Sunteți pe pagina 1din 16
Journal of Financial Economics 12 (1983) 263-278. North-Holland FRICTION IN THE TRADING PROCESS AND THE ESTIMATION OF SYSTEMATIC RISK* Kalman J. COHEN Duke University, Durham, NC 27706, USA Gabriel A. HAWAWINI INSEAD, Fontainebleau, France Steven F. MAIER Durham, NC 27706, USA Robert A. SCHWARTZ New York University, New York, NY 10006, USA David K. WHITCOMB Rutgers University, Newark, NJ 07102, USA Received December 1981, final version received February 1983 This paper considers how estimates of the market model beta parameter can be biased by friction in the trading process (information, decision, and transaction costs) that (a) leads to a distinction between observed and ‘true’ returns; (b) causes observed returns to be generated asynchronously for a set of interdependent securities; and (c) thereby introduces serial cross- correlation into security returns. Several propositions are derived from which consistent estimators of beta are obtained, and the effect of differencing interval length on beta estimates is specified. The formulation is contrasted with the related analyses of Scholes—Williams (1977) and Dimson (1979). 1. Introduction In recent years, a number of empirical studies have established that market model beta estimates vary systematically with the length of the measurement (or ‘differencing’) interval over which security returns are calculated.’ Since *The authors acknowledge helpful comments made on an earlier draft of this paper by Elroy Dimson, William Fung, Myron Scholes, Michael Theobald, and Richard Roll, the referee for the Journal of Financial Economics. \The studies include Pogue-Solnik (1974), Altman et al. (1974), Schwartz-Whitcomb (1974), Scholes-Williams (1977), Hawawini (1977,1980), Levhari-Levy (1977), Smith (1978), Dimson (1979), Fung et al. (1981), Ng (1981), and Cohen et al. (1983). Several of these papers contain theoretical analyses as well. 0304-405x/83/$3.00 © 1983, Elsevier Science Publishers B.V. (North-Holland) IPE E 264 KJ. Cohen et al., Trading friction and beta estimation theoretically (under the usual market model assumptions) beta should be invariant with the length of the measurement interval, attention has turned to explaining the observation and to developing procedures for correcting the implied bias in OLSE beta. Such is the purpose of this paper. The most closely related studies with similar purpose are those of Scholes-Williams (1977) and Dimson (1979). By and large, the intervalling-effect bias in beta has been believed to arise from very brief delays in the price-adjustment process. The most widely noted delay is the lag of transaction price adjustments behind quotation price adjustments [the so-called ‘Fisher (1966) effect’].? Building upon the Fisher effect, Scholes—Williams (1977) show that when some securities have their last transaction earlier in the measurement interval than others, an errors-in-variables-type bias occurs in observed beta. If the intervalling-effect bias in beta were due solely to the Fisher effect, either of two corrective procedures would in principle completely eliminate it. First, closing quotation prices could be used in place of last transaction prices for the computation of security and market index returns. The problem with applying this approach is that quotation prices are not readily available for this purpose. Alternatively, measurement times could be directly matched with trading times, as has been done by Schwert (1977), Franks et al. (1977), and Marsh (1979). The problem here is that the exact time of each trade must be known and a market index must be available that is updated continuously. There exists a less direct alternative that circumvents these problems of data availability: regress individual security returns on synchronous and non- synchronous market index returns. This corrective procedure has been used by Pogue-Solnik (1974), Ibbotson (1975), Scholes—Williams (1977), Schwert (1977), Dimson (1979), and Theobald (1980). Being less direct, this approach has the further advantage of encompassing any price-adjustment delay that impedes the movement of quotation prices as well as transaction prices. But this approach also suffers from a limitation: An efficiency loss due to measurement error restricts the number of lead-lag periods which can be used [see Dimson (1979)]. The seriousness of this limitation (and the conceptual limitation of the Fisher-effect oriented approaches) depends on just how protracted the price-adjustment process actually is. It is, of course, not possible to estimate directly the length of