HEALTH FINANCE

financialanalyst.pdf

Home >Business & Finance homework help >Economics homework help >HEALTH FINANCE

RAF 6,4

442

Review of Accounting and Finance Vol. 6 No. 4, 2007 pp. 442-459 # Emerald Group Publishing Limited 1475-7702 DOI 10.1108/14757700710835087

Alternative evidence on financial analysts’ use of financial statement information

Donal Byard Stan Ross Department of Accounting, Baruch College – CUNY,

New York, New York, USA, and

Fatma Cebenoyan Department of Economics, Hunter College – CUNY, New York, New York, USA

Abstract

Purpose – Financial analysts are frequently viewed as information intermediaries who process and interpret firms’ financial reports for other market participants. Much recent research, however, has cast doubts on analysts’ ability to fully utilize the information in firms’ financial reports. Using an alternative approach, this study aims to provide evidence on how sophisticated analysts are at using information in firms’ financial reports. Design/methodology/approach – The paper estimates different measures of firms’ operational efficiency, all of which are derived from financial statement data, and compares the strength of the association between these measures and analysts’ absolute forecast errors. It then compares a sophisticated frontier-based measure of firms’ operational efficiency that evaluates firms’ performance relative to their competitors with three more traditional efficiency measures; specifically the return on asset (ROA) ratio, industry-adjusted ROA, and the return on equity ratio. Findings – The results indicate that the more sophisticated frontier-based measure is more strongly negatively associated with analysts’ absolute forecast errors than the other three measures. The results thus suggest that analysts are capable of undertaking a sophisticated analysis of the information in firms’ financial reports, at least as it pertains to operational efficiency. Originality/value – To the extent that analysts serve as a key group of users of financial information, these results are likely to be of interest to accounting policy makers.

Keywords Financial reporting, Financial analysis, Accounting information

Paper type Research paper

1. Introduction Financial analysts play an important role in financial markets, a role that seems to have increased in importance in recent years. Analysts are frequently viewed as information intermediaries who gather, process, and disseminate firm information for investors (e.g. see Schipper, 1991). Indeed, much of the accounting literature views analysts as sophisticated agents who process or interpret firms’ disclosures for investors. Consistent with this view, Lang and Lundholm (1996) document that firms with higher levels of voluntary disclosure attract a larger analyst following.

The view of analysts as sophisticated information intermediaries can, however, be challenged. A large body of literature provides evidence that analysts do not efficiently use all the information contained in firms’ past financial reports. For example, DeBondt and Thaler (1990) and Abarbanell and Bernard (1992) both provide evidence indicating

The current issue and full text archive of this journal is available at www.emeraldinsight.com/1475-7702.htm

The authors thank Sinan Cebenoyan, Neal Galpin, Hongtao Guo, Ying Li, Devra Golbe, Kevin Sachs, and Ping Zhou. They also gratefully acknowledge the contribution of IBES International Inc. for providing earnings per share forecast data, available through the Institutional Brokers’ Estimate System. These data have been provided as part of a broad academic program to encourage earnings expectation research.

Financial statement

information

443

that, when forecasting earnings, analysts do not completely use all the information in past earnings and changes in earnings, and consequently make larger forecast errors. Similarly, Bradshaw et al. (2001) provide evidence that, when forecasting annual earnings analysts do not use all the information relating to working capital accruals from the most recent annual report.

It is difficult however to use such studies to draw inferences regarding how sophisticated analysts are as users of firms’ financial statements. First, studies that show the analysts underreaction to information in historic financial statements do not explicitly model any information processing costs. Second, much of the evidence in such studies may be attributable to: (1) a small number of extreme observations (Keane and Runkle, 1998; Abarbanell and Lehavy, 2003), or (2) the effect of such extreme observations on the econometric specification adopted by most studies (see Basu and Markov, 2004).

This paper takes an alternative approach to examine how sophisticated analysts are as users of firms’ financial statements. Specifically, we compare different measures of firms’ operational efficiency, all of which are calculated using information from firms’ financial statements. We compare these measures to determine which is more strongly negatively associated with analysts’ absolute forecast errors: a larger negative association indicates that the information in that variable is used more by analysts when forecasting earnings. Specifically, we compare a sophisticated frontier-based measure of operational efficiency with, in turn, (1) the return on assets (ROA) ratio, (2) industry-adjusted ROA (AROA), and (3) the return on equity (ROE) ratio. Our tests are based on the idea that, compared to the frontier-based measure, these benchmark ratios are all less sophisticated measures of firms’ operational efficiency derived from firms’ financial statements.

Our study is based on a simple result documented in prior studies: more efficient firms have more stable earnings (e.g. see Berger and Mester, 1997). Thus, if analysts can judge which firms are more efficient, this will help in their forecasting process, i.e. it will help them forecast more accurately. If analysts use the information in a particular efficiency measure (e.g. ROA) to determine which firms are more efficient, then we expect to find a negative association between this efficiency measure and analysts’ absolute forecast errors. We compare the strength of this negative association for the different efficiency measures; that is, we test if the more sophisticated frontier- based measure is more strongly negatively associated with analysts’ absolute forecast errors than each of the three (benchmark) accounting ratios. Our analysis includes controls for firm size, industry, and earnings variability, thus controlling for the inherent difficulty analysts face when forecasting earnings for different firms.

We use financial statement data from Compustat to calculate all of our measures of firms’ relative operational efficiency. Earnings forecasts and actual earnings data used to calculate forecast errors, and (historic) actual earnings data used to calculate earnings variability, are from the Institutional Brokerage Estimation System’s (IBES) database. Our sample period is 1993 through 1997, and consists of 1,216 firm-years, representing 611 firms. We document that analysts’ absolute forecast errors are more strongly (negatively) associated with the more sophisticated frontier-based measure than with any of the other three accounting ratio-based efficiency measures. Our analysis controls for firms’ industry, size (market capitalization), and firms’ level of earnings variability, i.e. the inherent difficulty forecasting earnings. Our results thus suggest that when forecasting earnings analysts use sophisticated tools to analyze the

RAF 6,4

444

information in financial statements, at least as it pertains to firms’ operational efficiency.

Our paper makes a number of contributions. Our results indicate that analysts’ forecasts reflect an understanding of firms’ operational efficiency based on information from firms’ financial statements that is more sophisticated than simply examining firms’ accounting ratios. These results support the idea that analysts are capable of sophisticated analysis of firms’ financial statements. In addition, it is noteworthy that the frontier-based measure that is more strongly (negatively) associated with analysts’ absolute earnings forecast errors assumes quite a detailed knowledge of industry production functions and the within-industry relationships between inputs and output. The results, thus, suggest a clear benefit to investors from analysts’ industry knowledge. While there is a marked industry-specialization among analysts, prior studies do not show any clear benefit to investors from this industry specialization of analysts[1].

