Information Asymmetry about the Firm and the Permanent Price Impact of Trades: Is there a Connection?

Transcription

1 Information Asymmetry about the Firm and the Permanent Price Impact of Trades: Is there a Connection? Gideon Saar Lei Yu 1 First Draft: July 2001 This Version: July Both authors are from the Stern School of Business, New York University, 44 West Fourth Street, Suite 9-190, New York, NY Saar can be reached at Tel: , Fax: , gsaar@stern.nyu.edu, and Yu can be reached at Tel: , Fax: , lyu@stern.nyu.edu. The authors thank I/B/E/S for providing the data on analysts earnings forecasts, Frank Russell Company for data on the 1999 reconstitution of the Russell 1000 index, and Foster Provost for helping with the matching procedure used in the paper. The authors are grateful for helpful comments from Hendrik Bessembinder, Ekkehart Boehmer, Joel Hasbrouck, and seminar participants at Babson College, University of Utah, and NYU.

2 Abstract We investigate whether permanent price impact measures describing intra-day price movements are helpful in characterizing information asymmetry about the fundamentals of firms. We conduct an event study of the Russell 1000 index reconstitution and find that the permanent price impact measures change around the event despite the fact that Russell 1000 membership is based on market capitalization and therefore the event is not associated with any change in private information about the firms. We also examine a large cross-sectional sample and find that the permanent price impact measures do not consistently reflect uncertainty about future earnings or relate to price informativeness about future earnings in a manner that is compatible with their use as proxies for information asymmetry about firms fundamentals.

3 InformationAsymmetryabouttheFirmandthePermanentPriceImpactofTrades:IsthereaConnection? The literature examining price movements on trades often makes a distinction between temporary and permanent price movements. Short-horizon temporary price movements have been attributed, for example, to inventory management on the part of market makers; the most common reason cited for the permanent price impact of trades has been the presence of investors with private information about the future cash flows of the firm who trade to profit from their information. A number of papers have put forward econometric specifications for estimating the permanent and temporary price impacts of trades using intra-day data comprised of trades and quotes (see, for example, Glosten, 1987; Glosten and Harris, 1988; Hasbrouck, 1988; Stoll, 1989; Hasbrouck, 1991a; Hasbrouck, 1991b; Madhavan and Smidt, 1991; Huang and Stoll, 1997; Madhavan, Richardson, and Roomans, 1997). 1 While the aforementioned papers focused on quantifying the relative importance of temporary and permanent factors in explaining price movements, researchers have soon found a rich set of applications for these new tools. Since theoretically the permanent price impact of trades was associated with the incorporation into prices of private information about fundamentals, measures of the permanent price impact were used to quantify the extent of information asymmetry about a firm. These measures proved attractive in areas outside market microstructure (e.g., corporate finance) where researchers have been interested in characterizing the information asymmetry environment of a firm, which is inherently unobservable. 2 There are two potential problems with using measures of permanent price impact estimated from trade and quote data to tell a story about the future cash flows, or fundamentals, 1 George, Kaul, and Nimalendran (1991) propose a methodology for estimating components of the spread that uses daily and weekly rather than intra-day data. 2 For example, these measures were used to investigate information asymmetry in the contexts of corporate events (Brooks, 1994; Jennings, 1994; Barclay and Dunbar, 1996; Krinsky and Lee, 1996; Desai, Nimalendran, and Venkataraman, 1998), the debt-equity mix (Illesy and Shastri, 2000), corporate diversification (Fee and Thomas, 1999), disclosure quality (Heflin, Shaw, and Wild, 2000), the opaqueness of banking firms assets (Flannery, Kwan, and Nimalendran, 2000), the size of the analyst following (Brennan and Subrahmanyam, 1995), and ownership structure (Heflin and Shaw, 2000a; Sarin, Shastri, and Shastri, 2000, Dennis and Weston, 2001; Dey and Radhakrishna, 2001). 1

4 of a firm. The first problem is that many elements in the trading environment of a stock can affect the estimation procedures and bias the estimates. For example, specific rulesthat govern trading, like the price continuity requirement at the NYSE, affect the adjustment of prices to order flow and therefore the measures of permanent price impact estimated from trades and quotes. A slowly moving target inventory level of NYSE specialists (see, for example, Hasbrouck and Sofianos, 1993) can also affect the estimates. This may partially explain the fact that, in our large cross-sectional sample, the magnitude of permanent price impact measures for NYSE stocks is on average three times that of NASDAQ stocks. Does it really mean that there is so much more information asymmetry about NYSE firms relative to Nasdaq firms? When elements that conceptually should not affect prices permanently get incorporated into the permanent price impact measures, it is unclear whether differences in the measures reflect differences in the information asymmetry environment or other influences. Even if we are certain that the permanent price impact measures indeed capture only things that should affect the price permanently, the second potential problem is that not only information asymmetry about the future cash flows of a firm can give rise to the permanent price impact of trades. The discount rate effect documented in the literature for longer horizon returns (e.g., Fama, 1990; Campbell and Ammer, 1993) demonstrates how return variation is partly caused by shocks to future expected returns, not just shocks to expected future cash flows. 3 Saar (2001) shows in a sequential trade model how uncertainty about the distribution of preferences and endowments of investors affects the risk premium and generates a permanent price impact for trades even in the absence of information asymmetry about future cash flows. Therefore, the permanent price impact measures may be picking up information about investors that is unrelated to the fundamentals of firms. 3 Fama (1990), for example, finds that the discount rate effect accounts for about 30% of the variance of annual real returns on a value-weighted portfolio of NYSE stocks. Campbell and Ammer (1993) use a vector autoregression (VAR) decomposition of monthly stock excess returns and find that only 15% of the variance of returns is attributed to the variance of news about future dividends, while 70% is attributed to the variance of news about future excess returns. 2

