Master Thesis Draft
University of Amsterdam | Business Economics: Finance
Overoptimism, Miscalibration and Innovation
By Wesley van der Vijgh (10012001) Under supervision of Florian Peters July 2016
Statement of Originality
This document is written by Student Wesley van der Vijgh, who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is 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.
Abstract
The evidence found in research on overconfidence and optimism in the field of finance is mixed. The concepts used to define these biased beliefs are used interchangeably. On the one hand, it is found that CEO overconfidence can be value enhancing and on the other, it is found that it can be value destroying. In contrast to previous research, this paper clearly defines overconfidence and its two separate elements, overoptimism and miscalibration. The separation of overconfidence allows this paper to examine the influence of the two elements of overconfidence on a risky type of investment (Research and Development) and its product, innovation. Previous research found that overconfidence leads to greater R&D and innovation. Unlike previous research, this paper uses Earnings per Share forecast data in combination with realized Earning per Share to measure the two elements of overconfidence. Weak evidence is found that shows that overoptimism has a negative influence on R&D investment. In addition to this, this paper finds robust evidence that miscalibration has a negative influence on R&D investment and innovation. This paper is an addition to the list of articles that contribute to the dilemma: why do certain firms hire biased managers when it could potentially harm firm value. Moreover, it induces more research on this topic to see if executives should consider hiring managers with distorted beliefs.
Table of Contents
I.#Introduction# 5" II.#Literature#review# 7" A.#Character#bias#and#Finance# 7" B.#Overconfidence#and#corporate#decisions# 8" C.#Degree#of#overconfidence# 11" D.#Measure#for#overconfidence# 11" III.#Data#and#empirical#method# 12" A.#Innovation#measures# 13" B.#Measures#for#CEO#Overconfidence# 15" C.#Control#variables# 16" D.#Descriptive#statistics# 17" E.#Empirical#Specification# 19" IV.#Regression#results# 21" A.#R&D#regression# 21" B.#Citation#count#regressions# 23" C.#Hypothesis#1# 25" D.#Hypothesis#2# 25" V.#Robustness#tests# 26" A.#Rolling#average# 26" B.#Average#per#CEO# 27" R&D"regressions" 28" Citation"count"regressions" 28" Separating"annual"and"quarterly"forecasts" 29" C.#Timing#of#announcement# 30" VI.#Conclusion#and#discussion# 31" Discussion# 33" References# 36" Appendix#I# 38" Appendix#II# 39" Appendix#III# 42" "I. Introduction
Overconfidence and optimism are character biases that are frequently researched in corporate finance. A lot of managers in highly ranked positions, such as the chief executive officer (CEO) position, seem to possess these traits (Larwood and Whittaker (1977)). Empirical evidence relating to these behavioral biases of CEOs is mixed. On the one hand, several papers like Hackbarth (2008), Goel and Thakor (2008) and Campbell, Gallmeyer, Johnson, Rutherford and Stanley (2011) show positive aspects of these biases such as the offsetting of risk aversion, which could cause more risk taking and could lead to more investment. On the other hand, papers like Malmendier and Tate (2005) and Malmendier and Tate (2008) show negative aspects such as taking on too much debt and taking value-destructing investment decisions. The question that arises is why certain firms still hire managers with certain distorted beliefs when it could harm firm value.
