Bias in European Anaiysts' Earnings Forecasts
Stan Beckers, Michael Steliarcs, and Alexander Thcmson
Forecasting company cnrniu^s /s ( difficult and hazardous task. In nn 7 efficient market where annly^^ts learn from ptist mistakes, there should be no persistent and systematic biases in consensus earnings accuracy. J^rcvious research has already established how some (single) individualcompany characteristics si/stematically influence forecast accuracy. So far, however, the effect on consensus eariiings biases of a comptmy's sector and country affiliatioti combined with a range of other fundameutal chanieteristics has remained largely unexplored, ilsiiig data for 19932002, this article diseiitangles and quantifies for a broad universe of European stocks how the number of analysts following a stock, the dispersion of their forecasts, the volatility of earnings, the sector and country classification of the covered conipamj, ami its nuirket capitalization influence the accuracy of the consensus earnings forecast.

5

ecurity analysts are considered to be the premier experts in the assessment of a company's prospects. Their research efforts are largely directed at producing accurate earnings forecasts, which are a key input into equity valuation models. Although the importance of financial analysis is beyond dispute, its quality has become the subject of much scrutiny and debate. Academic research has a long history of documenting systematic earnings forecast errors, but analysts' conflicts of interest and potential biases in their recommendations and earnings projections have now invaded even the headlines of the popular press. Forecast errors fall into two broad categories— optimism and herding. Despite the fact that optimism—a consistent tendency for analyst earnings forecasts to be positively biased—was documented as early as the 1970s {see, for instance, Crichfield, Dyckman, and Lakonishok 197S), systematic positive forecast errors persist to this day. Similarly, herding behavior—whereby analyst forecasts are less dispersed titan one would rationally expect—has been widely obser\ ed. This collective tendency not to deviate too much from the consensus is puzzling and, presumably, rooted in

Stan Beckers is visit ii\^ /jnj/i.'.'^sor in finance at Kaiholkkc Universiteit at Leuzvn, Belgium. Michael Steliiiros is an associate in quantitative equity research nt Barclays Clobal investors and visiting lecturer at CASS Biishiess School, London. Alexander Thomson holds an MSc in finance from CASS Bushtess School, Londo)i.

some deep human desire lo conform. The persistence and systematic nature of herding across time, firms, and markets suggest that analysts are willing to sacrifice prediction accuracy for conformity to the consensus. A straightforward and commonly applied procedure to mitigate these biases when valuing n company is to average the earnings forecasts of all the analysts following that company. To the extent that optimism and herding affect the typical analyst, however, they will also affect the consensus forecast. But although optimism and herding apply broadly to all stocks, they do not feature equally among all stocks, markets, or sectors. So, if the sign and magnitude of forecast errors vary systematically and persistently as a function of well-defined company characteristics, one should be able to improve upon the consensus forecast by correcting for these systematic effects. Previous studies of factors that influence forecast error and bias have concentrated on the effect of single variables. None have systematically explored sector and country effects in a multipleregression framework that also adjusts for other fundamental company characteristics (such as earnings forecast dispersion, market capitalization, earnings variability, or number of analysts following a stock). This article attempts to lielp fill this gap. In doing so, we hope to facilitate a more discriminate and intelligent use of analysts' work than in the past and to help practitioners identify those companies with the highest ('.v ante forecast error and, therefore, possible stock mispricing.

74

©2004, AIMR®

Bins in European Analysts' Earnings Forecasts

Previous Findings
A substantial bt)dy ot' literature has explored the phenomena of herding, optimism, and forecast errors. We summarize previous findings and discuss the practical implications of the inaccuracies in analyst consensus earnings forecasts. Herding, l-lcrding is a well-established social phenomenon, and the number of theories atteniptinj^ to explain il is almost overwhelming {see VVulch 2000 for a brief summary). Most of the research shows that the incentive to conform increases witli the number of previtnis adopters and with a lock of privacy {because the fear of being identified as a deviant stimulates herding). Herding also intensifies as the difficulty or ambiguity of the task increases. Academic research into herding among financial analysts has led to the following insights: • The tendency to herd to the consensus increases with the number oi estimates close to the consensus and with the inaccuracy of one's own past estimates (Stickel 1990; Crahcim 1999). • Older analysts are more likely to produce forecasts that deviate from the consensus, and conversely, younger analysts are typically less bold than their older counterparts (Hong, Kubik, and Solomon 2000). • The tendency to herd lias no relationship to the accuracy of the consensus forecast (Welch). • Herding increases with earnings unpredictability (Olsen 1996). Herding leads to an unusually narrow dispersion of individual analysts' forecasts. Although this narrow dispersion may be irrational, it should not affect the accuracy of the average forecast. The evidence is strong, however, that not only the dispersions of the forecasts but also their means are persistently biased. We review this evidence in the next paragraph. Optimism. The magnitude and significance of Ihe earnings surprise, defined as the difference between the consensus earnings forecast and the subsequent earnings realization, is the forecast bias. In an early study, Dreman and Berry (1995) established the persistent presence of an optimistic bias in analysts working in the U.S. market: Between 1972 and 1991, U.S. consensus earnings forecasts significantly exceeded actual earnings. The authors found no significant differences in the optimism among industries or economic cycles. The same persistent optimism has been documented for equity markets in the United Kingdom (Capstaff, Paudyal, and Rees 1995; De liondt and Porbcs 1999), Germany {Capstaff et al. 1998), and
March/April 2004

