Hedge Funds Are Not Destabilizing

Transcription

1 Hedge Funds Are Not Destabilizing Celso Brunetti Johns Hopkins University Michael S. Haigh Societe Generale This revision/print: November 30, 2007 First Draft (very preliminary) Abstract: The possibility that hedge fund trading destabilizes or creates a volatile market is frequently debated. This hypothesis is in stark contrast to the traditional speculative stabilizing theory that profitable speculation must involve buying when the price is low and selling when the price is high. To test the hypothesis that hedge fund trading is destabilizing we employ a unique dataset from the U.S. CFTC on individual positions of hedge funds in a heavily traded financial futures market, the 10-year T-notes. While others have used a more aggregated version of our data, here we test, for the first known time, whether hedge funds cause, in a forecasting sense, price movements and volatility in futures markets and, therefore, destabilize the market. Our findings confirm that hedge fund trading in futures markets is not destabilizing. In particular, hedge funds trading activity reduces volatility levels. Key Words: Hedge fund, volatility, price, Granger-causality JEL Codes: C3, G1 Acknowledgments: We would like to thank Frank Diebold, Jim Overdahl and Bahattin Buyuksahin for useful discussions and comments. This paper was written when Michael Haigh was Associate Chief Economist at U.S. Commodity Futures Trading Commission (CFTC). The authors would also like to thank the CFTC which provided a very fruitful research environment. The views expressed on this paper are those of the authors and do not, in any way, reflect the views or opinions of the Commodity Futures Trading Commission. The usual disclaimer applies. 1

2 1. Introduction Are hedge funds destabilizing financial markets? This question is of basic importance for regulators, investors and the general public, and has strong implication for understanding the fundamental role of speculation in financial markets. In recent years the term hedge fund has become somewhat of a contentious phrase in the financial industry s lexicon. It has been often suggested that the trading behavior of hedge funds and other large speculators can increase the fragility of financial markets leading to a potential destabilization of the broader market system. Indeed, hedge funds have been examined in several financial distresses, including the 1992 European Exchange Rate Mechanism (ERM) crisis, and the 1994 Mexican peso crisis (Fung and Hsieh, 2000); the 1997 Asian financial crisis (Brown, Goetzman and Park, 2000); and perhaps most famously the financial bailout of Long Term Capital Management (Edwards, 1999). In some episodes, hedge funds were deemed to have significant exposures and probably exerted market impact, whereas in other episodes they were unlikely to have contributed to destabilization. More recently, Brunnermeier and Nagel (2004) in their study of hedge funds and the technology bubble concluded that funds did not exert a correcting force on stock prices during the bubble and hence they question the efficient markets notion that rational speculators always stabilize prices. Significant accusations have been made on the role of hedge funds in the destabilization and volatility of markets. These allegations recently led the U.S. Security and Exchange Commission to propose a rule that would bar all but the wealthiest 1.3 percent of households from investing in hedge funds. The European Central Bank is also calling for new and more stringent regulation of hedge fund activity. Despite these accusations there has been surprisingly limited research on how hedge fund activity may impact prices and volatility. On the one hand this is particularly remarkable given the fact that this class of trader is controversial but on the other hand a lack of data stands in the way of a formal study of hedge fund trading in markets. In this paper, we use for the very first time, unique, highly disaggregated, position-level participant data on hedge funds (and other narrowly defined categories of traders) in 10-year T-note futures market collected by the U.S. Commodity Futures Trading Commission (CFTC). The analysis of futures data is of particular interest because hedge fund activity is often associated to derivative securities. Moreover, the 10-year T-note futures contract is a very attractive market to study as it is the second most heavily traded futures contract in the U.S. We provide empirical examination of the effect of hedge fund activity on both prices and volatility in the 10-year T-note futures contract. Contrary to the general wisdom, we find that hedge fund activity does not have any impact on prices. However, it does have some impact on volatility: hedge fund activity reduces volatility! In a simple multivariate framework we analyze Granger-causality between the daily rate of returns of the 10-year T-note futures contract and the daily positions of the five most important categories of market participants in this market. The analysis is conducted using several filtrations of the data. The results unambiguously show that hedge fund activity does not Granger-cause returns. In particular, hedge 2