price- adjustment delays, since these are not directly reflected in the available data. However, on the basis of other work we can obtain a rough sense of the order of magnitude of the delays. In Cohen et al. (1983), we observe that the 2Cohen et al. (1979) model transaction prices lagging quotation prices due to a Poisson order generating process, and show that this delay pattern causes observed market index returns to be autocorrelated even if the underlying ‘true’ returns generating process yields serially independent returns. K.J. Cohen et al., Trading friction and beta estimation 265 beta bias remains significant for differencing intervals extending up to at least several weeks. Furthermore, when we rank securities according to their market value, we observe that beta is biased upward for a few securities with a large value of shares outstanding; that beta is biased downward for the thinner issues; that across all issues there is a strong, monotonic relationship between the bias and a security’s market value; and that this cross-sectional relationship is not confined to very short differencing intervals. This suggests that the lag structure is not just a very short-period phenomenon. In Cohen et al. (1980), we consider how friction in the trading process can cause price- adjustment delays, and why these delays may be relatively protracted. Along with noting the Fisher effect, we also argue that somewhat longer delays might result from: (1) specialists/dealers impeding quotation price adjustments in the act of satisfying exchange stabilization obligations or redressing inventory imbalances, and (2) individual traders seeking to trade (and updating limit orders) only periodically, due to information, decision, and transaction costs. From these empirical observations and conceptual arguments, we conclude that price-adjustment delays longer than one trading day exist. Propositions demonstrated in this paper show that such delays give rise to non-zero serial cross-covariance, and it is trivial to show that the serial cross-covariance decays only gradually as the measurement interval is increased beyond the maximum length of the price-adjustment delays. Thus a correction procedure is needed that is not limited in its ability to encompass an extensive lead-lag structure (as is the regression approach that includes non-synchronous market index returns). In the present paper, we develop such a procedure. The essence of our approach is as follows. We show that the intervalling- effect bias in beta can be caused by friction in the trading process even when ‘true’ (but unobserved) returns are generated by a frictionless process. We use an equation from Cohen et al. (1980) relating true and observed returns via a price-adjustment delay structure to show how the price adjustments for a set of interdependent securities can be non-synchronous, and ‘that security returns, therefore, can be serially cross-correlated. It is this serial cross- correlation which leads to biased estimates of ‘true’ beta. We show how the bias in beta changes as the length of the measurement interval is varied, and that the bias monotonically approaches zero as the interval length is increased. Our results provide an analytical framework for empirically testing our model, and suggest implementable procedures for eliminating the intervalling-effect bias. In our empirical analysis [Cohen et al. (1983)], we use a two-pass regression analysis to estimate the asymptotic value that beta approaches as the differencing interval is increased. We show here that the asymptotic measure is a consistent estimator of beta even in the presence of long-lived price-adjustment delays. 266 K.J. Cohen et al., Trading friction and beta estimation The paper is organized as follows. In section 2, we use a simplified version of the frictions equation of Cohen et al. (1980) to obtain various propositions regarding the effect of differencing interval length and a security’s expected price-adjustment delay structure on its observed beta. From these propositions, consistent estimators of beta are obtained. Section 3 discusses the relationship of our work to that of Scholes—Williams (1977) and Dimson (1979). Section 4 contains our concluding remarks. 