An Empirical Evaluation of Accounting Income Numbers RAY BALL* and PHILIP BROWN{ Accounting theorists have generally evaluated the usefulness of aecount ing practices by the extent of their agreement with a particular analytic model. The model may consist of only a few assertions or it may be a rigorously developed argument. In each ease, the method of evaluation has ‘been to compare existing practices with the more preferable practices im- plied by the model or with some standard which the model implies all practices should possess. The shortcoming of this method is that it ignores significant source of knowledge of the world, namely, the extent to which the predictions of the model conform to observed behavior, It is not enough to defend an analytical inquiry on the basis that its assumptions are empirically supportable, for how is one to know that a theory embriees all of the relevant supportable assumptions? And how does ‘one explain the predictive powers of propositions which are based on un- verifiable assumptions such as the maximization of utility functions? ‘Further, how is one to resolve differences between propositions which arise from considering different aspects of the world? ‘The limitations of a completely analytical approach to usefulness are il- lustrated by the argument that income numbers cannot be defined sub- stantively, that they Inek “meaning” and are therefore of doubtful utilty.! ‘The argument stems in part from the patchwork development of aceount- * University of Chicago. University of Western Australia, The authors are indebted to the participants inthe Workshop in Accounting Research at the Univer: sity of Chicago, Profesor Myron Scholes, and Messrs. Owen Hewett and Ian Watte, * Versions ofthis particular argument sppesr in Canning (1920); Gilman (159) Paton and Litileton (160); Vatter (1947), Ch. 2; Edwards and Bell (1961), Ch. 1; ‘Chambers (100), pp 207-08; Chambers (1968), pp. and 102; Lim (1966), esp. pp. 45 ‘snd 649; Chambers (1967), pp. 745-85; Tjel (1967), Ch. 6, exp. pp. 120-31} and Sterling (4967, p65, 150100 JovENAL oF ACCOUNTING RESEARCH, AUTEMN, 1968 ing practices to meet new situations as they arise, Accountanta have had to deal with consolidations, leases, mergers, research and development, price- level changes, and taxation charges, to name just a few problem areas, Because accounting lacks an all-embracing theoretical framework, dissimi- latities in pructices have evolved. As a consequence, net ineome ia an ag- sregate of components which are not homogeneous. It is thus alleged to be ‘3 “meaningless” figure, not unlike the difference between twenty-seven, tables and eight chairs, Under this view, net income ean be defined only as the result ofthe application ofa se of procedures |X;, Xa, --+] to a sot of events {¥s, Ys, +++] with no other definitive substantive meaning at all Canning observes: ‘What i ot cat as a measure of net income ean never be sippoced tobe a fast ia ‘ny sense at ll except that i isthe figure that resulta whea the accountant hax Finished applying the proceduren which he adopts ‘The value of analytical attempts to develop messurements capable of definitive interpretation is not at issue. What is at isiue is the fact that an analytical medel does not itself assess the significance of departures from its ‘implied measurements, Hence it is dangerous to conclude, in the absence of further erpirical testing, that a lack of substantive meaning implies a lack of utili ‘An empirical evaluation of accounting incame numbers requires agree- ment as to what real-world outcome constitutes an appropriate test of use- fulness. Becanse net income is a number of particular interest to investors, the outcome we use as a predictive criterion ia the investment decision aa it is reflected in security prices. Both the content and the timing of existing annual net insome numbers will be evaluated since usefulness could be im- paired by defciencies in either. An Empirical Test Recent developments in capital theory provide justifontion for selecting the behavior of security prices as an operational test of usefulness, An im pressive body of theory supports the proposition that capital markets are both eficient and unbiased in that if information is useful in forming eapital asset prices, then the market will adjust asset prices to thet information 4quiekly and without leaving any opportunity for further abnormal gaint Af, 93 the evidence indicates, security prices do infact adjust rapidly to new information sa it becomes available, then changes in seewrity prices will re- * Cenning (129), p. 28. " Another approsch pursued by Beaver (168) ia to use the investment decision, siti refectoc in transactions volume, fora predictive eriterion, ‘For example, Smuelson (185) demonstrated that » market without biaa ia ite ‘evaluation of formation wil give rit to randomly uctaatng tne serie of prices, See also Cootner (od.) (1962; Fama (1906); Fama and Blame (190); Fama, e a (QGeTy; and dencen (068),EMPIRICAL EVALUATION OF ACCOUNTING INCOME NUMBERS 161 feot the flow of information to the market An observed revision of stock Drices associated with the release of the income report would thus provide ‘evidence that the information reflected in income numbers is useful. ‘Our method of relating accounting income to stoek prices builds on this theory and evidence by focusing on the information which is unique to a particular firm. Specifically, we construct two alternative models of what the market expeets income to be and then investigate the market's reac- tions when its expectations prove false EXPECTED AND UNEXPECTED INCOME CHANGES ‘Historically, the incomes of firms have tended to move together. One study found that about half of the variability in the level of an average firm's earnings per share (EPS) could be associated with economy-wide ‘effects? In light of this evidence, at least part of the change in a firm’ ‘come from oxe year to the next is to be expected. If, in prior years, the in- ‘come of a firm has been related to the incomes of other firms in a particular way, thea knowledge of that past relation, together with a knowledge of the reomes of tose other firms for the present year, yields a conditional e peetation forthe present income of the firm. Thus, apart from confirmation ‘effects, the amount of new information conveyed by the present ineome ‘number ean be approximated by the difference between the actual change in income ard its conditional expectation. ‘But not allof ths difference is necessarily new information. Some changes, {in income reat from financing and other poliey decisions made by the firm. ‘We ascume that, to a first approximation, such changes are reflected in the average change in ineome through time. ‘Singe the impacts of those two components of change—economy-wide and poliey eYects—are felt simultaneously, the relationship must be est ‘mated jointly, The statistical specification we adopt is first to estimate, by Ondinary Least Squares (OLS), the coefficients (aye, yj) from the linear regression of the change in firm j's income (AZj4-,) on the change in the average income of all firms (other than firm j) in the market (AM;,.