5 Both problems weaken the link between the permanent price impact measures and information asymmetry about the firm. To have confidence in using these measures to make an inference about the level of information asymmetry, it is important to investigate empirically how close is the association between them. The objective of this paper is to examine the question whether measures constructed from trade and quote data to describe intra-day price movements are helpful in characterizing information asymmetry about the fundamentals of firms. Four papers in particular are related to our investigation: Neal and Wheatley (1998) and Clarke and Shastri (2001) are two papers looking at close-end funds; Clarke and Shastri (2000) and Van Ness, Van Ness, and Warr (2001) conduct horse races among different permanent price impact measures. Neal and Wheatley postulate that since the net asset value of a closed-end fund is reported every week, there can be little information asymmetry about its current liquidation value. They find large permanent price impact measures, and conclude that the estimates may be unreliable. Clarke and Shastri (2001) reach a different conclusion, arguing that uncertainty about private benefits paid to blockholders creates adverse selection costs. Clarke and Shastri (2000) use six econometric methodologies to estimate permanent price impact measures for 320 NYSE firms. They look at how these measures relate to each other, and also at the correlations between the measures and certain firm characteristics. The correlations with firm characteristics are statistically significant only for net sales and a dummy for regulated firms. Similarly, Van Ness, Van Ness, and Warr (2001) conduct a horse race of five permanent price impact measures using 856 NYSE stocks and examine their relations with corporate finance variables, volatility measures, and measures that involve analysts and institutional holdings. They find mixed results, with no single econometric methodology producing estimates that are consistently correlated with the variables they consider as information proxies. The strongest relation they document is between the permanent price impact measures and volatility. 3

6 Both horse race papers mentioned above are subject to the same conceptual problem. Since information asymmetry is unobservable, they use a variety of proxies about which they have stories of how the proxies relate to information asymmetry about fundamentals. However, since the permanent price impact measures also have a story that relates them to information asymmetry, the exercise turns into looking at relations among proxies without the ability to determine whether lack of relation between the permanent price impact measures and the other information proxies is the fault of the measures or the other proxies. The design in the present paper is more powerful since we use an event study that allows for a characterization of changes (or their absence) in the information asymmetry environment without the need to rely on proxies. We additionally provide a cross-sectional analysis for three reasons. First, it allows us to examine whether a weaker link exists between the measures and other information asymmetry proxies that would justify using the measures under some conditions. Second, we look at a much larger sample (2,797 stocks) than the other two papers and use stocks from the NYSE, AMEX and Nasdaq as opposed to only NYSE in these studies. Third, we are able to test the measures even without information asymmetry proxies by examining the cross-sectional relation between the measures and price informativeness about future earnings. While the measures used in the aforementioned papers have different names, such as the permanent impact of the order flow innovation or the adverse selection component of the spread or the trade-correlated component of the efficient price variance, they all use intra-day data to estimate the permanent changes in price induced by trades. 4 Using any estimate of permanent price impact as a measure of information asymmetry about fundamentals is subject to the same two problems identified above. Thus instead of examining all possible measures, we choose two of the methodologies that impose different degree of structure on the trading process: a more structured trade-indicator model and a less structured variance decomposition procedure. The trade-indicator model, from Madhavan, Richard- 4 See Huang and Stoll (1997) for a discussion of the relationships among some of the measures. 4

7 son, and Roomans (1997), provides an estimate of the permanent price impact of the order flow innovation (henceforth, MRR). The variance decomposition procedure from Hasbrouck (1991b) provides two measures: (i) the trade-correlated efficient price variance component that is used as an absolute measure of information asymmetry (HASAB), and (ii) the ratio of the trade-correlated component to the total efficient price variance that is taken as a measure of the amount of private information relative to the total amount of information (HASR). We formulate the investigation of the association between the permanent price impact measures and information asymmetry about fundamentals in terms of necessary and sufficient links. The strongest possible association is when changes in information asymmetry about fundamentals are necessary to produce changes in the permanent price impact measures (holding constant the amount of uninformed trading). While such a claim is difficult to prove (as it requires considering all possible circumstances in which future cash flows can change), its converse is not. Showing that changes in private information about fundamentals are not necessary for the permanent price impact measures to change requires only one example. The Russell 1000 index reconstitution event provides a suitable environment in which we can test this issue. The composition of the index changes once a year and the new component stocks are determined solely based on their market capitalization at a single point in time (the last trading day of May), which is public information. Therefore, the event does not add information to what the market already knows, and we expect no change in information asymmetry about fundamentals. The permanent price impact measures should not change when controlling for the amount of uninformed trading if the measures indeed reflect information asymmetry about fundamentals. We look at how the permanent price impact measures change from the last trading week in May (before the new ranking of the Russell 1000 is established) to the first trading week in July (when the new index takes effect). We find that both MRR (the permanent price impact 5

