A Study of Mutual Fund Flow and Market Return Volatility

Charles Q. Cao Department of Finance The Smeal College of Business The Pennsylvania State University University Park, PA 16802, USA Tel: (814) 865-7891 Fax: (814) 865-3362 Email: charles@loki.smeal.psu.edu and Eric C. Chang School of Business The University of Hong Kong Pokfulam Road, Hong Kong Tel: (852) 2857-8347 Fax: (852) 2858-5614 Email: ecchang@business.hku.hk and Ying Wang Department of Finance The Smeal College of Business The Pennsylvania State University University Park, PA 16802, USA Tel: (814) 863-0486 Fax: (814) 865-3362 Email: yuw105@psu.edu

This Draft: April 2003 First Draft: May 2002

Comments are welcome. Please address all comments to the corresponding author: Eric C. Chang.

A Study of Mutual Fund Flow and Market Return Volatility Abstract

In this paper we investigate the impact of institutional trading on the market by examining the daily relationship between aggregate flow into U.S. equity funds and market volatility. We examine the relationships between market volatility and fund inflow and fund outflow, respectively. Our empirical results show that an asymmetric concurrent relationship between fund flow and market volatility exists: fund inflow is negatively correlated with market volatility, whereas fund outflow is positively correlated with market volatility. We discuss potential explanations for our results and suggest that they are consistent with the notion of information content asymmetry between buy and sell orders.

A Study of Mutual Fund Flow and Market Return Volatility

1. Introduction Stock market volatility has received a great deal of attention from investors, regulators, academics, and the press. Option traders, in particular, monitor it closely, since the value of an option depends largely on the volatility of its underlying asset. The existing literature provides evidence that volatility is time-varying1 and calls for a better understanding of the factors that contribute to its variability. The popular press often quotes practitioners to suggest that increased institutional participation may account for large market price fluctuations. One article, previously cited in Sias (1996), states [r]eviewing this weeks events, analysts are concluding that professional investors simply overreacted2 The quote highlights the perception that institutional traders contribute to stock market volatility. This view is echoed by another quote: Small investors, through their purchase of stock mutual funds, have emerged as the major driving force that has propelled the Dow Jones Industrial Average.3 Notwithstanding these perceptions, we know little empirically about the actual relationship between market volatility and institutional traders, especially mutual fund traders. Recently, academics have devoted considerable attention to the price impact of mutual fund trading. Chan and Lakonishok (1993, 1995, 1997), Keim and Madhavan (1997), and Jones and Lipson (1999), for example, have examined the price effect of money inflow into a mutual fund. In general, these studies suggest that institutional trading causes both permanent and temporary

price impacts.
Warther (1995, 1998) examines the relationship between aggregate cash flow into all mutual funds and market-wide returns. These studies document a strong positive relationship between unexpected mutual fund flow and contemporaneous stock returns. He also investigates the lead-lag relationship between flow and returns, but he rejects both sides of feedback trading,

arguing that security returns neither lag nor lead mutual fund flow. Although he cannot determine causality here, he suggests a plausible causal link from flow to returns. Edelen and Warner (2001) further address this issue using higher-frequency data. They report that aggregate unexpected mutual fund flow is positively correlated with concurrent market returns at a daily frequency. They also find causality from flow to returns within the day and the one-day lagged response of aggregate flow to market returns.4 While ample empirical evidence suggests that aggregate mutual fund flow is positively related to market returns, the relationship between aggregate mutual fund flow and market volatility is not clear. Some indirect inferences can be drawn from studies examining the impact of institutional trading on stock price volatility, but the evidence is inconclusive. For example, using quarterly data, Reilly and Wachowicz (1979) examine the relationship between institutional traders (e.g., pension funds, etc.) and stock price volatility during the period 1964-1976. They claim that institutional trading actually reduces stock price volatility. However, several studies suggest the opposite. For example, Sias (1996) performs cross-sectional analysis for all stocks listed on the New York Stock Exchange (NYSE) during the period 1977-1991. He documents a positive relationship between annual change of institutional ownership and contemporaneous security volatility after controlling for capitalization. Using a sample of 500 stocks in the S&P 500 index, Xu and Malkiel (2003) document a positive cross-sectional relationship between the idiosyncratic volatility of the stocks in the index and institutional ownership from 1989 to 1996. They demonstrate that most increased idiosyncratic volatility is attributable to institutional ownership. These studies, however, usually focus on the micro level effect. They examine the relationship between the level of institutional ownership of a stock and the stocks price volatility. It is still unclear how institutional trading is related to volatility on the aggregate level empirically.

The focus of this paper is on the daily relationship between aggregate mutual fund flow and market price volatility.5 We use a unique data set on daily aggregate net mutual fund flow from February 3, 1998 to December 29, 2000, for a period of 735 days. Our flow data are from the same source used by Edelen and Warner (2001), but over a longer period. To prevent our results from being sensitive to the particular volatility estimator used, we use three estimators of daily market volatility: (1) the high-frequency volatility estimated from the intraday return data of Standard & Poor (S&P) 500 index, using the method of Andersen, Bollerslev, Diebold and Ebens (hence ABDE, 2001), (2) the high-low volatility estimator developed by Parkinson (1980), and (3) the implied volatility index based on the option of the S&P 100 index, which is quoted by Chicago Board Options Exchange (CBOE). When we use aggregate net mutual fund flow across the whole flow range to examine the relationship, the empirical results indicate a significant negative contemporaneous volatility-flow relationship. This implies that increases in mutual fund flow are associated with a less volatile market. We further examine whether the negative relationship holds for both net fund inflow and outflow. Our further investigation is mainly motivated by two facts. First, we note the marked asymmetry between the effects on stock prices of institutional buying and selling (Kraus and Stoll, 1972; Holthausen, Leftwich and Mayers, 1987, 1990; Gemmill, 1996, Keim and Madhavan, 1996; and Chan and Lakonishok, 1993, 1995.) Specifically, purchases of a stock made by institutional traders are accompanied by an increase in its price, which continues to rise after the trade. But sells of a stock are accompanied by a drop in its price, which tends to revert to its prior level. Saar (2001) develops a theoretical model to show how the trading strategies of money managers create a difference between the information content of buys and sells, and hence produce a permanent price impact asymmetry between buys and sells. In general, existing evidence suggests that the permanent price impact of institutional trading depends on whether the initiator of a transaction is a buyer or a seller. Since unexpected inflow (outflow) should stand proxy for subsequent
5

unexpected institutional buys (sells) [Keim (1999), Edelen (1999), and Edelen and Warner (2001)], it is natural to ask whether the flow-volatility relationship holds for inflow and outflow. Our investigation is also motivated by the documented asymmetric relationship between price changes and trading volume. Karpoff (1987) shows that when prices go up, volume increases, but when prices go down, volume also increases. However, previous studies that do not consider this asymmetry usually document a positive contemporaneous relationship between volume and price changes. Karpoff argues that tests of the volume-price changes relationship that do not differentiate between positive and negative price changes are misspecified because they are based on the implicit false assumption that the relationship between volume and price changes is monotonic. Hence we take into account flow direction to ensure that a similar problem does not occur in our study. An interesting finding emerges after we consider the direction of aggregate net flow. The impact of net inflow and net outflow on the market is markedly asymmetric. We show that net fund inflow is negatively correlated with market volatility, whereas net fund outflow is positively correlated with market volatility. We argue that our findings are consistent with evidence in two lines of research. First, Harris and Raviv (1993), Shalen (1993) and Delong, Shleifer, Summers and Waldman (1990a) demonstrate that informed trades are useful in helping adjust prices to new information. However, the presence of a substantial portion of uninformed investors, and hence their trades, may increase market price volatility. Second, Chan and Lakonishok (1993, 1995), Keim and Madhavan (1995) and Saar (2001) all argue that there is a difference in the information content of mutual fund managers buying and selling orders: buys tend to convey more information than sells. They argue that mutual fund managers devote substantial resources to gathering and analyzing information and make decisions based on their private information. To maximize trading performance, they search for information about stocks regardless of whether they are in their portfolios, in order to buy stocks with favorable prospects. However, due to restrictions on the use
6

of leverage and short sells, their sells are limited to the assets they already own. If individual investors tend to redeem fund shares more in a down market, the liquidity demand would result in an asymmetry whereby mutual fund managers trading on selling stocks contains less information than their trading on buying stocks. The asymmetric relationship between market volatility and fund flow documented in our paper is consistent with the asymmetric information content argument. The remainder of the paper is organized as follows. In Section 2 we review stabilization and destabilization arguments pertaining to the impact of institutional trading on the stock market. In Section 3 we discuss our mutual flow data and volatility measures. In Section 4 we present our main results using daily data. In section 5 we discuss possible explanations for the asymmetric relationship. Section 6 concludes the paper. 2. Potential effects of institutional trading on market volatility Existing literature offers mixed views on whether the increasing presence of institutional trading in the financial market stabilizes or destabilizes financial asset prices. One of the difficulties in addressing this issue is the absence of a benchmark for normal price volatility. Most existing financial theories deal with asset price determination rather than volatility. In particular, the aim of valuation theories is to offer paradigms on how to rationally determine the fundamental value of financial assets, but taking asset price volatilities and/or co-variabilities as exogenously given. Since a normal level of volatility is neither empirically observable nor theoretically determinable, whether or not a particular market force has impacted price volatility can only be indirectly inferred. In general, any market forces that tend to drive a financial assets price temporarily away from its fundamental value have been viewed as destabilizing factors, and vice versa. The rationale is straightforward. In a competitive market, any disequilibrium cannot be sustainable in the long run. Therefore, in the subsequent market correction process, a relatively large price change in absolute value is bound to occur and results in a higher volatility than it
7

