An International Comparison of Analysts' Earnings Forecast Accuracy.
There is a substantial body of research that has examined the accuracy of financial analysts' earnings forecasts. To date, however, most of the empirical studies relate to analysts' forecasts in the U.S.  Recently, a few studies have emerged from Britain (for example, Bhaskar and Morris , Capstaff et al. , Cho , Forbes and Skerratt , and O'Hanlon and Whiddett ), Canada [Hennessey, 1994], France [Jacquillat and Grandin, 1994], Hong Kong [Lui, 1993; 1995], Japan [Conroy and Harris, 1995; Mande and Kwak, 1996], and Malaysia [Paudyal et al., 1996]. In contrast, there are very few studies that compare and contrast financial analyst forecasting across nations. The purpose of this paper is to provide some basic analyses on earnings forecast accuracy in 34 countries and to develop a simple model to explain differences among nations.
Research into the accuracy of financial analysts' earnings forecasts is valuable for a number of reasons. First, earnings forecasts and revisions in forecasts are major determinants of stock prices and stock price changes. Assessing the accuracy of forecasts can be used to help improve future forecasting and help investors make choices between analysts. Second, earnings forecasts may be used as inputs into the deliberations of regulators and policymakers. Third, profit forecasts are often used by researchers as a benchmark in studies on financial markets and accounting issues. Cross-national comparisons of earnings forecast accuracy are important because, as investors increase their exposure to foreign stocks, they need to learn more about global markets. The rapid internationalization of individual stock markets, where foreign firms list on domestic exchanges, also leads to a demand for more cross-border studies.
This sample data come from both the domestic (U.S.) and the international Institutional Brokers Estimate System (IBES) databases. Currently, the international database covers 48 [countries.sup.2] and incorporates data from as early as 1987. As of 1997, approximately 15,000 non-U.S. companies are represented in IBES. Because the sample sizes were somewhat small in the late 1980s and early 1990s, earnings forecasts were examined from 1994 through 1997. For the purpose of this research, only the forecasts recorded one month prior to the earnings announcement date were examined. These forecasts will be more accurate than those observed at earlier dates as virtually all studies find that forecast errors decrease as the forecast horizon shortens.
IBES records the forecasts made by individual analysts (often ten or more individual forecasts) and updates the records each month. The dates on which individuals make their earnings estimates, within each month, are reported. Consensus forecasts and dispersion of forecasts are also calculated each month. The latest forecast made in each month was chosen for this study. Prior research has shown that the latest forecast in a given month is the most accurate. Presumably, the latest forecast can impound knowledge of all previously made forecasts and can incorporate all publicly available information up to the date of the forecast.
In common with other studies in this area, this paper uses screening devices to detect and adjust for outliers. Forecast errors greater than 1.50 and less than -1.50 are recorded as 1.50 and -1.50, respectively. A similar rule is used for actual changes in profits. Thresholds of 1.00 or 1.50 are the most common in this type of research. Companies with negative earnings are omitted as it is difficult to interpret the results of calculations involving negative denominators.
The screening devices employed and the need for countries to have been in the database from 1994 onward results in a total of 34 nations being represented in the final sample. The overall sample size and the sample size for each country are presented in Table 1.
Two measures of forecast errors are computed for each firm and for each time period. They are the forecast error and the absolute forecast error. The forecast error (FE) is defined as:
[FE.sub.itm] = ([A.sub.it] - [F.sub.itm]/[A.sub.it] , (1)
where FE is the forecast error, [A.sub.it] is the actual earnings of firm i in year t, and [F.sub.itm] is the latest forecast in month m of the earnings for firm i in year t. The mean of the FEs for a given month m measures whether there is any systematic bias in earnings forecasts (that is, forecasts exceeding actual or vice versa). If there is no bias, then FE should equal zero.
The absolute forecast error (AFE) is:
[AFE.sub.itm] = /([A.sub.it] - [F.sub.itm])/[A.sub.it]/ , (2)
where AFE is a measure of the accuracy of the forecasts.