l-urope {Capstaff et al. 2001). Casual empiricism in this last study indicates a significant degree of heterogeneity in forecast bias among European countries—specifically, poor earnings forecasts for Spanish and Italian companies and better-thana\ erage forecasts for U.K. and Dutch companies. Other studies established that forecast bias increases witli earnings variability (Huberts and Fuller 1995; De Bondt and Forbes) and with a lack of earnings visibility (Das, Levine, and Sivaramakrishnan 1998) and that it varies inversely with the level of earnings (Francis and Philbrick 1993; Dowen 1996; Butlerand Saraoglu 1999; Easterwood and Nutt 1999). The presence of optimism has been linked to the way in which analysts react to new information. De Bondt and Thaler (1990) documented that analysts typically overreact to positive news and make ftirecasts that are more extreme than warranted by the available information. The overreaction effect is amplified by an underreaction to negative news. Olsen argued that forecasters who otherwise would tend toward a below-a\'ernge forecast are drawn toward the mean and are loath to fully reflect their (negative) views in their forecasts because tho market appreciates an optimistic outlook. The combined effect would help explain the positive bias in the overall consensus number. Another factor contributing to overly optimistic forecasts, according to Hong and Kubik (2003), is the compensation structure for financial analysts. The authors found—after controlling for accuracy—that analysts who are more optimistic than the consensus are more likely to experience positive career moves. McNichols and O'Brien {1997) similarly hypothesized that some analysts choose not to publish unfavorable forecasts. Finally, a well-documented assertion is that analysts from brokerage houses issue overly positive predictions for companies for whom their employer does other corporate work {Dugar and Nnthan 1995; Lin and McNichols 1998; Michaely nnd Womack 1999). Although tlie existence of optimism can be explained by and attributed to the workings of the human psyche and the labor markets, little research has so far been done to establish how this optimism varies by country, sector, or company. We address this subject later in the article. Forecast Error. The absolute value of the earnings surprise is ihcforccast error, and it captures how "wrong" the consensus number is. A considerable body of academic literature has reported on how various factors affect this forecast error. De Bondt and Forbes, for instance, found that the forecast error increases with forecast change {the
75

Financial Analysts journal

difference between next year's forecasted earnings and this year's realized earnings) and with analyst dispersion (cross-sectional standard deviation of analyst forecasts) whereas—unsurprisingly—it decreases with the forecast horizon (see also Capstaff etal. 1998). Sinha, Brown, and Das (1997) and Capstaff et al. (1999) showed that analysts' forecast errors are smaller for companies with comparatively large market capitalizations, for large absolute values of forecast earnings, and for companies followed by a large number of analysts. The study by Agrawal and Chadha (2002) highlights theimportanceof the level of disclosure to the accuracy of consensus earnings forecasts. They analyzed the impact of Regulation Fair Disclosure, which prohibits companies from selectively disclosing market-sensitive information. The introduction of Regulation FD has been associated by some buy-side and sell-side analysts with a significant reduction in the information quality of company statements (AIMR 2001), leading to a less transparent, moro confused forecasting environment than before Regulation FD. Agrawal and Chadha concluded that post-FD earnings forecasts become less accurate as their cross-scctionni dispersion increases.' Finally, Capstaff et ai. (1999, 2001) produced descriptive e\ idence that earnings forecast accuracy differs by the industry being covered (with the health care and public utility sectors ha\'ing, on average, more accurate estimates than the transportation sector), although they conceded that this situation may be a result of differences in earnings variability or visibility among the industries.

Practical Implications and Research Design
The evidence so far leads to the following insights: • Individual analysts herd when making their earnings forecasts. This behavitir leads to a narrower distribution of individual analyst forecasts than one would rationally expect. • Consensus earnings forecasts are optimistic; that is, they typically overestimate future realized earnings. • The accuracy of the consensus forecast— defined as the absolute value of the earnings surprise—may vary by the analysts' focus on sector or country and as a function of a range of company characteristics. In sum, systematic biases affecting earnings forecasts are well documented but analysts appear unable, on average, to correct for them. In fact, earnings surprises are the norm rather than the
76

exception and stock prices continue to react to forecast errors as though each is unique and has a low probability of occurring. Moreover, as long as market prices reflect the consensus earnings projections, they will be similarly biased. Diether, Malloy, and Scherbina (2002), for instance, argued that prices reflect the optimistic consensus view (possibly because the investors with the lower expectations are kept out of the market by high short-sale costs), leading to lower returns than for otherwise similar stocks. The implication is that the existence of systematic forecast bias can be used to establish a comparatis'o investment advantage. The identification of stocks with the highest t'.v ante consensus earnings bias and forecast error appears to open interesting investment opportunities. The picture from the preceding analyses, howe\'er, is confused: Bias and forecast error may be a function of the variability, visibility, and level of earnings, of the number of analysts following the stock, of the dispersion of their forecasts, and of the market capitalization, nationality, and industry classification of the followed company. Most of the academic studies have concentrated on the effect of a single stock characteristic on forecast error nnd bias. So, to infer specifically what the dominant and driving forces are that affect forecast error and bias remains difficult. In the study reptirted here, we endeavored to disentangle these \'arioLis effects. We used a multiple-regression framework to address the following questions related to consensus earnings forecasts for a universe of Furopean stocks for 1993-2002: • Does the quality of forecasts differ significantly from country to coimtry? • Which sectors typically have higher forecast errors and biases? • To what extent do the number of analysts following a stock, the variability of earnings, the market capitalization of the company, and the dispersion of analyst forecasts influence the accuracy of the consensus earnings forecast? In particular, the relative importance of country and industry effects on earnings forecast errors has so far been especially poorly documented. Although the waning impact of company nationality and increasingeffectof company sector on stock returns has been widely established (Cavaglia, Brightman, and Aked 2000), little effort has been made to disentangle the influence of country and sector on earnings forecasts and their accuracy. For country effects to ha\'e an impact on the accuracy and bias of earnings forecasts would be only natural (Pike, Meerjanssen, and Chadwick 1993; Higgins 1998): Significant geographical
©2004, AIMR®

Bias in European Anah/sts'

Ear)U)i^s Forecasts

differences in accounting practices and standards, in the frequency and detail of company reports, in government regulation, and in business and tax laws lead to widely divergent levels of corporate disclosure." Similarly, earnings variability and visibility typically differ significantly from industry to industiy, making it difficult to make any priina facie statements about the net effect of sector affiliation on earnings forecast accuracy.