would be the case without such a market force. In this section, we review several views that imply that institutional trading may either affect or correlate with stock price volatility. 2.1. Herding and positive-feedback trading Lakonishok, Shleifer, and Vishny (1992) argue that institutions may destabilize stock prices and increase market volatility through herding or positive-feedback trading. Herding occurs when money managers trade in the same direction at the same time without necessarily focusing on fundamental values. Positive-feedback trading occurs when money managers chase winners and sell losers without regard to fundamental values.6 Scharfstein and Stein (1990) favor the first view, arguing that money managers are typically evaluated against each other, not some absolute standard. To minimize the chance of becoming an outlier, money managers have incentives to follow the trading patterns of others rather than responding to their private information. Herding, motivated by this agency problem, may therefore amplify exogenous stock price shocks.7 However, Lakonishok, Shleifer, and Vishny (1992), Bikhchandani, Shilfer, and Welch (1992), and Hirshleifer, Subrahmanyam, and Titman (1994) argue that if, for example, institutional investors are better informed than individual investors, they will probably herd to undervalued stocks but away from overvalued stocks. Their trading actually moves prices toward rather than away from equilibrium values8. Chopra, Lakonishok, and Ritter (1992), Lakonishok, Shleifer, and Vishney (1994), and Brennan (1995) all suggest that institutional investors are more likely to behave rationally than individual investors. The empirical evidence in regard to this phenomenon is mixed. Sias and Starks (1997) demonstrate that the returns on portfolios dominated by institutional investors lead returns on portfolios dominated by individual investors. Wermers (1999) investigates the impact of mutual fund herding on stock prices, and shows that stocks bought by herds outperform stocks sold by herds. Both findings are consistent with the conjecture that managers herd on new information on fundamentals and help to speed up the information incorporation process. Nofsinger and Sias
8

(1999) show that analyses of post-herding returns provide no evidence that institutional herding is irrational. This is consistent with the hypothesis that institutional investors, at the margin, are better informed than other investors. However, Dennis and Strickland (2002) argue that earlier studies using quarterly or annual ownership data may not reveal herding that occurs over a shorter time interval. To circumvent this problem, they perform cross-sectional tests to investigate the relationship between ownership structure and returns of firms on event days when the absolute value of market return exceeds 2%. They find that a firms abnormal return on event days is related to the percentage of institutional ownership. They conclude that the evidence is consistent with the positive feedback behavior on the part of some institutions, particularly mutual and pension funds, which may drive asset prices away from their true values. 2.2. Investors preference Some scholars, although without asserting a causal relation, have inferred either a positive or a negative contemporaneous relationship between volatility and institutional trading based on institutional traders preferences. Kothare and Laux (1995), for example, predict a positive relation on the ground that institutional investors are attracted to more volatile stocks. Badrinath, Gay, and Kale (1989) and Arbel, Carvell, and Strebel (1983), however, suggest that institutional traders are prudent and more likely to avoid riskier (and typically smaller) stocks; they are thus associated with decreased volatility. Gompers and Metrick (2001) also suggest that large institutions, compared with other investors, prefer to invest in large, more liquid stocks. It is fair to say that existing theories offer no definitive conclusion about whether institutional trading is associated with either increased or reduced volatility. Therefore, we must appeal to empiricism to determine the relationship between institutional trading and market volatility. 3. Data and volatility measurement 3.1. Mutual fund flow data

We use data on daily net mutual fund flow from Trim Tabs (TT) financial services of Santa Rosa, CA. TT furnished us with the daily data on net asset values (NAVs) and total net assets (TNAs) for a sample of over 800 mutual funds9. The sampling period was from February 2, 1998, to December 29, 2000. The data include equity funds and bond funds, which represent approximately 15% and 12% respectively of the total funds covered by the Investment Company Institute (ICI). TT also provided us with daily net fund flow (new subscription less redemptions) based on the following formula:

Flowt = TNAt NAVt

TNAt 1 . NAVt 1

(1)

3.1.1. Fund classification and aggregation Since we want to explore the relationship between aggregate mutual fund flow and U.S. stock market volatility, our focus in this paper is on domestic equity mutual funds. Therefore, we isolate domestic equity funds from other funds. We then match the whole sample with those in the Center for Research in Security Prices (CRSP) survivor-bias free U.S. mutual fund database and classify the funds by the investment objectives. Like Warther (1995), we classify mutual funds by ICI category10 and include the funds with the following investment objectives in our sample: aggressive growth (AG), growth and income (GI), long-term growth (LG), sector funds (SF), total return (TR), utility funds (UT), income (IN) and precious metals (PM).11 The final sample contains 411 domestic equity mutual funds. 3.1.2. Data filter TT advised us that its data are prone to errors such as interchanged digits and digit transposition, because the data are collected by hand.12 Thus, we need to filter NAVs and TNAs before aggregating the daily U.S. equity mutual fund flow data. We use the same two filters as employed in Chalmers, Edelen, and Kadlec (hence CEK, 2001) to eliminate potential data error in

10

NAV and TNA series13. After the filtering, we calculate the net flow on the basis of these two series. Some funds may send one-day-old data or partially updated data (with updated NAVs but not reflecting the days fund-share transactions) to TT. However, Edelen and Warner (2001) suggest that TT include the funds in their sample only if the funds can reliably provide up-to-date daily NAVs. They employ various tests to reject the hypothesis that TT reports one-day-old data14. They caution that merely adjusting the fund's data by one day can result in severe classification errors. For this reason, Edelen and Warner (2001) make no adjustments to their data based on concern of data timeliness and argue that their decision would only strengthen their papers main conclusion: that flow-motivated trade has an aggregate price impact. Goetzmann, Ivkovic, and Rouwenhorst (hence GIR, 2001) also examine the reporting practices of each international fund in their sample. They compare TTs fund data with that in the CRSP mutual fund database. They identify 88 of 116 funds in their sample that report appropriate TNAs and 3 that report one-day-lagged TNAs. Data for the remaining 25 funds were either too noisy to allow for a determination or were not available for 1998. GIR conclude that the overwhelming majority of the funds in their sample follow the proper practice of reporting postflow TNAs. Moreover, they point out that the results obtained under the assumption that all 116 funds report timely data and those obtained from the 91 funds with clearly identifiable reporting practice are very similar. As a result, they only report results based on the whole sample. Taken together, previous studies suggest that in general TT reports timely data on mutual fund flow. Therefore, like Edelen and Warner (2001) and GIR (2001), we make no adjustments to fund flow data in this research. 3.1.3. Properties of daily aggregate mutual fund flow The dollar value of the TT asset base varied dramatically from 340 to 810 billion during our three-year sample period. Therefore, we normalize flow by expressing it as a percentage of

11

the previous days asset base. The normalized flow is defined as the one-day percentage change in TNAs, less the one-day percentage change in NAVs. Summary statistics for normalized flow data are presented in Table 1. Panel A describes the characteristics of aggregate flow by investment objectives. The daily flows of aggressive growth funds, growth and income funds, long-term growth funds, and sector funds are on average positive, while average flows of total return funds, utility funds, income growth funds, and precious metals funds are negative. Moreover, the mean of the aggregate U.S. equity mutual fund flow over the sample period is 2.94 basis points (0.0294%). There is substantial autocorrelation of the flows for all fund groups. In particular, there is statistically significant negative autocorrelation for aggregate flow at lags 1 and 2, but there is no significant autocorrelation at lag 5.15 Thus, there is no obvious week effect in aggregate flow data. The time series data of daily aggregate U.S. equity funds are depicted in Figure 1. Despite the average positive aggregate flow, we notice that among the 735 observations of aggregate net flow data, 416 are positive while 319 are negative. That is, during about 43% of trading days, aggregate cash flows out of rather than into equity mutual funds. We present the summary statistics of aggregate net inflow and outflow, respectively, in Panel B of Table 1. From the data, we observe that the aggregate mean on net inflow days (0.1183%) is greater than the absolute value of aggregate mean on net outflow days (0.0862%). 3.2. Measurement of daily volatility To examine the relationship between aggregate mutual fund flow and volatility, we next construct measures of daily market volatility. 3.2.1. Alternative volatility estimators We use three measures of daily market volatility in this study. The first is the highfrequency volatility estimator proposed by Andersen, Bollerslev, Diebold, and Ebens (hence ABDE, 2001). ABDE develop the daily volatility estimator by simply summing intraday squared

12

5-minute returns.16 They argue that the estimators thus constructed are, in theory, both free of measurement error and model free. Following ABDE (2001), We construct the five-minute return series of Standard and Poor (S&P) 500 index from the logarithmic difference between the prices recorded at or immediately before the corresponding five-minute interval. We obtain the intraday transactionlevel data from Tick Data, Inc.17 As pointed out by ABDE (2001), the use of fixed discrete time interval and the inherent bidask spread could induce the negative serial correlation in the return series and thus systematically bias the volatility measure. Therefore, much as ABDE (2001) did, we estimate an MA (1) model for the five-minute return series for the whole sample to remove the potential negative serial correlation. Consistent with ABDE (2001), we find that the estimated moving-average coefficient is 0.09818 with a t-value of 23.8. Then we construct the high-frequency volatility estimator