2, The model 2.1. Relationship between observed returns and true returns Defining returns to be the log of price relatives adjusted for dividends and splits, we assume: (A.1) The true return for security j in period t is generated by the market model,? it Bim t ee ()) We then use the following equation [see Cohen et al. (1980)] to distinguish between true returns and observed returns (r@,,):* Vit-naljt—m (2) where we further assume: (A.2) Yim and 7,,,, ate independent for all j£k,t,1,m,n. (A.3) 7j,, are independent of ry, and e,, for all j,k, t,t,n. (A.4) E0)j4m)=E()j.e.m) for all j,t,1,m. (AS) B(DMo72.9=1 for all j 2We make the usual assumptions about the process generating r,,,, that is: (i) ry, and ry,, are independent and identically distributed for all t#z; (ii) e,, and e,, are independent and identically distributed for all tt and E(e,, ,)=0 for all j and f; (iii) e, , and e,,, are independent for all j#k and all 1 and z; and (iv) ry,, and ¢,,, are independent for ali j, t, and’. Eq, (2) in the present paper is a special case of eq. (1) in Cohen et al. (1980). Here we have omitted random variables which reflect the direct impact of the bid-ask spread on observed returns because they do not enter into any of the cross-covariance and beta results derived in this paper. Dimson’s (1979) eq. (6) appears to be similar to our eq. (2). Note, however, that Dimson treats an observed price generated in period t as a true price, but he allows for the possibility that no price may be observed in a given period. Hence, Dimson’s expected return is a weighted average of returns generated in the current and prior periods. KJ. Cohen et al., Trading friction and beta estimation 267 The random variables ,,,, relate the observed return generated in one period to the true returns generated in that period and in N prior periods. Alternatively stated, the 7;.,0,7;,,1--»Vj1. Comprise a delay distribution that shows how the true return generated in period ¢ impacts on the returns actually observed during period t and the next N periods [eq. (2) implies that the true return generated at period t—N will have no remaining impact on returns observed after period 1]. Moreover, by Assumption A.5, on average each true return will be fully reflected in the future observed returns. We hypothesize that y,,,,#0 for n>0 results from friction in the trading process that causes price-adjustment delays. The structure of the yj, is of particular importance here because if the »;,,, are non-zero for n>O and differ across securities, the observed returns for a set of securities will adjust asynchronously to common movements in their true returns. We shall show that with such asynchronous adjustments across securities, cross-serial correlation is introduced into observed returns, and observed beta estimates are biased. Further, for any given differencing interval the bias in beta will be shown to vary systematically across securities, with its signed value being related to the relative magnitude of a security’s price-adjustment delays. Our model of the observed returns generation process can be utilized to establish the relationship between the systematic risk calculated from observed returns and the true systematic risk. This in turn allows us to examine the properties of the estimated beta coefficient and to derive a consistent estimator of a security’s systematic risk. We proceed as follows. In Proposition 1, we establish the relationship between the covariances of observed returns and the covariances of true returns for pairs of securities. From Proposition 1, we derive the relationship between the covariance of the observed returns of a security and the observed returns of a market index, stated in Corollary 1, and the relationship between the variance of observed returns of a market index and the variance of its true returns, stated in Corollary 2. Using Corollaries 1 and 2, we establish the relationship between a security’s beta coefficient calculated from observed returns and its true beta coefficient. This result is stated in Proposition 2. We then derive two alternative consistent estimators of true beta, one in Proposition 3 and the other in Proposition 4. In Propositions 5 and 6 we explore the distribution across securities of the bias in beta and how the bias depends on a security’s relative price-adjustment delays. 