3 fund activity does not Granger-cause any other variable in the system but it is Granger-caused by the other variables in the system. Our results suggest that, by taking the reverse positions of other market participants, hedge funds provide liquidity to the market. To assess the impact of hedge fund activity on risk we constructed realized volatility measures for the 10-year T-note futures contract and run again Granger-causality tests between volatility and positions of the five most important categories of market participants in the 10-year T-note futures contract. We find evidence that hedge fund activity Granger-causes volatility. We, therefore, analyze impulse response functions and find that hedge fund activity reduces volatility. The available evidence on whether hedge funds are destabilizing is mixed and does not permeate the literature. For example, Irwin and Yoshimaru (1999) find no significant relationship between what they call hedge fund positions and prices, and Irwin and Holt (2004) find a small but positive relationship between hedge fund trading volume and volatility but conclude that the relationship is entirely consistent with either the private information hypothesis or the noise trader hypothesis. The main issue is that earlier work used highly aggregated data which make difficult to assess the role of hedge funds in financial markets. In fact, data on hedge fund positions are difficult to obtain. To the extent that data on hedge fund trading patterns exist, it has usually been volunteered by cooperative organizations (Fung and Hsieh, 2000). 1 In the U.S futures and options on futures industry the Commodity Futures Trading Commission (CFTC) collects position level trading data on the composition of open interest across all futures and futures options contracts. These data are collected as part of the CFTC s market surveillance program. Through the years, several researchers have attempted to understand the role of large speculators using a highly aggregated public report produced by the CFTC based on market surveillance data. This report, called the Commitments of Traders (COT) report, classifies traders as either commercial or noncommercial. The non-commercial category includes participants who are not involved in the underlying cash business. This category has been heavily scrutinized by academics, practitioners and regulators alike to evaluate the role of large speculators on derivatives markets. Although studies that have employed COT data have made important contributions, researchers utilizing the broadly aggregated categories contained in the report have been forced to make assumptions about the composition of each classification. Because the data are so highly aggregated, these studies cannot precisely identify hedge fund activity and should be interpreted with caution. Despite the widespread use of COT data, only a 1 A major difficulty with this kind of study is the fact that hedge fund positions are virtually impossible to obtain. Except for very large positions in certain futures contracts, foreign currencies, U.S. Treasuries and public equities, hedge funds are not obligated to and generally do not report positions to regulators. Most funds do not regularly provide detailed exposure estimates to their own investors, except through annual reports and in a highly aggregated format. It is therefore nearly impossible to directly measure the impact of hedge funds in any given market. Fung and Hsieh (2000), page 3. Edwards and Caglayan (2001) and Brown, Goetztman and Park (2000) also describe the difficulty of obtaining data on hedge funds. 3

4 handful of studies (e.g., Bessembinder and Sequin, 1993) have suggested the importance and implications of identifying specific trader types rather than using aggregate classifications. In this paper we build on the hedge fund literature by employing a unique dataset on hedge fund (and other large trader category positions) that accurately identifies these participants in the financial futures trading in a way that other researchers have not been able to do. We proceed as follows. In section 2 we describe our data. In section 3 we analyze contemporaneous correlation between return, volatility and the five most important categories of market participants in the 10-year T-note futures. In section 4 we analyze Granger-causality tests for the return process while section 5 is devoted to Granger-causality tests for the volatility process. We conclude in section Data Throughout the paper we use data on daily rate of returns on the 10-year T-note futures, high frequency transaction data on the 10-year T-note futures which we employ for computing realized volatility measures, and data on daily positions (open interest) of the most important categories of market participants in the 10-year T-note futures. The data cover the period from January 2, 2002 through December 31, Futures contracts on the 10-year T-notes have maturity every quarter: March, June, September, December. The last day of trading for each contract is the seventh business day preceding the last business day of the delivery month. 2 Approximately a month before the expiration date, the open interest of the nearby contract decreases dramatically. This indicates that market participants roll over their positions from the nearby contract (March 2002, say) to the next-to-nearby contract (June 2002). This behavior generates some type of seasonality in the data. To mitigate these problems, the roll over strategy adopted in the paper is to switch to the new contract when the open interest of the nearby contract (March 2002) is lower than the open interest of the next-to-nearby contract (June 2002) 3 - see Gao and Wang (1999). Futures contracts are rarely executed. In fact, they are closed before maturity. The roll over strategy adopted in this paper may also solve the delivery distortion problems caused by the need of market participants of closing their position before the nearby contract expires. In what follows we describe the data in some detail. 2 In our sample expiration dates are March 19, 2002; June 19, 2002; September 19, 2002; December 19, 2002; March 20, 2003; June 19, 2003; September 19, 2003; December 19, 2003; March 22, 2004; June 21, 2004; September 21, 2004; December 21, Switch dates are February 27, 2002; May 30, 2002; August 29, 2002; November 27, 2002; February 26, 2003; May 29, 2003; August 28, 2003; November 26, 2003; February 26, 2003; May 27, 2004; August 30, 2004; November 30,

5 errors. 4 After filtering the data, the average number of transactions in the electronic platform from 7:20 10-year T-note Futures Daily Returns Futures contracts on the 10-year T-notes are traded on both an electronic platform and an open auction. Daily settlement prices refer to the prevailing price at 2:00 pm Central time, when the open auction closes. We only considered days when the market was open for at least 5 trading hours. Our sample is, therefore, composed by 746 trading days. Daily returns are constructed as r t = p( t) p( t 1), where p (t) is the natural logarithm of the closesettlement price in day t. When we switch contract from the nearby position to the next-to-nearby position, p (t) and p ( t 1) refer to the next-to-nearby contract. Table 1, row one, reports summary statistics for the return process. Daily returns on the 10-year T-note futures have a standard deviation that dominates the mean, negative skew and excess kurtosis. These results are standard for daily financial assets returns. Volatility To construct realized volatility measures, we obtained transaction data from the Chicago Board of Trade (CBOT). Data refer to transactions in the electronic platform which trades from 6:00 pm Central time the night before until 4:00 pm Central time that night, Sunday through Friday. There is also an open auction which trades from 7:20 am Central time until 2:00 pm Central time. The majority of trades (more than 90% of the total volume) take place in the electronic platform, but during the trading hours of the open auction (7:20 am - 2:00 pm Central time). These are the transactions we adopt for constructing realized volatility measures. Standard filters have been used to clean the data from outliers and recording am until 2:00 pm Central time is equal to 21,768 (median: 19,235). In our sample we have days where the number of transactions is close to 70,000, which corresponds to an average of 3 transactions per second. The median intertrade duration for the three years is a second. 5 Transaction data show that futures on the 10-year T-notes are heavily traded, in fact this is the second most liquid futures contract in the U.S. The use of high frequency data for constructing realized volatility measures could be problematic given the bias produced by market microstructure noise. Several solutions have been proposed to 4 We use the following filters (see Hansen and Lunde, 2004) 1) transactions with price equal to zero were deleted; 2) transactions with zero volume were deleted; 3) transactions that occurred outside the 7:20 am to 2:00 pm central time were deleted; 4) transactions with prices that were more than 1% deviation from a centered mean of 100 observations were deleted; 5) transactions with prices that increased/decreased more than 0.75% between i and i+1 and remain there 20 transactions or less, were deleted. 5 The CBOT provides data in seconds but actual transactions are recorded in hundredths of a second (centiseconds). 5