The paper of Hirshleifer, Low and Teoh (2012) finds that overconfident CEOs invest more in Research and Development (R&D) and achieve greater innovation. This paper defines overconfidence as the overestimation of expected cash flows or the underestimation of risk. In previous literature, the concepts (over)optimism and overconfidence are used interchangeably. The measures most papers use to empirically show the influence of overconfidence are based on the option based measure or stock ownership measure of Malmendier and Tate (2005), or on press portrayals of CEOs (Malmendier and Tate, 2008). These measures do not allow for a separation of overconfidence in an element that overestimates expected cash flows and an element that underestimates risk. However, Hribar and Yang (2015) divide overconfidence into two parts; overoptimism and miscalibration. An overoptimistic person is unrealistically optimistic about outcomes that are uncertain. A miscalibrated person underestimates uncertainty in the prediction of future events, which causes this type of person to underestimate the variance of future returns. Hribar and Yang (2015) find that: optimistic CEOs are more likely to miss their own earnings forecasts and miscalibrated managers issue earnings forecasts with a narrower range. def
This paper distinguishes itself from the research on this topic because it uses earnings per share (EPS) forecast data in combination with realized EPS to measure the two separate elements of overconfidence found by Hribar and Yang (2015). This
measure allows me to measure the influence of one of the elements of overconfidence by controlling for the other. In addition, the new measurement method allows me to disentangle the influence of overconfidence on R&D and innovation found by Hirshleifer et al. (2012). This leads to the following research question:
‘To what extent are the two separate components of overconfidence influencing the R&D investment and innovation of a firm’.
From this research question the following two hypothesis are drafted:
1) Overoptimism leads to more (wasteful) R&D spending and a lower value of innovation
2) Only a medium amount of miscalibration leads to greater innovation value
Answering the research question of this paper contributes to the big-picture puzzle of why certain firms hire managers with certain distorted beliefs when it could harm firm value.
The main regressions in this paper show remarkable results. The results show a negative influence of optimism and miscalibration on R&D and a negative influence of miscalibration on the value of innovation. These results contradict hypothesis 1 that is based on existing literature on this topic. In addition to this, the main regressions also show that a medium amount of miscalibration leads to lower innovation value. This finding contradicts hypothesis 2.
I conduct several robustness tests to see if the results found in my main regressions hold when controlling for possible measurement errors in the two separate elements of overconfidence. Firstly, I find that the results are unchanged when I control for possible fluctuations in individual forecasts that could interfere with the classification of miscalibration and optimism using a 3-year rolling average. Secondly, when I average the forecast data for a particular CEO tenure, which allows for a personal classification of a CEO on the scale of the two elements of overconfidence, the results also remain unchanged apart from the influence of optimism on R&D. Thirdly, when I control for the possible influence the two types of forecasts (quarterly and annual) may have on the classification of the forecasts of a CEO on the scale of the two elements of overconfidence I find that the negative influence of optimism on R&D does not hold when using only annual forecast data,
but does hold when using only quarterly forecast data. In addition to this, the negative influence of miscalibration on R&D and innovation remains unchanged. Fourthly, when I control for the influence the timing of a forecast could have on the forecasting behavior of a CEO, the influence of miscalibration on R&D and innovation remains. The negative influence of optimism on R&D also remains but is not significant in this case.
Lastly, I come up with shortcomings of this research that could have interfered with my results.
The remainder of this paper is structured as follows: section II gives a literature review of the research that is done on managers with biased beliefs in Finance. Section III describes the data and the empirical method used in this paper. Section IV describes and interprets the main regression results. Section V presents the performed robustness tests, and Section VI draws the final conclusions and discusses some shortcomings of this research along with ideas for further research. References and appendices can be found at the end of this paper.
II. Literature review
"
This section first provides background information on where the research on overconfidence in finance stems from. Furthermore, this section describes the research that shows the consequences the decision-making process of a biased manager could have on firms’ financing and strategy. Thereafter, this section provides some theoretical background information on the possible existence of multiple degrees of overconfidence and its corresponding implications. Lastly, this section elaborates on why a new measurement for overconfidence is valuable.
