The persistency of the earnings game mechanism in analyst earnings prediction accuracy and consistency through different periods of economic conjuncture

The Persistency of the Earnings Game

Mechanism in Analyst Earnings Prediction

Accuracy and Consistency through Different

Periods of Economic Conjuncture

N. Bleeker

B

ACHELOR

T

HESIS TO OBTAIN THE DEGREE IN

E

CONOMICS AND

B

USINESS

,

BS

C

ECB

A

T THE

U

NIVERSITY OF

A

MSTERDAM

(U

V

A)

F

ACULTY OF

E

CONOMICS AND

B

USINESS

A

MSTERDAM

B

USINESS

S

CHOOL

A

UTHOR

N

OORTJE

B

LEEKER

S

TUDENT NUMBER 11064439

E

MAIL

NOORTJE

.

BLEEKER

@

STUDENT

.

UVA

.

NL

D

ATE

26

TH OF

J

UNE

,

2018

S

TATEMENT OF

O

RIGINALITY

This document is written by Noortje Bleeker, who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

A

CKNOWLEDGEMENTS

During the process of writing my Bachelor Thesis I have received support and advice from many people. This has aided to the construction of my thesis to graduate my Bachelor degree with from the University of Amsterdam in Economics and Business, specialized in Finance and Organization. Therefore, I would like to take this opportunity to express my gratitude to them.

Firstly, I would like to thank my supervisor, Dhr. Pascal Golec MSc, for his consistent support for my subject. I would like to thank him for the regular meetings we had, in which he would provide me with helpful academic support. After learning from him in earlier Finance courses, I am happy I had the chance to receive his help once more while writing my Bachelor Thesis.

I would also like to thank my close friends and family, whom I cannot all name individually, for their support, advice and thoughts before and whilst writing my thesis. As many will experience, the writing process of a thesis can be a laborious task with many ups and downs, where the support of people close to you will be an important factor for success.

A

BSTRACT

Analyst earnings prediction accuracy appears to be affected by the state of the economy. This research will attempt to find proof for this economic conjuncture effect, arguing from the incentive and relational theory originating from the earnings game mechanism. In this paper, I use an additional measure of analyst precision compared to most existing literature’s use of analyst accuracy in testing the persistence of this effect, namely analyst consistency. This proposition is tested by establishing four empirical models of average and individual analyst data, testing for the economic conjuncture effect on forecast accuracy and consistency. Consistent with existing literature, support for the economic conjuncture effect is found, for both measures of analyst precision. Moreover, this result supports both the incentive and relational theorems, following the theory on the earnings game mechanism.

Keywords: Analyst Earnings Predictions, Accuracy, Consistency, Economic Conjuncture,

Earnings Game.

T

ABLE OF

C

ONTENTS

ACKNOWLEDGEMENTS

3

ABSTRACT

4

1

I

NTRODUCTION

6

2

RELATED LITERATURE

9

2.1

THE EARNINGS GAME INCENTIVE THEOREM

10

2.2

THE EARNINGS GAME RELATIONAL THEOREM

11

2.3

R

ESEARCH

I

MPLICATIONS

12

3

HYPOTHESES

14

4

DATA DESCRIPTION AND RESEARCH METHODOLOGY

16

4.1

DATA DESCRIPTION

16

4.2

RESEARCH MODEL

17

4.2.1

A

NALYST

A

CCURACY USING

A

VERAGE

A

NALYST

D

ATA

19

4.2.2

ANALYST CONSISTENCY USING AVERAGE ANALYST DATA

20

4.2.3

ANALYST ACCURACY USING INDIVIDUAL ANALYST DATA

20

4.2.4

ANALYST CONSISTENCY USING INDIVIDUAL ANALYST DATA

22

5

RESULTS

23

5.1

ANALYST FORECAST ACCURACY UNDER THE INCENTIVE THEOREM

24

5.2

ANALYST FORECAST BIAS UNDER THE RELATIONAL THEOREM

30

6

DISCUSSION AND CONCLUSION

34

7

REFERENCES

36

1

I

NTRODUCTION

Analyst earnings predictions remain in great demand by (institutional) investors and researchers for information on analyst reputation (Jackson, 2005), analyst incentives (Richardson, Teoh, and Wysocki, 2004), their ability to forecast prices and forecast accuracy (Hilary and Hsu, 2013). This factor of prediction accuracy itself has also gained a lot of attention with researchers since it has been proven to decrease in periods of economic downturn (Loh and Stultz, 2018; Das, Levine and Sivaramakrishnan, 1998; Arand and Kerl, 2012; Hope and Kang, 2005). This research will attempt to find similar results for the here fore mentioned economic conjuncture effect, but expand the informativeness of the accuracy measure by adding a dimension of precision; namely, analyst prediction consistency. This is used to eliminate the potential effect of merely a change of bias. Moreover, in combination with the conjuncture effect, this research will examine the informativeness of analyst earnings predictions making a new connection by using a theorem that is often linked to analyst earnings predictions; the earnings game mechanism.

The link between analyst earnings forecasts and the earnings game mechanism is twofold, but, to my best knowledge, has not been researched yet. The theory of the earnings game itself claims a continuous communication and coordination act between analysts and firms (Bradshaw, 2004; Richardson et al., 2004). This link is however challenged when put into the light of different periods of economic conjuncture. Firstly, recessions often prove to be detrimental to firm performance (Loh and Stultz, 2018; Lemmon and Lins, 2003), causing a drop in share prices. This drop in performance leads firm to be less persistent in their ability to consistently beat analyst earnings predictions (Conrad, Cornell and Lansman, 2002; Cohen and Zarowin, 2007). Moreover, in bad economic times, volatile firm performance and market circumstances make it more difficult for analysts to gather qualitative information towards generating accurate earnings predictions. Building on this proposition, Hayes (1998) finds that analyst remuneration policies are often based on variable incentives (see also Bradshaw, 2004). He claims that analysts often select their prospects and effort levels based on their expected profits. Following this line of argument, these variable incentives will affect analyst motivation in gathering (correct) information that will affect the quality and quantity of information reported. Meaning, in bad economic times, gathering this

information is more difficult and costly due to volatile firm performance and lower incentives, respectively, presumably resulting in less accurate forecasts, posing as the incentive theorem within the earnings game mechanism. Secondly, partially depending on analyst’ incentives, the communication and coordination act between analysts and firm management poses as the base of the relational theorem within the earnings game. The connection with firm management is an important source of private information for analysts to create accurate earnings forecasts with. Furthermore, positive forecasts are rewarded by the firm by providing the analyst with exclusive private information and shorter communication lines with management. Das et al. (1998) expand on this argument by asserting that analysts may engage in deliberate optimism (positively biased earnings forecasts) to obtain private information from managers of firms in times when public information is insufficient to accurately interpret firm performance. This might make analysts (deliberately) more inaccurate during periods of economic downturns in an attempt to obtain private information from firm management.

Note that this research will expand the general analysis of forecast precision. More specifically, this research will build upon the definition of accuracy adopted by most researchers, by introducing a more informative measure; forecast consistency. Using the definitions as proposed by Hilary and Hsu (2013). The concept of accuracy is determined as the absolute difference between realized earnings and analyst earnings forecasts. Analyst accuracy alone is however not informative enough to measure the actual change in forecast precision, as this variable does not dissect the effect of forecast difficulty (in economic bad times), and merely a change in analyst forecast bias. The measure of analyst consistency uses a different mathematical construction, as the variable is defined as the standard deviation of absolute forecast errors (accuracy), essentially eliminating the bias. This construction allows for the separation of the bias when comparing the two measures, presenting a more meaningful and informative analysis of the varying precision of analyst earnings predictions in economic bad times.