The paper is organized as follows. Section 2 outlines the background of our study. Section 3 outlines the study design, while section 4 provides details of the sample selection and methodology. The results are presented in section 5, followed by section 6, which describes our robustness tests. The paper concludes with a discussion in section 7.

2. Background 2.1 Analysts’ use of the information in firms’ accounting reports Financial analysts are widely viewed as important information intermediaries who process firms’ financial disclosures for investors (e.g. see Schipper, 1991). Consistent with this view, Lang and Lundholm (1996) show that firms with higher levels of disclosure attract larger analyst followings. Similarly, Stickel (1989) shows that there is an abnormal flurry of forecast revisions after interim earnings announcements, indicating that analysts use firms’ public accounting disclosures to revise their forecasts.

A large literature, however, suggests that analysts may not very good in their use of the information contained in firms’ financial statements. Many studies test if analysts’ earnings forecasts reflect all the information contained in some information variable (X) that is available to analysts when analysts make their forecasts. The typical specification of such studies is as follows:

FEtþ1 ¼ �0 þ �1Xt þ "tþ1 ð1Þ

where FEt+1 is the forecast error for the period t þ 1 earnings. FEt+1 is calculated as the difference between actual earnings for period t þ 1 (At+1), and the forecast of this earnings figure made prior to this announcement, at period t(Ft): FEt+1 ¼ At+1 � Ft. In these studies, X is an information variable known to analysts when they make their forecast. If analysts fully use all the information contained in X when making their forecast, then X should be unrelated to the subsequent realized forecast errors. In this case the forecasts are said to be ‘‘efficient’’ with respect to the information in the variable X.

Studies use variations of equation (1) to test the efficiency of analysts’ earnings forecasts with respect to different information variables that are known to analysts when they make their forecasts. For example, DeBondt and Thaler (1990) use the level of earnings for period t; that is, they set X in equation (1) above equal to the level of

Financial statement

information

445

earnings in period t. Using this approach, studies have documented inefficiencies in analysts’ earnings forecasts with respect to: earnings levels (DeBondt and Thaler, 1990); earnings changes (Abarbanell and Bernard, 1992); extreme past earnings changes (Easterwood and Nutt, 1999); and working capital accruals (Bradshaw et al., 2001). Broadly, these studies document that analysts underreact to information in firms’ historic financial statements.

These underreaction studies have been challenged, however, more recent studies raise methodological questions concerning such studies. Basu and Markov (2004) point out that estimating equation (1) above using ordinary least squared (OLS) implicitly assumes that analysts face a squared-error loss function – that analysts try to minimize the squared error in their forecasts. They show that the results from estimating equation (1) are sensitive to the specification used. Gu and Wu (2003) point out that because earnings are skewed, analysts may not, in fact, have an incentive to predict large negative earnings in some periods (see also Keane and Runkle, 1998). Finally, these studies do not explicitly model analysts’ information-processing costs.

Our approach to testing whether analysts use the information in firms’ financial statements differs from studies based upon equation (1) above. Such studies represent an absolute test of whether or not analysts’ forecasts reflect a particular information item. In contrast, we undertake a relative test: we compare two potential proxies for analysts’ information to see which one better reflects the information analysts seem to be using when forecasting earnings.

Our study is based on the fact that more efficient firms have more stable earnings (e.g. see Berger and Mester, 1997). Under the assumption that analysts seek to minimize their absolute forecast errors, we test if they seem to be able to use financial statement information to determine which firms have greater levels of operational efficiency[2]. If more efficient firms have more stable earnings, and analysts can determine which firms are more efficient, then analysts’ absolute forecast errors should be negatively related to measure of efficiency (Z ). If we have two efficiency measures (Z1 and Z2), we can test which of these measures better captures the information analysts are using when forecasting earnings using the following model:

jFEtþ1j ¼ �0 þ �1Z1t þ �2Z2t þ "tþ1 ð2Þ

where Z1 and Z2 are measures of efficiency based on data extracted from firms’ financial statements for period t. If we believe analysts are more likely to rely on the information in measure Z1 than the information in Z2, then we predict that �1 > �2. We use such an approach to compare different measures of operational efficiency to test which measure is more strongly associated with analysts’ forecast errors. If one efficiency measures is more strongly negatively associated with analysts’ absolute forecast errors than another efficiency measure, then this indicates that this efficiency measure better reflects the information analysts use when they are forecasting earnings.

2.2 Relative operational efficiency, earnings stability, and our research proposition The concept of relative efficiency estimated using a frontier approach is frequently employed to measure and compare firms’ operational efficiency in all types of industries from services (banking, auditing) to manufacturing to non-profit (schools, hospitals) (see, for example, van den Broeck, 1988; Allen and Rai, 1996; Berger and Mester, 1997; Rogers, 1998; Wheelock and Wilson, 2000; Dopuch et al., 2003; and Callen

RAF 6,4

446

et al., 2005). The widespread use of this frontier approach to measure firms’ operating efficiency results from this measures’ conceptual appeal, which derives from the fact that it captures firms’ relative performance within their industries. This approach yields a comprehensive benchmarking of a firm’s performance relative to that of its competitors[3]. In addition, in its estimation this methodology removes non- controllable random factors such as luck, climate, and machine performances that can affect the observed level of performance, yielding a direct firm-specific measure of the systematic factor(s) affecting a firm’s performance, such as managerial effectiveness (see Hughes et al., 2002 for technical details).

Research indicates that more efficient firms have more stable performance. Several studies within the banking industry, for example, report a negative correlation between banks’ level of inefficiency and the stability of their profitability (e.g. see Berger and Mester, 1997)[4]. Mills and Schumann (1985) conclude that more efficient firms have more stable output levels compared to less efficient firms. McGahan and Porter (1999) find that profits are more persistent for more efficient firms than for less efficient firms within the same industries. Collectively, these studies suggest that relatively more efficient firms have more stable earnings. If analysts can use firms’ financial statement data to determine which firms are more efficient than their competitors, then analysts should make smaller forecast errors. Consequently, we expect a negative association between analysts’ absolute (earnings) forecast errors and measures of firms’ operational efficiency. We compare the magnitude of this negative coefficient for different efficiency measures.

The primary measure of firms’ operational efficiency we use is a measure of firms’ efficiency developed from a frontier estimation that compares a firm’s operational efficiency to that of its competitors. We compare this sophisticated measure of firms’ operational efficiency with three (benchmark) accounting efficiency ratios. These three ratios are:

(1) ROA;

(2) AROA; and

(3) ROE.