In order to help assess the absolute forecast errors, compare the AFEs with the predictive accuracy of a simple time series extrapolation based on the random walk model. The random walk model forecasts this year's profit (RWF) as being the same as the immediately preceding year's profit. The absolute forecast error based on the random walk model is denoted AFE(RW) and is thus constructed:
AFE[(RW).sub.it] = /([A.sub.it] - [RWF.sub.it])/[A.sub.it]/ , (3)
where AFE(RW) is the absolute actual change in profit (as [A.sub.it-1] = [RWF.sub.it]) as well as being the absolute error from the random walk prediction model. A high value for AFE [(RW).sub.it] can be interpreted as a signal that corporate earnings for that company are more difficult to forecast. If the actual change in earnings is large, then analysts' forecasts will tend to be less accurate. It is hypothesized that financial analysts should be able to make far better forecasts than those from simple statistical extrapolations. Analysts use the simple no-change model as one input into their forecasts. In fact, the random walk prediction is probably the starting point in many analysts' forecasting processes. Financial analysts have advantages over time series models in terms of information used, knowledge of forecasts made by other analysts, and timing.
In this paper, a simple comparison is made between the mean AFEs and AFE(RW)s across countries. In addition, AFE(RW) is used as an explanatory variable in these cross-sectional models. Comparing analysts' forecasts to those from a random walk model has been used in other accuracy metrics such as the analyst superiority measure of Brown et al.  and the bias and rationality tests of Capstaff et al.  and De Bondt and Thaler .
There is a lot of variability in AFEs across companies, so a number of simple models are developed in an attempt to help explain the differences. AFEs are modeled as a function of size, the number of analysts actively following the company, the risk of a country, and the errors from a random walk prediction. Company size, measured by market capitalization at the date of the forecast, is hypothesized to have a negative relationship with AFE. Larger companies' profits may be easier to estimate as they have more control of their market settings and, therefore, have greater influence over the level of their profits. There tends to be more information on large companies, including greater direct disclosures from the firm itself. This leads to:
H1: There is a negative relationship between size (SIZE) and AFE.
Forecasting accuracy is expected to be a function of the number of financial analysts who actively follow a specific company. The forecasts of individual analysts become known to other analysts along with the reasons for those forecasts. This process leads to greater information sets being available, and it allows analysts to review and revise their own forecasts in light of the estimates made by others. As this paper uses the latest forecast in a month, this has the benefit of incorporating information from all previous estimates. IBES reports the number of analysts who revise their forecasts for a specific company within a particular month. The number of revisions proxies the number of analysts who actively follow the company. The hypothesis is:
H2: There is a negative relationship between the number of analysts who revise forecasts during a month (NOAN) and AFE.
Forecasting accuracy is likely to be a function of the risk class in which a company operates. There are many measures of risk but they are often difficult to operationalize in an international comparison setting. Company risk is partly dependent on the country where the company is located, and preliminary analyses indicated that forecasting accuracy varies across nations. Instead of using dummy variables for countries, country risk metrics are chosen here. The PRS Consultancy Group provides a comprehensive coverage of risk measures by country. In particular, they provide measures of political risk, economic risk, financial risk, and composite (political, economic, and financial) risk. The risk measures are on a scale of 0 to 100 with 100 being very low or no risk, and 0 is extremely risky. The financial risk score and the composite score is used here. It is hypothesized that high-risk countries will be associated with low forecast accuracy. Given the nature of the PRS ratings (high score equals low risk), t he following hypothesis is derived:
H3: There is a negative relationship between composite risk (or financial risk) and AFE.
It is more difficult to forecast the earnings of some companies than for others. Differences in product markets, technologies, exposures to interest rate and exchange rate changes, national economic policies and conditions, and other factors lead to differences in accuracies across companies and countries. It is difficult, however, to construct models to account for all of these factors. It is argued here that the accuracy of a no-change earnings forecasting model is an overall indicator of the forecasting difficulties faced by financial analysts. This leads to the following hypothesis:
H4: There is a positive relationship between change in actual profits (AFE(RW)) and AFE.
The hypotheses are tested via the following regression model:
AFE [[beta].sub.0] + [[beta].sub.1] SIZE + [[beta].sub.2] NOAN+ [[beta].sub.3] CRISK + [[beta].sub.4] AFE(RW), (4)
where: AFE = absolute forecast error; SIZE = market capitalization measured in U.S. dollars at the date of the forecast; NOAN = number of analysts who revise their forecasts in the month prior to the earnings announcement; CRISK = composite risk measure for the country in which the company is domiciled (the composite risk is obtained from PRS); and AFE(RW) = absolute forecasting error from a random walk model. It is the same as the absolute change in profits from year t - 1 to year t.