8 of the order flow innovation) and HASAB (the trade-correlated component of the randomwalk variance of quote midpoint changes) decrease significantly around the event. The design of the test also enables us to demonstrate that these results cannot be attributed just to changes in liquidity trading or analyst activity around the event. We test the robustness of our findings with respect to alternative estimation intervals, price level effects, and possible changes in market-wide liquidity, and find that these do not change our conclusions. The event study is a strong test of the relation between the permanent price impact measures and information asymmetry about the firm since it does not require us to use other variables to independently characterize information asymmetry. But while the MRR and HASAB measures fail to pass the necessary test, it is still possible that a weaker sufficient relation exists. More specifically, we would like to know whether differences across stocks in the degree of information asymmetry about fundamentals are sufficient to generate differences in the estimates of the permanent price impact measures that are used as proxies for information asymmetry. Generally speaking, it is difficult to find good proxies for information asymmetry about future cash flows, a reason that no doubt contributed to the popularity of the permanent price impact measures. We choose for our cross-sectional tests only proxies that are constructed using earnings or their forecasts. Intuitively, variables that explicitly incorporate information about the firm s future earnings should be more tightly connected to the firm s future cash flows than measures that do not use such direct information (like the estimated permanent price impact measures). To implement the test we use uncertainty about future cash flows to represent information asymmetry since the greater the uncertainty, the more room (and incentives) there is for investors to acquire private information and trade on it profitably. Twoofourproxiesfor uncertainty about future cash flowsarebasedondispersioninanalysts earningsforecastsand two others are based on past earnings variability. We find that MRR provides conflicting results since it is negatively related to dispersion in analysts earnings forecasts but also negatively related to earnings predictability (as it should be). HASAB is related to one 6

9 of the analysts forecasts variables in the predicted direction in univariate tests, but gives conflicting results when the two types of proxies are used together. HASR is not related to any of the proxies for future cash flows uncertainty. 5 We then conduct a cross-sectional test that bypasses the need for information asymmetry proxies by examining the relation between the permanent price impact measures and price informativeness about future earnings (i.e., how much of the information about future earnings is impounded into current prices). Ceteris paribus, the more trading on private information about the firm, the more informative should prices be with respect to future earnings (see, for example, Grossman and Stiglitz, 1980). We find that none of the permanent price impact measures relates to price informativeness in the predicted manner. By utilizing a variety of designs and tests, we offer a comprehensive assessment of the relation between the permanent price impact measures and information asymmetry about fundamentals. The results we present do not invalidate the use of measures estimated from intra-day data to examine the price impact of trades. However, they cast doubt on the use of very short-horizon price movements for making inferences concerning information asymmetry about the fundamentals of firms. The rest of the paper proceeds as follows. Section 1 describes the permanent price impact measures we use. Section 2 examines changes in the permanent price impact measures around the Russell 1000 index reconstitution. Section 3 presents cross-sectional analysis that relates the permanent price impact measures to proxies for uncertainty about future cash flows. Section 4 investigates the relation between the permanent price impact measures and price informativeness, and Section 5 is a conclusion. 5 For robustness, we also look at how the permanent price impact measures relate to analysts forecast errors, another common proxy for information asymmetry that uses earnings information. The analysis using this proxy points to the same conclusions as the analysis of uncertainty about future cash flows. 7

10 1 Permanent Price Impact Measures In this section we provide some details on the estimation of the permanent price impact measures. For a more comprehensive exposition of these methodologies we refer the readers to the original papers that developed them. 1.1 Madhavan, Richardson, and Roomans (1997) Let x t denote an indicator variable taking the value of 1 if the transaction at time t is buyer initiated and 1 if it is seller initiated, and let µ t denote the post-trade expected value of a stock. The revision of beliefs following a trade is the sum of the change in beliefs due to public information and the change in beliefs due to the order flow innovation: µ t = µ t 1 + θ (x t E[x t x t 1 ]) + t (1) where θ is the permanent impact of the order flow innovation and is a measure of the degree of information asymmetry, and t is the innovation in beliefs between times t 1 andt due to public information. Let p t denote the transaction price at time t, andφ denote the market makers cost per share of supplying liquidity (compensating them for order processing costs, inventory costs, and so on). The transaction price can then be expressed as: p t = µ t + φx t + ξ t (2) where ξ t captures the effects of stochastic rounding errors induced by price discreteness or possibly time-varying returns. Equations (1) and (2) can be used to obtain: u t = p t p t 1 (φ + θ)x t +(φ + ρθ)x t 1 (3) where ρ is the first-order autocorrelation of the trade initiation variable. Then, the measure of permanent price impact θ, alongsideφ, ρ, λ (the unconditional probability that a transaction occurs within the quoted spread), and a constant α can be estimated using GMM applied 8

