Beaver_1968_

The Information Content of AnnualEarnings Announcements

WILLIAM H. BEAVER*

The infonnation content of earnings is an issue of obvious import anee andis a focal point for many measurement controversies in accounting. Thispaper empirically examines the extent to which common stock investorsperceive earnings to posses infonnational value. The study directs its at-tention to investor reaction to earnings announcements, as refiected in thevolume and price movements of common stocks in the weeks surroundingthe announcement date.

Valuation theory has long posited a relationship between earnings andthe value of common stock. Miller and Modigliani postulate that one im-portant element in determining the value of common stock is the product ofearnings times the appropriate earnings multiplier for that risk class.Graham, Dodd, and Cottle take a similar position with respect to the com-putation of their "intrinsic value" of common stock securities. MM alsoprovide empirical evidence that si^gests if reported earnings are adjustedfor measiu'ement errors through the use of instrumental variables, the ad-justed earnings are useful in the prediction of the market value of electricutility firms. In fact, the evidence indicated that the earnings term was themost important explanatorj'- variable in the valuation equation.' I'he rela-tionship is a necessary condition for earnings to have information content,

* Assistant Professor, University of Chicago.^ Merton H. Miller and Franco Modigliani, "Some Estimates of the Cost of Capital

to the Electric Utility Industry, 1954-57," American Economic Review, LVI (June,1966), 341.

* Benjamin Graham, David L. Dodd, and Sidney Cottle, Security Analysis (NewYork: McGraw-Hill, 1962), 443ff.

' Miller and Modigliani, op. dt., p. 373. The thrust of their study was directed tovalue of the firm rather than value of common stock. However, the MM model con-tains Jin explicit relationship between the two, and the statements made above applywith equal force to the value of common stock.

67

68 EMPIRICAL RESEARCH IN ACCOUNTING: SELECTrED STUDIES, 1968

but the evidence does not preclude the possibility that the opposite maybe true.

Although there are many reasons for adopting the position that earningslack informational value, two are frequently offered. (1) Measurementerrors in earnings are so large that it would be bettei' to estimate the valueof common stock directly from the instmmental variables rather than useearnings as an intermediate step. (2) Even though earnings might convey

information, there are other sources available to investors that containessentially the same information but are more timely. By the time annualearnings are released, any potential infonnation content has already beenproces5>ed by investors and is impounded in the market price. The implica-tion of both arguments is earnings reports have little or no informationcontent.

The issue is of major concern to the accounting profession because itsoutcome directly reflects upon the utility of the accounting activity. Oneapproach to examining this issue is to specify an expectations model of howinvestors relate reported earnings to market prices. The paper presentedby Bejiston at last year's Conference followed such an approach,* Benstonfound price changes were largely insensitive to earnings, which taken atface value is unfavorable to the utility of earnings data. But such resultsare always difficult to interpret because the lack of an obsen'ed relationshipmay be due to either one or both of two factors. Either no relationshipexists or the expectations model was improperly specified. It is impossibleto determine the extent to which the negative findings are due to the latterrather than the former.

The approach taken here is to apply tests that require no assumptionabout the expectations models of investors. Xote that the issue under con-sideration is of a positive rather than a normative nature—that is, thequestion of concern is not whether investors should react to earnings butrather whether investors do react to eamings.

Definitions of Information Content

Information has been defined as a change in expectations about the out-come of an event.^ Within the context of this study, a firm's earnings report-is said to have information content if it leads to a change in investors'assessments of the probability distribution of future returns (or prices),such that there m a change in equilibrium value of the current market price.

* George J. Benston, "Published Corporate Accounting Data and Stock Prices,'Empirical Research in Accounting: Selected Studies, 1967, Supplement to Vol. 5,Journal of Accounting Research, pp. 1-54.

^ Henri Theil, Economics and Information Theory {Chicago and Amsterdam:Rand McXally and North Holland Publishing Company, 1967), Ch. 1.

' A further stipulation is often made that information concerns changes in expecta-tions about an event that is a parameter of a decision model. Defining earnings informa-tion in terms of its impact on future returns (or prices) is consistent with that further

CONTENT OF ANNUAL EARKIXGS ANNOTJNCEMENTS 69

Although neither the direction nor the magnitude of the price change canhe specified without knowing the expectations model(s) of investors, thev£.riability of price changes is likely to be greater when earnings are an-nounced than at other times during the year.'

Another definition of information states that not only must there be achange in expectations but the change must be sufficiently large to inducea change in the decision-maker's behavior. According to this definition, afirm's earnings report possesses informational value only if it leads to analtering of the optimal holding of that firm's stock in the portfolios of in-dividual inv^tors. The optimal adjustment might be to buy more sharesor to sell some or all of the shares already held. In either event, the shift inportfolio position would be reflected in the volume. If earnings reports haveinformation content, the number of shares traded is likely to be higher whentho earnings report is released than at other times during the year.

Relationships between Price and Volume TestsThe relationships posited above are consistent with the economist's

notion that volume refiects a lack of consensus regarding the price. Thelack of consensus is induced by a new piece of infonnation, the earningsreport. Since investors may differ in the way they interpret the report, sometime may elapse before a consensus is reached, during which time increasedvolume would be obser\''ed. If consensus were reached on the first transac-tion, there would be a price reaction but no volume reaction, assuniinghomogeneous risk preferences among investors. If risk preferences difFer,there still could be a volume reaction, even after the equilibrium price hadbeen reached.

An important distinction between the price and volume tests is that theformer refiects changes in the expectations of the market as a whole whilethe latt«r reflects changes in the exp«;tatlons of individual investors. Astipulation. For support, see the literature on portfolio theory, especially Harry M.Markowitz, Portfolio Selection: E_fficient Diversification of Investments Qtevr York;John Wiley & Sons, 1959).

' The change in equilibrium price is in addition to any price change that wouldnormally occur in the absence of any earnings announcement. The assumption isthat the two price changes are positively correlated, independent, or mildly corre-lated. If there were strong negative correlation, the price change variability mightnotoe greater at the announcement date. In the light of previous research in thebehavior of security prices, the assumption of independence is most likely to be thecorrect one. See Eugene F. Fama, "The Behavior of Stock Market Prices," Journalof Business, XXXVIII (January, 1965), 34-105.