A. Character bias and Finance
The character biases often researched in finance are overconfidence and (over)optimism, often used interchangeably. The research on the influence of biased managers is building on the “better than average” effect and the “illusion of control”, which is found in psychological research (Larwood and Whittaker, 1977). Larwood and Whittaker (1977) empirically show that management students tend to see themselves as more competent relative to their fellow students and tend to
overestimate future growth of their fictitious company. Additionally, they show that actual executives also tend to overpredict their firms’ performance relative to their competitors. Moreover, Weinstein (1980) finds that people on average tend to overestimate the chance of a positive event happening to them and underestimate the chance of negative events happening to them. This is attributed to the fact that people on average tend to overestimate the chance a desirable event is occurring and on average underestimate the chance an undesirable event is occurring. Furthermore, people on average tend to overestimate their chances in a certain event relative to others, people tend to overestimate the degree to which they can positively influence the outcome in a controllable event and people do not seem to identify with stereotypes to whom a negative event is likely to happen. In case of a perceived personal advantage, people underestimate the chance of others having such an advantage could be equal.
As mentioned before, this paper uses the definition of overconfidence of Hribar and Yang (2015). In contrast to most of the studies done on overconfidence in the field of finance, this paper disentangles overconfidence in two parts; overoptimism and miscalibration. Hribar and Yang (2015) define an overoptimistic person as being unrealistically optimistic about outcomes that are uncertain. This builds on the two psychological concepts mentioned in the previous section: the “better than average” effect and the “illusion of control”. A miscalibrated person underestimates uncertainty in the prediction of future events, which causes this type of person to underestimate the variance of future returns. Hribar and Yang (2015) use the option based and the press-based overconfidence measures of Malmendier and Tate (2005, 2008) and find that overconfident managers have a greater tendency to issue voluntary forecasts. Furthermore, they find that the overoptimism of an overconfident manager increases the likelihood of missing the announced forecasts and the miscalibration element of overconfidence causes managers to issue forecasts with a range that is 0.403 narrower compared to non-overconfident managers.
B. Overconfidence and corporate decisions
Based on the above-mentioned psychological concepts (the better than average effect and the illusion of control), much research is done on the influence of optimism and overconfidence on capital structure and corporate decisions. The papers below show the consequences of hiring an overconfident CEO. Malmendier and Tate (2005)
construct an option-based measure to empirically confirm their modeled theory. This option based measure defines a CEO as overconfident if the CEO holds on to a package of stock options that are more than 67% in the money for a period of 4 years or more, twice during his tenure. In addition to this measure, they use an option-based measure that defines a CEO as overconfident when he once during his tenure holds an option until the last year of its duration. Their modeled theory states that overconfident managers overestimate the returns to their investment projects. This causes the CEO to think that their shares are undervalued and that issuing shares will dilute shareholder value. Because of this perception, the CEO does not want to issue shares, which, especially for equity-dependent firms, causes the investments to be cashflow sensitive.
The paper of Malmendier and Tate (2008) uses an option-based measure similar to Malmendier and Tate (2005) and a press based measure to empirically prove the predictions of their model. According to the option-based measure, a CEO is overconfident when he holds on to a stock option until the last year before expiration during his tenure. The press based measure classifies a CEO as overconfident if the relative number of times the media refers to this person as confident and optimistic is higher than the amount of times this person is described as being conservative (the words reliable, conservative, cautious, steady, practical and frugal are counted). Using this model Malmendier and Tate (2008) argue that the overestimation of returns by overconfident CEOs has a positive impact on the acquisitiveness of a CEO. They argue that overconfident CEOs overestimate their ability to produce returns in general and also specifically when these CEOs estimate the combined returns of their firm and a potential target. Just as argued in Malmendier and Tate (2005), this overestimation of returns leads these CEOs to think their shares are undervalued. The reluctance to issue new shares causes the acquisitiveness of overconfident CEOs to be stronger for firms that are the least dependent on equity.
Landier and Thesmar (2009) use a sample of small French startups that are mostly financed with inside equity and bank debt to empirically show that entrepreneurial optimists self-select into short-term debt. Hackbarth (2008) has designed a model, which shows that optimistic managers are more likely to have a relatively large portion of debt. They argue that moderately optimistic managers can increase firm value because they choose a debt level that is closer to the optimal one. This is due to the fact that the optimistic bias causes a manager to invest in
expectation earlier, which offsets underinvestment. In addition to this, the disciplinary effect of this higher debt level offsets bad investments.