Table 1.

Descriptive Statistics of Company Sample, 1993-2002
12-Month Torecists 687 3,212 6,939 IS 1.46 1.894 29.7 2-1-Month Forecasts 632 2,370 7,728 13 1.53 2.157 41.0

Stalislic Tot.il number of companies ToUil iiiimtior of observations .•\vor.igc market ciip (S millions) A\'erage number of analysts per company Average realized liPS (S) Average forecast lil'S (S) Average dilTerencc between forecast and re.ilizntion ('';.)

The Study
The objective of our analysis was to identify which characteristics (all else being equal) are associated with higher-than-average consensus earnings forecast error and bias. This information will benefit sell-side analysts by enabling them to identify which consensus estimates are more likely to be wrong. It will also help investment managers identify the companies that have a higher likelihood of being mispriced. Data. All data used in this study come from Thomson 1/13/I:/S and cover nil relevant monthly information for cx'ory Euiopetin company for a 10year period from May 1993 to April 2002. This original dataset was trimmed down with the use of the following criteria:'* • Companies had to have a minimum market capitalization of $1 billion. • Only companies with a December year-end were considered. • A company had to have a minimum of seven analysts, on average, following it over the 10-year period. • For ail forecasts, matching actual earnings data were required. learnings forecasts wore taken 24, 12, 6, and 1 nu)ntli(s) before the corresponding earnings announcement date. Table 1 summari;^es some basic descriptive statistics for the resulting sample (more detailed descriptive statistics by country and sector are provided in Appendix A). Table 1 reveals that, on average, the forecasted earnings are between about 30 percent (at the 12month point) and 40 percent (at the 24-month horizon) above the subsequent realized earnings. This information is a strong first indication of optimism. Table 1 also reveals a somewhat puzzling characteristic of the European market in the high number of analysts that follow a stock and contribute to the 1/B/E/S database.-^ Forecast Accuracy: Naive Approach. A quick initial check on the relevance and accuracy of tho consensus oarnings forocasts is to compare the
March/April 2004

subsequent outcome with the original average, best-case, and worst-case forecasts. To that end. Table 2 summarizes some simple statistics: • The percentage of companies for which tho actual earnings fell within a range defined by the consensus estimate plus or minus 2 times tlio cross-analyst standard deviation of the oarnings forocasts (lR—for "in range"). • The percentage of companies for which the actual earnings fell within a range defined by the high and low estimates (HL). Individual analysts typically provido only point estimates, rather than a range, of expected futiue earnings. Tho HL range (tho rango defined by tho high and low estimates) will thus be narrowor than if analysts wore asked for an expected range and subsequently that range had been subjected to averaging.-^ Tablo 2 clearly indicatos that blind use of tho consonsus number—ovon adjusted up or down to what one could naively consider the best or worst possible outcome—will lead to significant disappointments: Ovorall, fewor than 60 percent (50 percent) of the actual earnings fell within the HL range (the IR 2 standard deviation range) when forocastod 12 months bofore the fiscal yearend. Even 1 month before tho fiscal yoar-ond, up to 25 percent of the actual oarnings fell completely outside the—nai\'oly defined—forecast ranges. Table 2 also reveals remarkable geographical and sectoral differences in accuracy. The low numbers for the United Kingdom are especially striking and may—at first glance—lead one to question the competence of U.K. company analysts (in comparison with analysts of Spanish or Portugueso companies, who consistently had the highest accuracy numbers). At tho sector level, the accuracy of the forecasts appears consistently high in the energy and public utilities sectors and low in the consumer durables and transportation sectors. Tho narrowly defined extreme values as estimated 24 and 12 montlisboforo the relevant vear-ond for tiio
77

Financial Analysts journal Table 2.
Category
Sortctl In/ci'iiiitrf/

European Analyst Forecast Accuracy, 1993-2002
24 Months IR 34"/:1. 38 29 36 41 38 30 41 41 25 III.
41%

12 Months
IK

6 Mlmths IR
52'^ 59 50 59 59 64 59 65 65 55

1 Month

HL
52';;, 62

IIL
65';;. 67 62 66 66 71 81 68 68 65

IR
70';;.

ML

.'\ustria Belgium Finland France Gertiiiiny lt.ily Netherlands Portugal Spain United Kingdom
Sorted bif sector

41';;.
51

78'^

53 38 46 44
51

47 48 53
55

62 54 57
61 74 64 64 47

73 71 78 85 S O 76 79
79 66

85 82 85
85 84 87 88 88

52 52 52
39

54 58
58 42

78

Basic industries Capital ^ooits Consumer durables Con.suniL'r nondiirables Consumer services Tinergy Finance Health care Public utilities Technology Transportation Overall

34';; 31 32 36 34 45 38
40

42'!:.