^ As a final parenthetical comment on definitions of information, note that reduc-iion of uncertainty was not one of tbe definitions chosen. It should be apparent thatin a dynamic situation (i.e., where probability distribution assessments are changingover time), a decision maker may be more uncertain about a given event after receiv-ing a message about the event than he was before he received the message. To useTheii's terminologj-, the entropj^ may increase as a result of a message, yet the mes-sage has information content. See Theil, op. dt., Ch. 2.

70 WILLIAM H. BEAVEE

piece of infonnation may be neutral in the sense of not changing the expec-tations of the market as a whole but it may greatly alter the expectationsof individuals. In this situation, there would be no price reaction, but therewould be shifts in portfolio positions refiected in the volume. Because theprice reflects the expectations of many investoi^, it may imply a very effi-cient forecast of eamings for several weeks prior to the announcement. Ifso, the price test may be less sensitive than volume to eamings reports.

The foregoing discussion suggests that a reaction may be obser\-ed inonly one of the tests or that the two tests may not respond equally. Ifneither test responds, the utility oi eamings data and the study's sampledesign will be su^ect.

Sample Design

Selection of Sample. The studj' is based upon a sample of annual earningsannouncements released by 143 firms during the years 1961 through 1965.Six criteria were used in the selection of the sample firms.

(1) The firm must be on the Compustat tape; (2) the firm must be amember of the New York Stock Exchange; (3) the fiscal year must end ona date other than December 31; (4) no dividends were announced in thesame week as the eamings announcement; (5) no stock splits wereannounced during the 17 week period surrounding the announcement ofeamings; and (6) there were less than 20 news announcements per yearappearing in the Wall Street JoumaL Table 1 indicates the extent to whicheach criterion affected the sample size.

Criterion (1) was selected because the Compustat population representsover 90 per cent of the total market value of the common stocks of publiclyheld corporations and hence is a relevant population for study. A secondaryreason is the ease with which finwicial statement items can be obtained forthe Compustat firms relative to firms not on the tapes. Although no finan-cial statement data are needed for the earlier phases of this study, even-tually the scope will be extended to relating market prices to the financialstatement items, namely the eamings numbers.

Criterion (2) was used because weekly price and volume data on NYSEfirms are relatively easy to obtain. The Center for Research in SecurityPrices (CRSP) provided tapes which contain daily price, volume, andtransaction information on all firms on the NYSE for the years 1961through I965.i«

Criterion. (3) was selected in order to avoid a ciiistering of announcement"EflBciency is defined in terms of E(£ — x*)^, where £ is the forecasted value of

reported earnings and z* is actual value. The closer the expectation is to zero, themore efficient the forecast is. Note that a forecast may be unbiased but very ineffi-cient. The distinction between efficiency and unbiasedness is more important to theinterpretation of the findings presented later.

" Without the cooperation of CHSP, the data collection chore would have beenoverwhelming.

CONTEST OF ANNUAL EARNINGS ANNOtTNCEMENTS 71

dates during any time period. Without this criterion, the sample datawould exhibit a lsirge clustering of announcements in the months of Feb-ruary, March, and April because two out of three Compustat firms are12/31 firms. In subsequent analysis, an. attempt will be made to removethe effects of market-wide events from the individual security's volumeand price data. When earnings announcements cluster, they become a formof market-wide price indexes and the volume statistics. Hence, any at-tempt to remove the effects of market-wide events would eliminate theeffects of the earnings report as well.

The purposes of criteria (4) and (5) are similar in that they attempt tominimize any ambiguity associated with an obser\'ed reaction in the weekof the earning annoimcement. If these criteria were not applied, therewould be a joint effect, and it would be extremely difficult to separate theannouncement effects of dividends or stock splits from those of the earningsreport. ^ Criterion (6) was chosen so that there would be weeks where few,if jmy, announcements were released. To the extent that news items areannounced in weeks other than the earnings announcement week, compar-ing those weeks with the earnings announcement week compares the infor-mation content of the earnings reports with that of other types of ne vsannouncements, which is not the issue under study.

Both the direction and magnitude of any potential bias introduced by theselection criteria are difficult to assess. Criteria (1) and (2) led to the selec-tion of the larger firms in the economy. The average total assets (per finan-cial statements) for the 143 firms in 1965 was 167 million dollars, and theiraverage market value of common stock outstanding in 1965 was 189 milliondollars. The effect of selecting larger firms would tend to induce a biasagainst earnings reports because the larger firms are generally associatedwith a greater flow of additional infomiation than smaller firms.

The effect of criterion (3) was twofold: (a) Out of the subpopulation ofCompustat, XYSE firms, the criterion tended to select the smaller firms,even though they are probably stiU larger than average for the economyas a whole. In 1965, the average total assets for Computsat, XYSE firmswere 441 million dollare, and the average market value of common stockoutstanding was 564 miUion dollars, (b) A greater proportion of retailersand food processors appears in the sample than would have been obtainedif firms had not been restricted non-12/31 firms. Retailers comprise 14.6per cent of the sample, while 17.5 per cent of the firms are food processors.The expected percentages would have been 6.8 and 10.0, respectively, if

' A pilot study with similar objectives did not exclude firms with dividend an-nouncements in the same week as earnings. The investor reaction in t erms of volumewas almost twice as large as the reaction observed in this study. Stock splits wereexcluded because previous research has found that stock splits possess informationcontent. See Eugene F. Fama, et al.^ "The Adjustment of Stock Prices to New In-formation," Report 6705 (Center for Mathematical Studies in Business and Eco-nOTQics, Graduate School of Business, University ol ^hisago, 1967), forthcoming inthe International Economic Review.

72 fliLlIAM H. BEAVER

a random sample were drawn from the Compustat, XYSE subpopulation.With respect to the implications for infonnation content, there are to beno obyious reasons why these firms would constitute a biased sample, withthe one exception that retailers tend to report financial statement datamonthly which would tend to induce a bias against finding informationalvalue in annual eamings reports. In fact, in the analysis described later,both the price and volume reactions were less dramatic for the retailers andfood processors than for the other firms ia the sample.

It is possible that the selection criteria, especially criterion (6), may in-duce some bias in the opposite direction. As long as the criteria are visibleex ante, the population for which the study's findings are relevant can beeasily identified. Also, the sample criteria can be relaxed in future studiesto discover the generality of the findings presented here for other popula-tions.

