THE INFORMATION CONTENT SIMULTANEITY OF RETURNS AND EARNINGS: THE CASE OF INDONESIA

Transcription

1 TH INFOMATION CONTNT SIMULTANITY OF TUNS AND ANINGS: TH CAS OF INDONSIA FUAD * Universas Dian Nuswantoro, Semarang DWI ATMONO Universas Diponegoro, Semarang ABSTACT This paper analyzes in detail the information content of returns and earnings, both linearly and simultaneously because they are jointly affected by information that is difficult to specify explicly. Based on empirical evidence, we challenge the assumption that earnings follow first order moving average process, while assuming that they actually follow first order autoregressive process. Using 9 firm-specific observations from 1996 to 24 we found that earnings (C) and returns response coefficients (C) provide similar estimates under joint estimation rather than linear one. We also found that 2-Stage Least Square (2SLS) reduces the bias of C and C in contrast to under Ordinary Least Square. Surprisingly, earnings and returns convey information larger in the short-term rather than in the long-term. In particular, we found that one year lag of return contain more information about future earning than two-or three years lag of return. Keywords : Joint stimation, arning esponse Coefficient, eturn esponse Coefficient, Information Content Bias. * Corresponding author: Fuad, Universas Dian Nuswantoro Semarang, Jl Nakula I No 5-11 Semarang, mail address: tofuad@gmail.com. Tel Fax No:

2 I. INTODUCTION One of the most compelling and intriguing research questions of our time is exploring the relationship between return and accounting earning, in which capal market equilibrium can be characterized by as a mapping from states into a set of secury returns. Similarly, earnings are signals from information system which is a mapping from states into signal. In general, there could be any relationship between returns and earnings depending upon the nature of the two mappings. If one assumes that return and earnings reflect a common set of events, is not unreasonable to assume that the two might be associated (Beaver, et al. 1979). Thus, can be argued that returns lead to earnings, while on the other occasions the direction is reversed. These two regressions have previously been estimated separately. Beaver et al. (1997), however, develop a new method, in the context of this accounting lerature, for exploring the bivariate relationship between secury returns and accounting earnings. The key innovation by Beaver et al. (1997) is to characterize the return-accounting earnings relationship as a system of simultaneous equations. They argue that earnings and returns can behave as if they are both endogenously determined because they are jointly affected by information that is difficult to specify explicly. Nevertheless, prior research provides some limed and mixed evidence of a posive association, even after controlling for various informational and environmental variables (see Kothari, 21 and Beaver, 22). Furthermore, prior researches have documented that the contemporaneous linear relation between annual stock returns and accounting earnings has declined over time (see, for example, yan and Zarowin 23, ly and Waymire, 1999) arising many speculations. One of the most popular is that earnings increasingly reflect news wh a lag relative to stock returns. Kothari and Sloan (1992), and Collins et al. (1994), among many others, further prove that annual returns predict future earnings for at least three subsequent years. Despe the fact that Beaver et al. (1997) ignore the consecutive lags for return-earnings relation, the method has been used extensively. Current study attempts to resolve the gap between Beaver et al. (1997) and Ali and Zarowin (1992), among others versus Kothari and Sloan (1992), Collins et al. (1994), among many 2

3 others by examining the time-series properties of the information content of earnings and return in Indonesia. Notwhstanding the fact that simultaneous equation approaches are fruful, prior approach used in explaining the information content of earnings and returns arose several limations. First, the approach ignores the established time-series properties of the data and ignores the role of measurement errors. Second, as previously stated, the simultaneous approach that was widely used ignores the consecutive lags for returnearnings relation. In fact, one-year lagged provide may be an arbrage estimated return response coefficients, whout letting the data speak for themselves. Third, prior studies use a naïve earnings expectations model (random walk) as the benchmark for the information content of lagged returns and earnings, while they do not adequately address the incremental information content of lagged returns. Few researches have empirically tested the un-root test of earnings and returns while at the same time also examine the nature of information content of earnings and returns (see Kothari, 21). Our research is greatly different from previous ones. First, we overcome the limations above. Second, inspired by the work of Ohlson and Juettner-Nauroth (25), we prefer to use abnormal return growth instead of price and differenced-price models because is econometrically less problematic. Third, we relax the assumption that earning response coefficients should be treated as cross-sectional and a temporal constant. We also test the sensivy of the C by reversed regression model, estimating the eturn esponse Coefficients (C). Fourth, following Collins and Kothari (1989), and Beaver et al. (1997), we assume that the firm s systematic risk (beta) act as a proxy for a broader information environment. We believe our paper greatly contributes to accounting lerature in some specific ways. First, prior research uses a naïve assumption that earnings and returns follow random walk (wh or whout drift). Based on empirical evidence and statistical perspective, we challenge the assumption, and examining that earnings and returns changes are in fact (weak) stationary. Second, we test the usefulness impact of earnings in short-term and long-term, wh regard to returns. At least based on our lerature review, no research has put a great concern about this. Third, we analyze the presence of whe noise in the earnings and returns change by examining that the information 3