2.2. Relationship between observed beta and true beta Proposition 1. The contemporaneous covariance between the observed returns of securities j and k plus the sum of the serial cross-covariances of their observed returns for all leads and lags up to N periods equals the 268 K.J. Cohen et al., Trading friction and beta estimation contemporaneous covariance between their true returns. For all j#k we have Ny Ny COV (FM + L COV Tew + Dy, COV (Tm Mey) =COV (1% e,0)- (3) n= The proof of Proposition 1 is given in the appendix. Note that the proof does not require that the price-adjustment delay variables y,,,, and y,,,,, be independent. As we shall discuss in section 3, recognition of this is the key to obtaining the appropriate form for Dimson’s model, and to relaxing the stationarity assumption of Scholes and Williams (1977) (which requires that a transaction be observed in each measurement period used in the estimation process). We now examine the covariance between the returns of securities and those of a market index (Corollary 1) and the variance of the returns of a market index (Corollary 2). We define an observed market index ri,,, and the true market index ry,, used in eq. (1) by Thaw = Lari and Ta t= DHT where the x, are appropriate weights for each security. Corollary 1. The contemporaneous covariance between the observed returns of any security j and a market index M, plus the sum of the serial cross- covariances of their observed returns for all leads and lags up to N periods, equals the contemporaneous covariance between the true returns of j and M. That is, N N COV (TS. The.) + X COV (79. 1s Tern) + d COV (TS, + ns Th4,2) = COV (Ty. sae, = The proof of Corollary 1 follows directly from Proposition 1. Summing eq. (3) over all securities with the appropriate weights yields Corollary 15 QED. Corollary 2. The variance of the observed returns of a market index M, plus twice the sum of the serial covariances of the observed returns of M for all lags SThis argument is somewhat inaccurate. The summation over all values of k in eq. (3) would include the term for which k=j, ie., the ‘own’ variance term. However, eq. (3) does not generally hold in this case because the variables 7,,,, and 7,,,._ ar¢ usually not independent (e.g, the existence of a sample period in which no trade takes place can cause complex interrelationships among the 7, variables). However, this is not of any real significance, since in a well-diversified portfolio such as a market index the ‘own’ variance terms becomes insignificant. K.J. Cohen et al. Trading friction and beta estimation 269 up to N periods, equals the variance of the true market index returns. That is, Nx var (rie) +2 D2 COV (PReses Tear 0) = Var (Pagaos net where we define var (he I= DEP COV (oho 5 COV Fhe Fhe =DE XPV Tha var (rig) =). > XjXk COV (N17, n)> 7 The proof of Corollary 2 follows directly from Corollary 1. Summing eq. (3) over the index j yields the desired result. QED. We can now derive the relationship between a security’s systematic risk calculated from observed returns (the observed beta Bf) and the security’s true systematic risk (f,). Define the following variables: Bi = COV ("5,1 Fh. 0/VAE (Phe, Bia, =COV (Pharm Pha )/VAr (Pha,e)> Bie, = COV (The, s—ms Pha )/VAE (Phe, Bi, = COV (TS, Pha, 2)/VAr (The, B3_, = COV (15,1 The, )/ Var (Teds By=COV(T sa, d/Val (ae.2)> where we call f} the observed security beta, fy, and f%__ the observed intertemporal lead and lag market betas, A}, and f}_, the observed intertemporal lead and lag security betas, and f, the true security beta. The relationship between these variables is expressed in Proposition 2. Proposition 2. The observed beta of security j is a function of the true beta of Again the argument is somewhat inaccurate in that the summation includes ‘own’ variance terms for the securities as well as all possible covariance terms. However, as explained in footnote 5 above, we believe this distortion can be safely ignored. 270 K.J. Cohen et al. Trading friction and beta estimation j, the observed intertemporal market beta, and the observed lead and lag betas of security j up to N periods; in particular w=ofi+2 3 u.,)- ¥ O..+8.) 4) To prove Proposition 2, note that by definition the right-hand side of eq. (4) is equal to COV (ers Mes 142 arr) SOV sem Pes) FCOV(M sm Fh) net var (Py,1) : Now by recognizing that our stationarity Assumptions A.1-A.5 imply that COV(T.15Tht—n)=COV (M4 :+ns7%,) and substituting Corollaries 1 and 2, we obtain COvOrj nT uae) VAE (Pugs) — COV 1+ ues) COV (ren) var(ry,.) var(rig,s) var (r44,.) A 00V (5. The ) _ pe var(ri) QED. Using Proposition 2, an estimator for true beta can easily be determined as shown in the following proposition: Proposition 3. A consistent estimator of true beta can be found by the Sormula + B+ By, pap, (5) 1+ Y dict bie, where bf, bf,,, bf_,5 bir, and bj_, are OLS regression estimators of B%, B?_,. B5_,, Bia, and Biy_,, respectively.” 