41 29 46 44 53 49 49 46 38 34
45'!i,

45% 45
40

53';:.

51% 58 57 57 59 63 61 59 72 53 53
59';:.

59';:, 64 58 68 68 68 66 66 76 63 63

70') ;>

80% 86 73 81 84 87 82 88 86 82 77 83'/..

51 43 56 56 61 58 58 64 52 47
58';:.

74 70 77
73 81 74 84 79 73 68

47 48 60 53 51 55 43 47
49';;,

35 23 18 35';:,

66';;.

75';';.

Notr: Studying llio timo p.ittern of these statistics, we could not detect anysignificant trend (i.e., we could not find any improvement or deterioration in analysts' forecasting performance over the 10-year period).

technology sector would lend to large surprises, but the sector converges to "average" accuracy as the time horizon shortens. The dnta in Table 2 arc difficult to interpret. Olsen, who found similarly low percentages in his study of the U.S. market, argued that they are a prime indicator of herding, but these accuracy numbers are affected by a possible bias in the anclioring point (i.e., the consensus number) as well as by the range itself. Because both the mean and the dispersion of the distribution influence the results, we will address the presence of herding as a possible contributing factor and will further separately analyze the bias in the consensus numbers. Herding among European Analysts. Hetding is defined as a narrower dispersion of analyst forecasts than one would rationally expect. The "rationally expected" range of forecasts is not obvious ( priori, however, especially because company7 specific factors, such as earnings visibility atid variability, should have an impact on the forecast.
78

(Most theories of herding imply that herding increases as the opacity of the forecasting environment increases.) Although the rational earnings forecast range for any individual stock is impossible to define, in a purely rational environment, the dispersion of the various analysts' point forecasts should increase as the forecast horizon lengthens. In fact, both the coefficient of variation across all analysts following a particular stock and the HL range (both expressed as a percentage of the consensus forecast) should be higher at the 24-month horizon (Fiscal Year 2) than at the 12-month horizon (FYl). Table 3 summarizes by sectors the ratio of PY2 to FYl for the coefficients of variation (RCF) and the (statidardized)HL range (RHL). liuropean analysts, on average, sliow more consensus the farther out into the future they forecast.*' Only the health care and public utilities analysts, as groups, behave rationally—in that the dispersion of their point estimates widens as the forecast horizon lengthens. Analyst agreement on
©2004. AIMR®

Bias in European A}iah/sts' Earnings Forecasts

Table 3. European Analysts' Earnings Forecast Dispersion: Ratio of FY2 to FY1,1993-2002 Sector RML ucr
IS.Tsic industries C.ipit.il goods CoiiMMiier diir.iLilos Consumer nondiir.ibles Consumer services Energy
Pin.iTice 1.07 l.tll

1.00 1.00 1.01
i).9fy

0.95 0.92 0.9S 0.86 1.10 1.16 1.0! lt.86 t).95

0.90 0.97 0.94 1.13 1.17 1.03 0.73 0.94

Me.illh cnre

Public utilities
TLv!inoloi;y Ir.mspt'rl.ition

Avenge

The forecast bias indicates Ihe persistent direction of differences (i.e., optimism or pessimism) between forecast and realization. The forecast error captures the average magnitude of the difference and can be tliought of as an o\'erall accuracy measure. Table 4 a\'L'rages these mcasuren for the entire sample of I:uropean estimates for various forecast horizons. It confirms the pre\'alence of optimism in the consensus numbers, i-orecasted consensus earnings exceeded subsequent actual nimibers by 39 percent two years before the rele\'ant year-end. Although this optimism decreased as the forecast time horizon shortened, one montli before yearend, it still averaged 3 percent. The (absolute) ftirecast error is obviously large: It avL'raged 76 percent at the two-year horizon, and one month before year-end, it still amounted to 28 percent.

earnings forecasts in the transportation and finance sectors increases as they forecast farther out into the future, l-or most sectors, howe\'er, Table3sho\s's no/?r////(j/(jf/('distinguishable difference in analyst dispersion for FY2 and FYl forecasts. Maybe mosl European analysts focus their attention on the i-Yl period, with the longer-term forecasts, on average, treated as an afterthought. Table 3 indicates the presence of herding among European financial analysts. This herding contributes to the HL range of earnings forecasts being an excessively conservati\'e definition of the possible outcomes, llerding is unlikely to be the sole explanation, however, for the low European analyst accuracy identified previously.

Table 4. Error and Bias in European Analysts' Forecasts, 1993-2002
Horizon 24 Months 12 Months 6 Months 1 Mnnth i'orec.isl Hrriir 7b".'.. 54 37
28

l\.rLV.ist lii.is
39';;.

23 13
3

Bias and Forecast Error among European
Analysts. When e\ akiating the accuracy of the consensus earnings forecast for lun-opean analysts, we used the following statistics: • forecast error (FE)—that is, absolute value of (Forecast earnings-Actual earnings)/Absolute value of actual earnings; • forecast bias (FI5)—that is, (Forecast earnings Actual eamings)/Absolute value of actual earnings. ^

Tlie relatively large forecast bias no doubt contributed to the low forecast accuracy statistics observed previously. Indeed, the percentage of earnings realizatitms "in range" is significantly negatively correlated with the forecast bias (-0.79). We can safely assume, therefore, that both herding and analyst optimism help explain the low forecast accuracy of Kin'opoan analysts. Although establishing the existence of forecast optimism as a possible contributor to forecast inaccm'acies is useful, it can be corrected only if we establish how the bias and error vary among the sectors, countries, and companies. Table 5 gives an initial indication of whether the forecast error and bias differ by country. The forecast error in this period was consistently lower in the United