4 content of earnings and returns change follows a first order autoregressive (A1). Prior research assumes that earnings follow a first-order moving average processes (Ali and Zarowin, 1992a). Present study contributes to accounting lerature in this way. The objectives of the study are in-line wh the contribution above. The remainder of the study proceeds as follows. Section 2 briefly describe the theoretical model and develops hypotheses. Section 3 presents the method of study describing the data analysis, research design and providing evidence of stationary. Section 4 briefly analyzes the results and discusses the findings. Last section provides conclusions, caveats, limations and future research. II. THOTICAL MODL AND HYPOTHSS The Information Content of arnings. Harvey (1993) and Buckberg (1993) pointed out that many emerging markets (including Indonesia) offer yields in excess of developed market returns suggesting that there might be unexploed prof opportunies to be made by investing in these markets. However, emerging markets exhib high volatile stock returns arising from the fact that they are small markets wh insufficient integration into the global economy and informational imperfections. They typically tend to be noisier wh fewer trades taking place, have limed reporting requirements and less information updates than in developed markets. In this context, understanding the relationship between stock returns (in advance of earnings announcements) and the published earnings figures is very important to investors, financial intermediaries and regulators. If the information contribution of earnings to investors is useful, then earnings should exhib considerable explanatory power wh respect to return revisions both cross-sectional and over time (Lev 1989). ecent research (Firth et al, 1995, Jaggi, 1997; Chen et al 21, Chen and Firth, 1999, Firth, 1998) argues that due to the absence of reliable information, investors rely on information disclosed in annual report and financial statements to make posive investment decisions. A related strand of lerature, reviewed in Verecchia (21), has dealt wh the theoretical modeling of how the disclosure of accounting information, particularly earning, affects investors as reflected in stock returns. Ball and Brown 4

5 (1968) and Beaver (1968) provide compelling evidence that there is information content in accounting earnings announcement. That is, at least a portion of the earnings increase experienced by the firms classified as good news firms was favorable to the market, which led to increased secury returns. arnings announcement do appear to contain information that is relevant to the stock market, and for the most part appears that the stock market reacts efficiently to this information (for a comprehensive view of value relevant research, see Beaver, 22). Algaebracally, can be stated as: Δ Δ α + e (1) = + α 1 i t The equation wh percentage change in return as dependent variable and percentage change in earning as explanatory variable is natural in the sense that accounting earnings are viewed as one, among many, information variables whose announcement can affect secury returns. α 1, which is defined as earning response coefficient, reflects a set of value-relevant variables (Beaver, 22) that are difficult to specify explicly (Beaver et al., 1997). The Information Content of eturn A cornerstone of market-based accounting research holds that market responds to signals from excess demand or supply to establish prices (returns) that reflect all available information. The relation between returns and future earnings is well documented (e.g., Beaver et al.198; Beaver et al.1987; Collins, Kothari, and ayburn 1987; Freeman 1987; and Collins and Kothari 1989). These studies conclude that because stock returns are condioned over broader information, they reflect changes in earnings on a more timely basis. They also develop the idea that the information set reflected in returns is richer than that in the contemporaneous accounting earnings. Furthermore, a number of researches (see, for instance, Ou and Penman, 1989) have also documented that stock returns lead accounting earnings. The results arose to the conclusions that returns provide information about earnings ahead of time and that earnings capture events that affect secury returns wh a lag. An idea behind the findings is that the content of information capture in the returns is richer than in the contemporaneous information (i.e., earnings). They argue that due to the nature of 5