Data Collection. The first step was the identification of firms that wouldcomprise the sample. Meeting the criteria in any one of the five years(1961-1965) was a sufficient condition for a firm's inclusion for that year.The result was a sample of 143 firms. Because all firms did not meet thecriteria in every year, the 143 firms gave rise to 506 annual earnings an-nouncements. The date of the earnings announcement wa« obtained fromthe W(dl Street Journal Index.

The distributions of financial statement dates and announcement datesappear in Table 2. Restricting the sample to non-12/31 firms was success-ful in reducing the clustering of dates. The most frequent month in whichthe fiscal year ended (June) represents only 23.8 per cent of the sample,while an unrestricted sample would have resulted in 67 per cent in a singlemonth (December). With respect to announcement dates, the highestthree-month period (September, October, and November) contains 37.6 percent of the announcements, while under an unrestricted sampling procedurethe highest percentage would have been approximately 67 per cent (duringFebruary, March, and April). The most frequent month of announcement(October) represents 13.4 per cent of the announcements, which is onlyslightly higher than the percentage that would be obtained under a com-pletely uniform distribution throughout the year (9.1 per cent).

One by-product of the data gathering was some insight into the timelag between the financial statement date and the announcement date (seeTable 3). The median lag was 9 weeks, only 3 per cent of the announce-ments were made by the end of 4 weeks, and 93 per cent of the eamingshad been reported by the end of 13 weeks. A possible avenue for futureresearch would be to study the information content of the time lag itself(e.g., is "bad" news reported less rapidly than "good" news?).

Definition of Variables. The next step was to compute the followingvariables for each firm on a weekly basis for the 261 weeks {irom January1, 1961 to December 31, 1965):

CONTENT OF ANNUAL EARNINGS ANNOUNCEMENTS 73

no. of shares of firm i„ _ traded in week ( 1

no. of shares outstanding no. of tradingfor firm i in week / days in week t

no. of shares tradedfor all NYSE firms

,- _ in week t 1no. of shares outstanding no. of trading

for all NYSE firms days in week tin week t

Dit = cash dividend "paid" on share of firm i in week t,Pit = closing price for share of firm i at end of week £,

Pi t-l = closing price at end of week t — 1, adjusted for capital changes(e.g., stock splits and stock dividends),

(&P) t = closing value of Standard and Poor^s Price Index at end of week t,-i = closing value at end of week t ~ 1.

Vit is a weekly average of the daily percentage of shares traded. Weeklyvohmie was divided by the number of shar^ outstanding so that the resultswould not be dominated by those firms with the largest number of sharesoutstanding. The percentage of shares traded per week were then di\adedby the number of trading days in order to adjust for the fact that not allweeks ha\'e the same number of trading days.

VMI reflects the level of volume for all NYSE firms. The weighting schemeimplicit in this voiume index assigns greater weight to percentage of sharestraded of firms with the larger number of shares outstanding. While thisfeature is not entirely satisfying, its use is defended on the grounds thatthis index is much easier to obtain than an index that assigns equal weightto all firms and because there is no reason to believe the use of this indexleads to either an upward or a downward bias in the findings regarding theinfonnation content of earnings reports.

Rit is the natural logarithm of the price relative and can be viewed as ameasure of price change or as the rate of return of the security assumingcontinuous compounding. ^ S.^, is a similar measure for 425 industrial

1 The properties of Rn are further described in Fama, op. dt.; Benjamin F. King,"Market and Indiistry Factors in St«ck Price Behavior," Journal of Business,XXXIX (January, 1966), 139-90-, James H. Loxie and Lawrence Fisher, "Rates ofReturn on Investment in Common Stocks," Journal of Business. XXXVII (January,1964), 1-21.

74 WILLIAM H. BEAVER

NYSE firms. This statistic has some limitations as a market-wide indexof price change and in many respects is less preferable than some recentlydeveloped indexes, notably Fisher's Link Relative.^' However, again itsuse is defended on the same grounds as those for the market-wdde index ofvolume. The Fisher Link Relative has been computed for monthly dataonly. To construct a similar index on a weekly basis is a research projectin itself. Not only is the S & P index easier to obtain but it was found inother studies that results were insensitive to which index is used. * Withinthe context of this study, there is no reason to believe that the use of theS & P based index vdU lead to an overstatement or understatement of theinformation content of eamings reports.

Volume Analysis—Unadjusted for Market InfluencesVit was computed for each week t in the report period for each of the 506

eamings announcement j . The report period is defined as the 17 week periodsurrounding the announcement date (8 weeks before the announcementweek, and 8 weeks after). Then the Vt (averaging across j) was computedfor each of the 17 weeks, and the results appear in Figure 1. The dotted linedenotes the value of Ft in the nonreport period (i.e., that portion of the261 weeks not included in the 17 week report periods).

The evidence indicates a rather dramatic increase in volume in the an-nouncement week (week 0). In fact, the mean volume in week 0 is 33 percent lai^er than the mean volume during the nonreport period, and it isby far the largest value observed during the 17 weeks. Investors do shiftportfolio positions at the time of the eamings annoimcement, and this shiftis consistent with the contention that eamings reports have infonnationcontent.

The contention is further supported by the behavior of investors in theother weeks. Eight weeks prior to the announcement, volume is belownormal, which suggests that investors may postpone their purchases andsales of the security until the eamings report is released. The four weeksafter the announcement, when the annual reports are received, exhibitslightly above normal volume and hence permit a more thorough evaluationof the eamings data.

The investor response appears to be very rapid, for almost all of theabove-normal activity occurs during week 0. This finding supports previousstudies that also show investors respond quickly (as refiected in pricechanges) to new pieces of information (see Fama, et al.).

Perhaps some comment is in order regarding the overall level of volumethroughout the year. The volume statistics reported in Figure 1 are muiti-

" For a discussion of the S & P Index vis-fi.-vis Fisher's Index, see LawrenceFisher, "Some New Stock Market Indices," Journal of Business, XXXIX (January,1966), 191-225.

'* Fama, et al., op. cit.

CONTENT OF ANNUAL EARNINGS ANNOUNCEMEISTTS /D

plied by the factor of lO''. The average volume in the nonreport period is.00112—that is, the average daily percentage of shares traded is slightlygreater than one-tenth of one per cent of the shares outstanding. This im-plies an annual turnover of approximately 25 per cent and a weekly turn-over of .5 of one per cent. If corporation X has 10 miUion shares outstanding,during a normal week 50,000 shares wUl be traded uith an expected volumeof 66,667 shares during the earnings report week.