11 to the following moment conditions: E x t x t 1 x 2 t 1ρ x t (1 λ) u t α (u t α)x t (u t α)x t 1 =0 (4) We use data from the TAQ database to estimate (4). 6 Following Madhavan et al., the classification into buys and sells is done as follows: (i) if a trade price is greater than or equal to the prevailing ask then x t = 1, (ii) if the trade price is less than or equal to the bid then x t = 1, and (iii) if the trade price falls between the bid and the ask then x t =0. We use the log transaction price for p t, and refer to the estimate of θ, the permanent impact of the order flow innovation, as the MRR measure Hasbrouck (1991b) Hasbrouck models the quote midpoint q t as the sum of two unobservable components: q t = m t + s t (5) where m t is the efficient price (i.e., expected end-of-trading security value conditional on all time-t public information) and s t is a residual discrepancy term that is assumed to incorporate inventory control, price discreteness, and other influences that cause the midquote to 6 We apply various filters to clean the data. We only use trades for which TAQ s CORR field is equal to either zero or one, and for which the COND field is either blank or equal to B, J, K, or S. We eliminate any trades with non-positive prices or where the change from the last trade price is greater than $5 ($20 for stocks with average price greater than $100). We also exclude a trade if its price is greater (less) than 150% (50%) of the price of the previous trade. We only use quotes from the exchange on which a stock is listed. We eliminate quotes for which TAQ s MODE field is equal to 4, 5, 7, 8, 9, 11, 13, 14, 15, 16, 17, 19, 20, 27, 28, or 29. We exclude quotes with non-positive ask or bid prices, where the bid is greater than the ask, and where the ask is more than $9999. We require that the difference between the bid and the ask be smaller than 25% of the sum of the bid and the ask. We also eliminate a quote if the bid or the ask are greater (less) than 150% (50%) of the bid or ask of the previous quote. When signing trades, we employ yet an additional filter that requires the difference between the price and the prevailing midquote to be smaller than $8. 7 Madhavan et al. estimate (4) using the transaction price, rather than its logarithmic transformation, and therefore the components they estimate are in dollar terms (i.e., components of the dollar spread). In the event study (Section 2), we estimate two versions of θ: one using dollar prices and the other using log prices to get components in percentage terms (i.e., components of the relative spread). The results from both specifications are similar. In the cross-sectional part of the paper (both in Section 3 and Section 4), we use log transaction prices since a component of the relative spread seems to be better suited for comparisons across stocks. 9

12 deviate from the efficient price. The efficient price evolves as a random walk: m t = m t 1 + w t (6) where the innovation w t reflects updates to the public information set including the information in the order flow. The market s signal of private information is the current trade innovation defined as x t E[x t Φ t 1 ], where x t is signed volume and Φ t 1 is the public information set prior to the trade. The impact of the trade innovation on the efficient price innovation is E [w t x t E[x t Φ t 1 ]]. Two measures of information asymmetry that Hasbrouck proposes are: σ 2 w,x Var (E [w t x t E[x t Φ t 1 ]]) (7) R 2 w σ2 w,x σ 2 w = Var (E [w t x t E[x t Φ t 1 ]]) Var(w t ) where σ 2 w,x is the trade-correlated component of the random-walk variance of quote midpoint changes representing an absolute measure of information asymmetry, and R 2 w is a measure of the amount of private information relative to the total amount of information. Let x 0 t be an indicator variable that takes the values { 1, 0, +1}, and define: x 1 t = log(x t + 1) (the log volume), and x t = {x0 t,x 1 t }. Let r t =logq t log q t 1. The absolute and relative permanent price impact measures are estimated using the vector autoregressive (VAR) model: (8) r t = x t = 5 i=1 5 a i r t i + c i r t i + 5 i=0 5 i=1 i=1 b i x t i + v 1,t (9) D i x t i + v 2,t (10) where b i is a 2 1vectorofcoefficients, D i is a 2 2matrixofcoefficients, and v 2,t is a 2 1 vector of error terms. 8 8 Hasbrouck (1991b) uses a specification with volume and volume squared instead of log volume. Applying his exact specification to our analysis did not materially affect the results. We chose to present the specification using log volume because it seems to fit better the concave relation between trade size and price impact. 10