7Note that eq. (5) contains estimators of both fy,, and fy, even though our assumptions would indicate that these values are identical. However, given’a finite sample of observations from which estimators can be constructed, the estimator by, would use a slightly different subset of the sample than would the estimator bt, . Thus in ofder to provide symmetry in the use of the available data, we have included both bi,” and biy_ K.J. Cohen et al., Trading friction and beta estimation am The proof of Proposition 3 follows immediately from eq. (4). Q.E.D. It is clear from Proposition 3 that b?, the OLS regression estimate of systematic risk, will generally be a biased estimator of true B;. 2.3. Intervalling effect and asymptotic estimator We now address the issue of what happens when we increase the differencing interval over which returns are measured. Let BAL) be the estimator defined by eq. (5) where the OLS estimators b4(L), bj, (L), b?_(L), by (L) and bj,_(L) are computed using non-overlapping periods consisting of L consecutive smaller measurement periods. Since BAL) is a consistent estimator of f,, we can also obtain the following result: Proposition 4. The asymptotic estimator, defined by + = lim b%(L), L>0 is a consistent estimator of the true beta B, where b%(L) is the observed OLS regression beta based on non-overlapping periods of length L. To prove Proposition 4, observe that the expected delay distribution can be defined for an arbitrary measurement interval of length L in terms of the original expected delay distribution as [hte Lte- Ey nDN=z SS Elan and [Ete Lint ttene-1 ED eA =Z » > Ebiadh I=Lnti-e where t’ covers the original time periods from t to L+t—1. As L+o, Ely;e,(LJ>1, while E[y,,,(L)]-+0 for all n21, yielding the desired result. Q.E.D. Propositions 3 and 4 provide two alternative methods by which consistent estimators of true beta (8;) can be calculated. The Proposition 3 estimator is similar to the methods used by Scholes—Williams (1977) and Dimson (1979) and is discussed in section 3. To implement the asymptotic estimator of Proposition 4 (since taking L to the limit is not practical with a finite set of m K.J. Cohen et al., Trading friction and beta estimation observations), one can regress the values of b%(L) for different L on an appropriate function of L, obtaining Bf as the asymptote. We turn next to the distribution across securities of the bias in fj and to those characteristics of a security that determine the bias. Proposition 5. The observed systematic risk, B} satisfies Lx,Aj=1, J where x, is the weight of security j in the market index. The proof of this proposition follows from the definition of Bj as an OLS estimator. QED. Proposition 5 states that on average (where average is taken to mean the weighted average as represented in the market index) the observed betas contain no bias. An important implication [noted first by Scholes—Williams (1977)] is that not all observed betas are underestimates; the bias is positive for some securities and negative for others, with the weighted average over all securities equal to zero. The implications of Propositions 4 and 5 are jointly testable empirically (see footnote 11 below). Note that the size of the bias in f¥ is related to the random variables y;,,,,, which are a direct measure of the price-adjustment delays. A comparative measure of price-adjustment delays can be defined as follows: Definition. A security j has cumulatively greater expected price-adjustment delays than security i if N x © EO4)> ¥ Evi.,s) for all delay periods p= This definition can now be used to prove Proposition 6. Proposition 6. If the price-adjustment delay variables are such that Ele, 1,0) 2 Elie,,1) «+» = ElYs,1,n) for all securities k with at least one inequality holding strictly for each security, and if security j has cumulatively greater expected price-adjustment delays than security i and is otherwise identical (x; =x;,B;=B,, and securities i and j have the same covariances with all other securities), and if securities i and j have positive covariances with all other securities, then BS < BF. KJ. Cohen et al., Trading friction and beta estimation 273 The proof of Proposition 6 is given in the appendix. Proposition 6 identifies securities with relatively lengthy price-adjustment delays (and positive betas) as being those securities for which beta will be underestimated by the usual OLS estimators. Conversely, securities with relatively short price-adjustment delays (and positive betas) are likely to have their betas overestimated by the usual OLS technique. 3. Relationship to the Scholes-Williams (1977) and Dimson (1979) models The original work on the impact of non-synchronous trading for the measurement of beta was by Scholes—Williams (1977). Their estimator of B; is Be=topb5 +1 +bj_/(1+2bi_,)s (6) which is identical to our estimator given in eq. (5) when N=1.