High,t =

(r
i =1

1/

t + i

)2

(2)

so that High,t denotes daily market volatility based on high-frequency data on day t, is the observation interval length (e.g., five minutes), 1/ is the number of observation intervals in one trading day (e.g., 79 five-minute intervals per day), and rt +i is the intraday de-meaned MA(1)filtered five-minute S&P 500 index returns in interval on day t. Two additional measures of daily market volatility are used to check the robustness of the results. The first is the extreme value estimator developed by Parkinson (1980), which is defined as

HL ,t = 0.601ln( H t / Lt )

(3)

13

where Ht and Lt, respectively, are the highest and lowest index prices on day t. The daily high and low S&P 500 prices are obtained from the Reuters Database. The second measure is the implied volatility of an option on a market index. The Chicago Board Options Exchange (CBOE) began to quote a daily implied volatility index ( VIX ) based on the option of the S&P 100 index in 1986. Our sample begins on February 2, 1998, and ends on December 29, 2000. 3.2.2. Properties and correlations of volatility estimators Summary statistics for alternative daily volatility estimators are shown in Table 2. Two observations merit discussion here. First, Panel A shows that all three volatility estimators exhibit substantial positive autocorrelation. Thus, we should control for this autocorrelation in our later tests. Second, implied volatility ( VIX ) is on average higher than volatility estimated from highfrequency intraday data ( High ) and high-low volatility ( HL ). But, from the correlations shown in Panel B, we can see that these three volatility estimators are highly correlated. The time series of three volatility estimators are depicted in Figure 2.19 4. Empirical Results 4.1. Daily flow-return relationships Edelen and Warner (2001) document a daily relationship between aggregate equity mutual fund flow and NYSE composite index returns during the period February 1998 to June 1999. Our data on flow come from the same source as theirs, but differ from theirs in that we do not have aggregate equity mutual fund flow. Instead, TT furnished us with daily data on NAVs, TNAs, and flow for a sample of about 800 funds. We then aggregated the domestic equity fund flow on our own and normalized the flow by dividing it by the previous days TNAs. Thus, in this section, we use data over a longer period to replicate their flow-return-relationship test. The purpose of this test is twofold: (1) to validate our self-constructed flow data and (2) to extend their tests to a more recent period.

14

We employ the same expected-unexpected flow decomposition that Edelen and Warner (2001) use to accomplish our replication flow-return-relationship tests. Edelen and Warner argue that, given the substantial persistence in the flow series20 and the strong dependence of flow on lagged returns21, it is necessary to explicitly separate expected and unexpected flow. Also, many previous studies [e.g., Warther (1995, 1998) and Edelen and Warner (2001)] have pointed out that returns correlate only with the unexpected component of flow but not with expected flow. We use three sample periods: (1) a whole period from February 1998 to December 2000, (2) the identical sample period (i.e., February 1998 to June 1999) used in Edelen and Warner (2001), and (3) a second sub-period from July 1999 to December 2000. All three sample periods produce qualitatively the same results, although the result based on the more recent sub-period is slightly weaker than that based on the first sub-period. For brevity, we report only the results in Table 3, based on the whole period. Table 3 is divided into two panels. Panel A reports the results of examining the dependence of fund flows on returns and past flows, and Panel B reports the dependence of returns on fund flows. In Panel A, daily flow is regressed on lagged flow and concurrent and lagged returns. As in Edelen and Warner (2001), usually several different lag values of independent variables are used in the regressions. Columns 1 and 2 of Panel A indicate that flow is closely related to lagged flow and lagged return, which is consistent with Edelen and Warner (2001). Their main focus is the concurrent flow-return relationship, which is shown in Column 3. We confirm that their results hold in an extended period and document a significantly positive concurrent flow-return relationship (coefficient 0.009 with a t-value of 2.01) after controlling for lagged flow and lagged returns. In Panel B, returns are regressed on concurrent and lagged flow using both the raw series and the expected-unexpected flow series. Column 4 presents the regression of returns on concurrent and lagged raw flow. Columns 5 and 6 show the regression of returns on expected and
15

unexpected flow. Expected daily flow is taken from the fitted values of model 2 in Panel A, and unexpected flow is actual minus expected. Again, we confirm the main finding of Edelen and Warner (2001) in this extended sample period. In particular, Column 6 shows that returns are positively related to contemporaneous unexpected flow (coefficient 0.629 with a t-value of 2.01) but not to expected flow (coefficient 0.459 with a t-value of 1.01). 4.2. Regression of volatility on aggregate flow The main focus of this paper is to investigate the concurrent daily volatility-flow relationship. In this subsection we first present our regression specifications, discuss the dependent and independent variables used, and then report the main empirical results. To examine the relationship between market volatility and aggregate mutual fund flows, our regression specifications take the following forms:

Ln( High,t ) = 0 + 1 Flowt + i Ln( High,t i ) + t

i =1

(4a)

Ln( High,t ) = 0 + 1 Flowt + 2Up _ down + i Ln( High,t i ) + t

i =1

(4b)

Ln( High,t ) = 0 + 1 Flowt + 2Up _ down + 3TVt + i Ln( High,t i ) + t

i =1

(4c)

where High,t is the daily volatility estimator based on high-frequency data on day t, Flowt is the aggregate net mutual fund flow on day t, Up _ down is a dummy variable equal to one when the market return is positive and equal to zero otherwise, and TVt is the market turnover rate on day t. The subscripts indicate the days lagged. The dependent variable in these regressions is the natural logarithm of the daily volatility estimator based on high-frequency data on day t ( Ln( High,t ) ). As mentioned, we have constructed three estimators of volatility. For brevity, we use the high-frequency volatility

16

estimator to present our main results and use the high-low-volatility and implied-volatility index as a robustness check. A discussion of the independent variables and their expected relationships with the dependent variable are as follows: Aggregate mutual fund flow ( Flowt ). The coefficient on Flowt is the main focus of this subsection, since this coefficient reflects the concurrent volatility-flow relationship. Lagged values of the natural logarithm of market volatility ( Ln( High,t i ) ). We include these values in our tests to account for persistence in market volatility (see, e.g., Bollerslev, Chou, and Kroner, 1992). Without including them as controlling regressors, there will be a positive bias on any included regressor that covaries with lagged volatilities. Thus, the estimation process will be biased and will incorrectly reflect the concurrent relationship between volatility and flow. Since existing evidence suggests that high volatility is often followed by high volatility and vice versa (see, e.g., Bollerslev, Chou, and Kroner, 1992), we expect the coefficients on these lag values to be significantly positive. The number of lagged differences to be included (i.e., the value of maximum i) is determined by the standard t-test of significance on the last lagged difference term. Up_down. This is a dummy variable equal to one when the market return is positive and equal to zero otherwise. Existing literature (see, e.g., Campbell, Koedijk, and Kofman, 2002) suggests that bear markets are associated with higher volatility than bull markets. Cox and Ross (1976) and Christie (1982), among others, document a negative correlation between market returns and volatility. Therefore, it is possible that the flow-volatility relationship thus gained in regression (4a) is only a spurious result of the positive flow-return relationship documented by Edelen and Warner

17

(2001). We include this dummy variable in regressions (4b and 4c) to determine whether meaningful changes occur after we control for changes in the volatility when the market is up that are not related to volatility to flow sensitivity. If the flowvolatility relationship still holds after we control for this dummy variable, we then have reason to believe that the concurrent daily flow-volatility relationship does not simply derive from the positive correlation between daily flow and returns. We expect the coefficient on this dummy variable to be negative. Market turnover rate ( TVt ). This rate is defined as the daily trading volume divided by shares outstanding at the end of the previous day. It is well known that trading volume is positively related to volatility [See, e.g., Karpoff (1987)]. We thus include TVt as a control variable in some versions of the tests22. Accordingly, we expect the estimated coefficient to be positively significant. Table 4 presents the main results of our volatility-to-flow-sensitivity tests. Results based on regressions (4a), (4b) and (4c) are reported in Columns 1, 2 and 3, respectively. Our discussion focuses on Model (4c). As indicated, the sign of estimated coefficients on the controlling variables are all consistent with the existing finance literature. First, the results confirm that market volatility persists over time [see, e.g., Bollerslev, Chou, and Kroner (1992)], with all coefficients on the lagged volatilities positively significant. Second, the coefficient of 0.11 on the Up_down dummy (with t-statistics of 6.1) suggests that bear markets are associated with higher volatility than bull markets. This result is consistent with the finding in Campbell, Koedijk, and Kofman (2002). Third, the coefficient on the turnover rate is significantly positive (with tstatistics of 10.32). More important for our analysis, however, is the coefficient on Flowt for all three columns. The coefficients remain negatively significant in all three cases (t-values = - 4.75, -4.50 and 5.60, respectively). The negative and significant coefficient on flow implies that increased