-.)* using data up to the end of the previous year (r = 1, 2, +++, 1): Mie At boydMpee tie PLE One well documented charscteriti of the security market is that utful sources of information are acted upon and tela sourees nr ignored. This is hardly surprise Ingsince the market consate of «large numberof competing setors who ean gain from ing upon better interpretations of the future than those of their rivals. See, for ‘xanple, Seholes (1067); and footnoted above. This evaluation of thesecuity mavket ‘fers sharply from that of Chambers (1968, pp. 272-73). "More presbely, we focus on information not common to all rms, since some in- tury facta are ot coeidered in thi paper. * Alternatively, 26 to 10 percent ould be awoclated with effects common to all ‘ira when inecme was dafined ua tax-adjusted Retura on Copital Employed. (Source: Ball and Brows (1967), Table 4) "We eall Ms “markot index” of income because itis constructed only from frm traded on the New York Stork Exchange,162 RAY BALL AND PUILIP BROWN ‘whore the hsts denote estimates. The expected income change for firm j in year tis thea given by the regression prediction using the change in the average income for the market in year t Ala = bie + do dM. ‘The unexpec‘ed income change, or foreeast error (2), is the actual ineome change minus expected: y= Aly — Aly C2) tis this forecast error which we assume to be the new information eon- veyed by the present income number. ‘THE MARKET'S REACTION It has also been demonstrated that stock prices, and therefore rates of retum from holding stocks, tend to move together. In one study, it was ‘estimated that about 30 to 40 per cont of the variability in astock’s monthly rate of retum over the period March, 1944 through December, 1960 could bbe associated with matket-wide effects. Market-vide variations in stock returns are triggered by the release of information which concerns all firms, ‘Since we are evaluating the income report as it relates to the individual firm, its contents and timing should be assossed relative to changes in the rate of return on the firm's stocks net of market-wide effets, ‘The impact of market-wide information on the monthly rate of return ‘rom investing one dollar in the stock of firm j may be estimated by its predicted valve from the linear regression of the monthly price relatives of firm ’s common stock on a market index of returns * King (900) The monthly price rel din) + cling PH (Pit ive of security for month m is defined a dividends divided by opening pele (pia): PR im = ions + din/P im: A monthly price relative is thus equal to the diserete monthly rate of return plus ‘nity; ita natural Togarithm is the monthly rte of retura compounded continuously. Tn this pape, mo aesume discrete compounding since the resulta are easier to inter. prot in that form * Fama, fl. (1907) conclude that “rogreasions of security on market returns over ‘ime are x saisnctory method for abstracting from the effcta of general market conditions on the moathly rates of return on individual securities.” Tn artiving at tele conshusion, they found that “sestter diagrams for the [returns oa) individual securities [vis-a-vis the market return] support very well the regression assumptions of linewity, Homoscodaetiity, and serial independence.” Pama, at wl. studied the ratural logarithnie transform of the price relatives, a did King (1068). Hower ‘Blume (1968) worked with equation (3). We also performed teata on the alternative specication: 1, PRjn) = bis + Bila, Cn) + iy co) there tn, denotes the natural logarithmic Funetion, The resulta correspond closely ‘with those repored below.EMPIRICAL EVALUATION OF ACCOUNTING INCOME NUMBERS 103 [PRim = 1) = by + balm — 1+ Dim @) where PRjq s the monthly price relative for firm j and month m, 1. is the link relative of Fisher's “Combination Investment Performance Index” [Fisher (1968)}, and #7 ia the stock return residual for firm j in month m. ‘The value of Zq ~ 1} an estimate of the market's monthly rate of return, ‘The m-subsespt in our sample assumes values forall months since January, 1946 for which data are available. "The residual from the OLS regression represented in equation (3) meas- ures the extent to which the realized return differs from the expected return conditional upon the estimated regression parameters (by, by) and the ‘market index [Za — 1. Thus, since the market has been found to adjust quickly and efciently to new information, the residual must represent the Som nooNOMEIMIC 155UES One assumption of the OLS income regression model” is that M; and ws are uncorrelated. Correlation between them can take at least two forms, ‘namely th inclusion of firm j in the market index of income (Mf), and the presence of industry effeete, The first has been eliminated by construction (denoted by che j-subseript on 1M), but no adjustment has been made for the presence of industry effects, It haa been estimated that industry effects probably account for only about 10 per cent of the variability in the level of a fims income." For this reason equation (1) has been adopted as the appropriate specification in the belief that any bias in the estimates ayy and ‘ty. Will not be significant. However, as a check on the statistical efficiency of tho model, we also present results for an alternative, naive model which predicts that income will be the same for this year as for last. Its forecast error is simply the change in income since the previous year. ‘Asis the eaze with the income regression model, the stock return model, as ‘presented, contains several obvious violations of the assumptions of the OLS regression medel. Fist, the