6 historical cost, accrual accounting system typically lags the occurrence of changes in the economic values of liabilies and assets. Abarbanell (1991) reports that analyst s earning forecast partially reflect the information in prior return changes. The information content of return and s lag is reflected algebraically as Δ Δ Δ β + μ = Nevertheless, researches has documented that the linear relation between annual stock returns and accounting earnings has declined over time (yan and Zarowin 23). At least there are two possible explanations for the lack of informativeness of earnings relative to returns. First, earnings relatively reflect news wh a lag relative to stock returns. Second, earnings increasingly reflect good and bad news in an asymmetric fashion. Wh regard to the later, Kothari and Sloan (1992) and Collins et al. (1994) maintained that annual returns predict future earnings for at least three subsequent years that can be expressed in a simple regression: Δ (2a) Δ Δ Δ Δ β + μ 2 3 = (2b) ather than estimating eq (2b), we use the adaptive expectations model to reducing both the number of lagged terms in the distributed lag models and eliminating the problem of multicollineary. Besides, the adaptive expectations model reduce the degrees of freedom, that is all lagged terms in eq (2b) are replaced by single lagged value of Δ. Algebraically, can be stated as: Δ Δ Δ β + ν = (2c). Unlike Beaver et al (1997), the equation do not arbrary set the number of lagged terms. In fact, the median lag and mean lag can be estimated by the data (as will be given in the next section). Simultaney Hypothesis 6

7 The simultaney of the information content of returns and earnings can be simply understood if equation (1) and (2a) are set side by side Δ Δ α + e (1) = + α 1 i t Δ Δ Δ β + μ = (2a) Prior researches maintains that e and μ as arising because the accounting earnings variable equals changes in ungarbled earnings (i.e. the impact of earnings on events that also affect prices) plus measurement error (Beaver et al. 198, 1987, 1997). Beaver et al. (198, 1997) further assumes that ungarbled earning is characterized as a first order moving average model. Ali and Zarowin (1992a, 1992b) also maintained that first order moving average process can be a reasonable description of the earnings process. So that measurement errors in eq (1) and (2a) become: e = θ. 1 + ε (3) e μ + t = θ 2. μ ε t (4) Where θ t is the parameter of a first order moving average model, or the portion of ungarbled earnings that is transory. According to Beaver et al. (198), α should equal to (1-θ 1 ) from eq. (1) and β should equal to 1/(1-θ 2 ) after estimating eq.(2). Thus, when α is multiplied by β (i.e. α β ), the product should equal to (or at least approaching to) 1. Nevertheless, prior research did not document this interesting phenomenon. For example, Beaver et al. (1987) found α to be.31 and β to be 1.69 that lead α β to be.523, which is far below 1; the theoretically product. Beaver et al. (1997) further argued that the inconsistent approach may be due to the bias of the estimated content of earnings and returns. However, we challenge the assumption that earnings process follows a first order moving average process. Instead, autoregressive is the more plausible description of the earnings process (see Dickey-Fuller test and Box- 7

8 Jenkins Method section) because of the noise in the Indonesian capal market. Yet prove that OLS estimation provides inconsistency due to measurement error bias is based on the assumption that earnings follow the first-order moving average process. Since earnings and returns are jointly endogenously determined, their parameters will be bias if the parameters are estimated by the OLS. Kindly recall that Δ while the endogenous variable in eq (2) may be correlated wh e in eq (1) because of measurement errors in earning (i.e. the omission of unrealized gains and Δ losses under historical cost accounting), and μ are also correlated. Beaver et al. (1997) in particular provide evidence that the behavior of reverse regression estimates is consistent wh the percentage change of return varying for factors unrelated to earnings. For example, changes in the risky discount rate may induce return changes. Thus can be further argued that the main source of inconsistency of 1-θ and α. β < 1 is the measurement error. everse regression (e.g. two stage least squares) could provide unbiased estimator (see Beaver et al. 1997, for a comprehensive explanation). Following Beaver et al. (1997), we therefore predict that the estimated earning response coefficients and return response coefficients from a system of equations (2SLS) will be greater than under single equation ordinary least square estimation because bias is reduced. As bias is reduced, the estimated coefficients on the return change and earnings variable are more accurate estimates of (1-θ ) and α. β 1 The simultaneous model of eq (1) and (2) has two endogenous variables; return and earning and one predetermined variables, lagged return, in eq (2). Thus, eq. (1) is underidentified and eq (2) is just-identified. To be identified, each of these equations must exclude at least 2 variables. Thus, we include four predetermined variables; which are beta and first-order autoregressive (A1) in eq (1) and one year-lagged of earning (i.e. Koyck transformation because of unidentified lag coefficient), and A(1) in eq. (2) Wh regard to the first two, the research to date generally evaluates the informativeness of lagged return changes compared to a naive expectation of annual earnings (e.g., random walk or random walk wh drift). Although these results are 8