Table 5. Error and Bias in European Analysts' Forecasts by Country, 1993-2002
24 .Months

l2Monlhs
I-H
Ul

6 .Months

1 Month VV. 25': 36 42 6 16 6 6 7 1 1 1
I-15

Ciiunlty rr.ince German V Iliilv Netlierlonds Sp.iiti United Kingdom

l-H

IU

I-U

n;
36';;. 50 45 31 30
IS

I-B

81'; ;.
83 7fi 60 86 61

46':;.
55 49 27 26 33

28';;.
31 33 16 14 IS

17';..
19 20 8

60 61 48 59 4S

S
9

March/April 2004

79

Financial Analysts journal

Kingdom nnd the Netherlands;^ Germany, France, and Italy showed, on average, very disappointing forecast accuracy. These countries also exhibited the highest degree of optimism, whereas the optimism bias was lowest in the United Kingdom, the Netherlands, and Spain. A similar analysis by sector is provided in Table 6. it shows particularly large forecasting errors for the basic industries, consumer durables, and energy sectors and persistent, large positive forecast bias (optimism) in the technology and consumer durables sectors. Forecast error and bias in Table 6 are consistently tlio lowest in the health cnre and public utilities sectors. Surprisingly, the public utilities sector is tlie only sector where, one month before fiscal year-end, the forecast nntierpretiictCii t!ie actual outcome. The data in Tables 5 and 6 are descriptive and do not disentangle country and sector effects or the impact of other company characteristics that may be associated with country or sector affiliation. Therefore, we carried out a multiple-regression analysis to help establish the relative importance of country and sector in explaining forecast error and bias. For this regression, in addition to the sector and country dummies, we also introduced the following variables that previous studies identified as possible important influences on error and bias: • Analyst dispersion (DISP), measured as the standardized cross-sectional standard deviation in eamings forecasts. We hypolliesized that forecast error and bias increase as analyst disagreement increases. • Thenumberofanalystscontributing to the consensus (AN). Naive intuition suggests that for a gi\'en level of analyst dispersion, the accuracy of the consensus number will increase with the number of contributing analysts.

Volatility [VOE). Large historical earnings volatility should increase the difficulty in forecasting future earnings, possibly increasing forecast errors. Unfortunately, estimating earnings \'olatility is not straightforward because European companies typically report earnings only on an nnntial basis. Tlius, the data points are few, and the result could be high estimation error in the earnings volatility number, which, in itself, may be nonstationary through time. As an alternative to earnings volatility, therefore, we used historical annualized daily stock return volatility during ihe one-year period preceding the earnings forecast. Stock return volatility can serve as a proxy for earnings \'olatility because a large proportion of the stock-specific risk results from the volatility of earnings. • Market capitalization. Although market cap is {weakly} positively correlated with number of analysts following a company, company size would still be expected to be inversely correlated with forecast error and bias because the amount (and quality) of corporate information provided by large companies is typically greater than that provided by small companies.'^ Table 7 stmimari/es the results of the regressions relating forecast error and forecast bias to these four variables and to the country and sector dummies. To avoid multicolinearity among the explanatory variables, we left out tlie country dummy for tlie United Kingdom and for the consumer durables sector. Therefore, the benchmarks for eventual country or sector effects are, respectively, the U.K. market and the consumer durables sector. We report the regression results for tlie 12-nionth and 1-month forecasts. Tlie 12-month number is most rele\'ant for asset managers because of their typical holding periods, whereas the 1-month forecasts provide

•

Table 6. Error and Bias in European Analysts' Forecasts by Sector, 1993-2002
24 Monliis Suctor
Basic iin.iiislrii,'.-<
[•i:

12 M o n t h s 1-1! 106';;, 33';^,

6 .Months FH Vli
!6';i,

1 Mcmth FE 32'K. 18 68 12 33 35 4S 11 8 25 10 FB
3"'!>

I-H
68':;,

147');.

Cnpitiil j;oods ConsiiniL'r

69 164 5i 66 217 77

46 113
4[J

46 113 32 48
131)

28 39 21 21 16

71' :i. 32 71 22 35 68

17 29 15 13 11 12 4 26 15
•Ol

3 7 4 3 4 4 2 -1 5 5

diimbles
C o n s II 111 L-r

nondiirdblus Consumer services tlnergy l-insions

1993-02

1993-97

0.309" -0.002 0.864" 0.000 -0,089 -0.067 0.206 0.212' 0.322" 0.351" 0.246' 0.008 0.227* -0.379 -0.558 -0,63S* -11.505' -0.058 -0.588' -0.763" -0,679* 0,44 -0,717*' 0.144

0.38S*" -0,001 1.294** 0.000 0.005 0.285* 0,880" 0,309" 0.266* 0.265* 0.237* -0.059 0.059 -0,244 -0,3114 -0,4BS* -0.506* -0.153 -0.392 -0.735** -0.571' 0.48 -0,439* 0.205

0,704*' -0,001 1.113" 0.001 0,033 -0,117 -0,054 -0,096 0.349** 0.109 0.430'* 0.015 0.039 -0.325 -0,594* -0,74" -0,632** -(1.2S9 -0.56' -0.788*' -0.536* 0.069 -0.471* 0.09-1

0.05* -0.004 1.411'* -0,001 0,02H 0.026 0.023 0.127* 0.198" 0.255" 0,123" 0.071 0,251" 0.076 0.106 0.047 0.081 0.052 -0,049 0,008 0,013 -0.01)3 0.046 0.074