market index of returns is correlated with the residual because the market index contains the return on firm j, and be- cause of industry effecta. Neither violation is serious, because Fisher's index is caleulatod over all stocks listed on the New York Stock Exchange (hence the return on security j is only a small part of the index), and beesuse in- dustry effects account for at most 10 per cent of the variability in the rate "8 That ia, at sarumption necemary for OTS to be the rinimum-variance, linea, unbiased estimator. "The maigaltuds assigned to industry afeote depends upon how bron an indus lay is defined, which in tar depend upon the particular empireal sppliation being considered. The estimate of 10 percent is bated on two-digit classification scheme ‘There in some evidence that industry ffeta might account for more than 10 percent shen the aagochtion a estimated in frst difeences [Brealey (168).164 RAY BALL AND PHILIP BROWN of return on the averuge stock." A second violation results from our predic- tion that, for certain months around the report dates, the expected values, of the v/s ar nonzero, Again, any bias should have little effect on the re sults, inasmuch as there is & low, observed autocorrelation in the 0s, and in no case was the stock return regression fitted over less than 100 observa- tions." Wo assume that in the unlikely absence of useful information about & particular firm over a period, its rate of return over that period would re- fleet only the presence of market-wide information which pertains to all firms. By abstracting from market effects [equation (3)] we identify the effect of information pertaining to individual firms. ‘Then, to determine if part ofthis effect can be associated with information contained in the firm's ceounting income number, we segregate the expected and unexpected. elements of izcome change. Ifthe income foreeast error is negative (that is, ifthe actual change in income is less than its conditional expectation), we define it as bad news and predict that if there is some association between accounting imome numbers and stock priees, thon release of the income ‘number would result in the return on that firm's securities being less than “Tho estimate of 10 per cent is due to King (186). Blume (1968) has recently ‘questioned the magnitude of industry effect, suggenting tht they could be somewhat Joa than 10 por cout His contention ia based onthe observation thatthe sigaifeance sttached to industry effects depends an the asmptions made about the parameters of the distributions underlying atock rates of return, 38ee Table, below. "Fama, ec. (1067) faced a similar situation. The expected values of the stock return residuale ware nonsero for some of the monthe in their study. Stock return regreasions were calculated soparatly for both exclusion and inclusion of the month for which tho stock return residuala were thought to be notaero, They report that Doth seta of reslts support the same conclusions. ‘An alternative to constraining the mean oto be zero isto employ the Sharpe Capi- tal Asset Pricing Model (Sharpe (164) to estimate (3b): Pig — RB ~ 1 = big Bi Um — RE — NF vin o) where RE isthe rak-foe ox ante rate of return for holding period m. Results from cstimating (Wb) ‘asing U.S. Government Bills to measure RF and defining the sbnor- smal return for frm jin month m now aa iy + vq) ave xsentally the same the renults from (2). TBquation (Bb) is still not entirely aeiafsctory, however, since the mean impact ‘fnew information is estimated over the whole history ofthe stock, which covers at least 100 months. f (3b) were fitted using monthly data, a vestor of dummy variables could be introduced to identify the faeal year covered by the annual report, thus permitting the mean residual to vary between Seal years. The impact of unusual information received in month m of year¢ would then be estimated by the eum of the constant, the dummy for year‘, and the ealeulted reidaal for month m and year t ‘afortunately, the eicfoney of estimating the stock return equation fa this partic. ular form has no: been investigated satisfactorily, hence our report will be confined to the realte fram estimating ().[BMPIRICAL EVALUATION OF ACCOUNTING INCOME NUMBERS — 165 TABLED Deciles of the Distributions of Squared Coafleients of Correlation, Changes in Firm ‘tnd Market Incoma® a0 | as | a | a | 36) | 2 ‘would otherwise have been expected.” Such a result ( < 0) would be evi- enced by negative behavior in the stock return residuals (0 < 0) around ‘the annual report announcement date. ‘The converse should hold for a positive foreeast error. ‘Two basic income expectations models have been defined, a regression model and a aaive model. We report in detail on two measures of income [net income and EDS, variables (1) and (2)] for the regression model, and ‘one measure (EPS, variable (3) for the naive model. ) Net income os | or | 10 | 15 | 28 Fy @ BPs. oo | 05 | at | 16 TBstimaved over the 21 years, 1HG-1008. Data ‘Three classes of data are of interest: the contents of income reports; the dates of the report snnouncements; and the movements of security prices ‘round the arnouncement dates, INCOME NUMBERS Income numbers for 1946 through 1966 were obtained from Standard and Poor's Compustat tapes. The distributions of the squared coefficients of correlation” between the changes in the incomes of the individual firms ‘and the changes in the market's income index* are summarized in Table 1. FFor the preseat sample, about one-fourth of the variability in the changes 7 We later divide the total retur into two para: « “normal return," defined by ‘the retarn which would have bees expected given the normal relationship betwoen a Stock aad the market index; and sn “abnormal return,” the diferene between the Stetual retra and the normal return. Formally, the two parts are given by: by + yj Ulm ~ 15284 8m os Tapes teed are dated 9/28/1965 and 7/07/1967. All corrlaioncoefisienta in tia paper are product-moment correlation coef St. "The market not income index was computed se the sample mean for each yout. "The market EPS index was computed at weighted average over thesample members, the number of stocks ontetanding (adjusted for stock spl providing the mghta, Note that when estims tf particular Som and the market, the income of that firm was excluded from the market index.166 nay BALL AND PHILIP BROWN TABLE? Deciten of