More formally, the aim of this research is to find out why analyst earnings prediction accuracy and consistency changes, as affected by the earnings game mechanisms, testing for different periods of economic conjuncture. Essentially, testing for the persistence of the incentive and relational theorems of the earnings game mechanism in analyst earnings predictions to aid investors in gaining factual, accurate knowledge of earnings forecasts, and give them tools to eliminate (systematic) biases.

The empirical study will test the hypotheses by adopting an OLS regression to examine the economic conjuncture effect separately for both the measures of accuracy and consistency. Both measures will be created following the variable construction inspired by Loh and Stultz (2018) and Hilary and Hsu (2013), additionally controlling for several firm- and analyst specific variables. All data will be retrieved from the WRDS Thomson Reuters I/B/E/S database, with data on detail earnings forecasts from 1970-2018, all from US analysts, combined with firm specific (accounting) data from the WRDS Compustat database. Moreover, following common literature, the definition of a recession is taken from the National Bureau of Economic Research (NBER).

Consistent with the main predictions, the results show that analyst accuracy and consistency decrease in economic downturns, replicating the results from the Loh and Stultz (2018) paper, and supporting the incentive theorem. Moreover, support is also found for the second hypothesis, where it is argued that a deliberate bias, arising from the coordination assumption of the relational theorem, also increases in bad economic times. The construction of both measures allows to compare regression results to identify a bias, for a more informative evaluation of analyst forecast precision and the economic conjuncture effect.

The remainder of this research paper will be structured as follows. The following chapter will outline the relevant theories and existing literature used for the design of this research. Following the theory, the hypotheses are presented in chapter 3. Subsequently, the research methodology will be presented in chapter 4. Here, a description of the data and the statistical methods are presented which will be used to find an answer to the proposed research question, at the hand of the hypotheses. The results of this regression analysis are presented in chapter 5. Lastly, the conclusion and further discussion are presented in chapter 6 to provide a conclusive and forward looking answer to the research question.

2

R

ELATED

L

ITERATURE

Existing literature offers many perspectives on different aspects of analyst earnings predictions and their effects. Bradshaw (2004) examines the link between analyst earnings predictions and their connection with valuation models. Richardson et al. (2004) analyze whether analyst forecasts in light of the earnings game are linked to managerial pay incentives, whereas Jackson (2005) introduces the link between analyst earnings predictions and their personal and brokerage firm reputation. Moreover, as supported by Brown, Call, Clement and Sharp (2015), this reputation can also be linked to their connection with firm management (see also Ke and Yu, 2006) and analyst’ personal incentives (Hayes, 1998). Another avenue of earnings predictions impact is researched by Hilary and Hsu (2013), where they test analyst precision for their impact on prices. They go beyond the scope of analyst precision generally applied in existing research, by expanding on the analysis on analyst accuracy with the concept of analyst consistency. The concept of accuracy is determined as the absolute difference between analyst earnings forecasts and realized earnings. However, consistency is considered more relevant for analyzing forecast usefulness, as it should lead to greater forecast informativeness. Consistency is explained using an example of two investors.

“Consider the forecasts of two analysts. Analyst A delivers forecasts that are consistently three cents below realized earnings, whereas Analyst B delivers forecasts that are two cents above realized earnings half of the time, and two cents below realized earnings the other half of the time. Investors should prefer the [more consistent] forecasts of Analyst A.” (Hilary and Hsu, 2013, p. 272).

The above mentioned existing research essentially proposes various empirical evaluations of different types of forecast informativeness, and the impact they can have on investors. These two factors of forecast informativeness and impact have also received attention from many researchers by testing for a time-related effect. More specifically, analyzing analyst precision by testing for different periods of economic conjuncture. Among others, Loh and Stultz (2018) (see also Arand and Kerl, 2012; Hope and Kang, 2005) empirically tested analyst accuracy through economic good and bad times, and found that analyst accuracy indeed decreases in periods of economic contraction. They developed an empirical model using average analyst data, and argued the decrease in

analyst accuracy to be partly dependent on the type or severity of an economic recession, but stated it was mainly dependent on the increased market uncertainty in economic bad times, resulting in more difficult prospect analyses for analysts.

This research will attempt to empirically replicate the research results of Loh and Stultz (2018), but expand the scope of their research by adding an informativeness factor of analyst earnings prediction precision as proposed by Hilary and Hsu (2013), namely, analyst consistency, to test why analysts are more imprecise during recessions. Moreover, this research focuses on testing a different cause for the economic conjuncture effect, namely, the earnings game mechanism, and the two underlying core theorems, creating a new link that, to my best knowledge, has not been made in existing literature yet.

Among others, Du and Shelley (2014) and Eccles, Herz, Keegan and Phillips (2001) explain that analyst earnings predictions can have a controlling effect on management and company strategy to steer firm performance towards. The market increases the pressure on firms to continuously meet or beat analyst predictions to send out a positive performance signal. Several business implications of this external investor perspective are criticized (Boot, 2009), however, it should be noted that this mechanism brings complications for the investor pool as well. Richardson et al. (2004) describe the earnings game as the process where analysts initially make optimistic earnings forecasts, to later ‘walk-down’ those estimates to a level set by the firm’s management that is in their range to meet or beat at the actual earningsannouncement. This anomaly places a question mark on the reliability of these forecasts and introduces concerns for a potential bias in analyst earnings predictions.

This research further develops two main theorems at the core of the earnings game mechanism; the incentive and relational theorem.

2.1 T

HE

E

ARNINGS

G

AME

I

NCENTIVE

T

HEOREM

The earnings game incentive theorem is mainly based on the construction of analyst incentive policies, and their effect on analyst performance. Hayes (1998) (see also Bradshaw, 2004; Shon and Young, 2011; Groysberg, Healy and Mayber, 2008; Alford

and Berger, 1999) finds that analysts are often compensated on a variable basis, taking the example of commission based remuneration policy. This means analysts are paid on their stock sales, which are related to (the accuracy of) their predictions and recommendations. Following this construction, Hayes (1998) finds that trading commission incentives impact analysts in such a way that they show different performance patterns than when working under a fixed compensation scheme. This behavior manifests itself in the way that it might affect analyst behavior in choosing their effort level in gathering the right information, or whether to follow a firm at all, based on the expected profits from following the firm. More specifically, these results suggest that the expected commission (variable) incentives affect analyst motivation towards gathering qualitative and quantitative information to establish accurate earnings predictions.

This monetary incentivized motivation is put under pressure in bad economic times. Increased uncertain and more volatile market circumstances in general make it more difficult for analysts to perform their jobs as they would in economic good times. In addition, Loh and Stultz (2018) argue that firms generally prove to be less profitable in times of economic recessions. This further decline in trading volume during recessions (Chordia, Roll and Subrahmanyam, 2001) and hence decreasing broker profits, impacts analysts’ variable performance rewards in a negative way. This leads to a decrease in analyst motivation to collect qualitative information, arguably resulting in less accurate analyst earnings forecasts in economic bad times.