13 We use data from the TAQ database for the estimation of the VAR model in (9) and (10). Following Hasbrouck, the trade classification into buys and sells is done as follows: (i) if a trade price is greater than the prevailing midquote then x 0 t = 1, (ii) if the trade price is less than the prevailing midquote then x 0 t = 1, and (iii) if the trade price is at the prevailing midquote then x 0 t =0. WeestimatetheVARsystemusingOLS.Afinite VAR generally possesses a vector moving average (VMA) representation of an infinite order. The random walk variance and the trade-correlated component are calculated from the coefficients of the VMA representation truncated after 100 lags. Throughout the paper, we refer to the estimate of the absolute measure, expressed in the form of standard deviation per hour, as HASAB and to the estimate of the relative measure (Rw) 2 as HASR. 9 2 Event Study: Russell 1000 Reconstitution In this section we investigate whether changes in information about the firm are necessary for the permanent price impact measures to change, holding constant the amount of liquidity trading. The Russell 1000 index reconstitution provides us with a natural experiment for testing the necessary connection. 10 The Russell 1000 index is comprised of the largest 1000firms in terms of market capitalization. 11 Membership in the Russell 1000 index is determined strictly by market capitalization at the end of the last trading day in May and reconstitution occurs once a year. During June, Frank Russell Company issues a list of firms to be added to and deleted from the indexes. The new indexes become effective on July 1. The more heavily studied changes to the S&P equity indexes, in contrast, involve a committee of nine people who make their decision taking into account attributes such as liquidity, profitability of the firm, market capitalization, 9 We also conduct the analysis using the raw estimates of the trade-correlated component of the randomwalk variance without conversion to standard deviation per hour. The results are very similar in both sign and significance level, and are therefore omitted from the presentation. 10 Madhavan (2001) examines return effects of the Russell reconstitution. 11 Excluded from all Russell indexes are foreign firms, limited partnerships, limited liability companies, royalty trusts, closed-ended investment management companies, ADRs, preferred stocks, pink-slipped companies, OTC bulletin board companies, and warrants and rights. The indexes also exclude stocks trading at below $1. 11

14 and sector representation of the stock. 12 Revisions to the S&P indexes also occur continuously throughout the year on a stock-by-stock basis. As such, the Russell reconstitution differs from changes to the S&P indexes in that the criterion used is public information (marketcapitalizationonmay28)andallchangesoccuratthesametime. Moreimportantly, additions to and deletions from the Russell 1000 index convey no new information about the firm, while S&P revisions might be viewed as expressing an opinion about a firm. Thus if the measures arise solely from private information about the fundamentals of firms, they should not change around the Russell 1000 reconstitution when controlling for the amount of uninformed or liquidity trading. If these measures are affected by other factors, like changes in the characteristics of investors that are unrelated to being informed or uninformed about the fundamentals of firms, we might detect changes around the reconstitution event. Controlling for the amount of uninformed trading is important because it has direct effect on the permanent price impact measures. However, it is difficult to identify trades as being initiated by liquidity traders (being uninformed is not an observable attribute). We use two variables in the regressions in an attempt to control for the amount of normal or liquidity trading. The analysis of both the addition and deletion samples, however, provides us with a more powerful robustness check on the effects of liquidity trading. The reason is that most stocks in our sample that were added to the Russell 1000 index (89 out of 105) were deleted from the Russell 2000 index, and all stocks that were deleted from the Russell 1000 index were added to the Russell Say we believe that a move from the Russell 1000 to the Russell 2000 index increases the amount of liquidity trading (perhaps more passive funds follow the Russell 2000 index). If the proxy for normal volume is not good enough, some of the increase in liquidity trading will be captured by the intercept in the regression we use to examine changes in the measures around the event, making it negative. At the same time, stocks that move from the Russell 2000 to the Russell 1000 should experience a decrease in liquidity trading. For them, insufficient control for liquidity trading will result in a positive 12 For papers examining changes to S&P equity indexes see Harris and Gurel (1986), Shleifer (1986), Jain (1987), Dhillon and Johnson (1991), and Lynch and Mendenhall (1997). 12

15 intercept. If the intercepts of the addition and the deletion samples have the same sign, however, the effect cannot be solely due to insufficient control for uninformed or liquidity trading. It is clear that the Russell reconstitution event is fairly transparent, and it is possible to predict prior to the last trading day in May which stocks are more likely to be added or deleted. In fact, firms like Investment Technology Group, Inc. offer the service of predicting the changes in order to assist fund managers who wish to begin rebalancing their portfolios in advance of the announced index changes. While some fund managers could be trading ahead of time based on a certain assessment as to which stocks will switch indexes, others could be trading based on a different assessment or waiting until the new composition is made known. Rebalancing before May 28 does not necessarily bias the results against finding any change between the permanent price impact measures estimated just before the end of May and immediately after July 1 if such behavior does not reduce the uncertainty about final membership. If the uncertainty is reduced, any effect that we detect is likely to be an underestimate of the real change in the measures. 2.1 Sample and Methodology The initial sample is constructed from a list of 157 stocks that were added to the Russell 1000 index in 1999 and 103 stocks that were deleted from the index. 13 We restrict the sample to common stocks (for compatibility with the cross-sectional tests in Section 3 and Section 4) with data from April 1 through mid-july. We use CRSP and Dow Jones Interactive to identify firms with corporate events including mergers and acquisitions, share offerings, earnings announcements, dividends distributions, and stock splits within the sample period. These firms are eliminated from the sample in order not to confound the results (if a corporate event occurs within the sample period, the information environment of a firm can change). 13 The preliminary list of the changes was published by the Frank Russell Company on June 11 and the final list was published on July 7 (though the new composition of the index took effect on July 1). To construct the sample, we eliminate 12 stocks that were not present in both the preliminary and final lists. Note that the number of additions and deletions can be different due to bankruptcies and mergers that decrease the number of firms in the Russell 1000 index throughout the year. 13