a Distributions of the Cosficiente of First-Order Autocorrelation in the Income Reqresion Residuals” Vase (Q) Net income (@) EPs. *"Rotinated over the 21 year, 104-1005, in the mediat firm’s income can be associated with changes in the market index. ‘The associstion between the levels of the earnings of firms was examined in the forerunner article (Ball and Brown (1967)}, At that time, we referred to the existence of autocorrelation in the disturbances when the levels of ‘net income aad EPS were regressed on the appropriate indexes. In this paper, the specification has been changed from levels to first differences because our method of analyzing the stock market's reaction to ineome ‘numbers presipposes the income forecast errors to be unpredictable at a ‘minimum of 12 months prior to the announcement dates. This supposition is inappropriate when the errors are autocorrelated. We tested the extent of autocorrelation in the residuals from the income regression molel after the variables had been changed from levels to first differences. The results are presented in Table 2. They indicate that the ‘supposition is not now unwarranted. ANVUAL REPORT ANNOUNCEMENT DATES ‘The Wall Sret Journal publishes three kinds of annual report announce- ‘ments: forecasts of the year’s income, as made, for example, by corporation executives shortly after the year end; preliminary reports; and the eom- plete annual report. While forecasts are often imprecise, the preliminary report is typically a condensed preview of the annual report. Because the preliminary report usually contains the same numbers for net ineome and EPS as are given later with the final report, the announcement date (ot, ‘effectively, the date on which the annual income number became generally available) ‘was assumed to be the date on which the preliminary report appeared in the Wall Street Journal, Table 3 revesls that the time lag between the end of the fiscal year and the release of the annual report has ‘een deolning steadily throughout the sample period. ‘stock Pnices Stock price relatives were obtained from the tapes constructed by the Center for Research in Security Prices (CRSP) at the University of Chi-BMPIRCAL EVALUATION OF ACCOUNTING INCOME NOWDERS 167 TABLES Time Diatibution of Annowncement Dater | el yar fe Tm [om Tome To 25 | 2am | 3/04 | 208 | 208 | 2/02 @ — |2/25 | 300 | ans | a7 | ans 7 [sao | 3/08 | 308 | aoe | 3/0 Tadicate tht 25 per cont ofthe income report 1067 had boen announced by 2/07/1058, TABLES Decilen ofthe Ditributions ofthe Squared Coeficient of Correlation forthe Stock Return Bapression, ard ofthe Coeicient of First Order “hudocorslation inthe Stock Return Resdvala® Tilimated over the 248 months, January, 166 through June, 1000. cago." The data used are monthly closing prices on the New York Stock ‘Exchange, adjusted for dividends and eapital changes, forthe period Janu- ary, 1946 through June, 1966. Table 4 presents the deciles ofthe distribu. tions of tho squared coelicient of correlation for the stock retur regression {equation (3), and of the coeficient of first-order autocorrelation in the stock rediduas. INCLYSION CRITERIA, ‘Firms included in the study met the following criteria: 1. earnings data available on the Compustat tapes for each of the years 1946-1966; 2, fiscal year ending December 31; 5, price dava available on the CRSP tapes for at least 100 months; and 4. Wall Steet Journal announcement dates available * Our analysis was limited to the nine fiseal years 1957-1965. By beginning the analysis with 1957, we were assured of at least 10 observations when Tha Cents for Research in Security Prices at the University of Cheago ia epon- sored by Merril Lynch, Perea, Fenner aad Smith Incorporated "* Announcement dates were taken initially from the Wall Street Journal Inder, then verified sgsinst tho Wall Street Journal168 RAY BALL AND PHILIP BROWN ‘estimating the income regression equations. The upper limit (the fiscal year 1905, the results of which are announced in 1066) is imposed because ‘the CRSP file terminatod in June, 1966, Our selection criteria may reduce the generality of the results. The sub- population does not include young firms, those which have failed, those which do not report on December 31, and those which are not represented (on Compustat, the CRSP tapes, and the Wall Street Journal. As a result, it may not be representative of all firma, However, note that (1) the 261 ‘remaining firms are significant in their own right, and (2) a replication of our study on a different sample produced resulta which conform closely to those reparted below. Results ‘Define month 0 as the month of the annual report announcement, and APD, the Abnormal Performance Index at month M, ae: 1a APLe = 3 TD, (1+ ‘Then API traces out the value of one dollar invested (in equal amounts) in all securities n (n = 1, 2, «++, N) at the end of month —12 (that is, 12 ‘months prior to the month of the annual report) and held to the end of ‘some arbitrary holding period (M = —11, ~10, --- , 7) after abstracting from market affects, An equivalent interpretation is at follows. Suppose two individuals A and B agree on the following proposition. B is to eon- struct a portfolio consisting of one dollar invested in equal amounts in. N securities. The securities are to be purchased at the end of month —12 and held until the end of month TT. For some price, B contracts with A to take (or make up), at the end of each month 3f, only the normal gains (or losses) and toreturn to A, at the end of month 7, one dollar plus or minus ‘any abnormal gains or losees. Then APT, is the Value of 4’s equity in the ‘mutual portfolio at the end of each month M2 ‘Numerical results are presented in two forms. Figure 1 plots API first for three portfolios constructed from all firms and years in whieh the income forecast errors, according to each of the three variables, were positive (the top half); second, for thrve portfolios of fims and years in which the income foreeast errors were negative (the bottom half); and third, for a Single portfolio consisting of all firms and years in the sample (the line which wanders just below the line dividing the two halves). Table 6 in- cludes the nurnbers on which Figure I is based, Duo to known error inthe data, not all ma could be included in all year. The fiscal year most affected was 194, whea three fer were excluded. "The replication investigated 