An alternative measure of risk, financial risk (FRISK), is also used in the regression. Thus, FRISK = financial risk measure for the country in which the company is domiciled (the financial risk is obtained from PRS).
Bias and Accuracy
Table 1 presents the basic results on bias and accuracy. The mean bias (derived from (1)), mean absolute forecast errors (from (2)), and the absolute forecast errors from the random walk model (see (3)) are shown. The means are computed for the overall sample by individual country, continent, and whether IBES classifies a country as a mature market or an emerging market.
The overall bias is 0.27 percent which is not statistically different from zero (t = 0.987). Although financial analysts make errors in their forecasts, there is no systematic over-or under-forecasting evident in the data here. The mean absolute forecast error for the whole sample is 11.51 percent. This error (11.51 percent) compares with the mean absolute forecast error from a no-change model of 40.41 percent. As expected, the analysts do a much better job of forecasting than does the random walk model.
The biases (FEs) range from 0.01 percent in Japan to 15.03 percent in Indonesia. Fourteen of 34 countries have negative bias where the actual earnings are less than the forecast earnings. The mean AFEs range from 1.74 percent in Austria to 33.45 percent in Brazil. In all countries, the mean AFEs are less than the mean APE (RW)s, therefore showing analyst superiority over simple time series models. Accuracy is particularly poor for South and Central American countries, and all four constituent countries (Argentina, Brazil, Chile, and Mexico) have high AFEs. South American companies also have the greatest positive bias metric. Scandinavia (Denmark, Finland, Norway, and Sweden) also has a high AFE, and the negative bias statistic shows that financial analysts' forecasts are too high. IBES categorizes some countries as emerging markets and others as mature markets. Although some of the selections may be debatable (for example, Hong Kong, Korea, and Singapore are treated as emerging markets), the bias and accurac y of emerging markets and mature markets are calculated. The emerging markets have lower accuracies than the mature markets, although the difference is not large.
The accuracy of earnings forecasts of U.S. companies may be used as a benchmark against which to compare other countries. Disclosure requirements and practices, economic stability, experience, and the intensity of financial analysis, may suggest that forecasting accuracy should be highest in the U.S. The data here, however, show that approximately one quarter of the countries examined in this study have smaller AFEs than the U.S. Cho  and Conroy and Harris  have previously concluded that forecasting accuracy may be better in the U.K. and Japan vis-a-vis the U.S. The study here finds very little difference, however, between the forecasting accuracies of the U.K. and U.S. and Japan and the U.S.
Cross-Sectional Explanatory Model
The results from the regression equation are shown in Table 2. Two proxies are used for risk. They are composite risk (CRISK) and financial risk (FRISK). These two variables are highly correlated, so they are included one at a time in the regressions. The [R.sup.2] s of 10.40 percent and 10.42 percent compare well with other cross-sectional models reported in the literature on analysts' forecasting accuracy.
SIZE has the expected negative sign, but statistical significance is weak. NOAN has negative coefficients as expected, although they are only significant at the 0.10 level. The evidence suggests that greater financial analyst scrutiny of a company leads to more accurate earnings forecasts. Therefore, Hypothesis 2 receives moderate support from the data here. Risk in the forms of CRISK and FRISK is significant at the 0.01 and 0.05 levels, respectively, and the negative directional signs are consistent with expectations. The evidence supports Hypothesis 3. The more risky a country is, the worse the earnings forecast accuracy is for companies in that country.
The change in the absolute profits variable, AFE(RW), is highly significant in explaining the absolute forecast errors. The positive sign indicates that large changes in profits from year t - 1 to year t make earnings forecasts by financial analysts less accurate. The results from Table 2 strongly support Hypothesis 4. Further analyses suggest that AFE(RW) is the most important of the independent variables in improving the explanatory power of the model. The coefficient on AFE(RW) is less than 1 which is consistent with analysts' forecasts being more accurate than the forecast given by the random walk model.
These analyses demonstrate significant differences in forecasting accuracies across nations, and there are a multitude of potential reasons for this. In this study, one potential dimension is investigated, financial risk and composite risk. Other reasons for differences between countries include disclosure regulations, accounting rules, tax regimes, corporate governance structures, and national economic policies and conditions. Future research will be directed toward investigating these factors.