All of these efficiency measures are calculated using information from firms’ financial statements. We compare if the frontier-based measure is more strongly associated with analysts’ absolute forecast errors. This tests whether the measure better captures analysts’ information, at least as it pertains to firms’ levels of operational efficiency. If this is the case, then this would suggest that analysts are quite sophisticated in their use of information from firms’ financial statements, at least with respect to firms’ operational efficiencies.

3. Study design All our measures of operational efficiency are calculated using data from firms’ financial statements. The frontier-based measure is a measure of firms’ relative operational efficiency and is modeled using a stochastic frontier methodology. We develop our frontier-based measure of operational efficiency in terms of a firm’s revenue (output) generating ability, given the resources (inputs) it expends relative to its competitors. To do so, we employ a stochastic frontier methodology that incorporates a two-component error structure. One component represents random, uncontrollable factors affecting a firm’s relative (in)efficiency, whereas the second

Financial statement

information

447

component measures the systematic firm-specific component of a firm’s relative (in)efficiency. This systematic component of the firm-specific error is then transformed to create a firm-specific measure of corporate efficiency, EFFIC. EFFIC measures how efficient a firm is relative to other firms in the same industry. As such, EFFIC captures a wide spectrum of data relating to a firms’ performance within a given industry.

We compare EFFIC with three other (benchmark) measures of operational efficiency:

(1) ROA ratio;

(2) AROA; and

(3) ROE ratio.

All three ratios are based on data from firms’ financial statements. We include two control variables in our analysis that control for the inherent difficulty of forecasting earnings for different firms. First, we include a measure of historic earnings variability: the standard deviation of actual earnings per share EPS over the prior five-year period (EPS_STD). Second, we include firm size: market capitalization deflated to constant 1993 CPI-dollars (SIZE). Larger firms have richer information environments which may make analysts’ forecasting process easier. In addition, larger firms may be more stable and, as a result, have earnings that are easier to predict. Finally, we also include industry dummy variables (IDp) as control variables for each two-digit standard industrial classification (SIC) code corresponding to industry p (two-digit SIC code) in our sample. Our analysis is, thus, based on an estimation of the following model:

jAVG FEjit ¼ �0 þ �1EFFICit þ �2Z1 þ �3EPS STDit þ �4SIZEi þ XP p¼2

�jIDp þ "it

ð3Þ

where Z1 is, alternatively, set equal to (1) ROA, (2) AROA, and (3) ROE. We expect that all of our efficiency measures will be negatively associated with analysts’ absolute forecast errors; that is, �1 < 0 and �2 < 0. In addition, following results from prior studies, we expect that analysts’ forecast errors will be larger for firms that have more variable earnings, so we expect that �3 > 0. In addition, following prior studies, we expect that analysts’ absolute forecast errors will be smaller for larger firms, so we expect that �4 < 0. To test our main research conjecture, we test the prediction that the frontier-based efficiency measure has a larger negative effect on analysts’ absolute forecast errors than each of the three (benchmark) efficiency ratios; that is, we test if �1 > �2.

4. Methodology and sample selection Our sample includes all firms from selected industries (see below) with sufficient Compustat data to estimate industry-specific production-functions in any year of our five-year sample period (1993-1997)[5]. We exclude financial services and other regulated industries, such as transportation and utilities: these regulated industries are likely to have different operating environments (e.g. regulated vs. competitive), which may affect firms’ behavioral goals. The remaining industry-years represented in the sample are those with sufficient data to estimate the production frontier for that industry-year. At the frontier estimation stage, the data are tested for possible outliers

RAF 6,4

448

using the standardized residuals method; observations with standardized residuals in excess of two are deleted from the final dataset (Belsley et al., 1980). Data for firm-years with the required efficiency scores are matched with Compustat data for market capitalization and SIC codes, and IBES data used to calculate both analysts’ average absolute forecast errors and the variability in actual earnings over the prior five-year period.

We calculate the average forecast errors for individual analysts using forecasts of annual earnings for year t þ 1 made during a 30-day forecast window immediately after the announcement of first quarter earnings for year t þ 1. This choice of forecast period ensures that the individual forecasts for each firm are conditioned on the same publicly available information – the year t annual financial statements[6]. Our efficiency measures (EFFICit, ROAit, AROAit, and ROEit) are calculated using accounting data for year t, while our analyst forecast errors are based on forecasts of earnings for year t þ 1. Our sample firm-years are from the sample period 1993 to 1997 that meet the following requirements:

(1) The quarterly earnings announcement date for first quarter earnings for year t þ 1 is available from either the active or research Compustat quarterly files.

(2) At least two forecasts of year t þ 1 annual earnings from two different individual analysts are issued in the 30-day period immediately following the first quarter earnings announcement. In addition, these forecasts must be updates of a previous forecast[7].

(3) Actual EPS data for year t þ 1 annual earnings are available from the IBES actual earnings file.

(4) Actual EPS data for the prior five years are available on the IBES actual earnings file – to calculate historic earnings variability.

(5) The necessary Compustat data needed to calculate the efficiency score, ROA, ROE, and data for the control variables, are available for firm i in year t[8].

This yields a final sample of 1,216 firm-years, representing 611 firms. Table I shows the industry composition of the sample firms. The sample firms are spread over 34 different industries, with concentrations in Oil and Gas Extraction (9 per cent of sample), Computer Services (mainly software) (7 per cent), Apparel and Accessory Stores (7 per cent), Computer and Office Equipment (5 per cent), and Electronic Components and Accessories (6 per cent).

We calculate EFFIC, the frontier-based efficiency measure, separately by industry- year. For each industry-year, we separately use a stochastic frontier methodology based on a translog production function. The firm-specific objective function employed can be expressed as:

Rev ¼ fðX; w; vÞ ð4Þ

where: Rev is sales revenue, X is a vector of operational inputs generating this revenue, w represents firm-specific deviations from the efficient frontier due to factors under managerial control, and v represents random uncontrollable factors that affect firm’s performance (see Berger and Mester, 1997). To separately estimate this production relation using data for each industry-year, we use a standard translog function with more than two inputs[9]. In the case of one output (Rev) and three inputs, the translog function can be expressed as follows (e.g. see Bairam, 1994):

Financial statement

information

449

lnðRevÞ ¼ �0 þ �1 ln X1 þ �2 ln X2 þ �3 ln X3 þ 0:5�11ðln X1Þ2

þ 0:5�22ðln X2Þ2 þ 0:5�33ðln X3Þ2

þ �12ðln X1 � ln X2Þ þ �13ðln X1 � ln X3Þ þ �23ðln X2 � ln X3Þ þ ln " ð5Þ

This is followed by the decomposition of the error term, ", from the estimates of equation (5) into its two components, v and w, as defined in equation (4). The decomposition is achieved by specifying the following distributional assumptions (see Jondrow et al., 1982):

v � iid Nð0; �2vÞ and w � jNð0; � 2 wÞj

This specification of the error terms is based on the intuition that the performance of firms differs as a result of: (1) random fluctuations such as luck, climate, machine performance, etc. (captured by v); and (2) a firm-specific components (w), capturing firms’ ability to follow their industry best practices. The firm-specific inefficiency scores (w) obtained through this error decomposition are then transformed into more intuitive efficiency scores (EFFIC) using the algorithm of Battese and Coelli (1988) (see Appendix for technical details of this estimation methodology).