Volume Analysis—Adjusted for Market InfluencesThe section will present an analysis which attempts to remove the effects

of market-wide events upon the individual security's volume. The motiva-tion for the analysis is two-fold. (1) It is possible that the abnormally highvolume may be caused in part by market-wide pieces of information thatare released at the same time as the earnings announcements. Since theearnings annoxmcements are released almost uniformly throughout the year,this is not a very plausible explanation of the finding. Nevertheless, re-moving the market mde effects should allay any fears that this unlikelysituation does account for the results. (2) More importantly, the analysiswill serve to reduce "noise" in the volume data. Noise is any movementsin volume due to unspecified factors, one of which is market-Ti'ide eventsthat would cause increases in the volume.

Analysis for Nonreport Period. The following model was used to abstractfrom market-uide factors:"

Vit = a* + biVm + en •

Estimates of a,- and fc,- were obtained from linear regressions based uponobservations from the nonreport period. The observations from the reportperiod were deleted from the regression because if earnings announcementshave infonnation content, the assumptions of the classical regression modelare \dolated during the report period (e.g., E(eit) ^ 0).

Some summary statistics relating to the regressions appear in Table 4.The mean volume of the sample firms is much higher than that of themarket index. One reason is the different weighting scheme implicit in eachmeasure. The market index assigns greater weight to firms with the greatermmiber of shares outstanding. If these firms have lower volume (expressedas a percentage of shares outstanding), then the market index would beexpected to have a lower mean. Another explanation is that the sample

Tationale for using this particular model is two-iold: (1) It is a simplerelationship, and there is no obvious reason why a more complex model would bemore appropriate. (2) It is analogous to the model that will be used to remove effectsof market-wide events upon the price changes of individual securities. The paper willlater indicate that such a model seems to be a reasonable way to characterize pricechanges. Hence, it would seem reasonable t-o assume a similar process is generatingvolume over time as well.

76 WILLLAM H, BEAVER

selection criterion implicitly favored higher tumover securities. But it is notobvious why that should be tme nor what implication it has for inferencesregarding information content.

The average correlation coefficient was low, implying that removing theinfiuence of TVf should have little effect upon the analysis. In spite of thelow association, the sign of the correlation coefficient was positive for 139firms and negative for only 4. These two findings taken together suggestthat the market infiuence on an individual firm's volume is significantlydifferent from zero but that its magnitude is small."

The residual, e,*, is that portion of an indi^ddual security's volume thatcannot be explained by market-^iide events as refiected in Vm . The meanof e, (averaging across time for given firm i) is forced to be zero by themechanics of the regression computations. However, the mean of et (aver-age across firms for a given week t) may be nonzero. An inspection of itsdistribution for the 261 weeks provides some interesting insights (seeFigure 2).

The distribution is skewed to the right, as indicated by the fact that 58per cent of et are negative and 42 per cent are positive. The median of etis — .02 and its mean is zero (again this must be true because of the me-chanics of the regression computations). The e./s are even more asyxQ-metrical, with 64.6 per cent negative and 35.4 per cent positive. One inter-pretation of the asymmetry is that infonnation is provided to investors indiscontinuous "lumps" rather than smoothly or continuously over time.

Residual Analysis for the Report Period. The residual, e^t, was computedfor each week t of the report period for each of the 506 eamings annoimce-ments^ in the following manner:

i= 1, • • • , 143eyt = Vit — at — biVnt j = 1, • • • , 506

i = _ 8 , ••• , + 8

where a; and 6 were obtained from the regressions in the nonreportThen the e* was computed for each of the 17 weeks, and the results appearin Figure 3. A positive residual imphes above normal volume; negative,below normal; and zero, normal volume.

The behavior of the volume residual is the same as that of the previousanalysis. There is a large peak in week 0, where the mean volume is ap-proximately 30 per cent higher than during the nonreport period (i.e.,.33/1.12, mean residual in week O/mean volume in the nonreport period)and is about 40 per cent higher than the mean volume in the weeks prior

" The probability that the expected value of the correlation coefficient is less thanor equal to zero is less than 1 chance in 100,000.

" Note that the subscript i refers to firm i or security i, but j refers to an earningsannouncement. Hence a; and 6,- may be used a maximum of five times; its frequencyof use will depend upon the number of earnings announcements of firm i or securityi included in the sample of 506 announcements.

CONTENT OF ANNUAL EARNINGS ANNOUNCEMENTS 77

to the announcement. Again the volume during these weeks is abnormallylow, while slightly above normal volume persists for four weeks followingthe announcement week. The interpretation of these findings is the same asthat of the previous analysis. In short, the results are verj- consistent with.r,he contention that earnings announcements possess information content.

Because a comparison of mean values can often be misleading, two addi-rional comparisons were made to see how imusual is an et of .33, whichwas the value observed in week 0. The first comparison examined thevalues of t in the report period and those in the nonreport period (seeMgure 2). Out of the 261 nonreport period values of e*, only 4 had valuesexceeding .33. Although such a comparison is admittedly a crude approxi-mation, it does suggest that the value in week 0 is unusually high.

Moreover this comparison tends to understate the unusual nature ofthe week 0 residual. The e* during the nonreport period is based upon amaximum of 143 obsen' ations per mean, while the et in week 0 (as \seilas the rest of the report, period) was based upon 506 observations. Sincethe ejj's are less than perfectly correlated, the dispersion of the distributionof et would decrease as the number of obsen^ations per mean increases.Hence, if the distribution in the nonreport period were also based upon506 observations per mean, its dispersion would be smaller and the numberof values above .33 would be fewer. Another factor leading to an under-statement is that Si in the nonreport period was based upon residuals takenfrom the same week, while the mean residual in week 0 was based uponobservations taken from different weeks. If contemporaneous residualsare more highly correlated than noneontemporaneous residuals (and theei'idence suggests they are), then a distribution of et in the nonreportperiod based upon noneontemporaneous observations would have a smallerdispersion and fewer values above .33.^' The major point is the (comparisonindicates that the mean residual in week 0 is uniisually high, in spite of thefact that the comparison tends to understate how unusual it really is.