9 important, they ignore the richer time-series properties of historical earnings and may exaggerate the incremental explanatory power of lagged returns. esearches commonly used a random walk or random walk wh drift model that provides the earnings change variable, which is then correlated wh prior returns. Although the random walk model has been widely used, may not adequately control for the autocorrelation in historical earnings. In the context of these studies, serves as a relatively weak benchmark to evaluate the informativeness of stock returns. Beta was hypothesized to affect the return change because beta reflects the environmental contingencies (Beaver et al. 1997). Furthermore, we challenge the assumption of Ali and Zarowin (1992a, 1992b) in which earnings follow first-order moving average process and emphasize on the first order autoregressive, because Indonesia market typically tend to be noisier wh fewer trades taking place, have limed reporting requirements and less information updates than in developed markets. Δ Δ = α beta + A(1) + e.. α 1 α α 3 (5) Δ Δ Δ Δ = β A(1) + μ 2 2 (6) III. MTHOD Sample Design and Data Collection merging market, such as Indonesian market, face a major shortcoming in not having appropriate data. Databases seldom exist and if they exist, they may often incomplete or inaccurate. However, computerized database from independent source (e.g. Indonesian Securies Market Data Base (ISMD) of Center for Accountancy Development of UGM) is found, at least for the abnormal return and corrected annual 9

10 beta variables. arnings were collected manually from firm s financial statement and Indonesian Capal Market Directory (and eher were other instrumental variables). [INST TABL 1 ABOUT H] The population of the study consists of manufacturing publicly-listed companies from 1996 to 24. We utilized Purposive sampling technique wh several requirements. First, sample firms must be manufacturing companies because homogeney is important given the assumption of this research design that coefficients are cross-sectional constant. Second, firms selected based on the completeness of firm s respective data. Final sample is 1 publicly traded manufacturing companies and we generate 9 pooled time series, cross section observations. However, the number of observations per year varies and is reported in table 1 due to our next restrictions of the sample: 1) no beta value exceeds absolute 3, 2) we deleted the univariate and event outliers to reduce the noise of the study; this is common in market based-accounting research. We measure the earning change as the difference of earnings per share before extraordinary ems and securies gains and losses for year t-1 and t divided by PS year t-1. Similarly wh percentage change in returns, which is defined over the interval from December, 31 of year t-1 to December 31, of year t. Annual betas are computed wh an individual manufacturer s companies daily return regresses on the IHSG return and corrected regarding non-synchronous trading problem in Indonesian capal market (see Hartono, 25). Annual corrected beta is retrieved from ISMD database. Data Analysis We, in a small part, follow the methodology proposed by Beaver et al. (1997) to test the simultaneous hypothesis of return and earning. We used the ordinary least square and two-stage least square on pooled time-series, cross section estimates and estimate them by cross-section each year. However, in contrast to Beaver et al. (1997) and given the panel nature of our sample, the return-earnings relations are tested for fixed effects because they have three potentially attractive features. First, combines the panel of time series and cross-section data into a single test of significance, in contrast to requiring the researcher to interpret a series of significance tests year-by-year regression (Al-Qenae, Li and Wearing, 22). Second, can lead to increased 1

11 efficiency of estimation and unbiased standard errors because fixed effects estimation can lead to regression residuals wh substantially lower intercorrelation. Third, the slope coefficients on the explanatory variables are less subject to bias from the effects of correlated omted variables (Beaver et al. 1989). Wh regard to 2SLS, we choose 12 instrumental variables (reported in table 2), which can increase the potential benefs of the simultaneous equations approach. Instrumental variables chosen are correlated wh the endogenous variables but independent of residuals. The instrumental variables were chosen because prior research has shown to be proxies for ungarbled earnings or expected returns (see Beaver et al. 1997). [INST TABL 2 ABOUT H] Dickey-Fuller Test and Box Jenkins Method Before estimating the causal and simultaneous relations between earnings and returns change, we draw attention to established time series properties of the data, which is unfortunately, prior research ignores. Particularly we estimate whether the data (e.g. earnings and return change) follow stationary processes. This test is important because if the data is non-stationary, we can only study s behavior only for the time period under consideration. As a consequence, is not possible to generalize the results to other time periods, and thus, for the purpose of forecasting, such non stationary time series may be of ltle practical value. In order to understand the nature of the un root process we test three alternatives of non-stationary: 1) random walk process whout drift, 2) random walk process wh drift, and 3) random walk wh both deterministic and stochastic trends (see equation 7a-9b, for these three types of random walk). The Dickey-Fuller Test is estimated using those three different forms, where the null hypotheses, δ =, mean that the time series are each non-stationary. 11