75 Srme with fatal year ending on dates other ‘han Deceber St, using the naive income foreeaating model, over the longer period 194745 That is the valuo expected atthe end of month Tin the abaence of further ab- normal gine and lowes,EMPIRICAL EVALUATION OF ACCOUNTING INCOME wuuBERS — 169 ae) Sues co nmareneae “Month Palative i Anal Report! Amourcemant Date Pio. 1 Absormal Performance Indexes for Various Portfolios Since the fst set of results may be sensitive to the distributions of the stock retura disturbances, a second set of results is presented. The third ‘column under each variable heading in Table 5 gives the chi-square statistic for a two-by-two classification of firms by the sign of the income forecast ‘error, and th sign of the stock return residual for that month. ovonviEw ‘Aa one would expect from a large sample, both sets of results convey cesientially the same picture. They demonstrate that the information con- tained in the annual income number is useful in that if setual income differs empirical distetbutins ofthe stock raturn residuals appear tobe decribed ‘wall by symmetric, stable distebutions that are characterized by tails longer than hose of the nara! distribution [Fama (106); Fama, eal (1067).170 may mat AND PuILIP DROWN TABLE 5s Summary Statistica by Mond Relatce to Anna Report Announcement Date 7 = ms c @ °°] @| 1.08 | oma | 16.8 | 1.007 | 292 | 90.4] 1.008 | 260 | 24.2 | 1000 toate | oss | 17.3 | 1.015 | ‘oa | 20.2 | 1-015 | ‘ova | 73.4 | 200 Lat | ‘or7| 7.0) torr | err | “3:7 | oie ‘oes | 20.4| “one vat] ort | 9.8) 1-022 | ‘art | 2°0| 1.022| ‘950 | “9:1 | “ons m3 | 090 | 21.8 | 1-027 | “260 | 27.1 | 1.0a¢| ‘ote | 9.0] “tos 11ms | ‘909 | 42.9 | 1.064 | ‘ous | a | 1007 | ‘sar | 9.4 | ow 4.0 | ‘oat 47-9 | 3.000 | ‘oat | 21-3] 1.082) ‘925 | 21.0 | “ove 1.60 | ‘980 | 40.0 1.080 | ago | 30:5 | 1.0m | “oia| a8 | “ona 1088 | [90 | 36.3 | 1.060 | ‘202 | 33.0 |] 1.000 | ‘009 | 37:2 | “ons x.om7 | at | “1-4 1.058 | “oto | “1's | 1:05] ‘oa | “0.1 | “ana 00 | 1914} 8.2 1-002] ‘s12] 8:2] 1.008 | 's00) 5.7 | “oor sort | ‘907 | 28.0 | 1.073 | ‘ons | 8.0 | 1056 | ‘ae |a5.8 | Sa 1.006 | ‘oot | “6-4 | 1.78 | sao | 5.5 | 1.057 ot} ‘one 1.076 | ‘890 2.7 1.078 | ‘aor | 1.9] 1.080] ‘era | 8.1] ‘ooo 4.ors| ‘306 | 0.6 | toro | 05] 1.2 | 060 |‘er8 | 0-1) oor ors | ‘s03 | 0-1 | :070 | a2 0.1 | 1-08r| ‘376 | 1.2] 90 4.075 ‘s03) 0:7 | 1.077 | ‘aot | 0.1] 1.055| ‘s70| 0:8 | a0 .o2| ‘50a 0.0 | 1.074 | ‘seo | 0.2] 1.0st| ‘sr7 | 0:1 | ‘as * Column bending (0) Abnormal Performance Index—firms and years in which the income forecsat error wa positive. {@) Abnormal Performance Index—firms end yoars in which the rrr wan nogative. {@) Chisquare statistic for two-by-two clussifation by sign of income foreenat crror (for th foal you) and sign of stonk return euidual (fo the indieated month). ‘Note: Probability (chi-square > 3.8 | x8 = 0) = .05, for degreo of freedom. ‘Probability (ehi-equare > 6.64 | x8 = 0) = 01, for I dogroo of freedom, come forecast from expected income, the markot typically has reacted in the same diree- tion. This contention is supported both by Figure 1 which reveals a marked, ‘positive association between the sign of tho error in forecasting income and the Abnormal Performance Index, and by the chi-square statistic (Table 5). ‘The latter shows it ia most unlikely that there is no relationship between ‘tho sign of the income forecast error and the sign of the rate of return re- sidual in most of the months up to that of the annual report announcement. However, most of the information contained in reported income is an- ticipated by the market before the annual report is released. In fact, an- ticipation is so accurate that the actual income number does not appear to cause any unusual jumps in the Abnormal Performance Index in the an- ‘nouncement month. To illustrate, the drifts upward and downward begin at least 12 morths before the report is released (when the portfolios are firstIMPIRICAL BVALUATION OF ACCOUNTING INCOME NUMBERS 171 TABLE 6 Contingency Table ofthe Sign ofthe Income Forecast Errore—by Variable Settee Vasiable (1) + Vasiablo @) + Vasiable @) constructed) and continue for approximately one month after. The per- sistenee of the drifts, as indicated by the constant signs of the indexes and by their almcet monotonic inereases in absolute value (Figure 1), suggests not only that the market begins to anticipate forecast errors early in the 12 ‘months preeeding the report, but also that it continues to do so with in- creasing success throughout the year. SPECIFIC RESULTS 1, There appears to be little difference between the results for the two regression motel variables. Table 6, which clasifies the sign of one variable’s {forecast error contingent upon the signs of the errors of the other two vari- ables, reveals the reason. For example, on the 1231 occasions on which the ingome forecast error was positive for variable (1), it was also positive on 11148 oocasions (out of a possible 1231) for variable (2). Similarly, on the 1109 oeeasions on which the income forecast error was negative for variable ((), it was also negative on 1026 occasions for variable (2). The fact that the results for variable (2) strictly dominate those for variable (1) suggests, however, that when the two variables disagreed on the sign of an income forecast error, variable (2) was more often correct. ‘While thers is litle to choose between variables (1) and (2), variable (3) (the naive model) is clearly best for the portfolio made up of firms with negative forezast errors, A contributing factor is the following. The naive ‘model gives the same forecast error as the regression model would give if "7 Note that Figure I containa averages over many fma and years and is not in- Aioative ofthe behavior ofthe securities of any particular firm in any ane year, While ‘here may be, sn average, a persistent and gradual anticipation of the contents of the report throughout the year, evidence on the extent of autocoreelation in the stork rtura rexdale would suggeat that the market's reaction to information about ‘ particular frm tends to oceur rapidly.172 RAY MALL AND PmILIP BROWN: (a) the change in market income were zero, and (b) there were no drift in the income of the firm. But historically there has been an increase in the market’s income, particularly during the latter part of the sample period, due to genera inerease in prices and