2.2 T

HE

E

ARNINGS

G

AME

R

ELATIONAL

T

HEOREM

The relational theorem emphasizes the relationship between the analyst and firm management, claiming a continuous communication and coordination act between the two parties (Cotter, Tuna and Wysocki, 2006; Fuller and Jensen, 2010; Bradshaw, 2004; Richardson et al., 2004). This process shows to be persistent since it benefits both parties. The direct relation with analysts benefits the firm in the sense that they can influence the analyst in his earnings forecast to ensure beatable earnings predictions to manage positive performance signals. The firm, in turn, rewards the analyst by providing the analyst with exclusive private information and shorter communication lines with

management. In this way, the analyst has a cheap, reliable information source to ensure accurate earnings predictions, resulting in higher performance related incentives. Additionally, analyst remuneration policies are often built to incentivize positive earnings forecasts. This again rests on the fact that firms are more likely to work with analyst companies that provide positive forecasts (Bradshaw, 2004). Moreover, non-monetary related incentives have to be considered from the analyst perspective, since Hong and Kubik (2003) find that analysts that predict positive earnings for companies are more likely to receive higher professional recognition and better career outcomes.

Following a similar argument as presented for the incentive theorem, in economic bad times publicly available information is often not of sufficient quality or quantity for analysts to generate accurate earnings forecasts, and firm performance proves more difficult to interpret. This creates a greater need for analysts to obtain more accurate private firm information in volatile, uncertain economic recessions. Das et al. (1998) use the relational assumption to expand on this argument by asserting that analysts may engage in deliberate optimism to obtain private information from firm management when public information is of insufficient quality to accurately assess firm performance, and firm earnings follow a less predictable time-series pattern. Francis and Philbrick (1993), in turn, find that analysts issue more optimistic forecasts, and interpret this as suggesting that forecast optimism is greater when analysts see the need for improving relations with firm management to obtain private information. In summary, the proposed relational theorem assumes economic recessions might make analysts (deliberately) more inconsistent during periods of economic downturns in an attempt to obtain private information from firm management.

2.3 R

ESEARCH

I

MPLICATIONS

Following the theory outlined above, the incentive and relational theorem are essentially two competing mechanisms for the explanation of the economic conjuncture effect. Firstly, the incentive theorem assumes the volatile, uncertain economic conditions to make it harder, and, more importantly, less profitable for analysts to generate accurate earnings forecasts. Secondly, the relational theorem uses the assumption of the communication and coordination act between analysts and firm management as the

main explanatory factor for a (deliberate) analyst bias during economic recessions. More specifically, following the proposed earnings game theorems, this research uses two measures of analyst imprecision that are assumed to change during periods of economic recession. Namely, analyst accuracy and consistency. The construction of analyst accuracy following common literature is however not a sophisticated enough measure on its own to test why analyst accuracy changes in periods of economic contraction, since it only tests for analyst imprecision in general. The statistical construction does not distinguish between pure analyst accuracy (as affected by forecast difficulty and lack of incentive in economic bad times) and a (change in) bias, as affected by the economic conjuncture effect. To provide a more informative analysis of the economic conjuncture effect on analyst imprecision, the measure of analyst consistency is added to the analysis as it proves to be a more informative measure since it eliminates analyst bias from the analysis by taking the standard deviation of absolute forecast errors of analyst i for firm j over a certain year t. The distinct (statistical) difference between the two measures of analyst precision is further elaborated on in section 4. Subsequently, separating the two measures over different models will allow to compare the results in order to find the pure economic conjuncture effect on analyst accuracy. Moreover, this separation will help interpret the regression results by analyzing the difference between the two coefficients of the economic conjuncture in the sense that it represents the size of the (positive) bias in times of economic recessions. More formally, if the empirical results for the regression analysis with analyst consistency show a significant but different relation with the economic conjuncture effect (in addition to finding significant results with analyst accuracy), this would mean that not all analyst imprecision is explained by analyst inaccuracy (incentive theorem), but that it might also be affected by the predicted deliberate analyst bias, finding proof for the relational theorem.

This separated construction is relevant since I believe only testing for analyst imprecision in general does not give conclusive evidence of the actual nature of the complete economic conjuncture effect. Moreover, this analysis will perform multiple regression analyses using both the individual as well as the average empirical analyst models, as proposed by Hilary and Hsu (2013) and Loh and Stultz (2018), respectively, with both measures of analyst precision.

3

H

YPOTHESES

This chapter will be dedicated to reformulating the literary assumptions into concrete hypotheses that the analysis of this research will attempt to find proof for. The hypotheses will be presented using the argumentation of the two core theorems of the earnings game mechanism, in combination with the economic conjuncture effect. Firstly, the incentive theorem that assumes tougher economic conditions in combination with deterring incentive effects to have a detrimental effect on analyst accuracy. Secondly, the proposed relational theorem that assumes a deliberate bias induced by relational desires for private firm information. The main explanatory variable that is researched is the economic conjuncture effect, which will be used to statistically formulate the hypotheses. Both hypotheses are researched using an average analyst model, controlling for additional firm specific variables, and an individual analyst model, where analyst specific controls are used. The hypotheses are based on multiple articles on the earnings game mechanism (Bradshaw, 2004; Richardson et al., 2004), and the economic conjuncture effect on general analyst accuracy (Loh and Stultz, 2018; Arand and Kerl, 2012; Hope and Kang, 2005), but specifically focuses on the incentive and relational theorems outlined by Loh and Stultz (2018) and Das et al. (1998), respectively. More formally:

H1: The informativeness factors Accuracy and Consistency are negatively related to the

economic conjuncture effect, following the Earnings Game Incentive Theorem.

Hypothesis 1 uses the proposed incentive theorem in expecting the detrimental effect of the crisis to negatively affect analyst commission incentives. Which makes analysts less (financially) motivated towards gathering correct and qualitative information to make accurate earnings forecasts. This decreased motivation would arguably lead to lower forecast accuracy.

H2: The informativeness factors Accuracy and Consistency have a different relation with

the economic conjuncture effect, following the Earnings Game Relational Theorem.

Hypothesis 2 uses the relational theorem, additionally following the proposition that in economic downturns, the publicly available information is of insufficient quality or quantity to establish accurate earnings predictions. To which the analysts act by

positively biasing their forecasts in an attempt to gain a stronger bond with firm management to obtain more qualitative private information.

4

D

ATA

D

ESCRIPTION AND

R

ESEARCH

M

ETHODOLOGY

In this section, the research methodology used to find results for the proposed hypotheses is elaborated upon. This research will develop several models for testing the economic conjuncture effect on analyst earnings predictions, separating the regressions for both measures of forecast informativeness; accuracy and consistency. Analyzing these measures separately helps to identify the pure economic conjuncture effect on analyst accuracy, and the potential systematic (positive) bias. Moreover, the analysis will be expanded further by looking at both individual analyst as well as average analyst data from US analysts.