6 overcome this problem. 6 In this paper we follow two approaches. Hansen and Lunde (2006) proposed a kernel estimator where the bias correction is achieved by taking into account the autocorrelation structure of high frequency returns. The second approach we follow is that of Andersen, Bollerslev, Diebold and Ebens (2001) where the bias correction is achieved by sampling at relatively lower frequencies, 5-minute, say. In total we use three measures of realized volatility. Here we describe them in some detail and to do so we need to introduce some notation. Let { p(τ )} τ t be the natural logarithm of the price process over the time interval t, and let [ a, b] t be a compact interval which is partitioned in m subintervals. In our setup the interval [ a, b] is a trading day. For a given m, the ith intraday subinterval is given by, τ ], where a =, [ τ i 1, m i, m τ 0, m < τ1, m <...τ m m = b, and the length of each intraday interval is given by Δi, m = τ i 1, m τ i, m. The intraday returns are defined as r i, m = p( τ i, m ) p( τ i 1, m ) where i = 1,2,..., m. Realized volatility in day t is the sum of squared intraday returns sampled at frequency m m 2 RV t, = r, (1) m i= 1 To correct for the bias generated by market microstrucutre noise, the kernel estimator of Hansen and Lunde (2006) adds an additional term to equation (1) which takes into account the autocorrelation structure of high frequency returns m q m h HL 2 m RVt, m = ri, m + 2 ri, mri + h, m (2) i= 1 h= 1m h i= 1 where q refers to the first q autocorrelations, and HL refers to Hansen and Lunde s (2006) volatility estimator. This realized volatility measure introduces a Newey-West type correction and is able to produce an unbiased measure of volatility. Intraday returns can be constructed using different sampling schemes. In our framework, when τ i, m denotes the time of a transaction, we refer to sampling in transaction time (for example, sampling every 10 transactions). On the other hand, when τ i, m denotes equidistant calendar time, we refer to sampling in calendar time, and Δ = ( a b) m (for example, i m i, m / sampling every 10 minutes). In order to implement equation (2), we need to choose m, the sampling frequency, and q, the length of the first intradaily returns autocorrelations. Volatility signature plots 7 provide valuable information about the bias in RV measures in (1) and about choosing m and q in order to correct that bias. We construct three measures of realized volatility. 1) RV kernel estimator with transaction time sampling. 6 See, among others, Bandi and Russell (2006), Hansen and Lunde (2006) and Zhang, Mykland and Aït-Sahalia (2005). 7 Volatility signature plots show average daily realized volatility measures, computed using equation (1), against sampling frequency m - see Andersen, Bollerslev, Diebold and Labys (2000). 6

7 Figures 1 depicts the volatility signature plot for realized volatility measures constructed using equation (1) with transaction sampling and m = 1,10,20, 300. In particular, figure 1 shows average daily realized volatility measures over the three year sample , against the sampling frequency m. RV in transaction time is heavily affected by market microstructure noise. When sampling at the highest possible frequency, RV exhibits a positive bias. However the bias dissipates rapidly as sampling frequency increases. Hansen and Lunde (2006) show that the best unbiased measure of realized volatility is achieved using the smallest possible value of q for which RV seems to be unbiased, and the largest possible value for m. Figure 1 shows that RV in transaction time stabilizes at a sampling frequency of 30. We, therefore, construct the RV kernel estimator, equation (2), in transaction time using the highest possible frequency (i.e. we use all transactions) and set q = 30. In the reminder of the paper we refer to this volatility measure as RVT. 2) RV kernel estimator with calendar time sampling. Figure 2 shows the volatility signature plot for realized volatility measures constructed using equation (1) with calendar time sampling and m = 1,10,20, 600. RV in calendar time is less affected (note that figures 1 and 2 have different scales) by market microstructure noise, but the effects of the noise take longer to dissipate. This is in line with previous literature. Following Hansen and Lunde (2006) the kernel estimator of realized volatility with calendar sampling, equation (2), is constructed by setting m at the highest possible frequency, a second, and by setting q=60. 8 In the reminder of the paper we refer to this volatility measure as RVC. 3) RV with five minutes sampling. The last measure of realized volatility we consider is constructed using five-minute equidistant intraday returns and equation (1). This is the classic RV measure introduced by Andersen et al. (2001). In this case, sampling at lower frequencies (5 minutes) mitigates the bias produced by market microstructure noise. In fact, figure 2 shows that the positive bias in realized volatility with calendar sampling disappears well before the 5-minute sampling frequency. In the reminder of the paper we refer to this volatility measure as RV5M. Figure 3 depicts the kernel estimator of RV with transaction time sampling (RVT). 9 In the first part of the sample RVT does not exhibit excess variability. At the end of July/beginning of August 2003 the level and the variability of RVT increase. In the last part of the sample volatility levels decrease although we observe some very high values. Picks of RVT correspond to U.S. economy data releases. For example, the two highest picks (August 1, 2003; and September 3, 2004) correspond to U.S. economic 8 For each day we construct an equidistant grid of one-second prices for a total of 24,000 observations per day. The 10 year T-notes futures market is very liquid. Therefore, within the same day, there are several transactions with the same time stamp (in seconds). In these cases we retain the first observation with that time stamp. In constructing the one-second grid, we need to assign a price to each time interval. Obviously, there are times when there are no transactions. In these cases we adopt the previous tick method. 9 Graphs of RVC and RV5M are similar to figure 3. 7