° The reason that their estimator does not contain the higher order lead and lag betas is related to two of their fundamental assumptions. First, they assumed that only non-synchronous trading leads to the oberved bias in beta, whereas our eq. (2) encompasses other possible reasons for observed prices lagging the underlying true returns process, some of which imply substantial lags. Second, they assumed that a transaction occurs in each measurement period, so any information lost due to an interval of non-trading will be observed in the immediately following period when the next trade is assumed to occur. Thus, their assumptions lead to a special form of our eq. (2), a= Vile t Pattie which in turn could only hold if N=1 in our formulation, which by our eq. (5) then yields eq. (6) above. Of special interest in the Scholes and Williams formulation is their assumption that the process generating the delay structure in observed betas was both independent and identically distributed over time. This assumption made it impossible to extend their basic results shown in eq. (6) to the case where securities did not trade at least once in each period.? Even with securities that trade relatively frequently, a day can arise where no trade occurs; Scholes—Williams had to discard such observations in their empirical "Theobald (1980) approaches the problem from a somewhat different analytical framework, but produces an estimator which (under certain conditions) approaches that of Scholes-Williams (1977) as the length of the differencing interval increases. Theobald notes that there is a cost, in terms of loss of degrees of freedom, for the lead and lag terms. °In commenting on an earlier draft of this paper, Scholes indicated that the Scholes-Williams (1977) estimator was developed for NYSE stocks, which typically trade every day. He said that they knew how to extend their model to incorporate periods of non-trading lasting more than one day, but decided not to do so for economic, rather than mathematical, reasons. 274 KJ. Cohen et al., Trading friction and beta estimation work. As pointed out by Dimson (1979), this in turn leads to a loss of information which can increase the size of the confidence interval around the Scholes-Williams beta estimator. Even more important, however, is the observation that the higher order lead and lag betas could be substantial, as we would expect for thinly traded securities; hence ignoring them can lead to an inconsistent beta estimator. Our assumption for frictions eq. (2) is only that the delay variables across securities are stationary and independent. There can, however, be complex dependency patterns over time for a given security. The beta estimator derived by Dimson (1979, p. 204) and given in his eq. (12), is (in our notation) wa B= b+ So. SB Unfortunately, Dimson’s estimator is incorrect.'° Dimson also suggests that the coefficients bj _, and bj, all be simultaneously estimated using multiple Tegression as opposed to independently estimated as suggested by Scholes— Williams. Such an approach would not yield estimators consistent with our definitions of B§, Bj,, and fj_,. For a further discussion of the problems inherent with the Dimson approach, see Fowler—Rorke (1983). Dimson brings up the important issue of the correct selection of the value for the parameter N. The appropriate selection of this parameter represents a direct conflict between model and statistical accuracy. On the one hand, the greater the number of lead and lag terms estimated, the better potentially is the model’s representation of reality. On the other hand, the more such terms included in eq. (5), the greater the potential noise introduced in the estimation process. In other words, as N increases, the model becomes both more realistic and less biased, but the statistical efficiency of the estimator declines. Dimson also proposes a further adjustment in the estimator b3, bj, and b3_, using the method of Vasicek (1973). A further discussion of the work of Scholes—Williams and Dimson can be found in Fung (1981). 