These papers point out that biased managers choose a capital structure that deviates from the capital structure that an unbiased manager would choose. This could cause firms with a biased manager to be more efficient.
In addition to the research of Landier and Thesmar (2009) and Hackbarth (2008), the research of Ben-David, Graham, and Harvey (2013), which is based on 10 years of quarterly surveys, also shows that overconfident managers tend to have more debt on average. These managers also tend to invest more. Ben-David et al. (2013) define overconfidence as, what this thesis defines as, miscalibration. The research of Goel and Thakor (2008) goes one step further in explaining the maximization of firm value through the investment channel. They theoretically show that the underestimation of project risk can offset risk-aversion, which causes a more optimal level of investment, higher firm value and lower risk of forced turnover.
The papers above show that overconfident CEOs tend to influence the decision making of the firm with respect to acquisitions, capital structure and investment. Some of the papers above ((Hackbarth, 2008), (Goel and Thakor, 2008)) even show that the distinctive decision-making process of overconfident managers with regard to investment could be beneficial for the firm. This thesis focuses on a specific type of investment (R&D) and the consequences of this type of investment. This type of investment is specifically interesting because R&D investments are risky investments that could lead to innovation as was found by Hirshleifer et al. (2012). Furthermore, Zachariadis (2002) finds that R&D investments have a positive influence on economic growth. Zacharis (2002) also finds evidence for technological spillovers across manufacturing industries. Moreover, Engelbrecht (1997) finds that R&D and international spillovers of R&D have a positive influence on economic growth. These findings show that individual R&D investments of firms not only increase the knowledge base of individual firms but increase the general knowledge intensity of an economy by serving as a catalyst for the creation of additional new knowledge through the spillover of knowledge across industries and countries. This induces economic growth. For R&D investments to be undertaken, confident CEOs who dare to take risk and see the benefits of investing in these risky projects could be beneficial for the firm. An overconfident CEO could have the tendency to invest relatively more in R&D (input), which could eventually lead to more innovative
knowledge (output) (Hirshleifer et al., 2012). This could lead to a steeper trend in the build up of industry knowledge and could illustrate the contribution of a biased CEO to a firm. In contrast to the paper of Hirshleifer et al. (2012), this thesis looks at the influence of the separate effects of overconfidence on the investment decision-making process.
C. Degree of overconfidence
In line with the predictions of the models of Hackbarth (2008) and Goel and Thakor (2008) described in the previous section, Campbell, Gallmeyer, Johnson, Rutherford and Stanley (2011) empirically prove that a non-linear relation exists between optimism and firm value. To measure this, what they define as, optimism, Campbell et al. (2011) use the option based measures of Malmendier and Tate (2005), the press based measure of Malmendier and Tate (2008) and firm investment levels. They show that a moderate amount of optimism offsets risk aversion in a way that causes them to choose the firms’ first-best investment level. This maximizes shareholder value. CEOs with an optimism level that is too low to offset the risk aversion or too high (overoptimism), cause a suboptimal investment level. This kind of CEO is more likely to face a forced turnover than the kind of CEO that is moderately optimistic.
This paper builds on this theory and uses the separation of overconfidence to see to what extent the degree of miscalibration influences R&D investment and innovation, which in the end could positively or negatively influence firm value. D. Measure for overconfidence
The empirical papers researching the overconfidence bias in Finance use measures based on stock options, stock ownership and press portrayals of CEOs to empirically show the influence of overconfidence. These measures do not allow for the separate measurement of the two concepts of overconfidence.
In addition to this, Sen and Tumarkin (2015) argue that the option-based measure that is usually used (a CEO is considered optimistic if he is holding on to stock options that are 67% in the money or more (Malmendier and Tate, 2005)) is affected by other factors than just the optimistic bias. The decision to exercise an option is influenced by the stock characteristics such as dividend yield, volatility and beta and characteristics from the manager such as wealth and risk aversion. For example, the optimal point for an optimistic manager to exercise his option could be
far before expiration in case he has low outside wealth. Besides this, they argue that managers can only be classified as overconfident if a firms’ stock reaches a certain threshold before option expiration. By using this option-based measure, optimistic managers are not classified as optimistic in case their firms’ stock significantly increased in value (which causes them to quickly exercise their option), or in case the firms’ stock did not increase in value enough to be above the optimism threshold.