8 data releases about GDP and employment, respectively, which caused the 10-year T-note yield to move more than 10 basis points. Table 1, rows 2-4, show descriptive statistics for the three measures of RV. RVT exhibits the lower sample mean. However, mean and median of these three measures are comparable. In fact, the main difference refers to the standard deviation. RVT has the lowest standard deviation which is half that of RVC and RV5M. This is a standard result in the RV literature (see Hansen and Lunde, 2006; Oomen, 2006). An additional standard result (e.g. Andersen et al., 2003) is that log-realized volatility is approximately Gaussian. Table 1, rows 5-7, shows descriptive statistics for logarithmic realized standard deviations. Skewness and kurtosis are much more in line with the assumption of normality although the Jarque-Bera test rejects the null of normality. Hedge Fund Positions The CFTC monitors U.S. futures and options on futures markets through its market surveillance program and since the 1920s, the CFTC (and its predecessors) has been utilizing the central tool of market surveillance known as the Large Trader Reporting System (LTRS). Following the Commodity Exchange Act (CEA), the CFTC collects and stores data from daily reports on market data and position information from the futures commission merchants (FCM s), foreign brokers, exchanges, clearing members, and also traders. These reports show the positions of traders that hold positions above specific levels set by the CFTC. These large trader reporting levels include, for example, Crude Oil (350 contracts), Euro (400 contracts) and for the contract market studied here: 10-year Treasury notes (1,000 contracts). The total amount of all traders positions reported to the CFTC represent approximately percent of total open interest in any market, while the remainder being traders who generally trade a small number of contracts, known as Non-Reportable Positions (NRP). 10 When a trader is identified to the CFTC, the trader is classified either as a commercial or noncommercial trader. A trader s reported futures position is determined to be commercial if the trader uses futures contracts for the purposes of hedging as defined by CFTC regulations. The non commercial category includes participants who are not involved in the underlying cash business otherwise known as speculators and include hedge funds, floor brokers, floor traders and so forth, see table The Commitment of Traders Report, which utilizes data from the LTRS, provides a summary of the percent of open interest (OI) held by the commercial, non-commercial and NRP categories as of each 10 Occasionally, the CFTC will raise or lower the reporting levels in specific markets with the objective of striking a balance between maximizing effective surveillance and minimizing the reporting burden on the futures industry. 11 Specifically, a reportable trader gets classified as commercial by filing a statement with the CFTC (using the CFTC Form 40) that he is commercially engaged in business activities hedged by the use of the futures and option markets. However, to ensure that the traders are classified consistently and with utmost accuracy, CFTC market surveillance staff in the regional offices checks the forms and re-classifies the trader if they have further information about the trader s involvement with the markets. 8