This research is a preliminary investigation of differences in financial analyst forecasting accuracies across companies and countries. The results indicate substantial inter-country variability in measures of forecasting bias and accuracy. South and Central America have the poorest accuracy, followed by the Scandinavian nations. Although the U.S. is usually argued to have the greatest amount of disclosure and the largest degree of investment analysis activity, the bias and accuracy measures are only average among the developed or mature nations. Cross-sectional regression results showed that the proxy for the inherent difficulty of forecasting, namely the absolute change in profits, is by far the most important variable in explaining forecast accuracy. Company risk is hypothesized to be another important determinant of accuracy, but it is difficult to operationalize a measure which is valid across countries. Because the national environment in which a company operates is likely to be a factor in a company's risk and because national risk ratings are available, a firm's risk by its country's risk is proxied here. The two measures of country risk, composite risk, and financial risk, are statistically significant and have the expected directional signs. Two other explanatory variables, size and number of analysts who actively research a company, have the hypothesized negative signs, although the statistical significances are weak to moderate.
The models in this paper explain about 10 percent of the variability in forecast accuracy, so there is considerable scope to investigate other factors that may influence or affect earnings estimation errors. These factors will include environmental and institutional features of the nation in which a company is domiciled. Further research on this topic will add to the growing literature on the globalization, internationalization, and cross-cultural differences of financial markets.
(1.) For review of this literature, see Brown , Brown et al. , and Givoly and Lakonishok .
(2.) Some of these countries have recently been added to the database so there is little or no data available as yet.
Bhaskar, K.; Morris, R. "The Accuracy of Brokers' Profits Forecasts in the U.K.," Accounting and Business Research, 3, Spring 1984, PP. 113-24.
Brown, L. D.; Richardson, G. D.; Schwager, S. J. "An Information Interpretation of Financial Analyst Superiority in Forecasting Earnings," Journal of Accounting Research, 25, 1987, pp. 49-67.
Brown, L. D. "Earnings Forecast Research: Its Implications for Capital Market Research," International Journal of Forecasting, 9, 1993, pp. 295-320.
Brown, P.; Foster, G.; Noreen, E. Security Analyst Multi-Year Earnings Forecasts and the Capital Market, Sarasota, FL: American Accounting Association, 1985.
Capstaff, J.; Paudyal, K.; Rees, W. "The Accuracy and Rationality of Earnings Forecasts by U.K. Analysts," Journal of Business Finance and Accounting, 22, January 1995, pp. 67-85.
Cho, J. Y. "Properties of Market Expectations of Accounting Earnings by Financial Analysts: U.K. versus U.S.," Accounting and Business Research, 24, Summer 1994, pp. 230-40.
Conroy, R. M.; Harris, R. S. "Analysts' Earnings Forecasts in Japan: Accuracy and Sell-Side Optimism," Pacific-Basin Finance Journal, 3, 1995, pp. 393-408.
De Bondt, W.; Thaler, R. "Do Security Analysts Overreact?, "American Economic Review: Papers and Proceedings, 80, 2, May 1990, pp. 52-7.
Forbes, W. P.; Skerratt L. C. L. "Analysts' Forecast Revisions and Stock Price Movements," Journal of Business Finance and Accounting, 19, June 1992, pp. 555-69.
Givoly, D.; Lakonishok, J. "Properties of Analysts' Forecasts of Earnings: A Review and Analysis," Journal of Accounting Literature, 1984, pp. 117-52.
Hennessey, S. M. "Are Revisions of Security Analysts' Earnings Forecasts Rational? Canadian Evidence," Journal of International Accounting: Auditing and Taxation, 3, 1994, pp. 205-20.
Jacquillat, B.; Grandin, P. "Performance Measures of Analysts' Forecasts," Journal of Portfolio Management, Fall 1994, pp. 94-102.
Lui, Y-H. "Market Reaction to Analysts' Multi-Year Forecast Revisions: A Non-Parametric Approach," British Accounting Review, 27, 1995, pp. 35-44.
__. "Revision Properties of Hong Kong Security Analysts' Earnings Forecasts, "British Accounting Review, 25, 1993, pp. 257-68.
Mande, V.; Kwak, W. "Do Japanese Analysts Overreact or Underreact to Earnings Announcements?," Abacus, 32, 1, 1996, pp. 81-101.
O'Hanlon, J.; Whiddett, R. "Do U.K. Security Analysts Over-React?," Accounting and Business Research, 22, Winter 1991, pp. 63-74.