4.1 D

ATA

D

ESCRIPTION

The analyses are conducted using data retrieved from multiple sources. Consistent with existing literature (Hilary and Hsu, 2013; Loh and Stultz, 2018; Richardson, Teoh and Wysocki, 2004 etc.), details data on analyst specific measures is obtained from the Thomson Reuters I/B/E/S database, retrieved from Wharton Research Data Services (WRDS). The unadjusted details file is used to mitigate the rounding effects problem in I/B/E/S, as pointed out by Diether, Malloy and Scherbina (2002). Note that quarterly earnings and forecast data is selected to follow the frequency of common earnings announcements, and to aid comparison with existing literature. In this dataset, observations sorted per analyst, firm, year, and quarter are included, containing further information on individual analyst forecasts, actual earnings numbers and publication dates. Observations for which there is no stated actual earnings numbers or analyst forecasts are omitted from the dataset. Analyst predictions for a certain quarter are not limited to a predetermined timeframe, since this factor will be controlled for in the model by the variable Horizon, which, similar to the other analyst specific control variables, follows the definition of Hilary and Hsu (2013). The construction of these variables will be elaborated upon in the succeeding section. In a similar fashion, details data from the I/B/E/S database is retrieved for average analyst forecasts, sorted per firm, year, and quarter. Note that the actuals and forecasted earnings numbers are downloaded in the Earnings per Share (EPS) format. This variable can induce some rather extreme

differences in (absolute) forecast errors, since the magnitude of earnings is partially dependent on firm size. To avoid a size-induced precision bias from the model itself, the EPS measures are scaled using total firm assets. This firm specific variable, along with data on firm age and their market to book ratio is downloaded from the WRDS Compustat database. Both datasets are downloaded using observations ranging from January 1970 through the first quarter of 2018. This broad time range aids the development of an extensive analysis of the economic conjuncture effect over time, including many different economic situations and analyst reactions.

Following the research methodology of Loh and Stultz (2018), this research constructs the main explanatory variable, the economic conjuncture effect, using the definition of a recession by the National Bureau of Economic Research (NBER).

“The NBER does not define a recession in terms of two consecutive quarters of decline in real GDP. Rather, a recession is a significant decline in economic activity spread across the economy, lasting more than a few months, normally visible in real GDP, real income, employment, industrial production, and wholesale-retail sales.” (Business Cycle Dating Committee of the NBER, 20/09/2010).

Following this definition, a dummy variable is created using NBER data to capture the existence of an economic recession that takes on the value of 1 if there is a recession, and 0 when there is not. Both datasets with individual and average analyst data are merged with this dataset to introduce the economic conjuncture data to the regression. All inclusion and elimination procedures described yield a final sample of 10,241,613 and 3,498,421 observations for the individual and average analyst datasets, respectively.

4.2 R

ESEARCH

M

ODEL

The design of the research models depends on the basic intuition of the research proposal. This research will mainly follow the research design and setup of Loh and Stultz (2018) for the average analyst analysis and Hilary and Hsu (2013) for the analyst specific analysis, since the basic intuition of this research largely overlaps with their proposals. Namely, to aid the information availability for all types of investors and to eliminate potential (systematic) economic conjuncture biases.

More specifically, this research will perform multiple regression analyses using both the individual as well as the average empirical analyst models, as proposed by Hilary and Hsu (2013) and Loh and Stultz (2018), respectively, with both measures of analyst precision. Using both types of datasets allows for the construction of firm specific as well as analyst specific control variables, and analyze their effects separately. Moreover, the average analyst model is included as a ‘reference model’ in order to test the first hypothesis and attempt to replicate the results of Loh and Stultz (2018) on decreasing analyst accuracy in economic recessions. Subsequently, consistency is applied to this regression, to test the second hypothesis, with the construction as provided by Hilary and Hsu (2013). The individual analyst model is also included to add explanatory power, in testing Hilary and Hsu’s (2013) original proposition of forecast impact, where they already included both measures of analyst accuracy and consistency. In this research design, the economic conjuncture variable is added in an attempt to test its significance, further using all analyst specific controls.

The regression models are built such that the potential effect on analyst accuracy and consistency is split over separate regressions. This separation will help analyzing the pure effects since I can compare regression results to find the magnitude of both effects individually. Essentially, it dissects both effects to understand the underlying nature of the economic conjuncture effect as theorized by the incentive and relational theorems. All regressions will be performed using the method of Ordinary Least Squares (OLS), additionally controlling for serial correlation and heteroscedasticity in standard errors by using a robust regression model. All results will be obtained by performing several t-test’s to test for the significance of included variables, and, most importantly, to test whether the main explanatory is of any significant influence on the informativeness measures of accuracy and consistency. The construction of the regression models will be outlined in the following subchapters, further separated in models for the different measures of analyst precision and individual and average analyst data. Descriptive statistics of all variables are included in table 5 and 6 included in the appendix.

4.2.1 A

NALYST

A

CCURACY USING

A

VERAGE

A

NALYST

D

ATA

To test the first hypothesis, the analyst precision measure of accuracy is regressed on the economic conjuncture effect, further controlling for several relevant firm specific variables. This is done to test the persistency of the earnings game effect in a firm controlled environment. The dependent variable in this regression, accuracy, follows a relatively straight forward construction. The dataset sorts all average forecasts and actuals per firm j, and quarter q, from which the dependent variable of accuracy is created by subtracting the mean, unrevised analyst forecast from the actual earnings per share. This results in a variable of average analyst accuracy for each firm j in each quarter q, denoted Accuracyj,q. This construction is inspired by the research methodology of Loh and Stultz (2018) (see also Hilary and Hsu, 2013; Senga, 2016). Note, as mentioned in the previous section, that this variable is scaled by total firm assets as recorded per firm j and quarter q, to prevent biasing the measure ex ante. As it essentially generates a return on assets measure. Additionally, the variable is transformed to take on only absolute values. Mathematically, this summarizes into the following formula:

𝟒. 𝟏 |𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦| 𝑗,𝑞 =

(𝐴𝑐𝑡𝑢𝑎𝑙 𝐸𝑃𝑆_𝑗,𝑞−𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝑀𝑒𝑎𝑛 𝐸𝑃𝑆_𝑖,𝑞)×#𝑆ℎ𝑎𝑟𝑒𝑠 𝑂𝑢𝑡𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔_𝑗,𝑞

𝐹𝑖𝑟𝑚 𝐴𝑠𝑠𝑒𝑡𝑠_𝑗,𝑞

As mentioned before, the main independent variable of the economic conjuncture effect (recession) takes the form of a dummy variable that takes on the value of 1 when, according to NBER definition, a recession is present, and 0 when there is not. The included firm specific control variables are constructed as follows. Firstly, following the research methodology from Senga (2016), firm age is controlled for by creating an age variable that is the log of the number of years that firm j has been present in the selected I/B/E/S database. Subsequently, the log of total assets in quarter q for firm j is recorded for each observation of average analyst accuracy. Lastly, the market-to-book ratio is included for each firm j at the end of each quarter q, representing another firm-specific control variable. Note that, to reduce the effect of potential outliers or extreme values, accuracy and the last control variable, market-to-book ratio, are winsorized at both extreme ends with a 1% cut. Finally, this setup will lead to regression model 1, including control variables and an error term, εj,q:

𝟒. 𝟐 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑗,𝑞

= 𝛽0+ 𝛽1× 𝑅𝑒𝑐𝑒𝑠𝑠𝑖𝑜𝑛 + 𝛽2𝐿𝑜𝑔(𝐴𝑔𝑒𝑗,𝑞) + 𝛽3𝐿𝑜𝑔(𝐴𝑠𝑠𝑒𝑡𝑠𝑗,𝑞) + 𝛽4𝑀𝑘𝑡𝐵𝑜𝑜𝑘𝑗,𝑞+ 𝜀𝑗,𝑞

4.2.2 A

NALYST

C

ONSISTENCY USING

A

VERAGE

A

NALYST

D

ATA

This second model will try to complement the results for hypothesis 1, and help in attempting to find an answer to the second hypotheses, where the predicted positive analyst bias is analyzed. This model will also depend on average analyst data, meaning the only difference between this model and model 1 will be the dependent variable. Here, analyst consistency is analyzed. The construction of this variable will build on the earlier analyst accuracy, by taking the standard deviation of (winsorized) absolute average analyst forecast errors. This results in a variable of average absolute analyst consistency for each firm j in each quarter q, denoted Consistencyj,q. The construction is based on the research of Hilary and Hsu (2013). Mathematically, this summarizes into the following formula:

𝟒. 𝟑 𝐶𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦 𝑗,𝑞 = 𝑆𝑡. 𝐷𝑒𝑣. (|𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦| 𝑗,𝑞)

= 𝑆𝑡. 𝐷𝑒𝑣. ((𝐴𝑐𝑡𝑢𝑎𝑙 𝐸𝑃𝑆𝑗,𝑞− 𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝑀𝑒𝑎𝑛 𝐸𝑃𝑆𝑖,𝑞) × #𝑆ℎ𝑎𝑟𝑒𝑠 𝑂𝑢𝑡𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔𝑗,𝑞 𝐹𝑖𝑟𝑚 𝐴𝑠𝑠𝑒𝑡𝑠𝑗,𝑞

)

The main independent variable as well as the control variables, and the construction of these variables, remain the same as in model 1. As a result, the regression equation for model 2, including control variables and an error term, εj,q, looks the following:

𝟒. 𝟒 𝐶𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦𝑗,𝑞

= 𝛽0+ 𝛽1× 𝑅𝑒𝑐𝑒𝑠𝑠𝑖𝑜𝑛 + 𝛽2𝐿𝑜𝑔(𝐴𝑔𝑒𝑗,𝑞) + 𝛽3𝐿𝑜𝑔(𝐴𝑠𝑠𝑒𝑡𝑠𝑗,𝑞) + 𝛽4𝑀𝑘𝑡𝐵𝑜𝑜𝑘𝑗,𝑞+ 𝜀𝑗,𝑞

4.2.3 A

NALYST

A

CCURACY USING

I

NDIVIDUAL

A

NALYST

D

ATA

The third model will attempt to complement the results of model 1 in testing the first hypothesis, where analyst precision measure of accuracy is regressed on the economic conjuncture effect. In this model, analyst specific controls are included to test for the persistency of the earnings game effect in the controlled analyst environment. In model

3, the dependent variable of individual analyst accuracy is included, which follows a similar construction as in model 1.

The dataset sorts all individual analyst earnings forecasts and actuals per analyst i, firm j, and quarter q, from which the dependent variable of accuracy is created by subtracting the unrevised individual analyst forecast from the actual earnings per share. This results in a variable of analyst i accuracy for each firm j in each quarter q, denoted Accuracyi,j,q. This construction is inspired by the research methodology of Loh and Stultz (2018) (see also Hilary and Hsu, 2013; Senga, 2016). Again, similar as in model 1, this variable is reformed to absolute values, and scaled by total firm assets as recorded per firm j and quarter q, to prevent biasing the measure ex ante. This variable is winsorized again, making cuts at both extreme 1% levels, to reduce the effect of potential outliers or extreme values. Mathematically, this summarizes into the following formula:

𝟒. 𝟓 |𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦| 𝑖,𝑗,𝑞 =

(𝐴𝑐𝑡𝑢𝑎𝑙 𝐸𝑃𝑆_𝑗,𝑞−𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝐸𝑃𝑆_𝑖,𝑞)×#𝑆ℎ𝑎𝑟𝑒𝑠 𝑂𝑢𝑡𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔_𝑗,𝑞 𝐹𝑖𝑟𝑚 𝐴𝑠𝑠𝑒𝑡𝑠_𝑗,𝑞

Similar as in model 1, the main independent variable of the economic conjuncture effect (recession) takes the form of a dummy variable that takes on the value of 1 when, according to NBER definition, a recession is present, and 0 when there is not.

This setup will lead to regression model 3, including a compound control variable Xki,j,q and an error term, εi,j,q:

𝟒. 𝟔 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑖,𝑗,𝑞= 𝛽0+ 𝛽1× 𝑅𝑒𝑐𝑒𝑠𝑠𝑖𝑜𝑛 + 𝛽𝑘𝑋𝑘𝑖,𝑗,𝑞+ 𝜀𝑖,𝑗,𝑞

This compound control variable, Xki,j,q, is composed of several analyst-specific control variables, based on the research of Hilary and Hsu (2013). The first control variable is Boldness, which is measured by the absolute value of the distance between analyst i’s forecast and the consensus forecast (defined as the average of all analyst forecasts on firm j in quarter q). Subsequently, the variable is winsorized with a 1% cut at both extreme ends to prevent biasing from extreme outliers. Secondly, Horizon is controlled for. This is measured using the number of days between the forecast date and the actual earnings announcement date. Following the argumentation of Clement (1999), forecasts closer to the actual earnings announcement date are more accurate due to the availability of more relevant, less volatile information. The same explanation holds for the inclusion of the following control variable. Furthermore, Experience is controlled for,

which is defined as the log of the number of quarters that the analyst has covered the firm. Here, the longer an analyst has followed the firm, the more familiar the analyst is with the firm performance and firm or market specific events. Also Breadth is included as a control variable, and consists of the number of firms that the analyst covers in a given year. Beyond these analyst specific control variables, an additional firm related control is included. Namely, Cover, which consists of the log of the number of analysts covering a specific firm in a certain year. Note that this control variable will denote the average value of the number of analysts covering a given firm over the entire period, here; quarter.

4.2.4 A

NALYST

C

ONSISTENCY USING

I

NDIVIDUAL

A

NALYST

D

ATA

The fourth and last model will attempt to find complementary support to the second hypothesis, by testing for the economic conjuncture effect on analyst consistency in the individual analyst environment. As this model will also depend on the individual analyst data from the I/B/E/S database, it is similar to model 3 except for using the dependent variable consistency rather than accuracy. This variable will follow a similar construction as in model 2. Where the variable consistency will be formed by taking the standard deviation of (winsorized) absolute individual analyst forecast errors. This results in a variable of analyst i consistency for each firm j in each quarter q, denoted Consistencyi,j,q. Again, the construction is based on the research of Hilary and Hsu (2013). Mathematically, this summarizes into the following formula:

𝟒. 𝟕 𝐶𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦 𝑖,𝑗,𝑞 = 𝑆𝑡. 𝐷𝑒𝑣. (|𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦| 𝑖,𝑗,𝑞)

= 𝑆𝑡. 𝐷𝑒𝑣. ((𝐴𝑐𝑡𝑢𝑎𝑙 𝐸𝑃𝑆𝑗,𝑞− 𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡𝑒𝑑 𝐸𝑃𝑆𝑖,𝑞) × #𝑆ℎ𝑎𝑟𝑒𝑠 𝑂𝑢𝑡𝑠𝑡𝑎𝑛𝑑𝑖𝑛𝑔𝑗,𝑞 𝐹𝑖𝑟𝑚 𝐴𝑠𝑠𝑒𝑡𝑠𝑗,𝑞

)

The main independent variable as well as the control variables, and the construction of these variables, remain the same as in model 3. As a result, the regression equation for model 4, including control variables and an error term, εi,j,q, looks the following:

5

R

ESULTS

Table 5 and 6 in the appendix provide descriptive statistics for the key variables included in all four models using average and individual analyst data. As mentioned in the variable description section in the tables, some variables have been reformed or winsorized to prevent biasing measures ex-ante. For more detail on these reformations I refer to chapter 4 or the previously mentioned tables in the appendix.