12 Δ Δ (7a) Δ Δ (8a) Δ = ϑ + 2 Δ Δ (9a) μ Δ = β + ϑ + μ 1 2 Δ = β. t + ϑ + μ t t t Δ Δ (7b) Δ Δ (8b) Δ = ϑ + 2 Δ Δ (9b) μ Δ = β + ϑ + μ 1 2 Δ = β. t + ϑ + μ t t t If return and earnings change are stationary, the value of ϑ should be negative and significant. Otherwise they are non-stationary. The results are presented in table 3. Surprisingly, the results strongly support that both the return and earnings change are stationary in all the three alternative models, since the null hypothesis that δ =, is rejected. Moreover, using the model selection creria (e.g. Akaike Information Crerion, AIC and Schwarz Srerion, SC), is likely that both earning change and return change are both stationary wh zero mean, which is the case of equation 7a and 7b. However, does not mean that future returns are easily predictable using prior returns data. The first difference method of earnings and returns change may transform the actually non-stationary (i.e. random-walk) into the stationary one. Besides, stationary model of earnings and returns provide strong evidence that spurious regressions are omted. [INST TABL 3 ABOUT H] We also estimate the first-order autoregressive pattern of the endogenous variables, earnings change and returns changes using the Box-Jenkins methodology from the following equation: (Y (),i, t - π () ) = α. (Y (),i, t-1 - π () ) + υ. (eq. 1) Where π () is the mean of earnings (return) change, Y (),i, t is the endogenous variable of earnings (return) change, for firm i and time t, and υ is uncorrelated random error term wh zero mean and constant variance. The equation is generated from the correlogram and partial correlogram in which due to space limations, are not reported. 12

13 As been predicted from the extent of noise in the Indonesian capal market, the above estimations of first order autoregressive (A(1)), is only statistically significant at.1% levels for the return change and the estimate is However, earning change does not follow the first order autoregressive process. IV. FINDINGS AND DISCUSSION OLS stimation This section provides evidence on the information content of earnings and returns using ordinary least squares. Table 4 provides the coefficients estimated under OLS for eq (1) and (2). arnings response coefficients fall between and 1 in all 9 years and are significant at the reasonable levels of 1, 5 and 1 percent; wh the exception of 1998 and 21. The average earning response coefficient value is.298, which is lower than the estimated coefficient of Beaver et al. (1997) (i.e..34) and Beaver et al. (1987) (i.e..31) in the similar setting of study. However, the sign of the coefficient on beta changes across year, and in fact only 2 out of 9 years are significant; indicating that the ex-post economy-wide excess returns would be expected to vary. [INST TABL 4 ABOUT H] The estimated return response coefficients are always posive and significant except in As would be expected, post-crisis trauma may eliminate the information power of return. The mean return response coefficient is.389 which is again lower than those reported by Beaver et al (1987, 198). The coefficient on the lagged return change term is posive in 7 of 9 years. Interestingly, the magnude of this coefficient is smaller than on the contemporaneous return term wh a mean of.154 (kindly compare to return mean,.389). Nevertheless, there are some inconsistencies in the parameter estimates. For example, as been discussed in the previous section, α 1 and β 1 are predicted to be equal to (1-θ ) and 1/(1-θ ), respectively. The product of α 1. β 1 should equal to 1. The results showed that the product of the mean value of α 1 and β 1 is.115 (.298 multiplied by.389), which is far below 1. Thus, as discussed earlier, the earning response coefficients and return response coefficients resulted from ordinary least squares are bias. 13