Paudyal, K.; Saadouni, B.; Briston, R. "Earnings Forecasts in Malaysia: An Empirical Analysis," working paper, University of Strathclyde, 1996.
(*.)Hong Kong Polytechnic University and Lingnan College--Hong Kong. The authors gratefully acknowledge the insightful comments made by the participants at the Forty-Fifth International Atlantic Economic Conference in Rome, Italy, March 14-21, 1998. Their helpful suggestions have been incorporated into this paper. The authors also acknowledge the contribution of the Institutional Brokers Estimate System, Inc. for providing earnings per share forecast data and is available through them. This data has been provided as part of a broad-based academic program to encourage earnings expectations research.
Earnings Forecast Bias and Accuracy by Country Mechanical Bias Accuracy Accuracy Sample Country (MFE) (MAFE) (MAFE(RW)) Size All Countries 0.0027 0.1151 0.4041 9,515 Argentina 0.0062 0.2210 0.5211 49 Australia -0.0065 0.0657 0.2775 290 Austria 0.0039 0.0174 0.4765 63 Brazil 0.0635 0.3345 0.6901 118 Chile 0.1397 0.2291 0.3947 66 China 0.0684 0.1509 0.5043 53 Denmark -0.0146 0.1292 0.3865 101 Finland -0.0647 0.2475 0.6147 68 France 0.0602 0.1590 0.3962 332 Germany -0.0004 0.1364 0.3923 232 Greece -0.0119 0.0961 0.4061 160 Hong Kong 0.0441 0.1156 0.2756 229 India 0.0732 0.1479 0.3651 130 Indonesia 0.1503 0.2004 0.4248 109 Ireland -0.0175 0.0840 0.3134 38 Israel 0.0038 0.1203 0.1298 3 Japan 0.0001 0.1194 0.5577 1,896 Malaysia -0.0131 0.1060 0.2511 243 Mexico -0.0373 0.3052 0.7021 77 Mature Markets -0.0053 0.1080 0.4054 7,335 Netherlands 0.0226 0.0808 0.2981 176 New Zealand 0.0098 0.1111 0.2709 88 Norway -0.0406 0.2129 0.4650 69 Pakistan 0.0090 0.2130 0.5083 19 Philippines -0.0235 0.1118 0.2716 85 Singapore 0.0163 0.0754 0.2616 167 South Africa 0.0238 0.0717 0.3337 122 Sri Lanka 0.0154 0.0648 0.4192 33 Sweden -0.0275 0.1553 0.4668 133 Switzerland -0.0511 0.1413 0.4374 138 Taiwan 0.0317 0.1638 0.4297 144 Thailand 0.0060 0.1135 0.4159 198 Turkey 0.0199 0.0799 0.5678 175 United Kingdom -0.0212 0.1000 0.2654 747 United States -0.0089 0.0927 0.3581 2,964 Asia 0.0132 0.1222 0.4649 3,276 Australasia -0.0027 0.0763 0.2759 378 Europe 0.0010 0.1099 0.3590 2,061 Scandinavia -0.0332 0.1758 0.4717 371 South America 0.0551 0.2869 0.6035 410 Emerging Markets 0.0299 0.1390 0.3998 2,180 Notes:M in MFE, MAFE, and MAFE(RW) denotesthan mean across all observations in a country or category. Cross-Sectional Regression Models of Absolute Forecast Errors (AFEs) Variables Coefficients t- Statistics Coefficients t-Statistics Intercept 0.1304 (4.18) [*] 0.0878 (4.41) [*] SIZE -0.0000 (-1.26) -0.0000 (-1.36) [***] NOAN -0.0015 (-1.49) [***] -0.0016 (-1.57) [***] CRISK -0.0010 (-2.71) [*] FRISK -0.0011 (-2.11) [**] AFE(RW) 0.1813 (32.53) [*] 0.1825 (32.59) [*] [R.sup.2] 0.1042 0.1040 Notes: (*.)denotes significance at the 0.01 level, (**.)denotes significance at the 0.05 level, and (***.)denotes significance at the 0.10 level.
|Printer friendly Cite/link Email Feedback|
|Author:||FIRTH, MICHAEL; GIFT [*], MICHAEL|
|Publication:||International Advances in Economic Research|
|Date:||Feb 1, 1999|
|Previous Article:||Fiscal Policy Effects, Convergence, and Growth.|
|Next Article:||Budget Deficits and Economic Activity.|