Table 1 provides the OLS regression results for model 1 and 2 of the analysis on the economic conjuncture effect using the average analyst data. In model 1, analyst accuracy is used as the dependent variable, whereas model 2 uses analyst consistency as the dependent variable. Table 2 provides OLS regression results for model 3 and 4 of the analysis on the economic conjuncture effect using individual analyst data. In model 3, the regression is based on analyst accuracy as the dependent variable, in model 4, analyst consistency is used as the dependent variable.

In short, results for all four (complete) models (1.4, 2.4, 3.6, and 4.6) provide statistically significant proof for the economic conjuncture effect, all at the 1% level, further supporting both hypotheses posed in chapter 3. Following the research implications proposed in chapter 2.3, this research finds significant but different results for both measures of analyst precision; accuracy and consistency. This supports the assumption made in hypothesis 2 in that not the complete economic conjuncture effect can be dedicated to analyst inaccuracy (incentive theorem), but that the proposed (deliberate) bias (relational theorem) is also found to be of significant influence. The statistical results require some reverse arguing, considering the regression coefficient of the economic conjuncture variable is positive in all four regressions. One could logically interpret this as a positive relation, where analyst accuracy increases where the economic conjuncture effect is present. This is however not the case. Due to the variable construction of analyst accuracy as being the absolute difference between the actual EPS and the analyst predicted EPS, it essentially represents the absolute forecast error. An increase of this difference (forecast error) implies that the analyst has actually become less accurate1_{. Similarly, the variable construction of analyst consistency, as}

1 _{The analyst predictions become farther away from the actual earnings numbers, meaning the}

being the standard deviation of absolute average analyst forecast errors for a certain firm j for a certain year t, requires the same reverse argument. An increase in this variable would mean that the absolute forecast errors increased, making the standard deviation of this variable increase as well, resulting in lower analyst prediction consistency. Additionally note that the dependent variables of accuracy and consistency have been (indirectly) scaled with firm j‘s assets in a certain quarter q to prevent the measure from being biased by the size of a firm. One could argue that larger firms (in terms of asset values) have relatively bigger EPS numbers, arguably also making the forecast errors more pronounced than for smaller firms with smaller EPS numbers. Following this scaled construction, the magnitude of the economic conjuncture effect can be interpreted as a percentage increase in analyst inaccuracy or inconsistency in terms of a percentage of return on assets. Intuitively, consider the coefficient for the economic conjuncture effect in model 1.4, being 0.28%. This number implies that an analyst making his earnings prediction in an NBER defined economic recession will be, on average, less accurate by 0.28% as a percentage of the specific firm j’s return on assets than when that estimate had been made in a period of economic growth.

5.1 A

NALYST

F

ORECAST

A

CCURACY UNDER THE

I

NCENTIVE

T

HEOREM

Recall hypothesis 1 being based on the proposed earnings game incentive theorem where the economic conjuncture effect would change analyst accuracy, arguing from Loh and Stultz (2018) and Hayes (1998). Following the proposed theory, the regression results show support for the detrimental effect of the crisis, assumed to negatively affect analyst commission incentives. As a result, analysts are less financially motivated towards gathering correct and qualitative information to make accurate earnings forecasts. This decreased motivation arguably leads to lower forecast accuracy. Statistically significant proof for this hypothesis has been found in all four models. The regression results are included in table 1 and 2, with both the dependent variables of analyst accuracy and consistency. Moreover, the difference between models 1 and 2, and 3 and 4 is the type of dataset used, where the first use average analyst data, and the latter models use individual analyst data.

As has been discussed in the research implications section, this research has expanded on the explanatory power for the economic conjuncture effect by including an additional measure of analyst forecast precision. Firstly, with the general measure of analyst precision, analyst accuracy, I find statistically significant proof for the economic conjuncture effect, replicating the results from the research of Loh and Stultz (2018). These regression results can be found in model 1 and 3, included in table 1 and 2, respectively. Support for the first hypothesis is further solidified by the statistically significant regression results for model 2 and 4, where I use the measure of analyst consistency. As stated, this measure proves more informative and relevant, since it eliminates a mere change in bias from the earnings predictions, to isolate a pure change in analyst forecast difficulty, and lack of incentives. These results further support the earnings game incentive theorem. Following, to underpin earlier claims, the regression results for each model will be discussed, to establish the theoretical (and economical) relevance of the models.

In table 1 model 1, the economic conjuncture effect shows a statistically significant effect at the one percent level with a coefficient varying from 0.0025 and 0.0028 between model 1.1 and 1.4. Which, following the intuition described above, represents an increasing analyst inaccuracy of 0.25% to 0.28% as a percentage of firm j’s assets in quarter q in periods of economic recession. The firm specific controls, age, assets and market-to-book ratio all show statistically significant coefficients at the 1% level throughout regression models 1.1 through 1.4. The negative coefficient for (the log of) firm age in model 1.4 implies that forecast accuracy is greater for older firms. This result is in line with common firm development theory, where it is argued that younger firms often follow less predictable earnings patterns, compared to their more mature peers (Chen, DeFond, and Park, 2002). Following a similar argument, the control variable assets supports the firm development theory, in which it is assumed that larger firms have more easily predictable earnings patterns (Eames and Glover, 2003), which should contribute to higher analyst accuracy. The market-to-book ratio follows a similar explanation as firm size measured by assets to increase analyst accuracy.

In table 1 model 2, the economic conjuncture effect shows a statistically significant effect at the one percent level with a coefficient varying from 0.0021 and 0.0022 between model 2.1 and 2.4. Which, following the intuition described earlier, represents an increasing analyst inconsistency of 0.21% to 0.22% as a percentage of

firm j’s assets in quarter q in periods of economic recession, supporting the incentive theorem. The firm specific controls, age and assets show statistically significant coefficients at the 1% level throughout regression models 2.1 through 2.4. The market-to-book ratio in model 2.4 is only significant at the 10% level. The control variable age is in model 2 negatively related with analyst consistency. This reasoning will follow a similar argumentation from model 1, in stating that forecast inconsistency is smaller for older firms. This is again consistent with common firm development theory, where it is argued that younger firms often follow less predictable earnings patterns, compared to their more mature peers (Chen et al., 2002). The control variable assets is similar to model 1, showing to be negatively related to analyst inconsistency, meaning it also follows the earlier mentioned firm development theory in which it is assumed that larger firms have more easily predictable earnings patterns (Eames and Glover, 2003), which should contribute to higher analyst accuracy and consistency. The market-to-book ratio has a negative coefficient, which means that the argumentation for assets does not hold here.