9 Tuesday s OI for markets in which 20 or more traders hold positions equal to or above the reporting levels established by the CFTC. The weekly COT reports are currently released every Friday at 3:30 pm Eastern time. Information released to the public in the form of the COT is highly aggregated, but the disaggregated LTRS enables the CFTC surveillance team to monitor the largest individual participants and/or a specific group of market participants in a market. It is these detailed groupings within the commercial and non-commercial categories that we analyze in this study. Active groupings of participants vary across contract markets but an exhaustive list of participants that hold positions in 10-year T-notes can be found in table 2. There is no consensus on the exact definition of a hedge fund in futures markets and there is nothing in the Commodity Exchange Act (CEA), the statute governing futures trading, to define hedge funds. 12 Accordingly, there is nothing that requires hedge funds to be categorized in the LTRS. Despite this, many hedge fund complexes are either advised or operated by CFTC registered Commodity Pool Operators (CPOs), 13 Commodity Trading Advisors (CTAs) 14 and/or and Associated Persons (APs) 15 who may control customer accounts. Therefore, through its LTRS, the CFTC obtains positions of both operators and advisors to hedge funds. In addition to these three categories of traders, market surveillance staff at the CFTC identifies other participants who are not registered in any of these three categories but are known to be managing money (MM). These four categories combined are defined as being the hedge fund category (see bottom of table 2). Despite not knowing the exact names of the funds trading in futures markets, it is clear that many of the large CPO s, CTA s and APs and MMs are generally considered to be hedge funds and hedge fund operators and so to conform to the academic literature and common financial parlance we refer to these four collectively as hedge funds in our study. The CFTC data contains daily open interest on 10-year T-note futures for 21 categories of market participants (see table 2). On average over the period analyzed, the open interest of the reportable 12 However, the SEC notes that a hedge fund is an entity that holds a pool of securities and perhaps other assets, whose interests are not sold in a registered public offering and which is not registered as an investment company under the Investment Company Act (p.3. SEC, 2003). 13 Commodity Pool Operator (CPO): A person engaged in a business similar to an investment trust or a syndicate and who solicits or accepts funds, securities, or property for the purpose of trading commodity futures contracts or commodity options. The commodity pool operator either itself makes trading decisions on behalf of the pool or engages a commodity trading advisor to do so. Source: Glossary of the CFTC ( 14 Commodity Trading Advisor (CTA): A person who, for pay, regularly engages in the business of advising others as to the value of commodity futures or options or the advisability of trading in commodity futures or options, or issues analyses or reports concerning commodity futures or options. Source: Glossary of the CFTC ( 15 Associated Person (AP): An individual who solicits or accepts (other than in a clerical capacity) orders, discretionary accounts, or participation in a commodity pool, or supervises any individual so engaged, on behalf of a futures commission merchant, an introducing broker, a commodity trading advisor, a commodity pool operator, or an agricultural trade option merchant. Source: Glossary of the CFTC ( 9

10 positions is equal to 75 percent of the total open interest, the remaining being small traders (NRP) who do not meet reporting thresholds (NRP). We are only considering 5 categories - the most important in terms of OI. These are commercial arbitrageur and broker/dealer (C_FA); commercial mutual fund (C_FE); commercial pension fund (C_FF); non-commercial floor broker/trader (N_FBT); and non-commercial hedge funds (N_HF). 16 Note the prefix C_ refers to commercial (hedgers), while the prefix N_ refers to non-commercial (speculators). The average open interest (including both short and long positions) of the 5 categories of market participants analyzed is 63 percent of the total open interest. Therefore, we are analyzing a good portion of the open positions in the 10-year T-note futures market. Over the sample analyzed, the open interest of commercial pension funds is non-stationary and presents a structural break. CFTC data reveal that the break was caused by a single trader who entered the 10-year T-note futures market with a large position and never left. We therefore filter the C_FA open interest with a dummy variable to account for the break. In the analysis below, for the open interests of the five most important categories of traders, we consider levels, first difference and Hodrick-Prescott (HP) filtered data. The level data represent the total position (number of contracts) held by each category. For example, on a given day a positive open interest for hedge funds signifies that this category of market participants holds a net long position. The first difference of the open interest represents the change in the levels and therefore is a measure of daily market activity for each single category analyzed. Finally, to isolate the cyclical component of the OI, we filtered the data using the standard HP filter. 17 Table 2 shows descriptive statistics for the level, first difference and HP filtered open interests. Data in levels have been rescaled by a factor of 10,000 (we have done so because the parameters in the estimations below are easier to interpret.) Open interest (levels) of commercial arbitrageurs or brokers/dealers and hedge funds is on average positive, which implies that these categories of market participants hold a long position on the 10-year T-note futures. The opposite is true for commercial mutual funds and non-commercial floor brokers/traders. Commercial arbitrageurs or brokers/dealers, commercial pension funds and non-commercial hedge funds exhibit high standard deviation, implying a high variability of their positions. Open interest is non-gaussian but stationary. Table 3 shows the participation rate of the five categories of traders analyzed. Over the three year period , the dominant category is commercial pension funds (C_FF) with 17% of the long side of total OI and 9% of the short side of OI. Interestingly, mutual funds (C_FE) only hold net short positions while commercial arbitrageurs or brokers/dealers (C_FA) hold mainly long positions. This is also true for 16 For a definition of broker, dealer, floor trader and floor broker, we refer the interested reader to the glossary of the CFTC, which can be found on line at 17 To compute the smooth parameter in the HP filter, we looked at the ratio of the variance of the cycle component and the variance of the second difference of the trend component. Under the assumption that the cycle component and the second difference of the trend component are normally and independently distributed, the smooth parameter is exactly equal to the ratio of the two variances. 10