The above-mentioned arguments illustrate the value of using an alternative proxy for overconfidence.
III. Data and empirical method
"
To construct the sample, several databases were used. The data regarding the EPS forecasts is retrieved from Thomson Reuters’ I/B/E/S guidance database. It consists of both quarterly (44% of the sample) and annual (56% of the sample) earnings per share forecast data. From this guidance database the following variables are extracted: the IBES ticker that allocates the individual forecast to a particular company, the date the forecast was announced, the point estimate in case the forecast was announced as a point estimate (15% of the sample), the upper bound and lower bound of the forecast range in case a forecast range is published (85% of the sample), a variable that indicates whether the forecast is annual or quarterly and the fiscal period end date that shows to which period the forecast applies to. This forecast data is then merged with the actual EPS for which the forecast was announced, extracted from the I/B/E/S database.
After the merging of the realized EPS data to the forecasted EPS, a linking dataset is used that links the IBES company identification ticker to the CRSP PERMNO company identification ticker. This allows me to merge the data on stock prices from the CRSP daily stock price database. This data on EPS and stock prices is used to construct measures for miscalibration and optimism, which is further detailed in the section below regarding measures for CEO overconfidence.
The accounting data that is used to construct the control variables used in the regressions are retrieved from the CRSP/COMPUSTAT merged database. Details on the construction of these variables are given in the control variables section below.
For the data on patents and citations a database is used that was created for the paper of Kogan, Papanikolaou, Seru and Stoffman (2012). Details on this database and the construction of innovation measures are given in the innovation measures section below.
The final sample consists of all the US firms with complete data input in the I/B/E/S guidance, CRSP, CRSP/COMPUSTAT merged and patent database. Firms with missing values on the control variables are deleted. Firm-years with missing citation data are not taken into account in the regressions on citation counts. The final sample consists of 76,887 firm-forecast observations, of these observations 16,608 firm-forecast observations have data on citation counts between 2001 and 2010. The section below explains in detail how the measures for the number of citations, miscalibration and optimism are constructed.
A. Innovation measures
The first type of regressions that are run in this paper take R&D expenses scaled by assets as the dependent variable. This dependent variable shows the resource input for innovation. The main goal of these regressions is to see the effect of the two elements of overconfidence (overoptimism and miscalibration) on the input of innovation.
As a measure of the output of innovation, the number of patent citations is calculated. After the grant date of a patent, the patent keeps on getting citations from other patents that build on the technology of the cited patent. Trajtenberg (1990) concludes in his paper that his findings suggest that patent citations are an indication of the value of innovations. In addition to this Hall, Jaffe and Trajtenberg (2005) empirically show that firms that have two to three times the median number of citations per patent show a 35% firm value premium. They conclude that patent citations show significant information on firm value, which they measure by Tobin’s Q. These findings on the value of patent counts makes it possible to run the second type of regressions, which shows the influence of the two elements of overconfidence on the output of R&D investment, innovation value.
The data on patents and the number of citations these patents receive is retrieved from the database that was used for the paper of Kogan et al. (2012). The authors used a matching algorithm to supplement the 2006 edition of the NBER patent database with data from 1926 up to and including 2010. The database consists of 1,928,123 matched patents to public firms, 27% of these matched patents weren’t
in the 2006 edition of the NBER database. In comparison, the NBER managed to match 32% of all the patents from 1976-2006 to firms and Kogan et al. (2012) managed to match 31% of all granted patents. This shows similar matching success and allows me to use the database from Kogan et al. (2012) paper that includes more recent data. After the matching of patents to firms, the number of patent citations per patents is calculated, which is then totaled for firm-years. This causes me to end up with 12,466 firm-year observations that show the total number of patents and patent citations per firm in the year the patent was applied for.