0.101* -0.003 3.166*" 0.000 0,045 -0,546' 0,213 0.118* 0.172* 0.309" 0.118 0.056 0.409" 0,452 -0,729 -0,069 0,020 t).0()7 0.066 -0.002 0.026 -0,003 -0,011 0,087

0,113* -0.006 0.638* 0.001

0.080 0.352* 0,012 0.015 0.111 0.046 0.061 -0.001 0.032

-0,117 0,.594 0,078 0,037 0.205 -0.096 0.008 0.020 -0.006 0,082 0,041

Company charncteristic DISP AN

VOL
Market cap

0.351-0.016" 1.II7" 0,000 -0.11 -0.09 0.169 0.183" 0.277' 0.283' 0,225' -0.015 0.2 -0.371 -0.53* -1).5'J4" -0.-171' -0.0S6 -0.559' -0.714" -0.648*' 0,46-1 -0,777" 0,183

0.394** -0.006 1.781" 0,000 0.006 0.399' 0,67.5* 0.287* 0.218-

0.621" -0.015'* 1.332" 0.000 0,034 -0.196 -0,03'J -0,077 0.278* 0.082 0.278* -0.029 0,031 -0.328 -0.607** •^).6]5** -0.494*

0.216" -0.009" 0,472** 0,001) 0.018 -0.008 -0.02 0,093* 0.148'* 0,194** 0,097* 0,049 0.227** 0.006 -0.033 -0.082 -0.056 -0.038 -0,044 -0,127 -0,082 -0.07 -0.101 0.092

0.515** -0.006'* 1.246" 0.000 0,036 0.145" -0.2 0.103* 0.133'* 0.253" 0.114* 0.037 0,411"* O.t)35 0.225 0.147 -0.017 -0.006 0.064 l),()40 -0,146 -0,071 0.025 0.105

0.653*' -0.010** 0.237* il.OOO

Country
Austria Uelj;iuin Finland France Germany Italy Netlierlands Portugal Spain Sector Basic industries Capital godds Consumer nondiimhles Consumer services Hnergy Financials Health care Puhlic utilities Teclinology Tmnsporlatioii R~ (adjusted)
0.065 -0,09 -0.012 0.012 0.09* 0.038 0.073 -0.001 0.033

a2120.1810.107 0.051 -0.242 -0.308 -0.479* -0.463* -0.285 -0.373 -0.643'* -0.510" 0.57 -0,567** 0,271

-0.009

-0.52 r
-tl.5O5' -0.621** -0.550" 0.087 -0,634*' 0,153

-0.173 -0.159 •4).i)M -0.129 -0.093 -0.168 -0.122 -0.141 -0,150 0,432

*Sigiiificant at the 10 percent level. "'Significant at the 5 percent level.

March/April 2004

81

Financial Anali/sts Journal

insight into the nature of the earnings surprise phenomenon. We report the regression results for the full 10-year period nnd for two 5-ycar subperiods. This division allowsevaiuation of possible intertemporal differences and a check of the robustness of our results. Because the forecasts were positi\^ely biased in most cases, forecast error and forecast bias are highly correlated. Thus, the forecast bias and the forecast error regressions yielded broadly similar results. The discussion concentrates on the (more interesting) forecast bias regressions. At every le\'el of forecast horizon, forecast optimism increases significantly with analyst dispersion and stock volatility whereas the number of analysts and the market cap of the company have no influence on the level of optimism.'2 All else being equal, stocks with high analyst dispersion and high historical volatility are more likely to have an upward bias in their earnings forecasts. For the entire 10-year history, the average forecast bias in the core "IZuroIand" countries (France, Germany, Italy, Spain, and the Netherlands) significantly exceeded that for the typical U.K. company. This observation holds at both the 12-month and 1-month forecast hori:^ons, but it seems to have been restricted almost exclusively to the first 5-year subperiod. One could argue, therefore, that the quality of forecasts leveled out across Europe in the more recent past, which may help explain the secular decline of country effects in stock returns that has been widely documented (Cavaglia et al. 2000). Country effects appear to have been ephemeral, but significant sector influences persisted through time at the 12-month horizon. The consumer nondurables, consumer services, health care, public utilities, and transportation sectors had significantly fewer optimistically biased forecasts than the other sectors.'-' These sector effects disappeared completely at the 1-month horizon. By far the consistently dominant influences on forecast error and forecast bias in Huropeareanalyst dispersion and stock return volatility. Secondary are certain strong sector effects, witli the 12-nionth forecast bias being consistently lower in the consumer cyclical, health care, and transportation sectors than in the indtistrial, energy, and technology sectors. Finally, we found evidence of Huropean integration, in that the country effects in forecast error and bias that were present in the mid-1 y90s have now virtually disappeared.

