), the log-transformed utilization-adjusted TFP is calculated as TFP_Util = TFP_Util(-1) + dtfp_util/400.

HFI The real quarterly time-series index calculated with the Credit Suisse Hedge Fund Index (HEDG)2), which is deflated by the Gross Domestic Product: Implicit Price Deflator (GDPDEF) obtained from the FRED website.3) The real Credit Suisse Hedge Fund Index is transformed to the per capita number by dividing it with the quarterly working age population from 15 to 64 (POP1564_NSA).4) The quarterly U.S. working age population data series is obtained from the OECD Statistics. HFI is calculated as HFI = (1000real HEDG)/(working age population)

UM_SENT The University of Michigan Consumer Sentiment Index (UMCSENT) obtained from the FRED website. For estimation, the sentiment index is detrended by the Hodrick-Prescott filter with lambda 1600.

SNP500 The real quarterly S&P 500 composite stock price index obtained from the Yahoo Finance website. The quarter-end close price adjusted for dividends and splits is deflated by the Gross Domestic Product: Implicit Price Deflator (GDPDEF) and transformed as per capital term with the quarterly working age population from 15 to 64 (POP1564_NSA). SNP500 is calculated as SNP500 = (1000 real S&P 500 Index)/(working age population) Then, SNP500 is detrended with the Hodrick-Prescott filter with lambda 1600 for estimation.

Note: 1) http://www.frbsf.org/economic-research/indicators-data/total-factor-productivity-tfp/.

2) https://lab.credit-suisse.com/#/en/index-nav.

3) https://fred.stlouisfed.org

4) OECD Labour Force Statistics; Short-Term Labour Market Statistics (http://stats.oecd.org).

<Figure 1>

Estimated Total Assets under Management (AUM) for the Hedge Fund Industry

This figure presents the historical growth of the assets managed by the hedge fund industry. The AUMs are voluntarily reported by contributing hedge fund managers to the Barclay Hedge Fund database.1) Individual AUMs were aggregated and provided as annual and quarter-end numbers from the website. Historical information is available after 1997. In this figure, all observations are annual (1997~1999) and 4th quarter-end AUMs (2000~ 2022). Only 2023 represents 2nd quarter-end AUM. As of 2nd quarter of 2023, the total AUM for the hedge fund industry was $5,138.7 billion, and the total AUM for the funds of funds was $335.2 billion according to the BarclayHedge website.

Note: 1) https://www.barclayhedge.com/solutions/assets-under-management/hedge-fund-as sets-under-management/hedge-fund-industry.

<Figure 2>

Time Series Plots over the Sample Period: From 1994Q1 to 2023Q2

Note: 1) The shaded areas represent the periods of US recessions provided by the NBER.

<Table 2>

Summary Statistics

<Table 2> presents summary statistics and the correlation coefficients among the variables in the SVAR and VECM system. HFI and S&P500 index are inflation-adjusted with the Gross Domestic Product: Implicit Price Deflator and transformed as per capita numbers. For estimation, exogenous variables such as UM Sentiment and S&P500 Index are detrended with the Hodrick–Prescott filter, but the correlation coefficients are calculated without detrending. The sample period is from the first quarter of 1994 to the second quarter of 2023 (118 observations).

Mean Median Maximum Minimum Std. Dev.
HFI 2.7008 2.8927 3.8605 1.0486 0.7754
ΔLn(HFI) 0.0097 0.0117 0.1447 -0.1191 0.0373
TFP 5.4057 5.4296 5.5240 5.2447 0.0755
ΔTFP 0.8121 0.8915 14.2378 -12.7854 3.0643
TFP_Util 5.4299 5.4585 5.5446 5.2489 0.0798
Output 6.7390 6.7654 7.0746 6.2722 0.2155
UM Sentiment 86.7009 90.6500 107.3000 50.0000 13.3253
HP_ UM Sentiment 0.0000 0.2108 10.5358 -17.9722 5.8565
S&P500 Index 11.4903 10.2463 25.0709 4.9984 4.4015
HP_S&P500 Index 0.0000 -0.0793 4.6698 -3.6392 1.3626

Correlations and t-Statistics (# of obs. = 118)

Variables (t-Statistics) (1) (2) (3) (4) (5) (6) (7) (9) (10)