In our selection of input and output variables used to estimate equation (5) we follow prior efficiency studies with some modification. We use the dollar value of net sales (Compustat item number A12) rather than the quantity of output as our measure of

Table I. Industry composition of

sample (sample size ¼ 611 firms)

SIC SIC Industry Code Firms Industry Code Firms

Oil and Gas Extraction 13 55 Measuring and Controlling Devices 382 11

Construction 15 2 Medical Instruments and Supplies 384 27

Food 20 21 Communications 48 15 Textiles 22 20 Durable Goods 50 12 Wood Products 24 6 Wholesale Trade 51 4 Paper and Allied Products 26 8 General Merchandise 53 9 Printing and Publishing 27 9 Food Retailers 54 9 Chemicals and Allied Products 28 37 Apparel and Accessory

Stores 56 47 Drugs 283 19 Eating and Drinking Places 58 20 Rubber, Leather and Glass 30 13 Miscellaneous Retail 59 5 Primary Metal Industries 33 25 Personal Services 72 2 Industrial Machinery 35 33 Business Services 73 17 Computer and Office Equipment 357 35

Entertainment Services 78 11

Electronic Equipment 36 9 Health Services 80 6 Communications Equipment 366 27 Educational Services 82 1 Electronic Components and Accessories 367 37

Engineering and Management Services 87 7

Transport Equipment 37 9 Computer Services 737 43

Total number of firms 611

RAF 6,4

450

output. When there are (even slight) differences in output quality across firms in the same industry, using the quantity of output can lead to a distortion in the measurement of output relative to costs, i.e. efficiency, across firms (Kolari and Zardkoohi, 1987).

The three inputs used in estimating equation (5) in our specification of firm’ production function are: Cost of goods sold, selling, general and administrative expenses, and physical capital cost (that consists of rent and depreciation expense) (Compustat item numbers A41, A189, and the total of A47 þ A14 � A163, respectively). Our choice of these three input categories also follows the choice of different types of costs in the literature regarding the evaluation of input choices in the production process, such as variable, semi-variable, and fixed costs (see Bairam, 1994). To avoid heteroskedasdicity and scale bias, we scale all the variables in our efficient frontier estimation by total assets (Berger and Mester, 1997)[10].

Our dependent variable is first calculated as the average absolute forecast error made by the individual analysts for each firm-year, i.e. |AVG_FE|i,t+1 ¼P

J j=1|Ai,t+1 � Fj,i,t+1|, where j individual analysts forecast year t þ 1 annual earning

for firm i. Important data and econometric issues arise in testing Equation (3) using our dataset. First, our data consists of observations for 1,216 firm-years, representing 611 firms. Our sample consists of an unbalanced set of panel data; some firms have more yearly observations than others. In addition, firm efficiency scores (and firm size) tend to be correlated across sample years (Kwan and Eisenbeis, 1996), introducing the possibility of cross-sectional dependence affecting a pooled cross- sectional time-series analysis of the data. As a result, we replace the firm-year observations with firm-specific mean values, computed across all the firm-years available for each firm. Using such across-panel means is a widely adopted procedure for such unbalanced panel datasets (Greene, 2000; p. 567)[11]. Our regression model is thus estimated as:

mjAVG FEji ¼ �0 þ �1mEFFICi þ �2mZ þ �3mEPS STDi þ �4mSIZEi þ XP p¼2

�jIDp þ "i

ð6Þ

where: m|AVG_FE|i is the average value of |AVG_FE|i,t+1 for firm i; mEFFICi is the average of frontier-based efficiency measure, EFFIC, for firm i; mZ is alternatively: mROAi, the average of the return on assets for firm i; mROEi, the average of the return on equity for firm i; mAROAi the average of the industry-adjusted return on assets for firm i; mEPS_STDi is the average of EPS_STDit: EPS_STDit is the standard deviation of annual EPS on the IBES database over the prior five years for firm i in year t; mSIZEi is average market capitalization, SIZE, of firm i; and IDp are a series of industry dummy variables equal to one for firms in industry p (each two-digit SIC code) and equal to zero otherwise.

Estimates of equation (6) test the association between a firms’ relative production efficiency (EFFIC or ROA) and the average of individual analysts’ absolute earnings forecast errors, controlling for earnings variability, firm size, and industry effects. We use equation (6) to test if �1 > �2.

Table II presents descriptive statistics for the sample of 611 firms. Our sample firms are quite large, with a mean market capitalization of approximately $3.4 billion. The descriptive statistics also reveal quite a lot of variation across our sample firms in the level of analysts’ average absolute forecast errors and efficiency (EFFIC, ROA, and ROE).

Financial statement

information

451

5. Results As the first step in our analysis, we estimate a production frontier for each industry- year in our sample using equation (5). The residuals from these industry year-specific frontier estimates are decomposed into random and firm-specific components for each firm-year. The firm-specific component is then used to generate EFFIC, the frontier- based efficiency measure. These efficiency scores are matched with IBES earnings forecast data to produce our final dataset of 1,216 firm-years, i.e. 611 firms, that is described in Tables I and II.

Table III presents Pearson correlations between our variables. As can be seen in Table III, as expected, we find a negative association between firms’ relative production efficiency (EFFIC) and analysts’ average absolute forecast errors (AVG_FE) (correlation coefficient equals �0.08; significant at p < 0.05, one-tailed test). Similarly, AVG_FE is also significantly negatively correlated with both ROA and ROE – the correlation coefficients are �0.18 and �0.17, both significant one-tailed at p < 0.05 and p < 0.01, respectively. The strength of the association between the ratios and AVG_FE, thus, seems to be stronger than the association between relative production efficiency (EFFIC) and AVG_FE. Consistent with prior research, we also find a negative association between firm size (SIZE) and AVG_FE (correlation coefficient equals � 0.09; significant at p < 0.01, one-tailed test). As we expect, we also find that analysts’ absolute forecast errors are positively correlated with the standard deviation of historic earnings; the correlation coefficient is 0.14, p < 0.01, one-tailed.