A second comparison involved the analysis of the frequency of positiveresiduals in each report period week as compared with the number duringthe nonreport period (see Figure 4). The behavior of the positive r^idualsis consistent with the pre\aous relationships obsen'ed in Figures 1 and 3.Prior to the announcement, the frequency of positive residuals is belowthat of the nonreport period, while the frequency' is slightly above normalafter the annoimcement. By far the largest frequency occurs in week 0, andthere is an extremely small probability that such a high number of positiveresiduals could have occurred by chance." This second comparison sug-gests the same inference drawn from the first—namely, the volume in week0 is an unusually high value. In sum, the behavior of volume uniformly

" T h e serial correlation is reflected in the positive autocorrelation coefficient ofthe residuals (see Table 4). Another indication is that the four values of ?; exceeding.33 occurred in a five-week period.

1- The probability is less than 1 chance in 100,(X)0.

78 -VnLLIAM H. BEAVER

supports the contention that earnings have information content for individ-ual investors.

In ' ome respects, these findings do not reflect the entnre extent to whichactivity is above nonnal in week 0. Not all of the eamings announcementsof the "l43 firms were used—in fact, only 506 out of a possible 715. The 17week periods sumDunding the remaining 209 are included in the nonreport.period This will tend to induce a bias against eamings reports since volumeactivity is increased in the nonreport. period by the inclusion of the 209"report periods." The extent of this bias could be senous because one ofthe reasons for placing a report in the 209 group was the announcementof eamings and dividends in the same week which would produce evenmore price and volume acti^dty than the 506 announcements studied. Ho •-ever there are also compensating factors. Although the activity m week 0was above normal, the activity in the weeks prior were below normal forthe 506 observ^ations. If this tends to be true of the deleted announcenientsas well, the bias may not be so great. If the 209 obsen-ations were deletedfrom the nonreport period, to be completely consistent, other types of newsannouncements would also have to be deleted for the same reasons. Theresult would be virtuaUy no observations in the nonreport. period, bmcethe nonreport period does include these events, it is important to stress thefact that comparing the eamings report periods ^dth the nonreport penodinvolves a comparison of the infonnation content of eammgs reports ^aththe average amount of infonnation being released dunng the nonreportperiod Bv necessity, this is a bias against eamings reports since the appro-priate coniparison would be a nonreport period mth no infonnation at all.

Price Analysis-Adjustedfar Influence of Market-Wide EventsIf eamings reports convey information in the sense of leading to chajiges

in the equUibrium value of the current market price the "^S^^ .ude « . heprice change (without respect to sign) should be larger m week 0 than^ the nonreport period. The first step in making this prediction opera-

is to remove the effect of market-wide events upon the mdividuaprice change. The reasons for wishing to abstract from the.esimilar to those cited in the volume analysis.- The model used

h^L^Lt suggested by Sharpe, and it provided the motivation for usmgan analogous model for volume. ^ The Sharpe model states:

Rit = a,- + hiRitt + Uii.

R,, is a measure of the price change of security .• during ^/^^f^^^L f ri h g e dunng time penod ( for 425 inis a measure of the p g

RL is a measure of average price change dunng time pd^trial NYSE firms. Both variables were defined earher. The residual,

" W i l W ^ F . Sharpe, "A Simplified Model for Portfolio Analysis," ManagementScience, IX (January, 1963), 277-93.

CONTENT OF ANNUAL EARNINGS ANNOUNCEMENTS 79

Uit , represents that portion of the individual security's price change thatcannot be accounted for by the effects of market-wide events as reflectedin Rut.

The Sharpe model has been investigated by Fama et al. and by Scholesand was helpful in abstracting from the infiuence of market-wide factors.King's study of monthly price changes foimd that, on the average, 31 percent of the variation in an individual security's price change can be ex-plained by market-\%'ide factors as refiected in a market-wide index of pricechange. ^ For these reasons, a price change analysis, unadjusted for theinfluence of market-wide factors, was not conducted. The e\ddence willlater indicate that if such an analysis had been conducted, the results wouldbe essentially the same as those reported here.

Since the direction of the price change cannot be specified, a knowledgeof the investors' expectation model(s), some transformation of Uit thatabstracts from its sign, is needed. One such transformation is the squareof the residual (i.e., uh). If earnings reports possess infonnation content,u]t should be greater during week 0 than during the nonreport period. Themean of u'a during the nonreport. period is simply the variance of thatvariable (si^).-^

The relationship between the squared residual in week 0 and the averagesquared residual during the nonreport period can be expressed in the formof the ratio, Uit, where the numerator is uit and the denominator is s^. Ifthe- ratio is greater than one, the residual price change is larger than normal,and conversely for a ratio of less than one. The prediction is the mean of XJ(a eraging across announcements) will be greater than one during week 0,if e-amings reports possess information content.

Analysis of Nonreport Period. Estimates of a,-, 6^, and s,- were obtainedfrom regressions based upon the nonreport period. The Gbser\ ations fromtho report period (i.e., the 17 weeks surrounding each announcement) weredeleted from the regression because if earnings have information content,the assumptions of the classical regression model are violated during thereport period (e.g., the variance of the r^iduals during the report periodis not equal to the variance during the nonreport period).

Some summary statistics relating to the regressions appear in Table 5.The mean price changes tend to be lower for the sample firms than for themarket index. Since the Rn can also be interpreted as a rate of return, thelower returns for the sample firms would sugg^t that they are less risky

" Fama, et al., op. dt.; Myron Scholes, "The Effects of Secondary Distributionsupon the Market Price" (paper presented at the November, 1967 session of theConference for the Study of Security Prices held at the Graduate School of Business,University of Chicago); and King, op. dt. The percentage refers to the period August,1952 through December, 1960.

" The variance o-,-* = E[uit — £(u,()]'. s,* is the estimat-e of o-, , computed fromsample data. Si^ = [^t-i{uiiy]/T, where T = number of weekly observations forthe nonreport period for security i.

80 WILLIAM H. BEAVEH

than the firms comprising the index. An inspection of the distribution of 6,-also lends support to that contention. Sharpe states that &,• can be viewedas an operational measure of a security's riskiness, with larger values of 6implying greater riskiness.^* A &,- of one denotes a security of "average"riskines3. The average 6,- for the sample firms is less than one (.89), whichsuggests that the sample firms are less risky. However, the discussion in thesection on definition of variables indicated that the definition of Rut basedupon the S & P index may be subject to measurement error. An errors-in-variables model suggests that measurement error in the independent varia-ble, even if it has a zero expectation, will induce a downward bias in theestimates of the regression coefficient associated with the independentvariable (i.e., hi).^° Efforts were undertaken to assess the extent of thedownward bias by computing 6 for the sample firms, using monthly dataand Fisher's Link Relative as a definition of Rut • The median bi was .993,suggesting the sample firms are of average riskiness relative to NYSEfirms (i.e,, the firms that comprise the Fisher Index).