18

aggregate fund flow is associated with decreased market volatility. The results also indicate that adding the dummy term and the turnover rate does not materially affect the flow coefficient. 4.3. Regression of volatility on aggregate net inflow and outflow The finding that the concurrent relationship between volatility and flow is negative suggests that the larger the fund cash flow, the less volatile the market. However, as Table 1 shows, of the 735 observations of aggregate net flow data, 416 are positive and 319 are negative. In other words, 43% of the trading days over Feb. 3, 1998 to Dec. 29, 2000, collectively, involve cash flowing out of rather than into the equity mutual funds. In this section, we further investigate the question of whether the above result holds for aggregate net inflow and outflow, respectively. Our further investigation is motivated by two facts. First, existing studies document that block purchases have a larger permanent price impact than block sales. Specifically, trades induced by block transactions and institutional traders impact prices in an asymmetric manner. Whereas prices go up on buys and down on sells of block trading, they tend to revert after sells but remain high after buys. In other words, the market seems to react differently to buy and sell orders23. Saar (2001) develops a theoretical model showing that trading strategies of mutual fund managers could make buy orders convey more information than sell orders. Therefore, equilibrium prices should adjust more for buys than for sells. Since unexpected inflow (outflow) could stand proxy for subsequent unexpected institutional buys (sells) [Keim (1999), Edelen (1999), and Edelen and Warner (2001)], it is natural to ask whether inflow and outflow will have different effects on market volatility. Second, Karpoff (1987) offers a review of the relationship between trading volume and price changes and presents a model that suggests an asymmetric concurrent relationship between volume and price changes in financial markets. He points out that, as the V shape in Figure 3 illustrates, there exists a positive relationship between volume and positive price changes and a negative relationship between volume and negative price changes. Moreover, he argues that tests on linear relationships between volume and price changes per se will yield positive correlations,
19

as indicated by the slope of the dotted line that connects the midpoints of the two sides of the V in Figure 3. Morgan (1976), Rogalski (1978), Harris (1986, 1987), and Richardson, Sefcik, and Thompson (1987) all document such a positive concurrent relationship between volume and price changes. However, Karpoff (1987) asserts that tests of the volume-price change relationship that do not differentiate between positive and negative price changes could be misspecified, since they may be based on the implicit false assumption that the relationship between volume and price changes is monotonic. Our investigation of the volatility-fund flow relationship bears some similarities with that of the volatility-flow relationship. While price changes can be either positive or negative, volume can never be negative. Similarly, while flow can be either positive (net inflow) or negative (net outflow), market volatility is always positive. Moreover, our previous models are also based on the implicit assumption that the volatility-flow relationship is functional and/or monotonic. Thus, it seems reasonable to take into account the direction of flow to see if any asymmetric volatilityflow relationship exists. To examine the impact of inflow and outflow respectively on market volatility, we introduce two dummy variables ( D1 and D2 ) to differentiate between inflow and outflow. D1 is defined as 1 when aggregate net flow is positive and as 0 otherwise, and D2 is defined as 1 when aggregate net flow is negative and as 0 otherwise. We then construct two new variable series, namely Inflowt and Outflowt , with these two dummy variables. Specifically, we multiply the Flowt series in the previous models (Regressions (4a)-(4c)) by D1 to derive the new

Inflowt series. Similarly, we multiply the Flowt series by D2 to derive the new Outflowt
series. Note, however, that we actually obtain the absolute value of outflow when Flowt is multiplied by D2 , but, for brevity, we refer to it as outflow henceforth.

20

Next, we replace the Flowt series in regressions (4a)-(4c) with these two new series,

Inflowt and Outflowt , on the right-hand sides of the regressions. In addition, we include the
lagged values of market volatility, the Up _ down dummy variable defined before, and market turnover rate ( TVt ) as controlling variables in our tests. Our regression specifications take the following forms:

Ln( High,t ) = 0 + 1 Inflowt + 2 Outflowt + i Ln( High,t i ) + t

i =1 3

(5a)

Ln( High,t ) = 0 + 1 Inflowt + 2 Outflowt + 3Up _ down + i Ln( High,t i ) + t (5b)

i =1

Ln( High,t ) = 0 + 1 Inflowt + 2 Outflowt + 3Up _ down + 4TVt + i Ln( High,t i ) + t

i =1 3

(5c)

Table 5 presents the results of estimating the revised models. In Column 1, the natural logarithm of daily volatility is regressed on inflow and outflow respectively after controlling for the persistence of volatility. In Column 2, we further control for the Up _ down dummy variable, as previously discussed. We also include the market turnover rate in Column 3 as an explanatory variable. The results in Table 5 confirm that relationships between volatility and control variables remain unchanged, as documented in Table 4. Specifically, significant and positive coefficients are found on lagged volatilities and market turnover rate while significant and negative coefficients on the Up _ down dummy variable are documented. The most important results, however, are the coefficients on Inflowt and Outflowt . The three columns in which these coefficients are shown display a consistent pattern of significantly negative coefficients on Inflowt and significantly positive coefficients on Outflowt . For

21

example, Column 1 shows that market volatility is negatively correlated with aggregate net cash flow into mutual funds with t-statistics of 3.32 and positively correlated with aggregate outflow (actually the absolute value of outflow, as mentioned) with t-statistics of 2.39. These results do not materially change even after controlling for the impact of market returns and trading volume on market volatility. Table 5 reveals an interesting finding: an asymmetric concurrent correlation between volatility and flow, depending on the direction of flow. We find that fund inflow is negatively related to market volatility, and that fund outflow is positively related to market volatility. That is, the larger the aggregate cash flow into the mutual funds, the less volatile the market. On the other hand, the larger the aggregate cash flow out of the mutual funds, the more volatile the market. We illustrate this asymmetry in Figure 4. The slopes of solid lines in Figure 4 roughly illustrate the asymmetric relationship between volatility and fund flows. Moreover, the slope of the dotted line that connects the midpoints of the two solid lines of V in the figure suggests that a negative correlation is likely to be detected when the volatility-flow relationship is examined across a whole spectrum of flow range. This is in fact the case; as indicated in Table 4, we detect significantly negative concurrent relationships between volatility and flow. 4.4. Robustness tests 4.4.1. Tests of outliers Our sample covers 411 U.S. equity mutual funds, including funds with various investment objectives such as aggressive growth and precious metals. The fund flow patterns are diversified. For example, whereas average flows of funds with the investment objectives of AG, GI, LG, and SF are positive, average flows of funds including TR, UT, IN, and PM are negative (see Panel A of Table 1). We perform robust tests to ensure that the results based on aggregate fund flows are not driven by a few outliers.

22

To test this possibility, we construct a new variable, dispersion rate ( DRt ), as follows:

DRt =

M t Nt Tt

(6)

where Mt refers to the number of funds whose flow on day t is positive, Nt is the number of funds whose flow on day t is negative, and Tt is the total number of funds on day t. Table 6 shows the summary statistics of DR. Among 735 observations of DR, 250 observations are positive with the mean of 0.119, and 485 are negative with the mean of 0.177. In addition, the correlation between flow and the DR variable across the whole time period is 0.63. We perform similar tests on DR to examine whether any outliers induce this asymmetric volatility-flow relationship. First, we introduce two new dummy variables: D3 is defined as 1 when DRt is positive and as 0 otherwise; D4 is defined as 1 when DRt is negative and as 0 otherwise. Next, we construct two new series, DRt _ p and DRt _ n , with these two dummy variables. Specifically, DRt _ p is DRt multiplied by D3 , and DRt _ n is DRt multiplied by

D4 . The models are specified as follows:

Ln( High,t ) = 0 + 1 DRt _ p + 2 DRt _ n + i Ln( High ,t i ) + t
i =1 3 3

(7a)

Ln( High,t ) = 0 + 1 DRt _ p + 2 DRt _ n + 3Up _ down + i Ln( High ,t i ) + t (7b)

i =1

Ln( High,t ) = 0 + 1 DRt _ p + 2 DRt _ n + 3Up _ down + 4TVt + i Ln( High ,t i ) + t

i =1 3

(7c)

The results are shown in Table 7. In Column 1, we regress natural logarithm of volatility on positive DRt and absolute value of negative DRt respectively, after controlling for the

23

persistence of volatility. In Column 2, we also control for the Up _ down dummy variable. Turnover rate is included in Column 3 as a controlling variable. The results confirm the asymmetric pattern reported in Table 5. Specifically, there is a negative relationship between concurrent volatility and positive dispersion rate, and there is a positive relationship between concurrent volatility and absolute value of negative dispersion rate. The results imply that a day where a higher number of mutual funds experience net inflow is more likely to be one of a lower market volatility. 4.4.2. Alternative volatility estimators In this section, we repeat our tests using two previously discussed alternative volatility estimators: high-low volatility and implied volatility index. We report the results in Table 8.24 Table 8 illustrates that the results using two alternative volatility estimators are similar to the previously presented results.25 In both cases, we find that fund inflow is negatively related to market volatility and fund outflow is positively related to market volatility. We conclude that our finding of an asymmetric relationship between volatility and flow is not driven by the particular volatility estimator used. 4.4.3. Sub-time periods tests We also divide our sample period into two sub-time periods: one from February 1998 to June 199926 and the other from July 1999 to December 2000. We then replicate all the tests for the two periods.27 Table 9 presents the results. The results are largely consistent with our previous results. For example, we obtain inflow estimates of 33.4 with the t statistic of 2.94 and 38.1 with the t statistic of 2.83 for two sub-time periods. At the same time, the outflow estimates during two sub-time periods are 25.5 with the t statistic of 1.93 and 63.1 with the t statistic of 2.33, respectively. 4.4.4. Other tests