The pairwise evidence presented in Table III suggests that analysts’ average absolute forecast errors (mAVG_FE) is negatively associated with both firms’ relative production efficiency, and the ratios ROA and ROE. As expected, we find a positive association between absolute forecast errors and earnings variability: the higher the variability in earnings, the larger the absolute forecast errors. These analyses are incomplete however, as we have not controlled for other known determinants of analysts’ information environment. We address this question next.

Variable Mean Standard deviation

25th percentile Median

75th percentile

mAVG_FEi 0.3838 0.4975 0.0825 0.2049 0.4769 mEFFICi 0.9071 0.0608 0.8743 0.9316 0.9508 mROAi 6.11 7.47 2.95 6.32 9.71 mROEi 11.60 14.23 7.18 12.54 18.26 mSIZEi 3,407 8,753 275 771 2,510 mEPS_STDi 0.7296 0.8641 0.2546 0.4552 0.8157

forecast of year t þ 1 annual earnings for firm i made by individual analyst j; mEFFICi is the average value of EFFICit for firm i: EFFICit is a measure of operational efficiency for firm i in year t calculated using an efficient production frontier calculated using data from year t financial statements for firm i and its competitors; mROAi is the average of ROAit: ROAit is the return on assets for firm i in year t; mROEi is the average of ROEit: ROEit is the return on equity for firm i in year t; mSIZEi is average of SIZEit: SIZEit is the market capitalization of firm i in year t; and mEPS_STDi is the average of EPS_STDit: EPS_STDit is the standard deviation of annual EPS on the IBES database over the prior five years for firm i in year t

Table II. Descriptive statistics

(sample size ¼ 611 firms)

RAF 6,4

452

Table IV presents the results from our OLS regression estimates of equation (6). Note that we do not display the coefficient estimates for the 34 industry dummy variables also included in the estimation. Shown in Table IV are three versions of equation (6), each estimated with a different benchmark accounting ratio (ROA, AROA, and ROE) we compared with EFFIC. First, note that the coefficient on all four efficiency measures are negative in the three versions of equation (6) estimated. This confirms a negative relationship between firms’ efficiency measures and analysts’ absolute forecast errors. Second, consistent with prior research that analysts make smaller forecast errors for larger firms, we find a significantly negative association between analysts’ average absolute forecast errors and SIZE (p < 0.01, one-tailed). Similarly, as expected, we find that analysts’ forecast errors are significantly positively related to EPS_STD, our measure of the variability of earnings (p < 0.01, one-tailed).

Our main research conjecture is tested by comparing the magnitude of the regression coefficient on EFFIC (�1) with, in turn, the magnitude of the regression coefficient on each of the benchmark ratios (�2) – ROA, AROA, and ROE. We use a Wald test to test our conjecture that �1 > �2 (see Greene, 2000, p. 273). As can be see in Table IV, the Wald test confirms that, as expected, EFFIC is significantly more negatively associated with analysts’ forecast errors than each of the benchmark efficiency ratios it is compared with (p < 0.05, one-tailed). This suggests that EFFIC better captures the information financial analysts use to judge firms’ operational efficiency than each of these benchmark accounting ratios. Since EFFIC is a broader

Table III. Correlation analysis (sample size ¼ 611 firms)

mAVG_FEi mEFFICi mROAi mROEi mSIZEi mEPS_STD

mAVG_FEi (Pred. sign) Corr. coeff. (p value)

(�) �0.0776 (0.03)

(�) �0.1824 (0.03)

(�) �0.1745 (<0.01)

(�) �0.0966 (0.01)

(þ) 0.1451 (<0.01)

mEFFICi (þ) 0.2611 (<0.01)

(þ) 0.1956 (<0.01)

(?) �0.0832 (0.04)

(?) 0.0098 (0.80)

mROAi (þ) 0.7803 (<0.01)

(?) 0.0551 (0.17)

(?) �0.1510 (<0.01)

mROEi (?) 0.1484 (<0.01)

(?) �

0.0738 (0.07)

mSIZEi (?) 0.1412 (<0.01)

J j=1

|Ai,t+1�Fj,i,t+1|, where Ai,t+1 is the actual earnings for firm i in year t þ 1 and Fj,i,t+1 is the forecast of year t þ 1 annual earnings for firm i made by individual analyst j; mEFFICi is the average value of EFFICit for firm i: EFFICit is a measure of operational efficiency for firm i in year t calculated using an efficient production frontier calculated using data from year t financial statements for firm i and its competitors; mROAi is the average of ROAit: ROAit is the return on assets for firm i in year t; mROEi is the average of ROEit: ROEit is the return on equity for firm i in year t; mSIZEi is average of SIZEit: SIZEit is the market capitalization of firm i in year t; and mEPS_STDi is the average of EPS_STDit: EPS_STDit is the standard deviation of annual EPS on the IBES database over the prior five years for firm i in year t. ap value for one-tailed tests of significance where sign is predicted, two-tailed otherwise

Financial statement

information

453

and more encompassing proxy for relative efficiency than ROA, AROA, or ROE, these results suggest that analysts act as if they use extensive within-industry knowledge regarding production processes and operational efficiency to forecast earnings more accurately. At the very least, analysts seem to be able to use financial statements to develop information on firms’ operational efficiencies that is more extensive sophisticated than that contained in simple accounting ratios.

6. Robustness We performed several sensitivity tests to establish the robustness of the results. Inferences from our results reported in Table IV are unchanged using any of these alternative specifications: (1) Weighted least squares (WLS): we re-estimated all the

Table IV. OLS regression analysis (sample size ¼ 611 firms)

Intercept mEFFIC mZ mEPS_STD mSIZE Test of �1 > �2

Variable Z set equal to:

Coeff. t-stat.

(p value) a

Coeff. t-stat.

(p value) a

Coeff. t-stat.

(p value) a

Coeff. t-stat.

(p value) a

Coeff. t-stat.

(p value) a

F-stat. (p value)

a Adj.