11 hedge funds although, hedge fund activity is more balanced between the long and the short side of the market. We now proceed with the correlation analysis. 3. Unconditional Contemporaneous Correlation Our preliminary analysis of the relationship between returns, volatility, and the traders positions begins by computing correlation coefficients. Table 4 reports correlation coefficients between returns, realized volatility measures, and the open interest in levels of the 5 categories of market traders analyzed. Commercial arbitrageurs or brokers/dealers have a significant negative correlation with returns while hedge funds have a positive correlation with returns. This implies that hedge fund positions move in the same direction as the market. It is also interesting to note that hedge fund activity is negatively linked to volatility (this is true for all three measures of realized volatility). An increase in hedge fund activity is associated with lower volatility levels. Finally, hedge fund activity is negatively linked to other traders positions. This may suggest that by taking the opposite position with respect to the other market traders, hedge funds may provide liquidity to the market. This is in line with the results of Haigh, Hranaiova and Overdahl (2005) who study the interaction between traders in the natural gas and crude oil futures markets. They find that hedge funds provide liquidity to hedgers. This is also in line with the theory of speculation as described in Keynes and Hicks which postulates that speculators positions should offset any imbalance of hedgers positions. Table 5 reports correlation coefficients between returns, realized volatility measures, and the first difference of open interest of the 5 market traders analyzed. Similar to the results for levels, hedge funds show a significant positive correlation with returns. However, in this case, hedge fund activity is not linked to volatility. Once again, open interest of hedge funds is significantly negatively correlated with the open interest of the other traders. This reinforces our conjecture that hedge funds may provide liquidity to the market. These results are confirmed by table 6, which considers HP filtered OI. The simple correlation analysis provides three main results. First, hedge fund activity is positively linked to returns. Second, hedge fund positions are negative correlated to volatility. Third, the correlation between hedge fund positions and positions of the other market traders is always negative. Correlation coefficients do not say anything about causation. In fact, we turn now to that issue. 4. Granger-Causality Analysis for the Return Process The concept of Granger-causality relates to linear predictions x t is Granger-causal for z t if x t contains useful information for (linearly) predicting z t. This definition of causality is practical but has limitations. In fact, the notion of causality is an old one and goes back to the ancient Greek philosophy - 11

12 Aristotle distinguishes between cause and effects: a cause is an event that produces its effects. Unfortunately, Granger-causality does not allow us to distinguish between causes and effects. Nonetheless, Granger-causality is easy to compute and provides useful information on whether hedge fund activity prompts, in a forecasting sense, price movements and/or vice versa. To investigate the relationship between hedge fund activity and returns we perform Grangercausality tests within the following (reduced form) Vector AutoRegression (VAR) framework. B B B I B B B ( L) B12( L) B13( L) B14( L) B15( L) B16( L) ( L) B22( L) B23( L) B24( L) B25( L) B26( L) ( L) B32( L) B33( L) B34( L) B35( L) B36( L) ( L) B42( L) B43( L) B44( L) B45( L) B46( L) ( L) B52( L) B53( L) B54( L) B55( L) B56( L) ( L) B ( L) B ( L) B ( L) B ( L) B ( L) yt C FA _ t C FE _ t = εt C _ FFt N FBT _ t N _ HFt where y t = r t (the return process) and B kj (L)s are lag polynomials. 18 A variable, hedge funds (N_HF) say, does not Granger-cause returns, r t, with respect to the full information set if B 16 = B 26 = B 36 = B 46 = B 56 = 0, or equivalently B 16 = B 15 = B 14 = B 13 = B 12 = 0. We estimate three different sets of VARs. In the first we adopt the level of the open interest; in the second we use the change in open interest (first difference); and in the third we consider HP filtered positions. We estimated VARs with 5, 4, 3, 2, and 1 lags. We first estimate VARs with OLS. However, errors are affected by heteroskedasticity and serial correlation. We, therefore, estimate VARs with GMM and use heteroskedastic and autocorrelation robust standard errors. 19 Tables 7, 8 and 9 report p-values for the null H : B kj = 0, k j, k,j = 1, 2,..., 6. The p-value in the last column of each table refers to the joint hypothesis H 0 : B 12 = B 13 = B 14 = B 15 = B 16 = 0 for the first row (returns); H 0 : B 21 = B 23 = B 24 = B 25 = B 26 = 0 for the second row (commercial arbitrageurs or brokers/dealers, C_FA); and so on. In other words, we are testing whether each variable is jointly Granger-caused by the other variables - i.e. are returns Granger-caused by positions? The last row reports p-values for the null H 0 : B 21 = B 31 = B 41 = B 51 = B 61 = 0 for the first column; H 0 : B 12 = B 32 = B 42 = B 52 = B 62 = 0 for the second column; and so on. In other words, we are testing whether each variable is jointly Granger-causing the remaining variables in the system. Here we are particularly interested in testing whether hedge fund activity is Granger-causal for returns. This corresponds to testing the null H 0 : B 16 = B 26 = B 36 = B 46 = B 56 = 0. To conserve space, we only report results for the VARs with 5, 2 and 1 lags. 20 (3) 18 All variables in (3) are stationary (see table 1). 19 VAR estimates have been performed in EViews The results for lags 3 and 4 are very similar to those of lags 5, 2, and 1. In our estimates, Bayesian Information Criterion (BIC) always selects VARs with a lag while Akaike Information Criterion (AIC) always selects VARs with 2 lags. 12