the strong influence of the protracted expansion sioce 1961. Thus, the naive model {variable (3)} typically identi- fies as firms with negative forecast errors those relatively few firms which showed a deerease in EPS when most firms showed an increase. Of the tree variables, one would be most confident that the incomes of those which showed negative forecast errors for variable (3) have in fact lost ground relative to the market, ‘This observation has interesting implications. For example, it points to a relationship Eetween the magnitudes of the ineome forecast errors and the ‘magnitudes cf the abnormal stock price adjustments, This conclusion is reinforeed by Figure 1 which shows that the results for positive forecast errors are weeker for variable (3) than for the other two, 2, The drift downward in the Abnormal Performance Index computed over all firms and years in the sample reflocta a computational bias The Dias arises because BLL + onl # IL + Blo, where H denctes the expected value. Tt can readily be seen that the bias ‘over K months is at least of anter (K — 1) times the covariance between ti and tq-1."Sinee this covariance is typically negative,” the bias is also negative, While the bias does not affect the tenor of our results in any way, it should be kept in mind when interpreting the values of the various API’. It helps explain, for example, why the absolute changes in the indexes in ‘the bottom panel of Figure I tend to be greator than those in the top panel; why the indexes in the top panel tend to turn down shortly after month 0; and finally, why the drifts in the indexes in the bottom panel tend to persist beyond the month of the report announcement, 83, We also computed results for the regression model using the additional definitions of income: (a) cash flow, as approximated by operating income,” and (b) net income before nonrecurring items, [Neither variable was as successful in predicting the signs of the stock return * The expected valu ofthe bias is of order minus one-half to minus one-quarter of one per cent por annum. The diflerence between the observed value of the APT computed over th total stnple and its expectation ism property ofthe particular sample (se footsote 2) Tn particular, the approximation neglects all permutations of the prod- 18= 1,20 RB, = 042, > Ras being of a second order of smallness ‘ee Table & ALI variable ditions are spocfiod in Standard and Poor's Compustat Manual [oe also Ball su Brown (1067), Appendix A.EMPIRICAL EVALUATION OF ACCOUNTING INCOME NUMBERS — 173 residuals as net income and EPS, For example, by month 0, the Abnormal Performance Indexes for forecast errors which were positive were 1.068 (net income, including nonrecurring items) and 1.070 (operating income). ‘These numbers compare with 1.071 for net ineome (Table 5, variable (1)]- ‘The respective numbers for firms and years with negative forecast errors ‘were 0.911, 0.917, and 0.907. 44. Both the API's and the chi-square test in Tablo 5 suggest that, at least for variable (8), the relationship between the sign of the income fore- cast error and that of the stock return residual may have persisted for as long'as two months beyond the month of the announcement of the annual report. One explanation might be that the market’s index of income was not Known for sure until after several firms had announced their income ‘numbers. The elimination of uncertainty about the market's ineome subse quent to some firms’ announcements might tend, when averaged over all firms in the sample, to be reflected in a persistence in the drifts in the APT's beyond the announcement month. This explanation can probably be ruled ut, however, since when thoce firms which made their announcements in January of any one year were exchided from the sample for that year, there ‘were no changee in the patterns of the overall APT's as presented in Figure 1, although generally there were reductions in the x statistics. ‘A.seoond explanation could be random errors in the announcement dates. rifts in the APT's would persist beyond the announcement month if errors resulted in our treating some firms aa if they had announced their income ‘numbers earlizr than in fact was the case. But this explanation ean also probably be ruled out, sinee all announcement dates taken from the Wall Street Journal Index were verified against the Wall Street Journal, ‘A third explanation could be that preliminary reports are not perceived by the market as being final. Unfortunately this issue cannot be resolved independently of an alternative hypothesis, namely that the market does take more time to adjust to information if the value of that information is Joss than the transactions costs that would be incurred by an investor who ‘wished to take advantage of the opportunity for abnormal gain. That is, even ifthe relationship tended to persist beyond the announcement month, it is clear tha; unless transactions costs were within about one per cent,” "The genera! reduction in the x*atatiti edu largely to the reduction in sample This result is obtained as follows, The ratio API_/AP In i equal to the mar- inal return in month m plus unity APla Pl = Ode Similarly, APly _ API, API, Pligg ~ APIs” AP =O tr tres174 may BALL AND PMILIP BROWN there was no opportunity for sbnormal profit onee the ineome information hhad become generally available. Our results are thus consistent with other ‘evidence thet the market tends to react to data without bias, at least to within transictions cost, [THE YAUUE OP ANNUAL NET ICOMD RELATIVE 70 OTHER SovUCES OF piroRMaTon* ‘The results demonstrate that the information coutained in the annual income number is useful in that itis related to stock pres, But annual accounting rports are only one of the many sourees of information availa- ble to investors. ‘The aim of this section isto assess the relative importance of information contained in net ineome, and at the same time to provide some insight nto the timeliness of the income report. It was suggested earlier that the impact of new information about an individual stock could be measured by the stoek’s return residual. For example, a negative residual would indicate that the actual return is les than what would have been expected had there been no bad information Equivalently, if an investor is able to take advantage of the information cither by seling or by taking a short position in advance of the