R 2 (%)

Prediction (?) (�) (�) (þ) (�) (1) ROA 1.54 �1.14 �0.01 0.09 <�0.01

3.25 �2.11 �2.13 3.91 �3.44 4.36 11.33 (<0.01) (0.02) (0.02) (<0.01) (<0.01) (0.02)

(2) AROA 1.51 �1.11 �0.01 0.09 <�0.01 3.20 �2.05 �2.07 3.88 �3.44 4.14 11.24

(<0.01) (0.02) (0.02) (<0.01) (<0.01) (0.02) (3) ROE 1.69 �1.32 �0.01 0.10 <�0.01

3.67 �2.49 �1.70 4.09 �3.50 6.17 10.26 (<0.01) (0.01) (0.04) (<0.01) (<0.01) (0.01)

Notes:

mjAVG FEji ¼ �0 þ �1mEFFICi þ �2mZ þ �3mEPS STDi þ �4mSIZEi þ XP p¼2

�jIDp þ "i

P J j=1|Ai,t+1 �

Fj,i,t+1|, where Ai,t+1 is the actual earnings for firm i in year t þ 1 and Fj,i,t+1 is the forecast of year t þ 1 annual earnings for firm i made by individual analyst j; mEFFICi is the average value of EFFICit for firm i: EFFICit is a measure of operational efficiency for firm i in year t calculated using an efficient production frontier calculated using data from year t financial statements for firm i and its competitors; mROAi is the average of ROAit: ROAit is the return on assets for firm i in year t; mAROAi is the average of industry-adjusted ROAit: ROAit is the return on assets for firm i in year t; mROEi is the average of ROEit: ROEit is the return on equity for firm i in year t; mSIZEi is average of SIZEit: SIZEit is the market capitalization of firm i in year t; mEPS_STDi is the average of EPS_STDit: EPS_STDit is the standard deviation of annual EPS on the IBES database over the prior five years for firm i in year t; and IDp are a series of industry dummy variables equal to one for firms in industry p (each two-digit SIC code) and equal to zero otherwise. ap Value for one-tailed tests of significance where sign is predicted, two-tailed otherwise.

RAF 6,4

454

regressions reported in this paper using WLS, where the number of observations available for each forms the weighting variable (Kmenta, 1997, p. 368); (2) Scaling variable: in our analysis AVG_FE is scaled by the absolute value of actual EPS. We also conduct our analysis scaling by stock price at the end of the fiscal year t � 1, and using unscaled AVG_FE; (3) Rank regression: we re-estimated our model using rank regression specification instead of OLS; (4) Different specifications of ROA: we also use a number of different specifications of ROA (for example, scaling by the average of total assets between the start and end of the fiscal year); (5) Alternative forecast sample periods: we also conduct our analysis using samples of forecast revisions made after the second and third quarterly earnings announcements for year t þ 1 and using forecasts that are not made immediately after an interim earnings announcement.

7. Conclusion This study compares the association between analysts’ absolute forecast errors and different measures of firms’ operational efficiency. We develop four measures of firms’ operations efficiency – a frontier-based measure (EFFIC) and three accounting ratio- based measures (ROA, AROA, and ROE). All four measures are developed using information from firms’ financial statements. We test, in turn, if the frontier-based measure of efficiency is more strongly negatively associated with analysts’ absolute forecast errors than each of the three benchmark ratios.

Our study is based upon the idea that more efficient firms have more stable earnings, so analysts’ should make smaller absolute errors if they can tell which firms are more efficient. Under the assumption that analysts seek to minimize their absolute forecast errors, we expect that analysts with better knowledge regarding firms’ operating efficiency will use this knowledge to forecast more accurately, i.e. to reduce their absolute forecast errors. We argue that the frontier-based measure is more sophisticated and proxies for a far more sophisticated level of industry knowledge. Our comparison of these different efficiency measures, thus, tests which one better proxies for the information analysts use when they forecast earnings. We argue that this comparison provides evidence regarding how sophisticated analysts are in their use of the information from firms’ financial statements.

Our results indicate that the stochastic frontier-based measure of efficiency is more strongly negatively associated with analysts’ absolute forecast errors than any of the three accounting ratio-based measures. This result suggests that analysts’ knowledge regarding firms’ operational efficiencies appears to be more sophisticated than just the information reflected in some simple accounting ratios (ROA and ROE), or even an industry-adjusted ratio like industry-adjusted ROA.

Our results support the view that analysts are capable of undertaking quite a sophisticated analysis of data contained in firms’ financial statement, at least as it relates to understanding firms’ operational efficiencies. In addition, it is noteworthy that the frontier-based efficiency measure that better describes analysts’ earnings forecast errors assumes quite a detailed knowledge of industry production functions and the within-industry relationships between inputs and output. The results, thus, point to a benefit to investors from analysts’ industry knowledge. While there is a marked industry-specialization among analysts, prior studies do not show any clear benefit to investors from this industry specialization of analysts.

Financial statement

information

455

Notes

1. For example, selecting only forecasts of annual earnings for 1997 from the IBES detail file, we find that in 1997 IBES tracked 4,753 analysts forecasting for 6,171 firms. The median (mean) number offirms each analyst followed was 10 (12.78). Although these 6,171 different firms were spread over 70 different two-digit SIC codes, the median numberof two-digit SIC codes representing the firms each individual analyst followed was just two. Furthermore, over 75 per cent of analysts only followed firms from four or less two-digit SIC codes.

2. Recent evidence supports the assumption that analysts seek to forecast accurately and have an incentive to minimize forecast errors (Mikhail et al., 1999). Annual rankings of analysts use forecast accuracy as one of the criteria for choosing the best analysts (see Institutional Investor, 2000; and Wall Street Journal, 2001). In the case of the Wall Street Journal’s annual survey ranking, the rankings are based solely on analysts’ earnings forecast accuracy.

3. Both individual firms and industry consultants use this within-industry relative performance evaluation approach to quantify objective performance rankings within industries (Berger and Humphrey, 1997).

4. Berger et al., (1993) argue that this relationship is consistent with more efficient firms shunning riskier projects in order to avoid the possibility of jeopardizing their relatively more profitable positions.

5. We use a five-year sample period as prior studies indicate that efficiency scores vary through time, and as a result, DeYoung (1997) recommends the use of a sample period of about six years in efficiency studies. If too long a sample period is selected, the concept of ‘‘average firm efficiency’’ loses its meaning. This arises because other factors such as management, technology, and/or regulatory environment may change over time and also affect the stability of the efficiency measures averaged. DeYoung (1997) shows that using a sample period of about six year alleviates this concern.

6. The short 30-day forecast window controls for the potential confounding effects of stale forecasts (Brown and Han, 1992). In addition, to target active analysts, we only select forecasts that are updates of previous forecasts, issued in the 60-day period before the announcement of first quarter earnings in year t þ 1 by the same individual analysts, because these analysts are less likely to be ‘‘herding’’ (Barron and Stuerke, 1998).

7. In rare cases where multiple forecasts are available from the same analyst in the 30-day forecast window after the first quarter earnings announcement, only the last forecast is selected.