14 Interestingly, when we assume that earnings follows first order autoregressive pattern, and returns change is used to predict three years ahead of earnings, as found by Collins et al. (1994) and Kothari and Sloan (1993), is likely that the information content of returns lose their abily to predict future earnings (see table 6). When returns are used to predict two or three years ahead of earnings, returns can not capture the information about future earnings. egarding the assumption of A(1) in returns change equation, earnings seems to convey information about return, which is similar under non-a(1) assumption. Two Stage Least Squares The results of 2SLS are reported in table 5. Almost similar to the results of OLS, the earning response coefficients generated by 2SLS are posive and significant, except those in 1998 and 23. The return response coefficients also provide identical results than those reported under OLS. Surprisingly, is likely that the information content of returns is reflected in the earning announcement during 1996 until 24 wh the exception, once again, The endogenous variable, beta varies greatly, not only in the sign of parameter estimates, but also in the magnude of the relationships. For example, only 2 relationships exist in the beta-return. Wh regard to lagged return variables, seems that prior return also contains information that is reflected in current earnings in which 8 of 9 relationships are significant. As expected, the product of α 1. β 1 is greater (thus approaching to 1) under 2- Stage Least Squares than under Ordinary Least Squares which is.728; more than six times higher under 2SLS than OLS (α 1. β 1 =.115). Moreover, (1-θ ) and 1/(1-θ ) is also closer,.64 and.878 (1/1.26). These parameter estimates showed that the information content of earnings and returns are simultaneous and the measurement errors in earning response coefficient and return response coefficients are in fact correlated. It also reveals that the product of the coefficients increases under a simultaneous equations approach. [INST TABL 5 ABOUT H] As predicted, the differences mainly because of the bias resulted under OLS estimation. The results may not directly comparable wh those of Beaver et al. (1997) 14

15 and Kothari and Zimmerman (1995) because of differences in deflators (e.g. lagged earnings and lagged prices, respectively versus lagged return) Moreover, under A(1) (see table 6), we find that earnings capture events that is reflected in returns. Both C and C are significant under conventional level. Similar to the results reported under A(1)-OLS, prior returns do not captures information that is reflected in earnings. [INST TABL 6 ABOUT H] Sensivy Test: Kocyk Method Note that equation (2a) may still not amenable to easy estimation since an arbrary (at least from the statistical perspective) parameter estimation of one year lagged return to the earnings change. Strictly speaking, although historical cost, accrual accounting system typically lags the occurrence and recognion of many assets and liabilies, how many period do the lag exist is still questionable and mixed (please compare the findings of Kothari and Sloan, 1992, Collins et all and BMS, 1997). In order to find out the most probable lag exist, we estimate eq. (2a) by pooled time series, cross section estimation using the adaptive expectations method. quation 1 is expressed in s inial equation: Δ Δ = α beta + e. α 1 α (1) Δ Δ Δ Δ = β A(1) + μ 2 2 (6) [INST TABL 7 ABOUT H] mploying the similar method, we estimate eq. (1) and (2a) using OLS and 2SLS pooled time series, cross section estimation. The results are presented in table 8. The results reveal that all the explanatory variables in eq. (2) are significant at the conventional levels of 1, 5, and 1 percent eher under OLS or 2SLS. The results also prove that the contemporaneous linear relations between returns and earnings are decline over time, or in simpler words, the information contents of returns in current 15

16 years is bigger than in the previous years (1.38 for current year and.258 for previous years, and and.17 for current year and.17 for previous year, under OLS and 2SLS, respectively). From table 8, is also interesting to note that the β 1 attached to return gives the short-run impact of a un change in return change on mean earning change. While the long-run impact of a sustained un change in return change on mean earning change can be generated by estimating β 1 /1- β 3, and thus s long run impact is.11 (1.38/1-.5). It is challenging to address that long run impact of return to earnings is smaller than one-year lagged return to earnings. There are in fact several speculations about this. First, there may be asymmetry wh respect to lagged return change. Givoly and Hayn (2) and Basu (1997) show that the association between contemporenous returns and earnings is stronger when returns are negative than when they are posive, reflecting the effect accounting conservatism. Second, return is losing s relevance and accuracy when is used to predict several year ahead of earnings. For example, when is used to estimate current earnings, one year lagged return may be posive, while on the two- and three year lagged return may be negative. [INST TABL 8 ABOUT H] In this test, we also show that the earnings adjust to lag return whin a relatively short time, both under OLS and 2SLS. For example, we find that the mean lag (β 3 /1- β 3 ) is.52 and.79, under OLS and 2SLS respectively. The median lag (log 2/log β 3 ), which is the time required for the first half of the total change in earning following a un sustained change in return, is.23 and.26, both for OLS and 2SLS, respectively. Interestingly, the results provide supportive evidence of previous results that earning changes are reflected in current and lagged return changes. V. CONCLUDING MAKS The study comprehensively analyzes the information content of earnings and returns, and put high consideration of the time series, cross section nature of data. We mainly find that earnings and returns are stationary and that both earnings and returns 16