24

We perform two additional tests. First, we include global equity (GE) and balanced funds (BL) in our sample to check whether the results are sensitive to these specific types of fund. Second, we repeat the analysis, excluding the December data from our sample, to check the dividend effect. Our data providers advised us that determining flow in the presence of dividends is pure guesswork because mutual funds do not handle distributions in a uniform manner. Since most distributions happen in December, for the sake of a robustness check, we re-estimate our models by discarding December data from the sample to see whether a nontrivial change occurs. In both cases, we find that the results remain qualitatively the same28. 5. Discussion In this section we attempt to offer an explanation for the documented asymmetric relationship between volatility and mutual fund flows. We first establish a link between market volatility and the information content of a trade. We then discuss a possible asymmetry in the information content between the trades that induce an average aggregate fund inflow and fund outflow. Harris and Raviv (1993) and Shalen (1993) demonstrate that the presence of a substantial portion of uninformed investors in the markets, and hence their trades, may increase market price volatility. They show that a wider dispersion of beliefs (that is, a wider dispersion of expectations in the current and preceding rounds of trade) creates excess price variability relative to equilibrium value. In particular, Shalen (1993) points out that uninformed (or less-informed) investors have difficulty interpreting the noisy signals of price change. They cannot differentiate between short-term random liquidity demand and overall fundamental changes in supply and demand. This could result in a general discrepancy in the true price embodied in revealed information. Moreover, uninformed investors tend to revise their beliefs more frequently than their informed counterparts after new information becomes publicly available, resulting in the slower disappearance of price fluctuations from their trading. In these ways, uninformed investors (and hence their trades) are more likely to overreact to fundamental price movements, which lead
25

to increased price volatility. Informed trades, however, are arguably useful in helping adjust prices to new fundamentals, and thus are immune from these problems. Next, we offer three arguments that suggest that, on average, mutual fund inflow may contain better information than outflow. The first argument comes from the joint nature of fund flow decisions by individual investors and fund managers. To buy or redeem mutual fund shares is basically an exclusive decision made by individual investors. However, mutual fund money managers proactive trading strategies (e.g., market timing, stock selection) may also play a crucial role in determining the timing of fund flow. Therefore, it is fair to say that the timing and the magnitude of fund flow into and out of the market are generally determined jointly by individual investors and fund managers. However, mutual funds are required by law to entertain share redemption requests by individual investors on a daily basis. The obligation to accommodate the liquidity, as pointed out by Edelen (1999), forces mutual fund managers to engage at times in a substantial amount of uninformed trading. A large fund outflow day could be a day characterized by substantial redemption on the part of individual investors. It is expected that the relative decision role fund managers play on such a day is less significant than that on a normal day. We hence argue that, on average, mutual fund inflow contains better information than outflow does. Chan and Lakonishok (1993) offer the second argument on the difference in institutional trades information content. They argue that since an institutional investor typically does not hold the market portfolio, the choice of a particular issue to sell, out of the limited alternatives in a portfolio, does not necessarily convey negative information. Rather, the stocks that are sold may already have met the portfolios objectives, or there may be other mechanical rules, unrelated to expectations about future performance, for reducing a position. As a result, there are many liquidity-motivated reasons to dispose of a stock. In contrast, the choice of one specific issue to buy, out of the numerous possibilities on the market, is likely to convey favorable firm-specific

26

news. Implied in this argument also is the suggestion that institutions buy orders convey more information than their sell orders Saar (2001) develops a model which provides the third argument that the trading strategy of mutual fund managers creates a difference between the information content of buys and sells. The asymmetry is mainly driven by two factors: (1) portfolio managers ability to gather, analyze, and optimally use private information and (2) a set of trading constraints (e.g., restriction on the use of leverage and short sells) that portfolio managers face. In particular, fund managers profitmaximizing trading strategies propel them to search for information about stocks not in their portfolios in order to buy stocks with favorable information, as well as about stocks already in their portfolios, in order to sell stocks whose price is expected to drop or on which there is no special information. Saar (2001) argues that such a dynamic strategy creates a difference between the information content of buys and sells, if the market knows that institutional investors are informed investors about the prospects of stocks29. Our findings of an asymmetric concurrent volatility-flow relationship are consistent with the differences of information content of mutual fund managers buying and selling behavior discussed above. When individual investors buy mutual fund shares, large sums of cash flow into mutual funds and induce mutual fund managers to invest money in a diversified portfolio. As pointed out by Saar (2001), mutual fund managers devote substantial resources to gathering and analyzing private information and to making selection and timing decisions based on their research departments predictions and recommendations. These trades on average contribute to adjusting prices to new fundamentals. We thus expect to see a negative relationship between market volatility and aggregate mutual fund flow. That is consistent with our finding that higher fund inflow is associated with a less-volatile market. In contrast, when individual investors redeem mutual fund shares, large sums of money exit the mutual funds. As mentioned previously, mutual funds are subject to some operational constraints (Saar, 2001). For example, use of leverage is restricted, and short sells are forbidden.
27

Hence fund managers are forced to sell stocks from their portfolios to satisfy the liquidity requirement of individual investors more often than they are forced to buy stocks. Such lessinformed trades may drive the prices away from the assets fundamental values. This is also consistent with our findings: the greater the aggregate mutual fund outflow, the more volatile the market. In a word, we link the impact of individual investors and mutual fund managers in a unified framework and provide a possible explanation for our findings. 6. Conclusion In this study, we use daily data to directly investigate the concurrent relationship between aggregate mutual fund flow and market volatility. Our high-frequency data enable us to conduct rigorous tests that offer new evidence. The study sheds light on the impact of institutional trading on the market and should be important to investors, practitioners, academics, and regulators. Our initial evidence suggests a significantly negative contemporaneous relationship between volatility and aggregate net mutual fund flow across the whole flow range. However, we gain additional insight into this issue by separating net inflow and outflow data. An important finding emerges after we take into account the direction of aggregate net flow: there is a marked asymmetry between the impact of inflow and the impact of outflow on the market. Increases in aggregate net inflow are associated with a less-volatile market, while increases in aggregate net outflow are associated with more-volatile market. We discuss the possible explanations for our findings of an asymmetric concurrent volatility-flow relationship and suggest the joint roles played by individual investors and mutual fund managers in the market. Our results are consistent with the differences of information content of mutual funds buys and sells.

28

Endnotes
For example, based on monthly observations, Schwert (1989) reports that stock volatility varied substantially during the period 18571987. Using daily return data, Haugen, Talmor, and Torous (1991) document a large variation in volatility during the period 18971988. Wood, McInish, and Ord (1985), among others, examine intraday market returns and show that market volatility is high at the beginning and the end of the trading day.
2 3 4

See Wall Street Journal (WSJ), July 21, 1995, p. A1. See WSJ, February 26, 1993, p. C1.

However, unlike Warther (1995), who finds a high correlation (R2=55%) between monthly flow and returns, they show that variation in aggregate flow accounts for only 3% of the variation of daily market index returns, thus providing limited evidence for the common view that mutual fund flow causes movement of security prices. Warther (1998) does question whether increased mutual fund flow leads to increased market instability, but he provides no empirical evidence that directly addresses the flow-volatility-relationship question.
6
5

See also De Long, Shleifer, Summers, and Waldmann (1990a).

Investors who follow either of these trading strategies can be included in the class of noise traders proposed by Delong, Shleifer, Summers and Waldmann (1990b). They show that the unpredictability of noise traders beliefs creates excessive risk, resulting, in turn, in a significant divergence of stock prices from fundamental values
8

Lakonishok, Shleifer, and Vishny (1992) also caution that herding and positive-feedback trading phenomena per se do not necessarily destabilize the market. Institutions might appear to irrationally herd if they all react to the same fundamental information in time or if they all counter the same irrational moves in individual investor sentiment.

TNA is the current market value of all the funds assets minus its liabilities. NAV is the price per share at which shares are redeemed; it is defined as TNA divided by the total number of shares outstanding.
10 11

Refer to ICI Mutual Fund Factbook, 2001, pp. 36.

Unlike Warther, we exclude international equities funds from our sample, given that we are concerned with U.S. domestic equity market volatility. This is also consistent with Edelen and Warner (2001). However, it is debatable whether global equity (GE) and balanced (BL) funds should be included in our sample, since GE funds invest in both U.S. and international equities, and BL funds invest in a mix of equity securities and bonds. Hence we will perform a robustness check by including GE and BL funds to see if the results are, in particular, sensitive to them. Greene and Hodges (2002), Goetzmann, Ivkovic, and Rouwenhorst (2001), and Chalmers, Edelen, and Kadlec (2001) also address this issue. The first is a five-standard-deviation filter intending to eliminate outliers due to typos. The second is a filter designed to catch potential false reversals. For details, please see CEK (2001).

12

13

14

Although they cannot reject the partially updated data hypothesis, they also note that the tests power is limited by the availability of semiannual Security Exchange Commission (SEC) reports. Furthermore, even

29

if a fund shows a one-day reporting lag based on comparisons of SEC and TT data, it is unclear whether there is also a one-day lag for all other days.
15 16

Statistics on flow data are consistent with Edelen and Warner (2001).

ABDL (2001) justify using a sampling frequency of five minutes, stating that this amount of time is long enough to avoid most measurement errors and short enough to avoid microstructure biases.

Tick Data, Inc. provided tick-by-tick data on 1500 U.S. equities, 60 futures, and 7 most popular cash indices. We refer the interested reader to www.tickdata.com for a more complete description of the actual data and the method of data capture.
18

17

The negative serial correlation may suggest the spurious dependence induced by non-synchronous trading and bid-ask bounce effects.

19

To save space, we present our results mainly by using high-frequency volatility. Also, to prevent our inferences from being sensitive to the particular estimators used, we repeat all the analyses with the highlow-volatility and implied-volatility index time series. Table 1, Panel A shows that flow is highly predictable. This is shown in Table 3, Panel A.