13 Table 7 reports Granger-causality tests between returns and trader positions in levels. Starting from panel 1 (VAR with 5 lags), we find that returns are Granger-caused by positions (p-value 2.390%) and returns Granger-cause positions (p-value 0.080%). However, when jointly testing whether hedge funds Granger-cause returns or any other variable in the system, we fail to reject the null. In fact, hedge fund activity is the only variable which is not jointly Granger-causing any other variable in the system (pvalue 12.48%). This implies that hedge fund positions do not provide any useful information for predicting returns or other positions. Moreover, hedge fund activity is Granger-caused by the other variable in the system (p-value 0.020%). If we pair this result with the negative correlation between hedge fund positions and positions of the other traders, it is reasonable to conjecture that by taking the opposite side in the market, hedge funds are providing liquidity to the market. Granger-causality strongly depends on the lag structure adopted. Panels 2 and 3 in table 7 report p-values for Granger-causality tests for VARs with two and one lag, respectively: all the above results are confirmed and reinforced by the Granger-causality tests reported in panels 2 and 3 of table 7. Interestingly, panel 2 shows that each single variable in the system is Granger-causing hedge fund activity but hedge fund activity is not Granger-causing any variable in the system. It seems that hedge funds are reacting to market conditions, and there is no indication that hedge fund activity is moving prices and/or positions of other trades. Table 8 reports Granger-causality tests between returns and changes in traders positions (first difference). Panel 1, which refers to VAR with 5 lags, shows some (weak) evidence that returns are Granger-caused by position (p-value 11.85%). This is a feedback relationship. In fact, returns Grangercause positions (p-value 0.000%). Hedge fund activity (N_HF) is not (jointly) Granger-causal for any other variable (p-value 19.72%), but hedge fund activity is Granger-caused by the other variables in the system (p-value 0.000%). In our analysis we consider positions of two non-commercial (speculators) traders: hedge funds and floor brokers/dealers. It is interesting to note that both positions are neither Granger-causal for returns nor for hedger positions (C_FA, C_FE and C_FF). Our results suggest that speculators in the 10-year T-note futures market are not destabilizing prices but are reacting to price movements and/or to trades of other market participants. In other words, the above results seem to provide support to the traditional Keynes-Hicks paradigm on the role of speculation in financial markets. These results are robust to the lag length in the VAR specification (see panels 2 and 3). In fact, there is no evidence that hedge fund activity is Granger-causing returns or other traders activities. Similar results are obtained when using HP filtered positions (see table 9). The analysis of Granger-causality between returns and traders activities yields three main results. First, returns are Granger-caused by positions, and, vice versa, returns Granger-cause positions. There is, therefore, a feedback effect between returns and traders positions. Second, hedge fund activity is Granger-caused by returns and/or by positions of the other market participants. Third, hedge fund activity is not Granger-causal for returns and/or positions of the other market traders. These last two results are 13

14 particularly important for the issue we analyze in this paper. In fact, they suggest that hedge funds are not destabilizing prices. Hedge fund activity seems to be responsive to market conditions but is not moving the market nor is generating trading activity from other traders. We are aware that Granger-causality tests have limitations. However, our results are very robust. Using different data filtrations and different VAR specifications we always find that hedge fund activity does not move prices and/or trading in the 10-year T-note futures market. These findings are in line with previous literature. Irwin and Yoshimaru (1999), and Fung and Hsieh (2000), for example, find no evidence that hedge fund activity has an impact on market prices. The previous literature, however, analyzes highly aggregated data while we are able to precisely identify hedge funs activity. In this respect, our results are noteworthy. Is hedge fund activity increasing risk? It is, in fact, possible that hedge fund activity may not have any impact on prices but it might have an impact on market volatility. This is the question we now consider in the following section. 5. Granger-Causality Analysis for the Volatility Process To study Granger-causality between volatility and traders positions, we consider the three measures of realized volatility developed in section 2. A stylized empirical finding in the realized volatility literature is that logarithm realized standard deviation is approximately Gaussian. Our realized volatility measures confirm this finding (see table 1). In fact, in modeling realized volatility measures in the context of VARs, it is customary to use logarithmic realized standard deviation e.g. Andersen et al. (2006). In the following analysis we use the three measures of realized volatility described in section 2 and their logarithm realized standard deviations counterparts. To conserve space, we only report results for logarithmic realized standard deviation in transaction time (LNRSDT) computed using the kernel estimator of Hansen and Lunde (2006). 21 The set up of the Granger-causality test is that of equation (3) with y t = LNRSDT t, the logarithm realized standard deviation in transaction time. As for the previous section, VARs have been estimated using GMM and heteroskedastic and autocorrelation robust standard errors. Table 10 reports Granger-causality tests (p-values) between volatility and traders positions in levels. Panel 1 refers to the VAR with 5 lags. Traders positions are not Granger-causal for volatility (pvalue 44.48%) but volatility is Granger-causal for positions (p-value 2.050%). Interestingly, hedge fund activity is Granger-causal for volatility (p-value 9.670%). Panels 2 and 3 refer to VAR with 2 and one lag, respectively. These two panels show that positions, including hedge funds, Granger-cause volatility but volatility does not Granger-cause positions. It seems that there is some causality, from hedge fund activity 21 Our results are robust to the different measures of realized volatility considered. 14