Conclusion
In a market where consensus earnings forecasts always accurately reflect future realizations, opportunities for active portfolio management would be severely limited. In fact, buy-sJde analysts already try to distinguish themselves, at least in part, by
82

improving upon the consensus forecast on which market prices are largely based. Asset managers need to be selective in the number and types of companies they have in their inx'estable universes, however, and would like to focus their efforts on the companies that are most likely to be mispriced. We hope to have contributed to the identification of the characteristics that are most likely to be associated with optimistic (positively biased) consensus earnings forecasts. Our study of European consensus earnings torecasts reveals significant, persistent, and systematic differences in accuracy. Broadly, we can conclude the following: • Forecast error and forecast optimism increase with dispersion in analyst forecasts and with (historical) stock return volatility. An increase in the number of analysts, all else being equal, decreases the forecast error (but has no impact on forecast optimism). Market cap has no significant impact on a company's consensus earnings forecast accuracy. • In the past, significant geographical differences existed in earnings forecast accuracy. In particular, analysts focusing on tiie core Huroland countries of France, Germany, and Italy showed significantly poorer accuracy in earnings forecasts than ciid analysts focusing on other European countries in the mid-1990s. These geographical differences have now broadly disappeared, and earnings forecast error no Ioiiger reflects any significant country effects. • Earnings forecast accuracy exhibits significant sector effects: On average, the forecasts for the consLimernondurables, healthcare, public utilities, and transportation sectors were significantly more correct in the period studied than were those for the other industries. These sector effects completely disappeared, however, as the forecast horizon shortened. • One month before the earnings announcement date, only analyst dispersion and stock price volatility remained significantly positively related to forecast bias. Although some of these insights have been documented elsewhere, this study is—to the best of our knowledge—the first that simultaneously tested the impact of country effects, sector influences, and otiier fundamental company characteristics on earnings forecasts. We hope that our findings raise awareness of the nature and magnitude of systematic biases and errors that affect consensus numbers for certain types of companies. The findings may even help buy-side analysts identify the companies that are more likely to be misvalued.
We would like to thank Michael Bilhnann, who provided excellent research assistance for parts of this study. ©2004, AIMR®

Bias in European Analysts' Earnings Forecasts

Appendix A. Descriptive Statistics of Company Sample by Country and Sector
Ciimp anies
12 24

Number of Observations
12 24 12

Analysts
24

Country/Sec lor
Countn/

.Months 10 33
34 146 98

Months 9 31 30
133 83 70

Months 46 147
149 701 376 337 260

Mcinths 29 114 96 511 264 240 216 46 223
631

Months 11 14 15 18

Mtintlis 7 9 14 14 22 10 20

Austria Uelgiiim

I-inland
Franco Germany Italy .Net her I.I lids Portugal Spain United Kingdom Salor Basic industries Capital goods Consumer durables Consumer nondiirables Consumer services

25
19 28 12 23 14

75 52 14
64 161

50 14 58 152

54
267 875

9 15
9

69 118 11 45 113 27 164

63
109 10 45 105 26 155

351 636 60 241 528 136

262 454 41 !S5 408 98

19 17 30 20
18

15 13
22 14 13 14 12

linergy i-inance Health care Public utilities Technology Transportation

752 154 174 120
60

559 114 125
80 44

20 18 20 21
13 12

35 51 43
11

31 42 36
10

15 16 10
9

This table should be read as follows: We included a total of, for instance, 146 French companies for which, at some point in the 10-year period, a 12-month forecast (and subsequent realization) was available. In total, 701 different 12-month earnings forecasts for French companies were included in the study. On overage, 18 analysts contributed to each 12-month consensus earnings forecast for a French company. These analysts constituted the entire universe of contributors of French company earningsforecasts to the 1/B/E/S database (and would represent a wide dispersion of analyst nationalities—see Note 4).

Notes
Tliese conclusions arc to some decree contr.idlcted by Lee, Uoscnthal, and Gleason (fortlicomiiij; 201.14), who could not detect iiny significani impact ot' Regiilatiim FD on stuck price\'olalility orcm tlie.idverseselection component of the bid-.isk spread. These effects would be niitij;.ited for large companies that are cross-lisled nn multiple exclianj;es and, therefore, have .idopted U.S, generally accepted accounting principles. The iUiropean a)iintries are Auslria, lk'lj;iuni, i^inland, |-rance, Ciermatn', Italy, the Netherlands, Portugal, Sp.iin, and llie United Kingdom. To facilitate comparisons amonj; these countries, .ill monetary .imoiints were converted lo U.S. dollars. A comparison in September 2002 of the a\ erago number of analysts contributing lo ihe l/Ii/I'VS consensus earnings niiniber.'iof the typical index constituent stuck provided the followinj; results: Hurnstoxx Index, 21 analysis; U.K, Sloxx Index, 13 analysts; S&P 500 Index, 15 analysts; Nikkei 225 Index, I3analysts,ThesedaUi illustrate Ihe remarkably large number of analysts, iin average, cuntributing to theconsensuse.irnings for the typical Continental European company, l-or instance, Philips Electronics has 42 analysis who contribute an earnings forecast to I/B/1-/S. Of these 42 conlributors, 10 are German, 12 Diilcii, 6 Frencli, 1 Spanish, 1 British, and 9 "global" brokers. This large number of contributors is obviou.sly a legacy offormer purely domestically oriented institutions now feeling obliged to add the btg luiropean companies to their followed list. We Ih.ink Craig Ruff for pointing this aspect out.

3.

4.