20 21 22

Edelen and Warner (2001) suggest that unexpected flow should stand proxy for subsequent unexpected institutional trading volume. Edelen (1999) documents a positive relationship between gross flow (half of the sum of inflow and outflow) and trading volume; however, he argues that a positive relationship between net flow (inflow minus outflow) and trading volume does not necessarily follow.

23

See, for example, Kraus and Stoll (1972), Holthausen, Leftwich and Mayers (1987, 1990), Chan and Lakonishok (1993), Gemmill(1996), and Keim and Madhavan (1996) For brevity, we only show the results for the most comprehensive models. Nevertheless, the results are slightly weaker than those using high-frequency volatility. The time period is consistent with that used in Edelen and Warner (2001). For brevity, we only show the most comprehensive models using high-frequency volatility. To save space, these results are not reported.

24 25 26 27 28 29

Keim and Madhavan (1995) also suggest that the information content of buys is greater than that of sells: Following a buy decision, institutional traders can choose among various potential assets; however, their sells are limited to the assets they already own, owing to the short sells restriction. Keim and Madhavan (1997) also show that trading costs for buyer-initiated trades exceed seller-initiated trades, and their findings are consistent with the differences between the information content of buys and sells.

30

References Andersen, T.G., Bollerslev, T., Diebold, F.X., Ebens, H., 2001. The distribution of realized stock return volatility. Journal of Financial Economics 61, 43-76. Andersen, T.G., Bollerslev, T., Diebold, F.X., Labys, P., 2001. The distribution of realized exchange rate volatility. Journal of the American Statistical Association 96, 42-55. Arbel, A., Carvell, S., Strebel, P., 1983. Giraffes, institutions and neglected firms. Financial Analysts Journal 39, 55-63. Badrinath, S.G., Gay, G.D., Kale, J.R., 1989. Patterns of institutional investment, prudence, and the managerial safety-net hypothesis. Journal of risk and insurance 55, 605-629. Bikhchandani, S., Hirshleifer, D., Welch, I., 1992. A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy 100, 992-1026. Bollerslev, T., Chou, R.Y., Kroner, K.F., 1992. ARCH modeling in finance: a review of the theory and empirical evidence. Journal of Econometrics 52, 5-59. Brennan, M.J., 1995. The individual investor. Journal of Financial Research 18, 59-74. Brown, K.C., Brooke, B.A., 1993. Institutional demand and security price pressure: the case of corporate spinoffs. Financial Analysts Journal 49, 53-62. Busse, J.A., 1999. Volatility timing in mutual funds: evidence from daily returns. Review of Financial Studies 12, 1009-1041. Campbell, J.Y., 1987. Stock returns and the term structure. Journal of Financial Economics 18, 373-399. Campbell, R., Koedijk, K., Kofman, P., 2002. Increased correlation in bear markets: a downside risk perspective. Financial Analysts Journal 58, 87-94. Chalmers, J.M.R., Edelen, R.M., Kadlec G.B., 2001. On the perils of financial intermediaries setting security prices: the mutual fund wild card option. Journal of finance 56, 2209-2236. Chan, L., Lakonishok, J., 1993. Institutional trades and intraday stock price behavior. Journal of Financial Economics 33, 173-200. Chan, L., Lakonishok, J., 1995. The behavior of stock prices around institutional trades. Journal of Finance 50, 1147-1174. Chan, L., Lakonishok, J., 1997. Institutional equity trading costs: NYSE versus NASDAQ. Journal of Finance 52, 713-735. Chopra, N., Lakonishok, J., Ritter, J., 1992. Measuring abnormal performance: do stocks overreact? Journal of Financial Economics 49, 235-268. Christie, A.A., 1982. The stochastic behavior of common stock variances: value, leverage and interest rate effects. Journal of Financial Economics 10, 407-432. 31

Cox, J.C., Ross, S.A., 1976. The valuation of options for alternative stochastic processes. Journal of Financial Economics 10, 407-432. DeLong, J.B., Shleifer, A., Summers, L.H., Waldmann, R.J., 1990a. Noise trader risk in financial markets. Journal of Political Economy 98, 703-739. DeLong, J.B., Shleifer, A., Summers, L.H., Waldmann, R.J., 1990b. Positive feedback investment strategies and destabilizing rational speculation. Journal of Finance 45, 379-395. Dennis, P.J., Strickland, D., 2002. Who blinks in volatile markets, individuals or institutions? Journal of Finance 57, 1923-1949. Edelen, R.M., 1999. Investor flows and the assessed performance of open-end mutual funds. Journal of Financial Economics 53, 439-466. Edelen, R.M., Warner, J.B., 2001. Aggregate price effects of institutional trading: a study of mutual fund flow and market returns. Journal of Financial Economics 59, 195-220. Ferson, W.E., Warther, V.A., 1996. Evaluating fund performance in a dynamic market. Financial Analysts Journal 52, 20-28. French, K.R., Schwert, G.W., Stambaugh, R.F., 1987. Expected stock returns and volatility. Journal of Financial Economics 19, 3-30. Gemmill, G., 1996. Transparency and liquidity: a study of block trades on the London Stock Exchange under different publication rules. Journal of Finance 51, 1765-1790. Glosten, L.R., Jagannathan, R., Runkle, D.E., 1993. On the relation between the expected value and the volatility of the nominal excess return on stocks. Journal of Finance 48, 1779-1801. Goetzmann, W.N., Ivkovic, Z., Rouwenhorst, K.G., 2001. Day trading international mutual funds: evidence and policy solutions. Journal of Financial and Quantitative Analysis 36,287-309. Gompers, P., Metrick, A., 2001. Institutional investors and equity prices. Quarterly Journal of Economics 116, 229-259. Greene, J.T., Hodges, C.W., 2002. The dilution impact of daily fund flows on open-end mutual funds. Journal of Financial Economics 65, 131-158. Harris, L., 1986. Cross-security tests of the mixture of distributions hypothesis. Journal of Financial and Quantitative Analysis 21, 39-46. Harris, L., 1987. Transactions data tests of the mixture of distributions hypothesis. Journal of Financial and Quantitative Analysis 22, 127-141. Harris, M., Artur, R., 1993. Differences of opinion make a horse race. Review of Financial Studies 6, 473-506. Haugen, R.A., Talmor, E., Torous, W.N., 1991. The effect of volatility changes on the level of stock prices and subsequent expected returns. Journal of Finance 46, 985-1007. 32

Hirshleifer, D., Subrahmanyam, A., Titman, S., 1994. Security analysis and trading patterns when some investors receive information before others. Journal of Finance 49, 1665-1698. Holthausen, R.W., Leftwich, R.W., Mayers, D., 1987. The effect of large block transactions on security prices. Journal of Financial Economics 19, 237-267. Holthausen, R.W., Leftwich, R.W., Mayers, D., 1990. Large-block transactions, the speed of response, and temporary and permanent stock-price effects. Journal of Financial Economics 26, 71-95. Jones, C., Lipson, M., 1999. Execution costs of institutional equity orders. Journal of Financial Intermediation 8, 123-140. Karpoff, J.M., 1987. The relation between price changes and trading volume: a survey. Journal of Financial and Quantitative Analysis 22, 109-126. Keim, D.B., Madhavan, A., 1995. Anatomy of the trading process: empirical evidence on the behavior of institutional trades. Journal of Financial Economics 37, 371-398. Keim, D.B., Madhavan, A., 1996. The upstairs market for large-block transactions: analysis and measurement of price effects. Review of Financial Studies 9, 1-36. Keim, D.B., Madhavan, A., 1997. Transactions costs and investment style: an inter-exchange analysis of institutional equity trades. Journal of Financial Economics 46, 103-131. Kothare, M., Laux, P.A., 1995. Trading costs and the trading systems for Nasdaq stocks. Financial Analysts Journal 51, 42-53. Kraus, A., Stoll, H.R., 1972. Price impacts of block trading on the New York Stock Exchange. Journal of Finance 27, 569-588. Lakonishok, J., Shleifer, A., Vishny, R.W., 1992. The impact of institutional trading on stock prices. Journal of Financial Economics 32, 23-44. Lakonishok, J., Shleifer, A., Vishny, R.W., 1994. Contrarian investment, extrapolation, and risk. Journal of Finance 49, 1541-1578. Morgan, I.G., 1976. Stock prices and heteroskedasticity. Journal of Business 49, 496-508. Nofsinger, J.R., Sias, R.W., 1999. Herding and feedback trading by institutional and individual investors. Journal of Finance 54, 2263-2295. Parkinson, M., 1980. The extreme value method for estimating the variance of the rate return. Journal of Business 53, 61-66. Reilly, F.K., Wachowicz, J.M., Jr., 1979. How institutional trading reduces market volatility. Journal of Portfolio Management 5, 11-17. Richardson, G., Sefcik, S.E., Thompson, R., 1987. A test of dividend irrelevance using volume reaction to a change in dividend policy. Journal of Financial Economics 17, 313-333. 33

Rogalski, R.J., 1978. The dependence of prices and volume. Review of Economics and Statistics 36, 268-274. Saar, G., 2001. Price impact asymmetry of block trades: an institutional trading explanation. Review of Financial Studies 14, 1153-1181. Scharfstein, D.S., Stein, J.C., 1990. Herd behavior and investment. American Economic Review 80, 465-479. Schwert, G.W., 1989. Why does stock market volatility change over time? Journal of Finance 44, 1115-1153. Shalen, C.T., 1993. Volume, volatility, and the dispersion of beliefs. Review of Financial Studies 6, 405-434. Sias, R.W., Starks, L.T., 1997. Return autocorrelation and institutional investors. Journal of Financial Economics 46, 103-131. Sias, R.W., 1996. Volatility and the institutional investor. Financial Analysts Journal 52, 13-20. Warther, V., 1998. Has the rise of mutual funds increased market instability? Brookings-Wharton Papers on Financial Services, 239-262. Warther, V., 1995. Aggregate mutual fund flows and security returns. Journal of Financial Economics 39, 209-235. Wermers, R., 1999. Mutual fund herding and the impact on stock prices. Journal of Finance 54, 581-622. Whitelaw, R.F., 1994. Time variations and covariations in the expectation and volatility of stock market returns. Journal of Finance 49, 515-541. Wood, R.A., McInish, T.H., Ord, J.K., 1985. An investigation of transactions data for NYSE stocks. Journal of Finance 40, 723-741. Xu, Y., Malkiel, B.G., 2003. Investigating the behavior of idiosyncratic volatility. Journal of Business, forthcoming. Zweig, M., 1973. An investor expectations stock price predictive model using closed-end fund premiums. Journal of Finance 28, 67-78.