16 The final sample consists of 105 additions and 70 deletions. 14 Table 1 presents summary statistics for the firms used in the event study. There seems to be a slight increase in mean price between May 28 (the last trading day in May) and July 1 (when the new composition of the index takes effect). Average turnover increases slightly from the last week of May to the first week of July for both samples, but the cross-sectional mean of the average number of trades decreases for the addition sample. Dollar spreads and percentage spreads decrease for both additions and deletions. Our event study methodology is an adaptation of standard specifications. 15 Let ψ i,τ denote the permanent price impact measure for stock i estimated over time interval τ {pre, post} (i.e., either the pre-event interval or the post-event interval). The simplest specification assumes that the permanent price impact measure can be described as the sum of astock-specific unconditional mean (µ i ), an event effect (α), and an error term ( i,τ ): ψ i,τ = µ i + αδ τ + i,τ (11) where δ τ is an indicator variable that takes the value zero in the pre-event interval and one in the post-event interval, and i,τ is assumed i.i.d and normally distributed with mean zero and variance σ 2 i /2. Theoretical models that show how information asymmetry causes trades to have permanent price impacts also demonstrate how the price impact of trades can be affected by changes to normal or liquidity trading (i.e., volume due to uninformed traders that is unrelated to information events). A formulation that takes normal trading into account can be written as: where V i,τ is a proxy for the amount of normal volume. ψ i,τ = µ i + αδ τ + βv i,τ + i,τ (12) By examining differences between the pre- and post-event intervals, we can eliminate the 14 Similar results are obtained without eliminating firms with major corporate events. 15 See, for example, Schipper and Thompson (1983) and Thompson (1985). 14

17 firm-specific mean. Equation 11 can be written in terms of differences as: ψ i ψ i,post ψ i,pre = α + i (13) where i is normally distributed with mean zero and variance σ 2 i. Similarly, the formulation in terms of differences of Equation 12 is: ψ i = α + β V i + i (14) where V i V i,post V i,pre. 16 Testing that the Russell 1000 reconstitution event affects the permanent price impact measures is therefore equivalent to testing whether the intercept (α) ofeitherequation13 or Equation 14 is statistically different from zero. Figure 1 shows the timeline for the event study. We use one week as the length of the interval over which the permanent price impact measures are estimated. Since the ranking of firms for the reconstitution is based on the closing prices on May 28, we use the week ending May 28 as the pre-event interval. Similarly, since the effective date of the changes is July 1, we use five trading days beginning July 1 as the post-event interval. 17 For a start, we assume that the error terms are identically distributed across stocks and Equation 13 is implemented as a simple t-test. We also use the non-parametric Wilcoxon signed rank test to perform the analysis under less restrictive assumptions. Equation 14 is estimated using OLS with White s heteroscedasticity-consistent standard errors. 16 We also used a specification with an interaction between the event dummy and the volume measure: Thus, ψ i,τ = µ i + αδ τ + βv i,τ + γδ τ V i,τ + i,τ ψ i = α + β V i + γv i,post + i,τ Tests under this formulation give similar results to those without the interaction. 17 Unlike the pre-event interval, the five days beginning July 1 do not include all days of the week (due to Independence Day). Since there may be a day-of-the-week effect in the estimated permanent price impact measures (see, for example, Foster and Viswanathan, 1993), we also conducted the analysis with the first five days in July that comprise a full set of the days of the week (dates 1, 2, 6, 7, and 12). The results obtained were very similar to the results using the first five trading days in July, and are therefore omitted for brevity of exposition. 15

18 Figure 1: Timeline for Event Study 5-day pre-event interval 5-day post-event interval May 28 July 1 Market capitalization calculated from closing prices is used to determine membership in the new Russell 1000 index. New Russell 1000 index takes effect. We use two proxies to control for changes in normal volume: the average daily turnover and the average daily number of trades in the interval. We fully acknowledge that these may not be perfect controls. However, the internal check discussed in Section 2 enables us to test whether our results are driven by insufficient adjustment for liquidity trading by examining the intercepts of the addition and deletion samples to see if they have opposite signs. While throughout the paper we use the MRR measure estimated using log prices to get a measure that can be interpreted as a component of the percentage spread, in the event study part of the paper we also use the same methodology with dollar prices to get the permanent price impact measure in terms of dollars, MRR$. This is done to examine the robustness of our results to possible changes in the level of prices induced by the event that are unrelated to the information asymmetry environment. 2.2 Results Panel A of Table 2 presents the results of the event study for the addition sample. The t-tests indicate that the decrease in the mean measure between the pre- and post-event intervals is statistically different from zero for MRR$, MRR, and HASAB. The non-parametric Wilcoxon statistics are also highly significant. Normalizing by the pre-event magnitude of the measures, the median percentage changes in MRR$, MRR, and HASAB are 27.41%, 45.78%, and 25.80%, respectively. The fourth column of Panel A shows the intercept from the regression with AvgTurn as a proxy for normal volume and the sixth column shows the intercept from the regression with AvgTrd as a proxy for normal volume. The intercepts in the regressions 16