In table 2 model 3, the economic conjuncture effect follows a similar, positive effect, by finding a constant effect of 0.31% of firm j’s assets in quarter q throughout regression 3.1 to 3.6. This result further complements the results of model 1 and 2 in supporting the incentive theorem proposed in hypothesis 1, with decreasing analyst accuracy in economic recessions. The control variable boldness is also positively related with analyst inaccuracy. Meaning bold analysts are more likely to make inaccurate forecasts. This is however inconsistent following the theory as proposed by Trueman (1994), Stickel (1990) and Hong, Kubik and Solomon (2000), who claim that more experienced, high quality analyst are more likely to deviate from or not rely on the consensus forecast. Assuming more experienced, high quality investors in general make more accurate forecasts, this is inconsistent with the findings in this study. Moreover, looking at the correlation matrix presented in table 7 in the appendix, it is shown that boldness and experience are negatively related, implying more experienced analysts are less bold. Whilst this definition of experience might follow a different construction, the inconsistency with existing literature remains. Moreover, the control variable experience is positively related with analyst accuracy in model 3.6, implying that more experienced analysts are more inaccurate, which is also inconsistent with the theory of Trueman (1994), Stickel (1990) and Hong et al. (2000). Looking at the less inclusive models of 3.4

Table 1

Regression Results Models 1 and 2, using Average Analyst Data

The dependent variables are accuracy and consistency in model 1 and 2, respectively. Accuracy is defined as the absolute difference between actual EPS minus average analyst forecasted EPS for a certain firm j and quarter q, i.e. absolute forecast errors. This measure is further scaled by firm assets to prevent bias from potential outliers. Following, this variable is winsorized with a 1% cut at both extreme ends. Then, this variable is reformed into absolute values. Consistency is constructed by taking the standard deviation of absolute forecast accuracy. The Economic Conjuncture is a dummy variable that takes on the value of 1 when, according to NBER definition, a recession is present, and 0 when there is not. The control variable Age is the log of the number of years that the firm had appeared in the selected Compustat database. Assets represents the log of total firm assets, defined for each firm at each quarter end. Retrieved from Compustat. The Market-to-Book ratio is computed by taking the book value of the firm as the Compustat reported common equity of a firm for a certain quarter. Following, the market value of the firm was determined by taking the reported quarterly share price times the quarterly number of shares. To compute the ratio, the market value was divided by the book value. Following, this variable is winsorized with a 1% cut at both extreme ends.

Model 1 (Absolute Forecast Error)

[1.1] [1.2] [1.3] [1.4] Economic Conjuncture 0.0025 (72.8)*** 0.0027 (78.12)*** 0.0027 (82.88)*** 0.0028 (84.53)*** (Log) Age -0.0031 (299.1)*** -0.0003 (26.66)*** -0.0029 (27.01)*** (Log) Assets -0.003 (487.2)*** -0.003 (490.83)*** Market-to-Book ratio -0.0003 (12.68)*** Constant 0.0078 (862.41)*** 0.0152 (508.73)*** 0.0282 (628.66)*** 0.028 (599.68)*** R-squared 0.0019 0.0306 0.1340 0.1341 Model 2 (Consistency) [2.1] [2.2] [2.3] [2.4] Economic Conjuncture 0.0021 (114.74)*** 0.0022 (122.79)*** 0.0022 (132.44)*** 0.0022 (132.40)*** (Log) Age -0.0021 (402.11)*** -0.0004 (67.24)*** -0.0004 (67.18)*** (Log) Assets -0.0018 (683.99)*** -0.0018 (687.47)*** Market-to-Book ratio 0.00002 (1.89)* Constant 0.0056 (1203.24)*** 0.0106 (712.85)*** 0.0186 (928.14)*** 0.0186 (883.37)*** R-squared 0.0048 0.0538 0.2017 0.2017

Constant included. Absolute value of t-statistics is denoted in parentheses.

***, **, *; Indicating significance at the one, five and ten percent level, respectively.

All levels of significance are determined based on a regression using a robust regression model, correcting for serial correlation and heteroscedasticity in standard errors.

Table 2

Regression Results Models 3 and 4, using Individual Analyst Data

The dependent variables are Accuracy and Consistency in model 3 and 4, respectively. Accuracy is defined as the absolute difference between actual EPS minus analyst forecasted EPS for a certain firm j and quarter q, i.e. absolute forecast errors. This measure is further scaled by firm Assets to prevent bias from potential outliers. Following, this variable is winsorized with a 1% cut at both extreme ends. Then, this variable is reformed into absolute values. Consistency is constructed by taking the standard deviation of absolute forecast accuracy. The Economic Conjuncture is a dummy variable that takes on the value of 1 when, according to NBER definition, a recession is present, and 0 when there is not. The control variable Boldness is defined as the absolute value of the difference between analyst i’s forecast and the consensus forecast (defined as the average of all analyst forecasts on firm j in quarter q). Following, this variable is winsorized with a 1% cut at both extreme ends. Horizon is created by using the number of days between the forecast date and the actual earnings announcement date. The control variable of Experience is defined as the log of the number of quarters that the analyst has covered the firm. The construction for Breadth consists of the number of firms that the analyst covers in a given year. Cover is defined as the log of the number of analysts covering a specific firm in a certain year.

Model 3 (Absolute Forecast Error)

[3.1] [3.2] [3.3] [3.4] [3.5] [3.6] Economic Conjuncture 0.0031 (176.43)*** 0.0031 (177.56)*** 0.0031 (177.47)*** 0.0031 (179.97)*** 0.0031 (180.30)*** 0.0031 (182.44)*** Boldness 0.0016 (263.65)*** 0.0014 (231.68)*** 0.0014 (224.09)*** 0.0014 (224.58)*** 0.0014 (221.25)*** Horizon 0.00002 (399.58)*** 0.00002 (399.87)*** 0.00002 (399.49)*** 0.00002 (404.93)*** Experience -0.0003 (42.45)*** -0.0001 (11.88)*** 0.0005 (47.85)*** Breadth -0.00001 (39.96)*** 0.00003 (63.49)*** Cover -0.0016 (126.11)*** Constant 0.0072 (1593.4)*** 0.0072 (1600.01)*** 0.0043 (565.45)*** 0.0048 (343.41)*** 0.0047 (335.34)*** 0.0073 (275.73)*** R-squared 0.0042 0.0127 0.0283 0.0285 0.0287 0.0304 and 3.5, the variable however shows negative coefficients, more in line with the theory. This switch in sign could arise from the addition of the control variable cover in model 3.6. The control variable horizon is positively related with analyst accuracy, which is consistent with the theory from Clement (1999), where he states that forecasts further away from the actual earnings announcement date are expected to be more inaccurate due to the lack of relevant, certain information availability. Breadth, the number of firms that an analyst i covers in a certain year t is positively related to analyst inaccuracy in model 3.6. This shows that, following intuitive arguing, analysts covering many firms

Table 2 – Continued Model 4 (Consistency) [4.1] [4.2] [4.3] [4.4] [4.5] [4.6] Economic Conjuncture 0.0025 (273.49)*** 0.0025 (273.49)*** 0.0025 (273.68)*** 0.0024 (266.26)*** 0.0024 (266.29)*** 0.0024 (265.89)*** Boldness 0.0001 (54.94)*** 0.0001 (57.33)*** 0.0002 (73.02)*** 0.0002 (73.01)*** 0.0002 (73.99)*** Horizon -0.0000 (24.75)*** -0.0000 (39.82)*** -0.0000 (39.83)*** -0.0000 (41.12)*** Experience 0.0005 (134.72)*** 0.0005 (106.52)*** 0.0004 (79.04)*** Breadth -0.0000 (1.37) -0.0000 (27.03)*** Cover 0.0002 (31.28)*** Constant 0.0054 (2308.8)*** 0.0054 (2309.31)*** 0.0055 (1215.44)*** 0.0048 (654.81)*** 0.0048 (647.88)*** 0.0045 (349.44)*** R-squared 0.0101 0.0103 0.0104 0.0122 0.0122 0.0123 Constant included. Absolute value of t-statistics is denoted in parentheses.