34

Table 1 Summary statistics for daily mutual fund flow Our data on daily flow (new subscriptions less redemptions), NAVs, and TNAs come from TT financial services of Santa Rosa, CA. We match the whole sample of about 850 mutual funds in TT with those in the CRSP survivor-bias-free U.S. mutual fund database and classify the mutual funds by the investment objectives defined by ICI. Included in our sample of all U.S. equity funds are funds from aggressive growth to precious metals. We apply two filters described in detail in the main text, the absolute-value filter and the reversal filter, to the TNA and NAV series, and we aggregate the two series. Flow is defined as the one-day percentage change in aggregate TNAs less the one-day percentage change in aggregate NAVs. Distributions are not accounted for in these data. Time period: 2/3/9812/29/00 (735 observations) Sample: 411 U.S. equity funds Panel A. Univariate statistics and autocorrelations of fund flow Fund investment objective Aggressive growth Growth and income Long-term growth Sector funds Total return Utility funds Income Precious metals All U.S. equity funds Mean (b.p.) 4.54 1.33 3.63 3.44 -4.44 -1.91 -1.68 -4.77 2.94 Median (b.p.) 3.54 1.11 2.57 1.33 -4.35 -2.97 -2.18 -21.3 1.63 Std. dev. (b.p.) 25.6 8.9 23.3 40.0 18.2 38.0 13.0 198.0 15.6 Std.err of mean (b.p.) 0.95 0.33 0.86 1.48 0.67 1.41 0.48 7.33 0.58 Autocorrelations Lag1 -0.023* -0.193* -0.065* -0.223* -0.076* -0.361* -0.076* -0.147* -0.091* Lag2 -0.144* 0.051 -0.310* -0.075* 0.031 0.009* 0.071 -0.237* -0.227* Lag5 0.094 0.129* 0.003 -0.012 0.122* 0.041 0.140* -0.034 0.060

Panel B. Univariate statistics of aggregate net inflow and outflow Obs. Aggregate net inflow 416 Mean (b.p.) 11.83 Median (b.p.) 9.00 -6.69 Std. dev. (b.p.) 12.8 10.5 Std. err of mean (b.p.) 0.63 0.59

Aggregate net 319 -8.62 outflow * Significant at 0.05 level, two-tailed test.

35

Table 2 Summary statistics for alternative volatility estimators Panel A shows the univariate statistics and autocorrelations of market high-frequency volatility ( High ), high-low volatility ( HL ), and implied volatility ( VIX ). the methods of Parkinson (1980).

High

is calculated from S&P 500 index five-

minute intraday returns using the estimator of ABDL (2001) and ABDE (2001).

VIX

HL

is calculated using

is the implied volatility index based on the option of the S&P 100

index quoted by the CBOE. Panel B shows the correlations of these three volatility estimators. t-statistics are in parentheses. Time period: 2/3/9812/29/00 (735 observations) Panel A. Univariate statistics and autocorrelations of volatility estimators Autocorrelations Lag 1 0.624* 0.368* 0.931* Lag 2 0.525* 0.349* 0.877 Lag 3 0.480* 0.286* 0.832

Mean (%)

Median (%) 15.2 14.0 24.7

Std.dev. (%) 6.8 8.2 5.1

Std.err. of mean (%) 0.25 0.30 0.19

High HL VIX

16.3 15.6 25.7

Panel B. Correlations of volatility estimators

High High HL VIX

1.0000 0.8059* (<0.0001) 0.6999* (<0.0001)

HL

VIX

1.0000 0.5909* (<0.0001) 1.0000

* Significant at 0.05 level, two-tailed test.

36

Table 3 Contemporaneous relationships between returns and flow In Panel A, daily flow (Flowt) is regressed on current and past observations of market returns of the NYSE index (Returnt-i) and past observations of flow (Flowt-i). In Panel B, daily returns of the NYSE index are regressed on concurrent and lagged daily flow in Column 4 and on concurrent and lagged unexpected daily flow (Uflowt-i) and concurrent expected daily flow (Eflowt) in Columns 5 and 6. Expected daily flow is taken from Panel A, Column 2. Unexpected flow is actual minus expected. The subscripts indicate the days lagged. t-statistics are in parentheses. Time Period: 2/3/9812/29/00(735 observations) Sample: 411 U.S. Equity funds Panel A. Flow dependence on returns and past flow 1 Coefficient on: Intercept Returnt Returnt-1 Returnt-2 Returnt-3 Flowt-1 Flowt-2 0.00031* (6.2) 0.068* (15.1) -0.036* (-5.9) -0.010 (-1.87) 0.00031* (8.2) 0.069* (15.6) -0.035* (-8.0) -0.009 (-1.67) -0.077* (-2.11) -0.220* (-6.03) 32.3% Exp.-Unexp. flow Coefficient on: 0.00012 (0.25) 0. 317* (1.97) 0.021 (0.08) -0.038 (-0.14) -0.00002 (-0.00) -0.119 (-0.44) 0.115 (0.43) 1.0% Intercept Uflowt Uflowt-1 Uflowt-2 Uflowt-3 Uflowt-4 Uflowt-5 Eflowt R
2

3 0.00030* (8.1) 0.009* (2.01) 0.066* (15.1) -0.038* (-8.5) -0.077* (-2.12) -0.225* (-6.18) 33.3% 5 0.00043 (0.01) 0. 663* (2.48) 0.006 (0.02) 0.273 (0.81) -0.325 (-1.03) -0.089 (-0.28) 0.166 (0.52) 0.483 (0.98) 2.3% 6 0.00013 (1.07) 0.629* (2.01) 0.459 (1.01) 1.9%

R2 28.5% Panel B. Returns dependence on flow Raw flow Coefficient on: Intercept Flowt Flowt-1 Flowt-2 Flowt-3 Flowt-4 Flowt-5 4

* Significant at 0.05 level, two-tailed test.

37

Table 4 Regressions of high-frequency volatility on flow Daily high-frequency volatility (ln( High )) is regressed on concurrent aggregate domestic equity fund
,t

flow in Column 1, after controlling for the persistence of volatility. Column 2 controls for a dummy variable ( Up _ down ). Column 3 also controls for market turnover rate ( TVt ) in addition to Up _ down .

High ,t

is calculated from S&P 500 index five-minute intraday returns using the estimator of ABDL (2001)

and ABDE (2001). Up _ down is defined as 1 when the S&P 500 daily return is positive and as 0 otherwise. TVt is the daily trading volume scaled by shares outstanding at the end of the previous day. The subscripts indicate the days lagged. t-statistics are in parentheses. Time period: 2/3/9812/29/00 (735 observations) Sample: 411 U.S. Equity funds 1 Coefficient on: Intercept -0.389* (-6.56) -31.530* (-4.75) ) 0.395* (10.94) 0.236* (6.10) 0.157* (4.36) 48.14% -0.318* (-5.38) -29.228* (-4.50) -0.120* (-6.00) 0.404* (11.44) 0.240* (6.33) 0.150* (4.27) 50.52% -1.049* (-11.68) -34.099* (-5.60) -0.114* (-6.08) 88.165* (10.32) 0.339* (10.09) 0.036* (5.88) 0.128* (3.89) 56.80% 2 3

Flowt

Up _ down
TVt Ln( High Ln( High Ln( High Adj R2

,t 1

,t 2

,t 3

* Significant at 0.05 level, two-tailed test.