17 are jointly determined by a larger set of publicly available information and difficult to specify explicly and not independently caused by other factors. Surprisingly, our results implicly indicate inefficiency in the Indonesian stock market and therefore contradict the general finding of efficiency. However, future research should examine the phenomenon comprehensively and consider the time series nature of data carefully. This study inherently has several limations. First, identification of instrumental variables is only based upon prior research. The assumption of independence among the instrument and their respective data is not directly testable. Future research should test the instrumental variables. Second, Non linear relation between earnings changes and earnings changes suggested by Hayn (1995) is very plausible, but unfortunately goes beyond the scope of this study. Future research may consider the non-linear relations of earnings and returns changes. Third, researches have shown that the timeliness of investor s reactions are different; depend upon the losses or profs. Future research may also extract the information content of losses and/or profs. 17

18 FNCS Abarbanell, J Do analyst s earnings forecast incorporate information in prior stock price changes? Journal of Accounting and conomics Ali, A. and Zarowin, P. 1992a. Permanent versus transory components of annual earnings and estimation error in earnings response coefficients. Journal of Accounting and conomics 15, and Zarowin, P. 1992b. The role of earnings level in annual earnings-returns studies. Journal of Accounting esearch 3, Al-Qenae,., Li, C., and Wearing, B. 22. The information content of earnings on stock prices: The Kuwa Stock xchange. Multinational Finance Journal 6, 197 Ball,. and Brown, P An empirical evaluation of accounting income numbers. Journal of Accounting esearch 6, Basu, S The conservatism principle and the asymmetric timeliness of earnings. Journal of Accounting and conomics 24, 3-37 Beaver, W.H The information content of annual earnings announcements. Journal of Accounting esearch 6, 67-92, Clarke,. and Wright, W the association between unsystematic secury percentage change in returns and the magnude of earnings forecast errors. Journal of Accounting esearch 17, , Lambert,. and Morse, D The information content of secury returns. Journal of Accounting and conomics 2, 3-28, Lambert,. and yan, S The information content of returns. A second look. Journal of Accounting and conomics , ger,c., yan, S. and Wolfson, M Financial eporting and the structure of bank share returns. Journal of Accounting esearch 27, , McAnnally, M., and Stinson, C The information content of earnings and returns: A simultaneous equations approach. Journal of Accounting and conomics 23, Perspectives on recent capal market research. The Accounting eview 77,

19 Chen, G., Firth, M. and Krishnan, G. 21. arnings forecast errors in IPO prospectuses and their association wh inial stock returns. Journal of Multinational Financial Management, 11: Collins, D., Kothari, S.P., Shanken, J. and Sloan, Lack of timeliness versus noise as explanations for low contemporeneous return earnings relations. Journal of Accounting and conomics 18, , Kothari, S.P. and ayburn, J Firm size and the information content of returns wh respect to earnings. Journal of Accounting and conomics 9, and Kothari, S.P An analysis of intertemporal and cross sectional determinants of earnings response coefficients. Journal of Accounting and conomics 11, ly, K. and Waymire, G Accounting standard setting organizations and earnings relevance: Longudinal evidence from NYS common stocks, Journal of Accounting esearch 37, Firth, M IPO prof forecasts and their role in signaling firm value and explaining post listing returns. Applied Financial conomics. 8: 29-39, Branson, K., Chee, K. Liaou, T., and Yeo, G Accuracy of prof forecasts contained in IPO prospectuses. Accounting and Business eview. 2: Freeman, The association between accounting earnings and secury returns for large and small firms. Journal of Accounting and conomics 9, Givoly, D. and Hayn, C. 2. The changing time series properties of earnings, cash flows and accruals: Has financial reporting become more conservative? Journal of Accounting and conomics 29, Hartono, J. 25. Pasar fisien Secara Keputusan. (PT Gramedia Pustaka Utama, Jakarta) Jaggi, B Accuracy of forecasts information disclosed in the IPO prospectuses of Hong Kong companies. The International Journal of Accounting 32, Jindrikovska, I. 21. The relationship between accounting numbers and returns: Some empirical evidenced from the emerging market of the Czech epublic. The uropean Accounting eview 1, Kothari, S.P., and Sloan, Information in returns about future earnings. Journal of Accounting and conomics 15,