***, **, *; Indicating significance at the one, five and ten percent level, respectively.

All levels of significance are determined based on a regression using a robust regression model, correcting for serial correlation and heteroscedasticity in standard errors.

will produce less accurate forecasts than analysts following less firms in the same year t. Cover is negatively related to analyst accuracy, which also follows an intuitive argumentation, that the more analysts following a firm, the more collective power in collecting correct information can be used. This would arguably result in a more accurate consensus forecast, which would be more informative for individual analysts to take into account and use as a guiding mechanism for their own forecast.

In table 2 model 4, I find similar results for the economic conjuncture effect, for which I find a positive coefficient of 0.0025 to 0.0024 throughout model 4.1 to 4.6. Intuitively, this represents an effect of analyst consistency decreasing with 0.24% to 0.25% of firm j’s assets in quarter q. This result complements the results of previous models, additionally supporting the incentive theorem proposed in hypothesis 1, with decreasing analyst accuracy in economic recessions. The control variable boldness shows a positive coefficient, which means it follows the same argumentation as provided for model 3. The control variable horizon is positive in this model, presenting an inconsistency with the earlier presented article of Clement (1999). Experience has, similar as in model 3.6, a positive coefficient in model 4.6, implying that more experienced analysts are more inaccurate. This conclusion is also inconsistent with the

theory of Trueman (1994), Stickel (1990) and Hong et al. (2000). The explanations for the control variables breadth and cover follow similar but reverse reasoning as in model 3. Breadth, the number of firms that an analyst i covers in a certain year t is negatively related to analyst inconsistency in model 4.6. Note that it is only a very small coefficient. But the negative coefficient counters the argument presented for model 3.6. For the control variable cover, I find a positive coefficient, which would indicate an increasing inconsistency the more analysts following a firm in a certain year, the more inconsistent the analyst predictions will get. One could argue that when the consensus forecast becomes less accurate and also less consistent, following this forecast would also make the individual analyst more inconsistent in his earnings forecasts.

5.2 A

NALYST

F

ORECAST

B

IAS UNDER THE

R

ELATIONAL

T

HEOREM

Hypothesis 2 focuses on the earnings game relational theorem, following the proposition that in economic downturns, publicly available information is of insufficient quality or quantity to establish accurate earnings predictions. To which the analysts act by (positively) biasing their forecasts in an attempt to gain a stronger bond with firm management to obtain more qualitative private information, arguing from Das et al. (1998). Consistent with the theorized relational effect, this research finds support for the hypothesis that analysts positively bias their earnings predictions in economic recessions. Proof for this hypothesis has been found by comparing all models researching this effect, by looking at the difference in regression coefficients between the measures of analyst accuracy and consistency, since the difference represents the absolute analyst bias. The difference between models 1 and 2, and 3 and 4 is the type of dataset used, where the first use average analyst data, and the latter models use individual analyst data. I refer to subchapter 5.1 for the general explanation of the statistical significance of the models.

To provide numerical support for the second hypothesis, in table 1, model 1.4 and 2.4, I find regression coefficients for the economic conjuncture effect of 0.28% and 0.22%, respectively. Following the intuition described earlier, the difference between these coefficients can be interpreted as the absolute analyst bias in economic recessions,

Table 3

Regression Results with Non-Absolute Analyst Accuracy, using Average Analyst Data.

Sample size: 3,498,421

The following regression results present the coefficients for a complete model using average analyst data (similar as in model 1 and 2). Note that the non-absolute measure of analyst accuracy has been used. The results are used to interpret the direction of the implied analyst bias. Following the relational theorem (hypothesis 2), I expect a positive bias, which implies that analysts overestimate the actual predicted earnings numbers for a particular quarter. The variable analyst accuracy is constructed as being the difference between actual EPS minus estimated EPS (forecast error). This means that a negative outcome, where estimations exceed the actual earnings, implies a positive bias. In the regression results as presented below, I find support for the proposed relational theorem. The negative direction of the economic conjuncture (recession) coefficient implies that, in times of economic contraction, analysts are indeed overestimating earnings predictions. This supports the hypothesized positive analyst bias.

Model 1 (Non-Absolute Forecast Errors) [1.4] Economic Conjuncture -0.0019 _(56.23)*** (Log) Age -0.00008 _(7.23)*** (Log) Assets 0.0017 _(313.53)*** Market-to-Book ratio -0.0024 _(115.64)*** Constant -0.0126 (342.86)*** R-squared 0.0361

Constant included. Absolute value of t-statistics is denoted in parentheses.

***, **, *; Indicating significance at the one, five and ten percent level, respectively.

All levels of significance are determined based on a regression using a robust regression model, correcting for serial correlation and heteroscedasticity in standard errors.

(partially) supporting the earnings game relational theorem by proving that there is indeed a bias present. More specifically, this can be interpreted as the absolute bias increasing by 0.06% as a percentage of firm j’s assets in quarter q in periods of economic recession. Similar for model 3.6 and 4.6, the regression coefficients for the economic conjuncture effect are 0.31% and 0.24%, where the difference represents an increase in the absolute analyst bias by 0.07% as a percentage of firm j’s assets in quarter q in periods of economic recession. Note that with these analyses, I can only make claims on the absolute increase of the bias, not the direction, since these models use absolute

analyst accuracy and consistency as dependent variables. To further develop the argument of the relational theorem, the direction of the bias is analyzed using the same regression model as in models 1 and 3, however, now using non-absolute measure of analyst accuracy, or, forecast errors. This construction is included to help identify the direction of the bias. Regression results for these non-absolute models are included in

Table 4

Regression Results with Non-Absolute Analyst Accuracy, using Individual Analyst Data.

Sample size: 10,241,613

The following regression results present the coefficients for a complete model using individual analyst data (similar as in model 3 and 4). Note that the non-absolute measure of analyst accuracy has been used. The results are used to interpret the direction of the implied analyst bias. Following the relational theorem (hypothesis 2), I expect a positive bias, which implies that analysts overestimate the actual predicted earnings numbers for a particular quarter. The variable analyst accuracy is constructed as being the difference between actual EPS minus estimated EPS (forecast error). This means that a negative outcome, where estimations exceed the actual earnings, implies a positive bias. In the regression results as presented below, I find support for the proposed relational theorem. The negative direction of the economic conjuncture (recession) coefficient implies that, in times of economic contraction, analysts are indeed overestimating earnings predictions. This supports the hypothesized positive analyst bias.

Model 3 (Non-Absolute Forecast Errors) [3.6] Economic Conjuncture -0.0028 (142.23)*** Boldness -0.0009 (130.87)*** Horizon -0.00001 (233.51)*** Experience -0.0003 (28.27)*** Breadth -0.00002 (40.09)*** Cover 0.0013 (88.15)*** Constant -0.0026 (86.87)*** R-squared 0.0120

Constant included. Absolute value of t-statistics is denoted in parentheses.

***, **, *; Indicating significance at the one, five and ten percent level, respectively. All levels of significance are determined based on a regression using a robust regression model, correcting for serial correlation and heteroscedasticity in standard errors.