38

Table 5 Regressions of high-frequency volatility on aggregate net inflow and outflow Daily high-frequency volatility (ln( High )) is regressed on concurrent aggregate net inflow ( Inflowt )
,t

and outflow ( Outflowt ) in Column 1, after controlling for the persistence of volatility. Column 2 controls for a dummy variable ( Up _ down ). Column 3 also controls for market turnover rate ( TVt ) in addition to

Up _ down . High ,t is calculated from S&P 500 index five-minute intraday returns using the estimator of
ABDL (2001) and ABDE (2001). Inflowt and Outflowt are two series representing net inflow and net outflow, respectively. Up _ down is defined as 1 when the S&P 500 daily return is positive and as 0 otherwise. TVt is the daily trading volume scaled by shares outstanding at the end of the previous day. The subscripts indicate the days lagged. t-statistics are in parentheses. Time period: 2/3/9812/29/00 (735 observations) Sample: 411 U.S. Equity funds 1 Coefficient on: Intercept -0.389* (-6.45) -31.647* (-3.32) 31.335* (2.39) ) 0.395* (10.93) 0.236* (6.09) 0.157* (4.36) 48.07% -0.315* (-5.23) -31.215* (-3.35) 25.934* (2.02) -0.121* (-6.01) 0.404* (11.43) 0.240* (6.33) 0.150* (4.27) 50.46% -1.296* (-13.83) -36.819* (-4.37) 28.035* (2.42) -0.115* (-6.32) 110.740* (12.85) 0.317* (9.70) 0.200* (5.82) 0.127* (3.99) 59.61% 2 3

Inflowt Outflowt Up _ down

TVt Ln( High Ln( High Ln( High Adj R2

,t 1

,t 2

,t 3

* Significant at 0.05 level, two-tailed test.

39

Table 6 Dispersion rate statistics Dispersion rate ( DRt ) is calculated by dividing the difference between the number of funds with positive flow and the number of funds with negative flow on day t by the total number of funds on the same day. Statistics of DRt _ P ( DRt > 0 ) and DRt _ N ( DRt < 0 ) are also shown in this table. Time period: 2/3/9812/29/00 (735 observations) Obs. Mean Median Std. dev. Std. err of mean 0.0064

DRt DRt _ P DRt _ N

735

-0.077

-0.091

0.173

250

0.119

0.113

0.081

0.0051

485

-0.177

-0.165

0.110

0.0050

40

Table 7 Regression of high-frequency volatility on dispersion rate Daily high-frequency volatility (ln( High )) is regressed on positive ( DRt _ P ) and negative dispersion
,t

rate ( DRt _ N ) in Column 1, after controlling for the persistence of volatility. Column 2 controls for a dummy variable ( Up _ down ). Column 3 also controls for market turnover rate ( TVt ) in addition to

Up _ down . High ,t is calculated from S&P 500 index five-minute intraday returns using the estimator of
ABDL (2001) and ABDE (2001). DRt _ P and DRt _ N are positive and negative DRt , respectively, where DRt is calculated by dividing the difference between the number of funds with positive flow and the number of funds with negative flow on day t by the total number of funds on the same day. Up _ down is defined as 1 when the daily S&P 500 return is positive and as 0 otherwise. TVt is the daily trading volume scaled by shares outstanding at the end of the previous day. The subscripts indicate the days lagged. t-statistics are in parentheses. Time period: 2/3/9812/29/00 (735 observations) 1 Coefficient on: Intercept -0.517* (-8.24) -0.380* (-2.36) 0.493* (4.99) ) 0.357* (9.90) 0.233* (6.20) 0.159* (4.52) 50.42% -0.443* (-7.13) -0.362* (-2.31) 0.500* (5.20) -0.125* (-6.40) 0.365* (10.39) 0.238* (6.51) 0.152* (4.43) 53.01% -1.306* (-13.72) -0.235# (-1.63) 0.496* (5.44) -0.121* (-6.72) 98.612* (11.35) 0.298* (9.04) 0.195* (5.73) 0.134* (4.24) 60.08% 2 3

DRt _ P DRt _ N Up _ down

TVt Ln( High Ln( High Ln( High Adj R2

,t 1

,t 2

,t 3

* Significant at 0.05 level, two-tailed test # Significant at 0.10 level, two-tailed test

41

Table 8 Dependence of high-low volatility and implied volatility on aggregate net inflow and outflow Controlling for the persistence of volatility, a dummy variable ( Up _ down ), and market turnover rate ( TVt ), daily high-low volatility (Ln( HL ,t )) and implied volatility (Ln( VIX ,t )) are regressed on concurrent aggregate net inflow ( Inflowt ) and outflow ( Outflowt ) in Column 1 and Column 2, respectively.

VIX ,t

is calculated using the methods of Parkinson (1980).

VIX ,t is the implied volatility

index based on the option of the S&P 100 index quoted by the CBOE. Inflowt and Outflowt are two series representing net inflow and outflow, respectively. Up _ down is defined as 1 when the S&P 500 daily return is positive and as 0 otherwise. TVt is the daily trading volume scaled by shares outstanding at the end of the previous day. The subscripts indicate the days lagged. Time period: 2/3/9812/29/00 (735 observations) Sample: 411 U.S. Equity funds 1 Ln( HL ,t ) Estimate Intercept -1.535* -39.671* 48.481* -0.160* 114.305* 0.151* 0.237* 0.133* 34.70% t-value -11.97 -3.03 2.71 -5.68 9.20 4.43 7.00 3.94 2 Ln( VIX ,t ) Estimate -0.034# -4.835* 4.066# -0.085* 3.885* 0.950* 94.02% t-value -1.67 -2.69 1.82 -22.96 2.34 90.32 -

Inflowt Outflowt Up _ down

TVt Ln( t 1 ) Ln( t 2 ) Ln( t 3 ) Adj R2

* Significant at 0.05 level, two-tailed test. # Significant at 0.10 level, two-tailed test.

42

Table 9 Sub-period check of high-frequency volatility on aggregate net inflow and outflow Controlling for the persistence of volatility, a dummy variable ( Up _ down ), and market turnover rate ( TVt ), daily high-frequency volatility (ln( High )) is regressed on concurrent aggregate net inflow
,t

( Inflowt ) and outflow ( Outflowt ) during two sub-time periods in Column 1 and Column 2.

High ,t

is

calculated from five-minute intraday S&P 500 index returns using the estimators of ABDL (2001) and ABDE (2001). Inflowt and Outflowt are two series representing net inflow and net outflow, respectively.

Up _ down is defined as 1 when the S&P 500 daily return is positive and as 0 otherwise. TVt is the daily
trading volume scaled by shares outstanding at the end of the previous day. The subscripts indicate the days lagged. t-statistics are in parentheses. Time period 1: 2/3/986/30/99 (355 observations) Time period 2: 7/1/9912/29/00 (380 observations) Sample: 411 U.S. Equity funds 1 Time period 1 Estimate Intercept -1.064* -33.404* 25.459# -0.137* 132.201* ) ) ) 0.357* 0.236* 0.166* 59.61% T-value -7.80 -2.94 1.93 -5.13 7.12 7.42 4.54 3.47 2 Time period 2 Estimate -1.442* -38.113* 63.096* -0.095* 107.138* 0.300* 0.166* 0.086* 52.69% T-value -10.56 -2.83 2.33 -3.69 9.48 6.53 3.51 1.96

Inflowt Outflowt Up _ down

TVt Ln( High Ln( High Ln( High Adj R2

,t 1

,t 2

,t 3

* Significant at 0.05 level, two-tailed test. # Significant at 0.10 level, two-tailed test.

43

0 .020 0 .015 0 .010 0 .005

FLOW

0 .000 -0 .005 -0 .010 -0 .015 -0 .020 98 0203

98082 2

99 0310

99092 6

00 0413

0010 30

Fig. 1. Time series of daily aggregate U.S. equity mutual fund flow. The figure shows the time series of daily aggregate domestic equity mutual fund flow. Our data on flow (new subscriptions less redemptions) come from TT financial services of Santa Rosa, CA. Flow is defined as the one-day percentage change in aggregate TNAs, less the one-day percentage change in aggregate NAVs. The sample covers the period February 3, 1998, through December 29, 2000, for a total of 735 observations.

44

0.6

High

0.5

Volatility

0.4

0.3

0.2

0.1

0.0 980203 980630 981124 990420 990914 000208 000704 001128

Figure 2 (a) High-freq. volatility

VIX
0.50 0.45
0.5

0.6

HL

0.40

Volatility

0.4

0.3

Volatility
980630 981124 990420 990914 000208 000704 001128

0.35 0.30 0.25 0.20 0.15 980203 980630 981124 990420 990914 000208 000704 001128

0.2

0.1

0.0 980203

Figure 2 (b) High-low volatility

Figure 2 (c) Implied volatility

Fig. 2. Time series of different volatility estimators. The figure shows time series of three different market volatility estimators: high-frequency volatility ( High,t ) in (a), high-low volatility ( HL,t ) in (b), and implied volatility ( VIX ,t ) in (c). Parkinson (1980).

High,t

is calculated from S&P 500 index five-minute intraday

returns using the estimator of ABDL (2001) and ABDE (2001).

HL,t

is calculated using the methods of

VIX ,t

is the implied volatility index based on the option of the S&P 100 index quoted

by the CBOE. The sample covers the period February 3, 1998, through December 29, 2000, for a total of 735 observations.

45

Volume

Price < 0

Price > 0

Fig. 3. Illustration of an asymmetric volume-price change relationship (Karpoff, 1987). The figure illustrates the asymmetric concurrent volume-price change relationship. The solid lines represent the asymmetric volume-price change relationship depending on the direction of the price changes: Volume is positively correlated with positive price changes and negatively correlated with negative price changes. The dotted line represents the positive correlation between volume and price changes per se.

Volatility

Flow<0

Flow>0

Fig. 4. Illustration of an asymmetric volatility-flow relationship. The figure illustrates the asymmetric concurrent volatility-flow relationship. The solid lines represent the asymmetric volatility-flow relationship depending on the direction of flow: Volatility is negatively correlated with inflow and positively correlated with outflow (ignore the sign). The dotted line represents the negative correlation between volatility and flow per se.

46