19 with MRR$, MRR, and HASAB are all negative and statistically significant. The mean, median, and intercepts in the HASR regressions are all positive but insignificant. The results for the deletion sample are presented in Panel B of Table 2. The means and medians of differences between the pre- and post-event measures are negative and statistically different from zero for the two MRR measures. HASAB also decreases after the event, though the t-test is only marginally significant (with a p-value of ); the Wilcoxon signed rank test is significant. Normalizing by the pre-event magnitude of the measures, the median percentage changes in MRR$, MRR, and HASAB are 21.50%, 17.44%, and 9.55%, respectively. The intercepts of the regressions with the proxies for normal volume tell the same story: significant negative intercepts for MRR$, MRR and HASAB. Again, HASR does not show any significant change around the event. The results from both the addition and deletion samples point to the same conclusion: the Russell 1000 reconstitution event is associated with a decrease in the permanent price impact measures (with the exception of HASR). This cannot be due to a change in private information about the fundamentals of the firms since the only criterion used for the reconstitution market capitalization on May 28 is public information. The finding of a decrease in the measures for both the addition and deletion samples also rules out an explanation based on insufficient control for changes in liquidity trading. As the discussion in Section 2 points out, even if the proxies for normal volume are not perfect, a change in normal trading would produce opposite results in the addition and deletion samples. The reason is that if a membership in a certain index is associated with a particular amount of uninformed trading, we would expect that being added to the Russell 1000 and deleted from the Russell 2000 would produce the opposite effect to being added to the Russell 2000 and deleted from the Russell Finding that the permanent price impact measures decrease in both addition and deletion samples is evidence that the result is not being driven by changes in liquidity trading. This feature of the test also enables us to rule out that the effects on the permanent 17

20 price impact measures are due to changes in the production of fundamental information by analysts induced by changes in index membership. It is possible that analysts are more willing to follow stocks that are included in a certain index, say the Russell This can be due to more trading activity in these stocks by index funds, or just because of a larger market capitalization that induces more mutual funds to own the stock. If this is the case, it could be that the amount of fundamental information production would increase after inclusion in the Russell 1000 index. But at the same time, a stock leaving the Russell 1000 index should experience a decrease in fundamental information production. The fact that we get a similar decrease in the permanent price impact measures for both addition and deletion samples means that our results cannot be solely due to a change in fundamental information production by analysts. We conducted several tests to evaluate the robustness of our results. First, we used interval lengths other than five days. Two-day intervals displayed a similar significant decrease in the permanent price impact measures MRR$, MRR, and HASAB. 18 We also used onemonth intervals and obtained results similar to those presented in Table 2 (i.e., a statistically significant decrease in MRR$, MRR, and HASAB). 19 Since MRR$ and MRR behave in a similar fashion, it seems unlikely that the results are due to possible changes in the price level that may affect the estimated measures due to discreteness of the price gird. Nonetheless, we added the change in AvgPrc as a regressor to the specification in Equation 14 to account for changes in the price level, and found that it did not alter our results. We also used a dummy variable to examine whether the 16 stocks that were added to the Russell 1000 index but were not previously in the Russell 2000 index behave differently from the rest of the stocks. The coefficient of the dummy variable was not significant and similar results were obtained. For MRR$ and MRR, we repeated the tests 18 We used May 27 and 28 for the pre-event interval and July 1 and 2 for the post-event interval. Both intervals are comprised of the days Thursday and Friday, and hence the results are not due to a day-of-theweek effect. 19 The only difference was with respect to the HASR measure, which was significantly negative in the deletion sample and significantly positive in the addition sample. 18

21 using only positive estimates and only estimates that were statistically different from zero. This was done to eliminate the influence of stocks that did not have sufficient amount of trading to produce reliable estimates of the permanent price impact measures. The results were similar to those reported in Table 2. Lastly, we constructed a matched sample of stocks based on both market capitalization and average turnover in April The motivation behind constructing the matched sample was to investigate whether the Russell 1000 reconstitution event coincided with a market-wide change in liquidity that could have affected the estimates of the permanent price impact measures of all stocks. Unlike the results for the addition and deletion samples, all permanent price impact measures except one showed no statistically significant decrease in the matched sample. The decrease in HASAB in the addition-matched sample was significant but smaller in magnitude than the decrease in the addition sample, and a paired t-test (as well as a Wilcoxon signed rank test) showed that the difference between them was statistically significant. To summarize the results so far: the event study shows that the necessary link between information asymmetry about future cash flows and the permanent price impact measures does not exist. The estimated measures should not have changed around the Russell 1000 reconstitution event if they only reflect the information asymmetry environment, but we have found a significant decrease in all measures (except for HASR). The decrease in the measures is consistent with changes to the extent of uncertainty about the investor population in the spirit of Saar (2000, 2001). The Russell indexes are popular benchmarks of stock market performance in the United States. In 2001, an estimated 117 billion dollars in investment were indexed to the Russell 1000, 2000, and 3000 indexes. 21 Therefore, the annual reconstitution event can generate buying and selling pressures on 20 The universe of stocks that was used for the matching consisted of all common stocks in the CRSP database with information from April 1 to July 12 that were not added to or deleted from the Russell 1000 index. 21 This estimate is taken from a report by Investment Technology Group, Inc. on the 2001 Russell index reconstitution. 19