March/April 2004

83

Financial Analysts journal
6. A comparison vvilh U.S. HY2 and FYl forecasts for the S&P 500 constituents .IS of September 20tl2 reveals RCF and KHL statistics Ihat average, respectively, 1.75 and 1.77, with not a single sector showing a ratio below 1,2. Tliis result implies that herding may hu more pronounced in tlurope than in the United Stales, To avoid having the results of our analyses dominated by a relatively small number of extreme observations (companies with a low absolute wilue of actual earnings), we trimmed the samples by eliminating all tibser\'.itions for wliich the forecast error or forecast bi.is was more than 3 .sample (cross-sectional) standard deviations from the sample mean. The result in each cast? was .T reduction of about 7 percent in the sample. This finding confirms the results from Capsl.iffetal. (2001), whn also found that the United Kingdom .ind the Netlierlands had lower-tliaii-a\'erage forecast errors. Across al! regions, there was in our study a correlation of 0,4 on average between the number of analysts following .i company and its market cap.'Ilie tendenc}'for more analysts to follow bigger companies is somewhat bemusing because such companies offer few opportunities to add value through analysis. One way for .inalysts to differentiate themselves would be by researching neglected companies and industries, yet the evidence suggests that analysts prefer to work in areas where their potential for adding value is low. 10. [(eterosced.isticity-consistent standard errors were used in estimating all regressions. Also, the inclusion of country and sector dummy wiriables eliminated the possible problem of cross-sectional fixed effects inducing heterosced.isticity in the residuals of our regression output. 11. We also ran the regressions for the 24- and 6-monlh time horizons, bul the results were not materially different from those presented for the 12-and 1-month periods and are not presented here. Those results a re available from the authors upon request. 12. Interestingly, however, an increase in the number of analysts does significantly reduce the forecast error. 13. The small numberofoptimisticalK'biased forecasts in those sectors partially confirms the preliminary insights of Capstaff et al. (1999, 2001), who documented similar sector effects in forecast biases.

7.

8.

9.

References
Agrawal, A., and S. Chadha. 2002, "Who Is Afraid of Reg FD? The Behavior and Performance of Sell-Side Analysts Following the SliC's Fair Disclosure Rules." Working paper, Culverhouse College of Commerce and Business Administration, University of Alabama, AIMK. 2001. "Fair but Not Full Disclosure: Fver\'one Flas Less Information under Regulation FD, Analysts, Portfolio M.inagers Say in AIMR Survey." AIMR pres.s release (18 October): www,ainir,org/pressroom/01 releases/01 RegFD.html. Butler, K.C, and 11. Saraoglu. 1999. "Improving Analysts' Negative Karnings Forecasts," I'inaiicinl Aunlpts Imtrnal, vol. 55, no. 3 (May/June);48-56. Capstaff, John, Krishna Paudyal, and William Rees. 1995. "The Accuracy and Rationality of Earnings Forecasts by UK Analysts." joiirttii! of Bu$incfs Finnncc and Accountiu}:,, voi, 22, no. 1 0nnuary):67-85, . 1998. "Analysts' Forecasts of German Firms' Earnings: A Comparative Analysis." journal of liitcrnntionul Financial Miwo^cniciit ami Aa-oiinling, vol. 9, no. 2 yunL'):K3-l]6, . 1999. "llie Relative Forecast Accuracy of UK Brokers." Accounting imd Bushwss Research, vo\. 30, no. 1 (VVinter):3-16, De Bondt, W.F.M,, and R.H. Thaler. 1990, "Do Security Analysts O\'erreact?" Aniaicnn Econoiitic Rn'ifw Papers and Procei-dings, vol.80, no. 2(May):52-57. Diether, K.B., C), Malloy, and A. Scherbina, 2002. "Differences of Opinion and the Cross-Section of Slock Returns." journal of flntmcc. voL 57, no. 5 (October):2113-11. Dowen, R.j. 1996. "Analyst Reaction to Negative Earnings for L,irge Weil-Known Firms." \ournal of Portfolio Maiuigemcnl, vol. 23, no, 1 (Fall):49-55. Dreman, D.N., and M.A. Berry. 1995. "Analyst Forecasting F-rrors and Their Implications for Security Analysis." I'inancinI Analysis loiiriuil, vol. 51, no. 3 (May/June):3lMl, Dugar, A., and S, Nathan. 1995. "The Effects of ln\'estment Banking Relationships on Financial Analysts' Earnings Forecasts and hnestment Recommendations," Conteiiiporur\/ Acamnliii^i Kcstvird;, vol. 12, no. 1 (Fall):131-160. Easterwood, J.C, and S.R, Nutt. 1999. "Inefficiency in Analysts' Earnings Forecasts: Systematic Misreaction or Systematic Optimism?" journal of Finance, vol. 56, no. 5 (October):1777-97. Francis, J., and D, Philbrick. 1993. "Analysts' Decisions as Products of a Multi-Task Environment." lournn! of Acautnling Ri'M'tireh, vol. 31, no. 2 (Autumn):143-163.

. 2(101, "A Comparative Analysis of Harnings Forecasts in liurope," Journal of Business Fiiuuuvand Account In;^, vol. 28, no. 5 Graham, J.R. 1999. "Merding among ln\'estment Newsletters: Theory and Evidence." journal of ['iiutncc, vol, 54, no. 1 yuly):531-562, (February): 237-268, Cavagtia, S., C. Brightman, and M, Aked, 2000. "The Increasing 1-Iiggins, H.N. 1998. "Analyst Forecasting Performance in Seven Importance of Industry Factors." Financial Annh/sts \ourntil, Countries," Financial Aiiah/st^ Journal, vol. 54, no, 4 (Julv/ vol. 56, no. 5 (SL'ptL'mber/October):38-51. ALigust):38-62. Crichfield, T., T. Dyckman, and ]. Lakonishok. 1978. "An l-v.iluation of Security Analysts' I'orecasts." Acaninting Rfvicw, vol. 53, no. 3 (|uly):651-668. Das, Somnath, Carolyn B. Levine, and K. Sivaramnkrislinan. 1998. "Famings Predictability and Bias in .Analysts' Forecasts." Accounting Review, \'ol. 73, nn, 2 (April):277-294. De Bondt, Werner, and William P. Forbes. 1999. "Herding in Analyst Karnings Forecasts: Evidence from tlie United Kingdom." Zinro/Ji'dii Financial M