15 to volatility. In other words, it seems that hedge fund trading is causing volatility in a Granger sense. To further investigate this issue we compute impulse responses. Pesaran and Shin (1998), proposed a technique, termed generalized impulse responses, which is invariant to the ordering of the variable in the VAR and does not require shocks to be orthogonal. Our analysis is merely empirical. We would like to investigate whether hedge fund activity is producing higher risk. We are agnostic about the theoretical role of hedge funds in financial markets. We let the data guide our steps. The Pesaran and Shin s (1998) technique is very helpful in this respect. Figure 4 depicts generalized impulse responses for the VAR with a lag. 22 We are particularly interested in the response of volatility to a standard deviation shock to hedge fund activity. The last graph in row 1 shows a statistically significant reduction in volatility. In other words, a shock in hedge fund activity reduces volatility. On the other hand, volatility does not have any statistically significant effect on hedge fund activity. We also compute impulse responses using the classic Cholesky decomposition which is very sensitive to the order of the variable in the VAR. To mitigate this problem we consider several ordering of the variables. Our VAR estimates imply a total of 6! different combinations of the variables. If we also consider that we estimate 5 different VARs specifications (with lags from 5 to 1), we end up with 6! 5= 3,600 impulse responses. Given that our focus is on hedge fund activity and volatility, we only consider a subset of all the possible combinations. In particular, we consider the remaining traders positions (C_FA, C_FE, C_FF and N_FBT) as a unique block. Figure 5 depicts standard Cholesky decomposition 23 for a VAR with 2 lags and with the following order of variables: volatility, commercial arbitrageurs or brokers/dealers (C_FA), commercial mutual funds (C_FE), commercial pension funds (C_FF), noncommercial floor brokers/traders (N_FBT) and hedge funds (N_HF). Figure 5 shows that hedge fund positions have an impact on volatility. In particular, hedge fund activity reduces volatility. A category of market traders is unlikely to cause volatility if this category does not trade or does not trade often. Ordering the 5 categories of traders in terms of their trading activity, might be a reasonable approach for computing impulse responses using the Cholesky decomposition. The daily change (first difference) of open interest for each category of market participants is a measure of daily market activity for these market traders. We are interested in measuring how active each category of traders is in the 10-year T-notes futures market. Therefore, we measure trading activity of each market participants over the sample period by the arithmetic mean of the absolute value of the change in open interest. We need to use the absolute value in order to avoid that short and long trading activities cancel out. We are aware that this is only an approximation of trading activity. In fact, this measure of trading activity is not able to account for intradaily transactions. For example, if a market trader goes long 10, This is the lag length selected by the BIC. Generalize impulse responses for VARs with 5, 4, 3 and 2 lags are very similar to those reported in figure 4. Response standard errors are computed with 10,000 Monte Carlo replications in EView 5.1. On the horizontal axis of each graph is the number of days after the shock, here 20 days. 23 Standard errors are computed with 10,000 Monte Carlo replications in EView

16 contracts at the opening of the market and then shorts the same amount of contracts before the market closes, the net daily open interest for this market trader will be unchanged even if the trader has been actively transacting. Such a trading strategy is typical of daily traders : traders that do not carry any positions overnight. The categories of market participants we are analyzing do not fit the definition of daily traders ; nevertheless, we need to be aware of this shortfall. Using the mean of the absolute value of the change in positions as measure of trading activity, we find that commercial arbitrageurs or brokers/dealers (C_FA) is the most active category of traders, followed by hedge funds (N_HF), commercial pension funds (C_FF), non-commercial floor brokers/traders (N_FBT), and commercial mutual funds (C_FE). 24 Therefore, hedge funds are the second most active category of market traders. Panels 2 and 1 in table 10 show that volatility is Granger-caused by positions but not vice versa; therefore, in our VAR estimates, we position volatility at the bottom. Once again, we find that hedge fund activity reduces volatility. In fact, figure 6 shows impulse responses for VAR with 2 lags where the ordering of the variables is given by the trading activity of each market category. The second graph in the last row clearly shows that a shock in hedge fund activity has a negative impact on volatility, i.e. hedge fund activity drives volatility down. 25 We now return to the Granger-causality analysis and consider volatility and change (first difference) in market positions. Results are reported in table 11. The VAR with 5 lags shows that positions do not Granger-cause volatility (p-value 65.01%) but volatility Granger-causes positions (pvalue 2.980%). Hedge fund activity does not Granger-cause volatility or any other variable in the system (p-value 27.43%) but it is Granger-caused by all the other variables (p-value 0.292%). The same applies to N_FBT. There is evidence that speculation activity is not Granger-causal for volatility and hedger positions. These results are robust to the VAR specification (see panels 2 and 3). When analyzing the Granger-causality relationship between volatility and positions in first difference, we find different results than those obtained using the levels. In general, we find that the change in traders positions is not Granger-causal for volatility. In particular, we do not find any evidence that volatility is linked, in a forecasting sense, to hedge fund activity. Table 12 contains p-values for Granger-causality tests between volatility and HP filtered trader positions. Panel 1 (VAR with 5 lags) shows that volatility Granger-causes positions (p-value 7.300%) but positions do not Granger-cause volatility (p-value 96.25%). There is also evidence that hedge fund 24 The measure of trading activity is given by 1 TAs = N N t = 1 abs ( OI OI ) where s = C_FA, C_FE, C_FF, N_BFT, N_HF; OI s,t is the open interest in day t of trading category s; N = 746, the number of days in our sample. TA C_FA = 11,055; TA N_HF = 8,910; TA C_FF = 7,599; TA N_FBT = 3,960; TA C_FE = 2, Our results contrast those of Irwin and Holt (2004) who find a positive relationship between trading volume of large hedge funds and market volatility. s, t s, t 1 16