8. In addition, we only include firms from industries with sufficient data to calculate the efficiency score needed. Some industries in certain years could not be included due to the inability of the industry-specific production function estimates to converge, especially when there is a small number of firms in the industry. Instead of overriding the stringent converging rules, all non-converging industries are left out of the sample to avoid possible model specification problems.

9. The translog specification is used in preference to the Cobb-Douglas specification. Unlike the Cobb-Douglas specification, the translog specification is flexible in that it allows for non-uniform scale characteristics, and a degree of substitution among inputs that is not limited to unity.

10. Since the costs and revenues of large firms are expected to be larger than for small firms, the random errors of larger firms would have larger variances without any normalization. This is an important consideration since the (in)efficiency terms are derived from the combined residuals. Without an appropriate normalization, this effect may cause the variances in these terms to depend on firm size.

11. We also use a WLS estimator using the number of firm-year observations per firm as a weighting variable. Our inferences are unchanged using this alternative specification.

RAF 6,4

456

12. This is the biggest advantage of stochastic frontier approach. None of the other technologies make any accommodations for (random) non-controllable factors. Furthermore, this approach is stochastic which allows the researcher to make statistical inference based on the results.

13. These distributional assumptions can be relaxed if one can use panel data to estimate the inefficiency measures for each producer (Greene, 2000).

References

Abarbanell, J. and Bernard, V. (1992), ‘‘Tests of analysts’ overreaction/underreaction to earnings information as an explanation for anomalous stock price behavior’’, The Journal of Finance, Vol. 47 No. 3, pp. 1181-207.

Abarbanell, J. and Lehavy, R. (2003), ‘‘Biased forecasts or biased earnings? The role of reported earnings in explaining apparent bias and over/underreaction in analysts’ earnings forecasts’’, Journal of Accounting and Economics, Vol. 36 Nos. 1-3, pp. 105-46.

Aigner, D.J., Lovell, C.A. and Schmidt, P. (1977), ‘‘Formulation and estimation of stochastic frontier production function models’’, Journal of Econometrics, Vol. 6, pp. 21-37.

Allen, L. and Rai, A. (1996), ‘‘Operational efficiency in banking: an international comparison’’, Journal of Banking and Finance, Vol. 20, pp. 655-72.

Bairam, E. (1994), Homogeneous and Nonhomogeneous Production Functions, Avebury Ashgate Publishing Ltd, Aldershot.

Barron, O. and Stuerke, P. (1998), ‘‘Dispersion in analysts’ earnings forecasts as a measure of uncertainty’’, Journal of Accounting, Auditing and Finance, Vol. 13, Summer, pp. 243-68.

Basu, S. and Markov, S. (2004), ‘‘Loss function assumptions in rational expectations tests on financial analysts’ earnings forecasts’’, Journal of Accounting and Economics, Vol. 38 Nos. 1-3, pp. 171-204.

Battese, G.E. and Coelli, T.J. (1988), ‘‘Prediction of firm-level technical efficiencies with a generalized frontier production function and panel data’’, Journal of Econometrics, Vol. 38, pp. 387-99.

Belsley, D., Kuh, E. and Welsch, R. (1980), Regression Diagnostics, John Wiley & Sons, New York, NY.

Berger, A. and Humphrey, D. (1997), ‘‘Efficiency of financial institutions: international survey and directions for future research’’, European Journal of Operational Research, Vol. 98 No. 2 (special issue on ‘‘New approaches in evaluating the performance of financial institutions’’), pp. 175-212.

Berger, A. and Mester, L. (1997), ‘‘Inside the black box: what explains differences in the efficiencies of financial institutions?’’, Journal of Banking and Finance, Vol. 21, pp. 895-947.

Berger, A., Hancock, D. and Humphrey, D. (1993), ‘‘Bank efficiency derived from the profit function’’, Journal of Banking and Finance, Vol. 17, pp. 317-47.

Bradshaw, M., Richardson, S. and Sloan, R. (2001), ‘‘Do analysts and auditors use information in accruals?’’, Journal of Accounting Research, Vol. 39 No. 1, pp. 45-74.

Brown, L. and Han, J. (1992), ‘‘The impact of annual earnings announcement on convergence of beliefs’’, The Accounting Review, Vol. 67 No. 4, pp. 862-75.

Callen, J.L., Morel, M. and Fader, C. (2005), ‘‘Productivity measurement and the relationship between plant performance and JIT intensity’’, Contemporary Accounting Research, Vol. 22 No. 2, pp. 271-309.

DeBondt, W. and Thaler, R. (1990), ‘‘Do security analysts overreact?’’, The American Economic Review, Vol. 80 No. 2, pp. 52-7.

DeYoung, R. (1997), ‘‘A diagnostic test for the distribution-free efficiency estimator: an example using US commercial bank data’’, European Journal of Operational Research, Vol. 98, pp. 243-9.

Financial statement

information

457

Dopuch, N., Gupta, M. and Simunic, D. (2003), ‘‘Production efficiency and the pricing of audit services’’, Contemporary Accounting Research, Vol. 20 No. 1, pp. 47-77.

Easterwood, J. and Nutt, S. (1999), ‘‘Inefficiency in analysts’ earnings forecasts: systematic misreaction or systematic optimism?’’, The Journal of Finance, Vol. 54 No. 5, pp. 1777-97.

Greene, W. (2000), Econometric Analysis, 4th ed., Prentice-Hall, Englewood Cliffs, NJ.

Gu, Z. and Wu, J. (2003), ‘‘Earnings skewness and analyst forecast bias’’, Journal of Accounting and Economics, Vol. 35 No. 1, pp. 5-30.

Hughes, J., Lang, W., Mester, L., Moon, C. and Pagano, M.S. (2002), ‘‘Do bankers sacrifice value to build empires? Managerial incentives, industry consolidation, and financial performance’’, Working Paper No. 02-2, Federal Reserve Bank of Philadelphia, Philadelphia, PA.

Institutional Investor (2000), ‘‘The 2000 all-American research team’’, Institutional Investor, October.

Jondrow, J.C., Lovell, A.K. Materov, I.S. and Schmidt, P. (1982), ‘‘On the estimation of technical inefficiency in the stochastic frontier production function model’’, Journal of Econometrics, Vol. 19, pp. 233-8.

Keane, M. and Runkle, D. (1998), ‘‘Are financial analysts’ forecasts of corporate profits rational?, Journal of Political Economy, Vol. 106 No. 4, pp. 768-805.

Kmenta, J. (1997), Elements of Econometrics, 2nd ed., The University of Michigan Press, Ann Arbor, MI.

Kolari, J. and Zardkoohi, A. (1987), Bank Costs, Structure, and Performance, Lexington Books, New York, NY.