4. Conclusion In this paper we have developed an analytical model that shows how friction in the trading process can cause estimates of beta to be biased even 1°The error appears to be in the derivation of eq. (11) of Dimson (1979, p. 203). Dimson attempts to equate the appropriate terms in his eq. (9) with those in his eq. (10) to derive eq. (11). The basic problem is that all terms that are dependent upon M,,, are not equivalenced. This is because f;,, is itself dependent upon M,,,, therefore eq. (11) omitted at least one term. In addition, the equivalence used to derive eq. (11) requires only that all terms associated with M, in eq, (9) must equal all terms associated with M, in eq. (10), not that each subterm be equal. Thus a corrected form of Dimson’s estimator is our eq. (5). Dimson (1982) argues that the bias in his earlier estimator is empirically very small, and presents an alternative form of it. K.J. Cohen et al., Trading friction and beta estimation 215 when ‘true’ returns are generated by an underlying process that is frictionless. Our most important results are: (1) in the presence of finite price-adjustment delays, OLSE beta becomes a consistent estimator of true beta in the limit as the measurement interval increases without bound; and (2) for any measurement interval, the bias in OLSE beta is cross-sectionally distributed around zero and depends on the relative magnitude of a security's price- adjustment delays. These results provide the framework for two alternative empirical procedures for adjusting OLSE beta for intervalling-effect bias. The first is an extension of the work of Scholes—Williams (1977) to allow for price- adjustment delays of more than one day’s duration. The second procedure is to estimate OLS betas for various measurement intervals and then to regress these estimates on an asymptotic function of the measurement interval. This procedure makes it unnecessary to specify a lead-lag structure (which would involve choosing between bias and loss of efficiency, as in the first method). Our result that the bias in a security’s OLSE beta depends on the relative magnitude of its price-adjustment delays suggests an avenue for further research. Since price-adjustment delays are not directly observable, it would be desirable to identify an observable variable which is related to price- adjustment delays. Such a variable could be used, with short-interval OLSE betas, to provide an adjustment procedure that would require a shorter calendar span of data than the asymptotic beta estimator.!! Appendix Proof of Proposition 1 By the definition of r2, from eq. (2) we have ) By Assumption A.2, the y delay variables are mutually independent and independent of r, yielding A r cov namoov( S Daath De Yharm mt NON COV (TS, Th,0) Ds ok, E07), DE, COV (P),1—15 7a. ~m)s ™Based on the argument of Cohen et al. (1980) that price-adjustment delays are greater for thinner securities, papers by Fung et al. (1981), Ng (1981) and Cohen et al. (1983) have shown (for French, British and American data, respectively) that the bias in OLSE beta is significantly related to a security's market value of shares outstanding. [Based on a different theoretical framework, a similar empirical result has been obtained (for British data) by Dimson (1979).] As a proxy for the effect of price-adjustment delays, however, market value might introduce too much measurement error; whether a better proxy exists and whether the relationships are stationary over time remains to be determined. 216 K.J. Cohen et al, Trading friction and beta estimation where we have dropped the middle subscript on E(y) because of Assumption A44 in order to simplify notation. The assumption that r, and r, are stationary with cov(r,,,,7,,,)=0 for all t and t non-overlapping yields N COV (3.51, D=COV(T 1 nd De Ey, DE(,0- A similar analysis for the second and third terms on the left-hand side of eq. (3) yields N N cov (rr) + x COV (Ms Thr + 2, Cov (rh e-mth) N NN=n =0V(rn nad 2) E+ YS) Bl dE 0) NoON-n +2 D2 Eby ,sdBtn | Nn x =Cov(r.15 nae ru) are -) 5 mao By Assumption A.5 that E()}.0 7,1.) =! for all j, we obtain Nn x COV (TS. That 2, 0OV otha) + 2 COV (Th. ns Ths) =cov(r, nt): QED. Proof of Proposition 6 Rewriting eq. (4) we obtain B-B= 26) Bru. 3 Bint Bh» Expanding the last summation by substituting the definitions of Bj, and 3_, and r’,, from eq. (2) we obtain N x Bj—Bj=2B; y Bia. X4COV (5. 157e.e) x EQ») var (14,2) y +, YB JLB. 0) Ebi. dh K.J. Cohen et al., Trading friction and beta estimation 277 where we have again dropped the middle subscript on E(y) to simplify notation. Using this relationship, we can now write an expression for the difference between B}—B; and f?—8;, 1 P= eta FE COM od x{ § te, 0—Bly dE.) Bl Ih (ad) where all identical terms have been cancelled. [Note: By assumption B;—B,; and COV (Pr; 157.2) =COV(;,157r,1) for all k.] Since we assume cov(r;,,,7;,,) is positive for all k, the first half of eq. (A.1) is obviously negative. If we can show that the term { } is always positive, we will have established our result. Rewriting { } from eq, (A.1) we have i N B}— B= E%,0) D [EQ -E(i 0] -2 Em DEO; - EQ) Substituting the identities ) , E(y,,,)=1—E(y;,o) for securities i and j in the above expression yields N Bi —BP aoe EQ, MEQ -EQ01, which by algebraic manipulation is equal to Bi P= Bln.g) ¥ (Et..)