20 , and Zimmerman, J eturn versus eturn Models. Journal of Accounting and conomics 2, Capal markets research in accounting. Journal of Accounting and conomics 31, Lev, B On the usefulness of earnings and earnings research: lessons and directions from two decades of empirical research. Journal of Accounting esearch 27, Ohlson, A. and Juettner-Nauroth, B xpected PS and PS growth as determinants of value. eview of Accounting Studies 1, Ou, J. and Penman, S Financial statement analysis and the prediction of stock percentage changes in returns. Journal of Accounting and conomics 11, yan, S.G., and Zarowin, P.A. 23. Why has the contemporaneous linear returnsearnings relations declined? The Accounting eview 78, Verecchia, ssays on disclosure. Journal of Accounting and conomics 32,

21 Table 1: mean and standard deviation of % change in return, %change in earning and beta, % change in return % change in earning Beta Year N Mean Std. dev. Mean Std. dev. Mean Std. dev Pooledtime series Table 2: List of Instrumental Variables Percentage change in current assets Percentage change in total revenues Lagged of return on equy Lagged of eturn on total assets Percentage change in total assets Percentage change in total liabilies Percentage change in long term liabilies Operating income to total assets Total revenues to total assets Percentage change of total revenues to total assets Total revenue to cash Lagged of earnings price ratio 21

22 Table 3: Dickey-Fuller test results of earnings and returns change Un root test for return change equation q. (5a) q (6a) q (7a) stimates τ-values stimates τ-values stimates τ-values ϑ * * * β β AIC SC Un root test for earning change equation q. (5b) q. (6b) q. (7b) stimates τ-values stimates τ-values stimates τ-values ϑ * * * β β AIC SC * significant at.1 level of significance 22

23 23

24 Table 4: Coefficients estimated by Ordinary Least Square for earnings-returns relations each year, eturn change equation Δ = Δ α + α 1 + α beta + 2. e arning change equation Δ Δ Δ = β + + μ β1 2 2 Year N 2 α t-stat α 1 t-stat α 2 t-stat 2 β t-stat β 1 t-stat β 2 t-stat * * * * *** *** * *** ** *** * *** * ** * ** * *** * * * *** * ** * ** * ** * * * Mean SD * Significant at.1 level of significance ** Significant at.5 level of significance *** Significant at.1 level of significance 24

25 Table 5: Coefficients estimated by 2-Stage Least Square for earnings-returns relations each year, eturn change equation arning change equation Δ Δ = + + beta + e Δ Δ Δ α α 1 α 2 = β μ 2 Year N 2 α t-stat α 1 t-stat α 2 t-stat β t-stat β 1 t-stat β 2 t-stat * ** ** *** * * * ** ** ** *** * *** * *** * * * * * * *** * * * * * * *** *** * * * Mean SD * Significant at.1 level of significance ** Significant at.5 level of significance *** Significant at.1 level of significance 25

26 Table 6: Pooled time series, cross-section estimation of earnings-returns relations, eturn change equation Δ = Δ α + α 1 + α beta + 2. e arning change equation Δ Δ Δ = β + + μ β1 2 2 N 2 α t-stat α 1 t-stat α 2 t-stat β t-stat β 1 t-stat β 2 t-stat Ordinary Least Squares * * *** Two-Stage Least Squares ** ** *** * Significant at.1 level of significance ** Significant at.5 level of significance *** Significant at.1 level of significance 26

27 Table 7: Pooled time series, cross-section estimation of earnings-return relations, assuming both earnings and returns follow first order autoregressive process Δ Δ = α + α 1 + α beta+ α A(1) + e 2. 3 Δ Δ Δ Δ Δ 2 3 = β A + μ stimates t-values stimates t-values OLS 2SLS OLS 2SLS OLS 2SLS OLS 2SLS α α * 2.18** α α * * β β * 2.25** β *** 2.872* β β ** β *** *** * Significant at.1 level of significance ** Significant at.5 level of significance *** Significant at.1 level of significance Padang, Agustus 26 K-INT 1 27

28 Table 8: Pooled time series, cross sectional estimations of sensivy tests using Koyck, Adaptive xpectations Method, , OLS and 2SLS quation (1) a quation (2) Adaptive expectations, OLS Pooled time series, cross-section estimations stimates t-values stimates t-values α ( β ) α 1 ( β 1 ) * α 2 ( β 2 ) β Adaptive expectations, 2SLS Pooled time series, cross-section estimations stimates t-values stimates t-values α ( β ) * α 1 ( β 1 ) ** ** α 2 ( β 2 ) * β ** a. the results of equation 1a is similar to those resulted in table 6, because the sample and variables used are similar. * Significant at.1 level of significance ** Significant at.5 level of significance 28