Kwan, S. and Eisenbeis, R. (1996), ‘‘An analysis of inefficiencies in banking: a stochastic frontier approach’’, Economic Review, No. 2, Federal Reserve Bank of San Francisco, San Francisco, CA, pp. 16-26.

Lang, M. and Lundholm, R. (1996), ‘‘Corporate disclosure policy and analyst behavior’’, The Accounting Review, Vol. 71 No. 4, pp. 467-92.

McGahan, A. and Porter, M. (1999), ‘‘The persistence of shocks to profitability’’, The Review of Economics and Statistics, Vol. 81 No. 1, pp. 143-53.

Maddala, G.S. (1977), Econometrics, John Wiley & Sons, New York, NY.

Mikhail, M., Walther, B. and Willis, R. (1999), ‘‘Does forecast accuracy matter to security analysts?’’, The Accounting Review, Vol. 74 No. 2, pp. 185-200.

Mills, D. and Schumann, L. (1985), ‘‘Industry structure and fluctuating demand’’, The American Economic Review, Vol. 75 No. 4, pp. 758-67.

Rogers, K.E. (1998), ‘‘Nontraditonal activities and the efficiency of US commercial banks’’, Journal of Banking and Finance, Vol. 22, pp. 467-82.

Schipper, K. (1991), ‘‘Analysts’ forecasts’’, Accounting Horizons, Vol. 5 No. 4, pp. 105-21.

Stickel, S. (1989), ‘‘The timing of and incentives for annual earnings forecasts near interim earnings announcements’’, Journal of Accounting and Economics, Vol. 11 Nos. 2-3, pp. 275-92.

Van den Broeck, J. (1988), ‘‘Stochastic frontier inefficiency and firm size for selected industries of the Belgian manufacturing sector: some new evidence’’, in Dogramaci, A. and Fare, R. (Eds), Applications of Modern Production Theory: Efficiency and Productivity, Kluwer, Boston, MA, pp. 59-101.

Wall Street Journal (2001), ‘‘A talk with the top stock pickers – industry by industry’’, Wall Street Journal, 26 June, p. R2.

Wheelock, D. and Wilson, P. (2000), ‘‘Why do banks disappear? The determinants of US bank failures and acquisitions’’, The Review of Economics and Statistics, Vol. 82 No. 1, pp. 127-38.

RAF 6,4

458

Appendix: Stochastic frontier methodology The stochastic frontier methodology is derived from the econometric efficiency model developed by Aigner et al. (1977) and further developed by Jondrow et al. (1982). This methodology takes any objective function, estimates a best practice frontier based on the chosen variables, and then provides a measure of firm-specific inefficiencies from a decomposition of the residual term. The objective function can be written, in short form, as:

Y ¼ fðX; �Þ þ e ðA1Þ

where Y is the maximum attainable output from a given vector of inputs (X), � is the vector of parameters estimated, and e a composite error term made up of two components. The justification for this specification of the error term is that firms differ from each other in terms of their objective due to: random fluctuations such as luck, climate, machine performance, etc. and their ability to follow the industry best practice[12]. Hence e decomposes as follows:

e ¼ v þ w ðA2Þ where v represents the disturbances due to uncontrollable events, and w the deviations caused by the controllable factors. Let Y* be the maximum output (sales in this study) that can be produced given the inputs of X1, X2, and X3 (cost of goods sold, general and administrative expenses, and physical capital, respectively). Y* is the frontier objective function of which its parameters, vector �, are estimated from the given observations of Y, and X’s. Since the frontier function is assumed to be stochastic, rather than deterministic, Equation (A1) becomes:

Y ¼ Y� þ w ðA3Þ where:

Y� ¼ fðX; �Þ þ v ðA4Þ

In the above equation, v assumes random fluctuations (both negative and positive), but w assumes only negative values. Since the external events can be both favorable and unfavorable, they can increase or decrease output and are consequently assumed to be drawn from a two-sided distribution (usually normal). On the other hand, inefficiencies only decrease the output or increase costs and are assumed to drawn from a one-sided distribution (usually half-normal). Since the inefficiency component, w, cannot be observed directly, it has to be obtained from the estimated e, which obviously contains information on w. The solution considers the conditional distribution of w, given e (w|v þ w). To do this, first, the joint density of w and � is written as the product of their individual densities. Since e is defined as the sum of these two error terms, this joint density is transformed initially into the joint density of e and w, and subsequently into the density of e by integrating w. The conditional distribution of w, then, is given as the ratio of the joint density of e and w to the density of e (see Maddala, 1977; the Appendix of Jondrow et al., 1982, for detailed calculations). The mean or the mode of this distribution is used as a point estimate of firm-specific inefficiency measure w. Jondrow et al. (1982) gives expressions for these point estimates assuming a half-normal distribution for wi as follows [13]:

EðwjeÞ ¼ �� þ �� fð��=��Þ

1 � Fð��=��Þ ðA5Þ

where:

�� ¼ ��2we=� 2; �2 ¼ �2w þ �

2 v; and �

2 � ¼ �

2 w�

2 v=�

and f and F represent the standard normal density and cumulative density functions respectively. Since the likelihood function used to decompose the error term evaluates f and F at the point where �=�w/�v, ��*/�* in equation (A5) becomes e�/�, and we get:

Financial statement

information

459

EðwjeÞ ¼ �� fðe�=�Þ

1 � Fðe�=�Þ �

e�

�

� �� ðA6Þ

The production function, defined in equation (5) is estimated for each industry-year in our sample giving firm-specific residuals (the e’s estimated using equation (5)). These errors are then decomposed into firm-specific (w) and random (v) components using the maximum likelihood estimator (A6). These firm-specific inefficiency measure (the w) is then transformed into a more appealing firm-specific efficiency measure (which we label EFFIC). This transformation is accomplished using Battese and Coelli’s (1988) algorithm for the case of logarithmic production functions, as is used here. Specifically, for the case of logarithmic production functions, Battese and Coelli specify the firm-specific efficiency measure (EFFIC) as:

EFFICi ¼ expðxi� þ vi � wiÞ

expðxi� þ viÞ ðA7Þ

where a firm’s estimated production function at its own inefficient state is compared to the firm’s estimated production function if the firm-specific level of inefficiency (wi) is zero. This measure, by design, has values between zero and one. For example, if a firm’s efficiency measure is 0.70, this implies that the firm realizes 70 per cent of the production possible for a fully efficient firm operating under the same conditions (Battese and Coelli, 1988).

Corresponding author Donal Byard can be contacted at: [email protected]

To purchase reprints of this article please e-mail: [email protected] Or visit our web site for further details: www.emeraldinsight.com/reprints

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.