22 stocks that are added to or deleted from the indexes. Uncertainty about whether a stock will end up in the Russell 1000 index or the Russell 2000 index during the last week of May creates uncertainty with respect to the characteristics of the investors (e.g., the amount of money in index funds) who trade the stock. This increased uncertainty for stocks that may be added to or deleted from the index increases the price impact of trades and is picked up by the permanent price impact measures. In the post-event interval, there is less uncertainty about the investor population since the changes to the index are announced and therefore the permanent price impact of trades decreases. 3 Uncertainty about Future Cash Flows In this section we conduct a cross-sectional investigation of the relation between the permanent price impact measures and proxies for uncertainty about future cash flows that we use to represent information asymmetry about the firm. We test whether differences across stocks in the degree of information asymmetry about future earnings are sufficient to generate differences in the estimates of the permanent price impact measures. 3.1 Sample The initial sample for the cross-sectional analysis is constructed from all common stocks with information in the CRSP database for the entire year of 1999 that had, on average, more than one analyst covering them in the I/B/E/S database. The reason for requiring more than one analyst is to allow the calculation of the standard deviation of analysts earnings forecasts. This screen leaves 3,144 stocks. The sample is further restricted to firms with information in TAQ, COMPUSTAT, and Value Line Investment Survey in order to allow the computation of the permanent price impact measures and the construction of other variables that will be explained in greater detail below. This requirement eliminates 268 firms from the sample. We also eliminate firms that switched exchanges during the sample period, leaving a final sample of 2,797 firms. 20

23 Table 3 presents summary statistics for the sample. The average daily market capitalization (AvgCap) over the sample period (January 4 through December 31, 1999) for firms in our sample ranges from $6.05M to $446.16B, testifying to its heterogeneous nature. The sample also spans a range of trading activity and price levels. The most active firm has on average 29, daily trades (AvgTrd), while the average firm has daily trades, and the least actively traded firm in the sample has 3.28 trades per day. Average daily closing prices (AvgPrc) range from $0.24 to $ The sample is also diverse with respect to analyst following and institutional holdings. NumEst is the average of 12 monthly observations (January through December, 1999) of the number of analysts with end-of-fiscal-year earnings forecasts for the current year in I/B/E/S. While the mean of NumEst is 8.80, it ranges from a minimum of 2 to a maximum of The percentage institutional holdings from Value Line Investment Survey (InstHol) ranges from 0.23% to 99.59%. For eight stocks in the sample, the MRR estimates came out negative. While there is nothing in the estimation to constrain the MRR measure from being negative, it is unclear how to interpret these estimates. The negative measures do not affect our results as our findings are practically identical with and without these eight stocks. The estimation of the Hasbrouck (1991b) specification proved infeasible for two stocks in the sample for which the VMA could not converge. These stocks are eliminated from the sample for the tests involving the HASAB and HASR measures. 3.2 Methodology The first test we run examines whether cross-sectional differences in MRR, HASAB, and HASR reflect differences in information asymmetry about future cash flows. Sincewedo not have a way to directly measure information asymmetry, we focus instead on uncertainty about future cash flows. We use two definitions of dispersion of analysts forecasts to represent uncertainty about future cash flows: the standard deviation of analysts forecasts (StdEst) and the coefficient 21

24 of variation (CoefEst). 22 The logic behind the test is that the greater the uncertainty about future cash flows, the more room (and incentives) there is for investors to acquire private information and trade on it profitably. A theoretical treatment of the information environment of analysts is provided by Barron, Kim, Lim, and Stevens (1998). They show that an increase in private information about the firm, ceteris paribus, will increase the dispersion of analysts forecasts. Monthly observations for the standard deviation and the mean of analysts earnings forecasts for the upcoming end-of-fiscal-year are taken from I/B/E/S. StdEst is the simple average of the monthly standard deviation observations for all months in 1999, and CoefEst is the ratio of StdEst to the absolute value of the average monthly mean observations. Time left until the end-of-fiscal-year seems to exert an influence on the monthly observations of the standard deviation of the forecasts in I/B/E/S (as well as the number of estimates). However, since we use twelve consecutive months of data, the average should have similar properties across firms with different end-of-fiscal-year dates. We also use two proxies for uncertainty about future cash flows constructed from past earnings information. The first is the earning predictability measure provided by Value Line Investment Survey (VLPred). This variable, available for 2016 stocks in our sample, is derived from the standard deviation of percentage changes in quarterly earnings over an eight-year period, with special adjustments for negative observations and observations around zero. VLPred takes the values 1 to 100, where higher values are associated with less variability of past earnings and hence higher predictability of future earnings. We also compute the standard deviation of past annual earnings from COMPUSTAT using the five years prior to 1999 (StdEPS). 23 The idea behind both measures is that higher past earnings variability makes it more likely that future earnings cannot be predicted with great precision (allowing more room for private information production and profitable informed trading). 22 We also used the difference between the high and the low forecasts as an additional definition of uncertainty about future cash flows. The results using this definition were very similar to the results using the standard deviation of earnings forecasts, and are therefore omitted for brevity. 23 We repeated the analysis with the standard deviation of annual earnings computed using five years that include The results were very similar to those obtained using StdEPS. 22