On the average, the association between Rn and Rut was low. Only 6per cent of the variation in Rit can be explained by the variation in RMI,as meastired by the square of the average correlation coefficient. The im-plication is two-fold: (1) Removing the infiuence of R^t should have littleeffect upon the results, relative to what would have been obtained if Ritwere analyzed rather than «i(. (2) The explanatory power is much lowerthan that obtained by King, suggesting that either weekly data have morenoise than monthly data or that Rut was not properly defined, or both.The presence of either factor vnR make it more difficult to detect any priceeffects of the earnings reports.

The distribution of tJt (averaging across 143 firms, t = 2, •••,261)during the nonreport period is shown in Figure 5. It will be used as a basisfor assessing the significance of the U^s obser\'ed during the report period. ^

Price Residual Analysis for Report Period. The residual, un, was com-puted for each week t of the report period and for each of the 506 eamingsannouncements J in the following manner:

i^ 1, • • • , 143Uit = Rn — a, — hiRut j = 1, • • • . 506

i = - 8 , • • • , + 8 .

The residual was then squared and divided by the variance of the residualsfor its firm during the nonreport period, as follows:

"William F. Sharpe, "Capital Asset Prices: A Theory of Market Equilibriumunder Conditions of Risk," Journal of Finance, XIX (September, 1964), 425-42,

" J. Johnston, Econometric Methods (New York: McGraw-Hill, 1963), 148fF." The distribution is skewed to the right. One explanation for this phenomenon

is the leptokurtic nature of the underlying un's (see Fama, op. cit.). The distribdtionof «i( is also skewed in the same direction. Alttiough the mean of «« is one for eachsecurity during the nonreport period, only 26 per cent of the olwervations exceed one.

CONTENT OF AJHs'UAL EARNING ANKOUNCEMEmB 81

^ =j =t =

1,1,

- 8 ,

• • • , 1 4 3

••• ,506• • • , + 8 .

L't (averaging across j) was computed for each of the 17 weeks of the report])eriod, and the r^ults appear in Figure 6.

The magnitude of the price changes in week 0 is much larger (67 per centiiigher) than the average during the nonreport period. The above normalprice activity is what would be expected if changes in eqiuUbrium pricesare more likely to occur when earnings reports were released, and hencethe evidence is very consistent with earnings reports possessing informa-tional value.

Although the price activity is highest in week 0, the next largest valuesoccur in the weeks immediately contingent to week 0. Price changes areabo%'e average in the week immediately prior to the announcement, whichmay reflect information leakage or the fact that the Wall Street JournalV'Rs not the first source to report, the earnings in some cases. Above normalacti\"ity is also present for two weeks after the announcement, during whichtime the annual reports are released and are evaluated by investors.

The below price activity in weeks —8 through —2 is open to at leasttwo interpretations: (1) There is a below nonnal amount of informationcoming onto the market at this time. (2) The below normal price acti'v^tyis a result of the below normal volume also obser\ ed during the sameperiod. More will be said about both (1) and (2) later.

The beha^dor of the mean residual, u*, also indicates greater price activ-ity in week 0 (see Table 6). The mean in week 0 is .00500, which is the largestvalue obsen'ed during the 17 weeks and is four times larger thaa the aver-age value of Rit during the nonreport period (.00125, see Table 5). Themeans give the impression that serial correlation may be present in thedata. However, the average autocorrelation of the price residuals was quitelow (—.08) during the report period. The low degree of autocorrelationsupports the similar findings of Fama and his conclusion that the marketmoves to new equilibrium positions quickly." Further evidence of this isrefiected in the fact that the bulk of the price reaction does occur in week 0(see Figure 6). The low autocorrelation also suggests that the price changeswore permanent in nature and were not reversed in subsequent weeks. Infact, the autocorrelation of the residuals in the weeks immediately afterthe announcement week was slightly positive.^

Two additional comparisons (analogous to those made in the volimieanalysis) were conducted to see how unusual an l ( of 1.67 is. The first com-parison examined Ut in the nonreport period (see Figure 5). Out of 260values, only 11 exceeded 1.67. The comparison suggests that tho price ac-

" Fama, op. dt.» The autocorrelation was examined on a week-by-week, cross-sectional basis,

i . e . . , = \T.T'^ (e:<ei<-O]/CLPi (e,,)'], f = - 8 , ••• , + S .

82 WILLIAM H, BEAVEB

tivity in week 0 is unusually high, in spite of the fact that such a compari-son tends to understate how unusual it really is. * The second comparisonexamined the frequency of Ujt's larger than one relative t-o the frequencythat occurred during the nonreport period (see Figure 7). The frequencyof values above one is greatest in week 0, with the next highest values oc-curring in the weeks adjacent to the announcement week. There is an ex-tremely small probability that such a high number (181 in week 0) couldhave occurred by chance.'" The interpretation is the same as that of themean analysis—namely, there is above nonnal price activity when eamingsreports are released. What this analysis reveals that the mean analysis doesnot is the fa«t that the abnormally high mean is not caused by a few ob-servations dominating the results but rather by a substantial proportionof the sample data.

In summary, the behavior of the price changes uniformly supports thecontention that earnings reports possess information content. Observinga price reaction as well as a volume reaction indicates that not only areexpectations of individual investors altered by the eamings report but alsothe expectations of the market as a whole, as reflected in the changes inequilibrium prices.

Relationship between Ihe Volume and iJie Price Findings. The previoussentence raises the issue, "how much of the increased price activity can beattributed merely to the fact that there is more 'action' in the security,rather than to changes in equilibrium prices?"

One way to approach this question is to view the price change duringa given time period as a sum of price changes on each transaction that oc-curred during that period. In a worid of uncertainty, the price change fromeach transaction can be treated as an observation from a probability dis-tribution of the investor's assessment of what the price change should be.The price change per period, then, is a sum of random variables. If trans-actions occur as if they are independent over time (evidence on daily,weekly, and monthly price changes suggest they do), the variance of theweekly price change will increase in direct proportion to the number oftransactions that occur during the time period.'