-B,0] Ny y + [E(.1-)- EQ) Pe EQ; — Ein]. (A.2) The first term of this expression is zero since xN, o Ey.) = Yio E(y;,)=1. By the assumption of the proposition, the left half of the second term is non- negative for all securities and strictly positive for at least one value of | for each security k. The right half of the second term is positive for all values of I, since security j is assumed to have cumulatively greater expected price- adjustment delays than security i. Thus, eq. (A.2) is positive for all values of k. QED. 278 K.J. Cohen et al., Trading friction and beta estimation References Altman, EL, B. Jacquillat and M. Levasseur, 1974, Comparative analysis of risk measures: France and the United States, Journal of Finance 29, 1495-1511. Cohen, K.J., SF. Maier, R.A. Schwartz and D.K. Whitcomb, 1979, On the existence of serial correlation in an efficient securities market, TIMS Studies in the Management Sciences 11, 151-168, Cohen, K.J., G.A. Hawawini, S.F. Maier, R.A. Schwartz and D.K. Whitcomb, 1980, Implications of microstructure theory for empirical research on stock price behavior, Journal of Finance 35, 249-257, Cohen, K.J., G.A. Hawawini, S.F. Maier, R.A. Schwartz and D.K. Whitcomb, 1983, Estimating and adjusting for the intervalling-effect bias in beta, Management Science 29, 135-148. Dimson, E., 1979, Risk measurement when shares are subject to infrequent trading, Journal of Financial Economics 7, 197-226. Dimson, E., 1982, Friction in the trading process and risk measurement: Response, Working paper (London Business School, London). Fisher, L., 1966, Some new stock market indexes, Journal of Business 39, 191-225, Fowler, DJ. and CH. Rorke, 1983, Risk measurement when shares are subject to infrequent trading: Comment, Journal of Financial Economics, this issue. Franks, JR, J.E. Broyles and M.J. Hecht, 1977, An industry study of the profitability of mergers in the United Kingdom, Journal of Finance 32, 1513-1525. Fung, W.K.H,, 1981, The empirical properties of alternative procedures for estimating betas with nonsychronous data, Working paper (Rutgers University, New Brunswick, NJ). Fung, W.K.H, R.A. Schwartz and D.K. Whitcomb, 1981, Testing alternative adjustment procedures for the intervalling-effect bias in beta estimates using Paris Bourse data, Working paper (Rutgers University, New Brunswick, NJ). Hawawini, G.A,, 1977, On the time behavior of financial parameters: An investigation of the intervaling effect, Ph.D. dissertation (New York University, New York), Hawawini, G.A., 1980, Intertemporal cross dependence in securities daily returns and the short- Tun intervalling effect on systematic risk, Journal of Financial and Quantitative Analysis 15, 139-150. Ibbotson, F.G., 1975, Price performance of common stock new issues, Journal of Financial Economics 3, 235-272. Levhari, D. and H. Levy, 1977, The capital asset pricing model and the investment horizon, Review of Economics and Statistics 59, 92-104. Marsh, P.R., 1979, Equity rights and the efficiency of the UK stock market, Journal of Finance 34, 839-863. Ng, K.W., 1981, Theoretical and empirical evaluation of alternative procedures for estimating betas on nonsychronous data, M.S. thesis (University of Manchester Institute of Science and Technology, Manchester). Pogue, G. and B. Solnik, 1974, The market model applied to European common stocks: Some ‘empirical results, Journal of Financial and Quantitative Analysis 9, 917-944. Scholes, M. and J. Williams, 1977, Estimating beta from non-synchronous data, Journal of Financial Economics 5, 309-327. Schwartz, R.A. and D.K. Whitcomb, 1974, Can betas be estimated from daily data?, in: Proceedings of the seminar on the analysis of security prices, Vol. 19, no. 1 (University of Chicago, Chicago, IL). Schwert, G.W., 1977, Stock exchange seats as capital assets, Journal of Financial Economics 5, 51-78. Smith, K.V., 1978, The effect of intervalling on estimating parameters of the capital asset pricing model, Journal of Financial and Quantitative Analysis 13, 313-332. Theobald, M., 1980, The analytic relationship between intervalling and non-trading effects in continuous time, Working paper (University of Manchester, Manchester). Vasicek, O. 1973, A note on using cross-sectional information in Bayesian estimation of security betas, Journal of Finance 28, 1233-1239,

S-ar putea să vă placă și