market adjustment, shen the residual will represent, ignoring transactions costs, the extent fo which his return is greater than would normally be expected. If the difference between the realized and expected return is accepted as also indicating the value of new information, then it is elear that the value of new, mon:hly information, good or bad, about an individual stock is given by the absolute value of that stock's return residual for the given ‘month, It folows that the value of all monthly information concerning the average firm, received in the 12 months preceding the report, i given by: ELL, + lewd] - 100, API aPh, ‘Thus, the marginal return forthe two monthe after the announcement date on the portolio consiting of firma for which HPS decease would have been 0878/0887 14 — 010; sivlany, tho marginal retura on the portfolio of frm for which EPS Increased would have been 1050/1086 ~ 1 Se 03. After allowing forthe computer ‘ional bias it vould appear that transactions costs must have been within one Der eat for opportunities to have existe for abaormal proBt from applying some aechan- eal trading rae "This analysis does not consider tho marginal contribution of information eon ‘ined in the anaual income qumber. I would be isterenting to analyze dividends in a ‘vay similar to that we have uted for ineome announcenets, We expect there would me overlap. To the extent that theres an overlap, we atribute the information to the income auzaber and consider the dividend nanouncesnent tobe the media by which the mask loans about income. Thi asnumption is highly artificial in that Astoria! income numbers and dividend payments might both simply be rections of tho sme, more fundamental informational determinants of stock prices Th ‘od, in general Ctra) oO treEMPIRICAL EVALUATION OF ACCOUNTING INCOME NMBENS 175 where TZ denotes total information.” For our sample, averaged over all firms and yeers, this sum was 0.731. For any one particular stock, some of the information between months will be offsetting.* The value of net information (received in the 12 monthe receding the report) about the average stock is given by: (Fem) = 100 Hl, where NZ denotes net information. This sum wa 0.165. ‘The impact of the anaual income aumber is also» net number in that net income isthe result of both ineome-inereasing snd ineome-deeressing event. If one accepts the forecast error model," then the value of informa- tion contained in the anausl income number may be estimated by the average of the value inerementa from month —11 to month 0, where the increments se averaged over the two portfolios constructed from (buying ‘or salling short all frzas and year as classified by the signs of the income forecast erro. ‘That i, NCAP — 1.00) — N2(APEP - WF NA) where ZT denotes income information, and 1 and N2 the number of os- casions on waich the ineome foreeast error was postive and negative re- spectively. This number was 0.081 for variable (1), 0.088 for variable (2), and 0.077 for variable @). From the shove numbers we conclude (1) shout 15 per cont [(731 ~ -165)/-731] of the value of all information appears to be offsetting, which in tum implies that about 25 per cent per- sists; and (2) of the % per cent which persists, about half [49% 50%, and 47 %— caloulated as 081/165, 088/.165, and .077/.165—for variables (1)-()] ‘ean be assocsted with the information contained in reported income. ‘Two further conclusions, nt directly evident, ae: {@) of the value of information coutained in reported income, than about 16 to 15 percent (12%, 11%, and 18%) has not been ant by the month of the report and ii Me= eS 100) We “Note that the information is reflected in a valu increment; thu, the original ‘$1.0 is deducted frm the terminal value "This amertion ie supported by the obwerved low autocorrelation inthe stock re- turn residuals 5 Note that since we are interested in the “avorage frm,” an investment strategy mut bo adopto on every sample member. Beonuso there are only two relevant rat- ‘pie involved, in culclont to kaow whether one la better of to buy oto sell short. [Note alao the the analyais ascumes the strategy is fst wlopted 12 months prioe to the aanouncemant date. Sr The average monthly yield from » paieyof buying a portfolio consisting of all firms with postive foreenat errrm and adopting a aor position on the rest would have resulted ian average moathly abaormalrate of return, from —I1 to —1, of176 RAY BALL AND PHILIP BROWN (2) the value of information conveyed by the income number at the time of its release constitutes, on average, only 20 per cent (19%, 18%, and 19%) of the value ofall information coming to the market in that month.” The second conclusion indicates that accounting ineome numbers eap- ture about half of the net effect of all information available throughout the 12 months preceding their relense; yet the fourth conclusion suggests that not income contributes only about 20 per cent of the value of all informa- tion in the month of its release. The apparent paradox is presumably due to the fact that: (a) many other bits of information are usually released in the same menth as reported income (for example, via dividend announce- ‘ments, or perhaps other items in the financial reports); (b) 85 to 90 per cont of the net effect of information about annual income is already re- flected in security prices by the month of ite announcement; and (c) the period of the annual report is already one-and-one-half months into history. Ours is perhaps the first attempt to assess empirieally the relative im- ortance of the annual income number, but it does have limitations, For example, our resulta are systematically biased against findings in favor of saceounting rports due to: 1, the assumption that stock prices are from transactions wi taken phe simultaneously atthe end of the month; the assumption that there are no errors in the data; the discrete nature of stock price quotations; ‘the presumed validity of the “errors in forecast” model; and ‘the regression estimates of the income forecast errors being random variables, which implies that some misclassifcations of the “true” earnings forecast errors are inevitable, Concluding Remarka ‘The initial objective was to assess the usefulness of existing accounting, income numbers by examining their information content and timeliness, ‘The mode of analysis permitted some definite conclusions which we shall