*' The reasons for understatement are similar to those stated in the volume analy-sis. See p. 77.

»> The probability is less than 1 chance in 100,000.' The evidence regarding serial correlation of daily and monthly price changes

can be found in Fama, op. cit. and Fama, et al., op. cit., respectively. The averageautocorrelation coefficient for weekly changes in this sample was —.08, which wouldcause the variance to increase less than proportionately with the number of transac-tions. Within a given trading day, the autocorrelation may be higher (e.g., becauseof certain institutional factors, such as clustering of limit orders or stop loss orders).However, the existence of arbitragers should prevent the autocorrelation from beingvery large. In order for the price activity to be explained entirely by increased trans-action activity, the autocorrelation would have to be one. This would be highlyunlikely because oi the empirical evidence cited and the opportunities for arbitrage.

CONTENT OF ANNUAL EARNINGS ANNOTINCEMENTS 83

The issue now is what is the appropriate measure of th(! number oftransactions occurring during a given period. If the volume is used as ameasure, then Ut in week 0 would be expected to be 1.30 merely becauseof more action in the security. The remaining portion would be attributedto changes in the equihbrium prices of the securities. Howe\'er, it is notat all clear that volume (or even number of transactions) is the appropriatemeasure, because it refiects only the explicit transactions that occur.

It could be argued, with considerable support from economic theory,•:hat the expectations of aU investors influence the market price, whetheror not they engage in a purchase or a sale. If the market acts in this manner,i:he total number of transactions, explicit and implicit, are the same pertime period. Hence all of the above average price activity can be attributed1:0 changes in equilibrium prices.

Additional empirical research is needed before this issue will be resolved.The research would consist of studjnng increased volume activity due toI'easons other than information coming onto the market. An initial analysisof the seasonal variation in volume (Vut) from 1946 through 1966 revealedthat the volxmae is greatest during the months December and Januarj'. The(xplanation seems to stan from tax considerations rather than an abovenormal flow of information. R^earch also indicated that the price variabil-ity of RMI during these months (i.e., Ut) was only .996, indicating no aboveaverage price variability during these months. This finding lends supportto the position that none of the price activity in week 0 is due merely tomore motion.

Before leaving this topic, note that isolating the volume effects on pricechanges is of concern only to the extent one wishes to distinguish betweeninformation that alters the expectations of the market as a whole from in-formation that alters only the expectations of individual investors. All ofthe price activity can be attributed to infonnation in the latter sense.

Frequency of Othsr News Announcements during Report Period

The purpose of this analysis was to discover if there was any clusteringof other news announcements aroimd week 0 that might possibly accountfor the volume and price reactions. As indicated earlier, the sample designexcluded any firms that announced dividends in the same week as earningsor any flrms that split their stock during the report period. However, it isconceivable that di^adends announcements might cluster in weeks immedi-ately prior to and after week 0 or that other types of announcements (e.g.,management eamings forecasts) might cluster in week 0. To examine thispossibility, the occurrence of other news announcements in the Wall StreetJournal during the 506 report periods was examined (see Table 7).

By far the most frequent type of announcement was di\adends, whichexceeded the frequency of all other types of announcements by a factor of

84 WILLL4.M H. BEAVER

9 to 1. With respect to the purpose of this analysis, there is no clustering ofdividend annoimcements in weeks —1 or 4-1; in fact the opposite seems tobe tme. Also there is no clustering of any other type of announcements atany time during the period, including week 0. The volume and price reac-tion in week 0 does not appear to be attributable to the clustering of othernews announcements.^-

Suggestions for Future Research

The dramatic price and volume reaction indicates that investors do lookdirectly at reported eamings and do not use other i ariables to the exclusionof reported eamings. The e\ddence also indicates that news announcementsoccurring prior to the eamings report do not entirely preempt the informa-tion content of reported eamings. Given these findings, one of the firstextensions of the study will be to explore the possibihty of constructingexpectations models that will permit a prediction of the direction andmagnitude of the price residual.

The results of a recent study by Ball and Brown in this area are veryencouraging.'' They used an earnings model similar in form to the priceand volume models described in this study (e.g., changes in the earnings ofan individual security were viewed as a linear function of market-wide indexof eamings changes). The sample was divided into two groups: instanceswhere the earnings residual was positive (actual eamings were higher than"expected") and instances where the eamings residual was negative (actualeamings lower than "expect.ed"). The behavior of the price residuals forthese two groups was examined, and the findings were: (1) The sign of thecumulative price residual (summed over a 12 month period including theannouncement month) was highly associated with the sign of the eamingsresidual. (2) There was a persistent upward drift in the cumulative meanprice residuals for the positive eamings residual group. This drift started 11months prior to the earnings announcement, and over 90 per cent of thedrift had taken place by the beginning of the announcement month. Thenegative earning group exhibited an analogous behavior pattern.

The findings indicate that reported eamings are associated mth underly-ing events that are perceived by investors to affect the market price. Be-cause earlier news announcements convey some of the same information asthe eamings reports, investors are able to use this information to revisetheir forecasts of eamings and to adjust the price accordingly. In fact, bythe beginning of the announcement month, investors form largely unbiasedforecasts of reported earnings, even though the reported earnings are above

'^ As measured in terms of number of news announcements per week, the flow ofinformation during the weeks prior to the announcement does not appear to be belownormal and hence wouJd not account for the below normal price activity during w eeks- 2 through - 8 .

" Hay Ball and Philip Brown, "An Empirical Evaluation of Accounting IncomeNumbers," Journal of Accounting Research, 6 (Autumn, 1968). pp. 159-78.

C<:)NTENT OF ANNUAL EARNINGS ANNOUNCEMENTS 85

or below nonnal relative to their historical relationship with market-wddeea;mings.

Although the forecasts are unbiased, they are not very efficient, for ifthey were, there would be no volume or price reaction when earnings reportswere released.** The Ball and BroTsu findings and the findings pr^entedhere are mutually supportive with respect to the infonnation content ofearnings reports and also are uniformly consistent with the iindings ofprevious studies in the behavior of security prices. One extension of the re-search presented here will be to replicate the Ball and Brown study on thissample of non-12/31 firms (the Ball and Brown study dealt exclusively ^ t h12/31 firms) and then to attempt to predict the magnitude, as well as thesign, of the price residual.