(1) HFI 1.00
-
(2) ΔLn(HFI) -0.13 1.00
(-1.42) -
(3) TFP 0.98 -0.18 1.00
(51.89) (-1.96) -
(4) ΔTFP -0.10 0.27 -0.07 1.00
(-1.10) (2.99) (-0.74) -
(5) TFP_Util 0.95 -0.16 0.98 -0.07 1.00
(32.85) (-1.79) (54.95) (-0.80) -
(6) Output 0.98 -0.20 0.98 -0.11 0.96 1.00
(54.02) (-2.16) (56.70) (-1.23) (36.11) -
(7) UM Sentiment -0.42 0.21 -0.44 0.25 -0.49 -0.41 1.00
(-5.04) (2.35) (-5.22) (2.82) (-5.98) (-4.78) -
(8) HP_UM Sent 0.06 0.19 0.07 0.24 0.01 0.05 0.61 1.00
(0.63) (2.12) (0.79) (2.68) (0.12) (0.52) 8.31) -
(9) S&P500 Index 0.69 -0.03 0.66 -0.01 0.61 0.74 -0.06 0.18 1.00
(10.14) (-0.33) (9.46) (-0.15) (8.18) (11.70) (-0.65) (1.94) -
(10) HP_S&P500 0.14 0.23 0.09 0.15 0.00 0.06 0.25 0.43 0.38
(1.49) (2.50) (1.00) (1.60) (0.01) (0.64) (2.79) (5.10) (4.41)

<Table 3>

VAR System with a Lag Order of 1

<Table 3> reports the estimation results with the VAR system. A lag order of 1 is chosen based on LR, FPE, AIC, SC, and HQ tests. TFP and Ln(HFI) are included as endogenous variables and lagged exogeneous variables such as HP-detrended UM Sentiment and S&P500 Index. T-statistics are in the parentheses.

Variables ΔTFP ΔLn(HFI)

Coefficient Coefficient

(T-Stat) (T-Stat)
ΔTFP(-1) -0.0962 0.0004
(-1.0852) (0.4274)
ΔLn(HFI)(-1) 29.6770 0.1804
(3.9036) (2.4227)
UM_SENT 0.0869 -0.0009
(1.4824) (-1.6445)
SNP500 0.6877 0.0246
(2.2568) (8.2522)
UM_SENT(-1) -0.0989 0.0008
(-1.7778) (1.5034)
SNP500(-1) -0.6309 -0.0261
(-2.0902) (-8.8352)
Constant 1.1322 0.0204
(2.0802) (3.8329)
Trend -0.0085 -0.0002
(-1.1105) (-2.6692)
R-squared 0.2548 0.4985
Adj. R-squared 0.2070 0.4663
# of observations: Q2:1994 ~ Q2:2023 117 117

<Table 4>

Granger Causality Test with a Lag Order of 1

<Table 4> presents the result of the VAR Granger Causality/ Block Exogeneity Wald Tests between the two endogenous variables. This test examines whether one variable contains useful information for predicting the other variable in the VAR system in <Table 3>.

Null Hypothesis: Obs. Chi-Sq. d.f. P-value
ΔLn(HFI) does not Granger Cause ΔTFP 117 15.2379 1 0.0001
ΔTFP does not Granger Cause ΔLn(HFI) 0.1827 1 0.6691

<Table 5>

SVAR Impulse Responses and Variance Decompositions

<Table 5> reports the results of impulse responses and forecast error variance decompositions over the next 5 quarters in the SVAR system with a lag order of 1. A long-run restriction is imposed to identify the effect of a news shock on productivity. The left (right) panel presents a Structural (Cholesky) decomposition for the factorization matrix of the covariance matrix.

Factorization: Structural Factorization:Cholesky

Impulse Responses

Response of ΔTFP Response of ΔTFP

Period ΔTFP ΔLn(HFI) Period ΔTFP ΔLn(HFI)
1 2.57701 -0.90160 1 2.73017 0.00000
2 0.04205 0.82564 2 -0.23297 0.79321
3 0.07665 0.04398 3 0.05783 0.06683
4 0.00765 0.02713 4 -0.00174 0.02813
5 0.00282 0.00353 5 0.00149 0.00427

Response of ΔLn(HFI) Response of ΔLn(HFI)

Period ΔTFP ΔLn(HFI) Period ΔTFP ΔLn(HFI)

1 0.00977 0.02490 1 0.00100 0.02673
2 0.00272 0.00416 2 0.00119 0.00482
3 0.00051 0.00106 3 0.00013 0.00116
4 0.00012 0.00021 4 0.00004 0.00023
5 0.00002 0.00005 5 0.00001 0.00005

Variance Decompositions

Variance Decomposition of ΔTFP Variance Decomposition of ΔTFP

Period ΔTFP ΔLn(HFI) Period ΔTFP ΔLn(HFI)

1 89.094 10.906 1 100.000 0.000
2 81.633 18.367 2 92.268 7.732
3 81.627 18.373 3 92.221 7.779
4 81.620 18.380 4 92.212 7.788
5 81.620 18.380 5 92.211 7.789

Variance Decomposition of ΔLn(HFI) Variance Decomposition of ΔLn(HFI)

Period ΔTFP ΔLn(HFI) Period ΔTFP ΔLn(HFI)

1 13.341 86.659 1 0.139 99.861
2 13.895 86.105 2 0.327 99.673
3 13.904 86.096 3 0.329 99.671
4 13.904 86.096 4 0.329 99.671
5 13.904 86.096 5 0.329 99.671

<Figure 3>

SVAR Impulse Responses over the Next 5 Quarters(Structural Decomposition)

<Figure 3> presents the orthogonalized impulse responses over the next 5 quarters in the SVAR system with a lag order of 1. To make one underlying shock to be uncorrelated with the other shock, Structural decomposition methods of the covariance matrix in error terms were used, and thus the shocks are orthogonalized and instantaneously uncorrelated each other. Δ TFP is ordered first.