briefly restate, Of all the information about an individual firm which be- ‘comes availatle during a year, one-half or more is captured in that year's income number. Its content is therefore considerable. However, the annual income report does not rate highly as a timely medium, since most ofits content (about 85 to 90 per cent) is captured by more prompt media which perhaps include interim reports. Since the efficiency of the eapital market have 0.88%, 0.60%, and 0.00% for variables (1), @), and (2) rapectively. The marginal rate of return ic month O for that same strategy would have been 0.27, 0.80%, and (0.94% respeetivaly. However, relatively much more information is conveyed inthe month of the rport announcement than in ether ofthe two months immedialy preceding the announcement month or in the two months immediately following It ‘This result is ecasistent with those nbtained by Beaver (048. ‘An optimun policy (thats, one which takes wivantage ofall information) would havo yialded an abnormal rate of return of 4.2% in month 0.EMPIRICAL EVALUATION OF ACCOUNTING INCOME NUMBERS 177 is largely determined by the adequacy of ita data sources, we do not find it disconcerting that the market hae tured to other sources which can be acted upon rrore promptly than annual net income. "This study raises several iamues for further investigation. For example, there remains the task of identifying the media by which the market is able to anticipate net income: of what help are interim reports and dividend snnouncements? For secountants, there is the problem of assessing the cost of preparing annual income reports relative to that of the more timely interim reports. ‘The relationship betwoen the magnitude (and not merely the sign) of the unexpected income change and the associated stock price adjustment could also be investigated.® This would offer a diferent way of measuring, the value of information about ineome changes, and might, in addition, furnish insight into the statistical nature of the ineome process, a process little understood but of considerable interest to accounting researchers. Finally, a mechanism has been provided for an empirical approach to a restricted clase of the controversial choices in external reporting. ‘REvERENcES Bats, Ray xn Purse Buows (1997) "Some Prliminary Findings onthe Association botmena the Karaings of a Firm, Its Industry and the Keonomy,” Hmpiical Re- ‘swatch in Accounting: Selected Stes, 867, Supplement to Volume B of the Jur ral of Acounting Research pp. 65-TT. Baran, Witttin I (198), “The Taformation Content of Annual Barnings An- ‘oueemeets,"frthooming in Empirical Reearch in Accounting: Selected Stes 41008, Supplement to Volume 6 ofthe Journal of Accounting Research Buows, Manauats B. (008). “Tho Assecament of Portfolio Performance” (wspub- Tiebed Ph.D. diasertation, University of Chieago). Banazr, Rroesxo A. (1968). "The Talagnce of the Reonomy on the Earnings of the ‘Firm’ (unpublished paper presented at the Sloane School of Finance Seminar, ‘Massachuantts Inritute of Technology, My, 168). Brows, Puruir axo Vicron Nrepzasovven (1068). ‘The Predictive Content of ‘Quarterly Earnings,” Journal of Busines. Ccanmina, foun B, (929), The Beonomicn of Accountancy (Now York: The Ronald Press Co.) CCmaunnns, Rarwono J. 0064) “Messurement and Objectivity in Accounting,” The Accounting Review, XXXIX (April, 1964), 20-74 — (inte). Aesounting, Evaluation, and Beonamic Behavior (aglewood Cif, N..: Prontice-Hal) 067). “Gintinuossly Contemporary AecountingAdaltivty and Action,” The “Accounting Review, XLII (October, 167), T5137. Corsa, Pav Hed. (194). The Rondom Character of Stock Market Prices (Cam ‘bridge, Masa: "The MLLT. Pres). ‘There are ome dificalt econometric problema asgociated with tha relationship, including apec'ying the appropriate funetinal form, the expected statistical dstri- Duions af the uiderlying parameters, the expected behavior ofthe regression resd- unla, and the etent and effects of measurement errors in both dependent and independent variables. (The funetional form need not necessarily be linear, if only ‘because income ummbers convey information about the covarabilty ofthe income proces.)178 may Baty AND ruLiP BROWN Epwanos, Bonan O. axo Panuir W. Bu. (1061). The Theory ond Mearurement of ‘Busines facie (Berksley, Cal.: The University of California Press) aus, Evan P. (108). “The Behavior of Stock Mackot Prices," Jounal of Buei- ‘nee, XXXVETT Ganuary, 1968), 34-105 ss Mansnatt B, Bienes 0968). “Flter Rules and Stock Masket Trading,” Journal of Busines, XXXIX (Supplement, January, 1966), 25-41 —, Lawman Pistny, Sicnant C. Jaxx, 8 R2cHd Rott (907). "The Ad- jastmont af Stock Prices to New Information,” Report No, 6718 (University of Chicago: Center for Mathematical Studies in Businew and Reonomie; forthe coming i the International Beoomie Review) Frsmmx, Lasrmxen (1960). “Some Now Stock Market Indiees,” Jourel of Businses, XXXIX Supplement, January, 1060), 191-235, Grnata, Sousa (19). Accounting Concept of Prot (New York: The Ronald Prove 0.) Im, Yom (07). The Foundations of Accounting Measurement (Englewood ‘NI. Prectice all). Jnnsns, Micusas C. (1966). “Riek, the Preingof Capital Ase, and the Evaluation ‘of Investment Portfolios” (onpublished Ph.D. disertation, University of Chicago. ‘Kino, Bassa F. (1000). “Market and Industry Factors in Stock Pree Behavior, Journal of Businest, XXXIX (Supplement, January, 1968), 189-20 Law, Roxaua 8. (1960). “The Mathematical Propriety of Accounting Mesaurements ‘nd Caleu ations," The Accounting Review, XLI (October, 1068), 642-51 Paron, W. A. snp A. C, Lavzuzron (140). Introduction to Corporate Aosounting ‘Standards (American Accounting Association Monograph No.3). Sawomuzox, Pav A. (106). "Proof That Properly Anticipated Prices Fluctunte “Randomly” Zaduarial Management Resiew, 7 (Spring 108), 41-40. Scnous, Miraex J. 0087). "The Bet of Secondary Distributions on Price” (an- ‘published paper presented at the Seminar on the Analysis of Security Prices, University of Chicago). ‘Suanrs, WasiuseF. (108) “Capital Asgt Prices: A Theory of Masket Equilibrium ‘under Coniitions of Risk," Journal of Finance, XIX (September, 1964), 125- 2 ‘Seunuing, Rossa R. (1067) "Elements of Pure Accounting Theory,” The Account- ing Review, XII Gunary, 1907), 62-73 ‘Varna, Wiasst J. (140). The Pund Theory of Accounting (Chicago: The University ‘of Chiengo Press) its,