A second area of further research is the application of this methodologyto other types of news announcements. At an earlier meeting of the Con-ference, Green and Segall explored the information content of interim re-ports. An analysis of volume and price changes during the announcement ofinterim earnings would provide a different approach to this same issue. Theinformation content of dividend announcements is another topic that hasreceived much attention and still is in need of additional empirical investiga-tion. Such research "noil indicate the importance of annual earnings an-nouncements relative to other kinds of infonnation.

Perhaps the most important extension of this study would be dealingwith the normative issue, "Should decision makers perceive earnings reportsto possess informational value?" The normative question can be approachedby selecting an event of interest to decision makers (preferably as free aspossible from the influence of their perceptions) and by investigating theability of earnings data to predict that event. A few studies of this type havebeen presented at earlier meetings of the Conference, but much more workis needed in this area." Hopefully, the findings presented here with respectto the positive question will provide greater insight into the normativeissue as well.

'* The distinction between unbiasedness and efficiency was discussed in footnote 9.'* James O. Horrigan, "The Determination of Long-Term Credit Standing with

Financial Ratios," Empirical Research in Accounting: Selected Studies, 1966, Supple-ment to Vol. 4, Journal of Accounting Research, pp. 44r-62, and William Beaver,"Financial Ratios as Predictors of Failure," ibid, pp. 71-102.

86 WILLIAM H. BEAVEB

TABLE 1of Selection Criteria upon Sample Size

Criteria No. of firms

CompuBtat firms (step 1)'Less: 12/31 firms

Non-12/31 firms (step 2).Less: Non-NYSE firms .

NYSE and non-12/31 (st-ep 3)Less:

More than 20 announcements per year 48Dividends in earnings announcement week 39

, Stock split during report period 7Other'' 5

Sample size (step 4)

896599

29755

242

143

' Sample criteria were applied sequentially in four stages. The sample size aftereach stage is denoted by parenthetical comment (e.g., steps 1, etc.).

*> Miscellaneous reasons such as firm's earnings were not reported in Wall StreetJournal.

TABLE 2Distribution of Financial Statement and Announcement Dates

JanuaryFebruaryMarehAprilMayJuneJulyAugustSeptemberOctoberNovemberDecember

Total 100.0

• Total number of firmB equals 143, and total number of announcements equals 506.

COKTEKT OF ANSTTAL EABSIKGB ANNOTJNCEMENTS 87

TABLE 3Number of Weeks between Fiscal Year-End and Date of Announcement

No. of wedcs

Leas than 4456789

101112131415

More than 15

Total*

Percentage of ansouncemeiiU

1.71.54.1

11.614.013.811.211.08.68.66.93.02.21.8

100.0

Cumulative percentage

1.73.27.3

18.932.946.757.968.977.5S6.193.096.098.2

100.0

Total number of announc-ements is 506.

T A B L E 4Summary of Regression Statistics Volume Analysis

Item

Fractile.10.25.50.75.90

No. of obser-vations perfirm in non-

report period

165176193210227

Mean of depend-ent vaiiftbteiVi) X 10»

.33

.53

.881.562.36

Mean of inde-pendent variable

iTi) X 10»

.577

.583

.588

.595

.60S

Correlationcoefficient

.06

.16

.28

.39

.46

Autocorrelationcoefficient of

residuals

.21

.29

.39

.50

.62

T A B L E 5Sum-mary of Regression StaHstica Price Analysis

Item

Mean

Fractile.10.25.50.75.90

No. of obser-vations perfirm in non-report penod

187

165176193210227

Mean of dcmnd-ent variable0?i) X 10'

1.25

-2.13- . 2 61.512.883.98

iitxa <si inde-pendent variable

(Ril) X 10»

1.73

.961.251.512.042.96

Regression coef-

.89

.42

.62

.871.131.44

Correlation

.26

.13

.22

.27

.32

.37

•WILLIAM H. BEAVEE

TABLE 6Analysis of Mean Price Residual

Week

- 8- 7- 6- 5- 4- 3- 2- 1

01234g678

Mean residnal

.00183- .00105- .00029- .00064- .00096

.00019- .00047

.00229

.00500

.00204

.00163

.00120

.00109

.00354- .00040

.00257

.00343

TABLE 7Occurrence of Other News Announcements

Week

- 4- 3- 2- 1

01234

Total

No. of dividendannouncements

433942160

16333241

282

Ko. of all othertypes of

announcements

324544432

31

CONTEXT OF ANNUAL EARNINGS ANNOUNCEMENTS 89

7, X lO'

Average V, X Kf during non-report period = 1J2.

- 8 -6 - 4 -2 0 -¥2Weeks after announcement

YiG. I. Volume Analvsis

+6

Itelativefreqiiencj-

1 olraervatioii 1 observatioa

-.40 - 5 0 -20 -.10 0 .10 .20 .30Value of ?;

FIG. 2. Distribution of e, in the Nonreport Period

9 0 WILLIAM H. BEAVER

e, X 10'

.40

-.10

-.20

No. of

Mean e, X Uf during non.report period = 0.

- 8 - 4 -2 0 +2 +4Vfeek after annouticemait

FIG. 8. Residual Volume Analysis

+6 +8

260

240

220

200

180

160

Expected na of pontive e^'abased on relative frequency inQonxeport p^iod.

-8 -6 -4 -2 0 +2 +4 +6 +8Weeks

Fia. 4. Frequency of Positive e,('s—Residual Volume Analysis

CONTEXT OF ANNUAL EABNINQS ANNOtTNCEMENTS 91

Rslativefrequency

.10

.on

.OB

0.4 0.6 0.8 LO 12 hi 1.6Value of Ui

FIG. 5. Ut in Nonreport Period

5 obaervatima

2X)

1.?

1.1

506 506 -,( 8,

, +8

s? = varianne of residual in theQooreport period.

Mean Ut during nonrqwrtperiod-UOO.

- 8 - 6 - 4 - 2 0 -H2 -H4Week after amunmcement

FiQ. 6. Price Residual Analysis

+6

9 2 WILLIAM H. BEAVEB

180

160

140

120

100

E]g>ected na of Va'a > IJOO basedon relative frequency in nonreportperiod

- 8 - 6 - 4 - 2Fia. 7, Frequency of Cj

0 + 2 +4 +6 +8> 1.00—Residual Price Analysis