<Figure 4>

SVAR Impulse Responses over the Next 5 Quarters(Cholesky Decomposition)

<Figure 4> presents the orthogonalized impulse responses over the next 5 quarters in the SVAR system with a lag order of 1. To make one underlying shock to be uncorrelated with the other shock, Cholesky decomposition methods of the covariance matrix in error terms were used, and thus the shocks are orthogonalized and instantaneously uncorrelated each other. ΔTFP is ordered first.

<Table 6>

VECM system with a Lag Order of 1

<Table 6> reports the estimation results with the VECM system. Since the Trace and Maximum Eigenvalue tests indicate a statistically significant cointegrating relation between differenced TFP and Ln (HFI), I estimate the VECM system with a lag order of 1 as a robustness test. T-statistics are in the parentheses.

Cointegrating Eq: CointEq1
ΔTFP(-1) 1.0000
ΔLn(HFI)(-1) -55.3108
(-6.10502)
Const. -0.2700

Variables Δ (ΔTFP) Δ (ΔLn(HFI))

CointEq1 -1.0266 0.0036
(-7.69962) (2.24409)
Δ (ΔTFP(-1)) -0.0101 -0.0016
(-0.11510) (-1.54156)
Δ (ΔLn(HFI)(-1)) -10.6156 -0.2502
(-1.45900) (-2.89451)
Const. -0.0100 0.0004
(-0.03809) (0.13420)
UM_SENT 0.0443 -0.0020
(0.74049) (-2.78867)
SNP500 0.6993 0.0249
(2.18811) (6.54980)
UM_SENT(-1) -0.1173 0.0002
(-1.98977) (0.33076)
SNP500(-1) -0.7693 -0.0295
(-2.45599) (-7.91730)
R-squared 0.5895 0.5321
Adj. R-squared 0.5629 0.5017
# of observations: Q3:1994 ~ Q2:2023 116 116

<Table 7>

VECM Impulse Responses and Variance Decompositions

<Table 7> reports the results of impulse responses and forecast error variance decompositions in the VECM system with a lag order of 1. I impose a Cholesky decomposition for the factorization matrix of the covariance matrix. ΔTFP is ordered first.

Impulse Responses

Response of ΔTFP Response of ΔLn(HFI)

Period ΔTFP ΔLn(HFI) Period ΔTFP ΔLn(HFI)
1 2.82994 0.00000 1 0.00642 0.03300
2 0.19263 1.52356 2 0.00908 0.01826
3 0.50876 1.13732 3 0.01154 0.02133
4 0.61222 1.15232 4 0.00996 0.02103
5 0.56475 1.16655 5 0.01040 0.02104
6 0.57146 1.16357 6 0.01033 0.02103
7 0.57207 1.16319 7 0.01034 0.02104
8 0.57171 1.16351 8 0.01034 0.02103
9 0.57172 1.16344 9 0.01034 0.02103
10 0.57174 1.16344 10 0.01034 0.02103

Variance Decompositions

Variance Decomposition of ΔTFP Variance Decomposition of ΔLn(HFI)

Period ΔTFP ΔLn(HFI) Period ΔTFP ΔLn(HFI)

1 100.000 0.000 1 3.645 96.355
2 77.609 22.391 2 7.998 92.002
3 69.673 30.327 3 12.029 87.971
4 63.716 36.284 4 13.299 86.701
5 58.806 41.194 5 14.383 85.617
6 54.910 45.090 6 15.118 84.882
7 51.719 48.281 7 15.669 84.331
8 49.053 50.947 8 16.095 83.905
9 46.795 53.205 9 16.435 83.565
10 44.856 55.144 10 16.712 83.288
11 43.174 56.826 11 16.943 83.057
12 41.701 58.299 12 17.138 82.862
13 40.401 59.599 13 17.305 82.695
14 39.244 60.756 14 17.449 82.551
15 38.208 61.792 15 17.575 82.425
16 37.275 62.725 16 17.686 82.314
17 36.430 63.570 17 17.785 82.215
18 35.662 64.338 18 17.873 82.127
19 34.961 65.039 19 17.953 82.047
20 34.317 65.683 20 18.025 81.975

<Figure 5>

VECM Impulse Responses over the Next 5 Quarters (Cholesky Decomposition)

<Figure 5> presents the orthogonalized impulse responses over the next 5 quarters in the VECM system with a lag order of 1. To make one underlying shock to be uncorrelated with the other shock, Cholesky decomposition methods of the covariance matrix in error terms were used, and thus the shocks are orthogonalized and instantaneously uncorrelated each other. ΔTFP is ordered first.