Hall, Jeffe and Trajtenberg (2001) find that simple citation counts suffer from a truncation bias. This is especially the case for younger patents. Older patents structurally have more citations because these patents receive citations over a longer period. To be able to compare the citations of younger patents with the citations of older patents, the younger patents should be compensated for the shorter time span in which they were able to receive citations from patents that are building on the knowledge registered in the patent. In addition to the truncation bias, Hall et al. (2001) also mention the difference in the propensity to cite in different industries and the difference in the propensity to cite over years. They argue that this influences the comparability of individual simple patent counts. To control for these factors influencing the comparability of patent counts, the patent counts are adjusted using two variations of the fixed effect approach (Hall et al. 2001). As was done in the paper of Kogan et al. (2012), the truncation bias and the difference in the propensity to cite over years are addressed by dividing the number of citations per patent by the average number of citations in a year. Then a variable Qcitation count is constructed that sums up the number of citations across all patents applied for by a firm in a year. As was done in the paper of Hirshleifer et al. (2012), the truncation bias and the difference in the propensity to cite in different industries is addressed by dividing the number of citations per patent by the average number of citations in the same industry in the same year. Then a variable TTcitation count is constructed that sums up the number of citations across all patents applied for by a firm in a year.
The TTcitation count variable is winsorized at the 2.5% level in both tails. The Qcitation count variable is winsorized at the 10% level in the right tail to control for the large outliers in the right tail of the distribution. This is explained in more detail in the descriptive statistics section.
B. Measures for CEO Overconfidence
The way overconfidence and its two elements overoptimism and miscalibration are measured in this paper is based on the empirical findings of Hribar and Yang (2015), who show that miscalibrated CEOs tend to issue EPS forecasts with narrower ranges. In addition to this, they also find that overconfidence causes forecast optimism, shown in the likelihood these CEOs tend to miss their forecasted EPS. Besides the use of forecast data coupled to realized values to infer something about the biased beliefs of an individual by Hribar and Yang (2015), there are other studies that have used similar methods. Otto (2014) uses a measure of optimism where EPS forecasts were compared with realized EPS to show that optimistic CEOs receive a compensation package that is tailored to their bias. To come to the findings regarding startups and short-term debt mentioned in the section above, Landier and Thesmar (2009) used a questionnaire with which they compare expectations on growth and the hiring of personnel with actual accounting data to classify a CEO as optimistic. In addition to this, Ben-David et al. (2013) used forecast errors to come to the above-mentioned findings on overconfident managers and their tendency to select a high level of debt.
To measure the miscalibration element of optimism the findings of Hribar and Yang (2015) are used to build a variable called range. This variable is constructed by dividing the range (upper bound minus lower bound of the forecast) of the EPS forecasts (that are already stock-split adjusted) by the stock-split adjusted stock price of 5 days before the announcement of the earnings forecast. A period of 5 days is chosen to control for the possible stock price reaction that could take place when information is incorporated in the stock price before the forecast announcement took place. When CEOs announce a point estimate, this variable takes on the value 0. To be able to run a regression on several degrees of miscalibration, three miscalibration dummies are constructed: low miscalibration, medium miscalibration and high miscalibration. The dummy low miscalibration takes on value 1 if the range of the earnings forecast is larger than the industry mean forecast range. The dummy medium miscalibration takes on value 1 if the range of the earnings forecast is smaller than the industry mean forecast range. The dummy high miscalibration takes on value 1 when a CEO announced a point estimate.
To measure the optimism element of overconfidence, first, the EPS forecast announcements are coupled to the realized EPS using the variable that indicates to which period the forecast applies. After the merging of the EPS forecasts to the
realized EPS, two measures are constructed. The first measure of optimism divides the difference between the midpoint of the publicized EPS forecast range and the actual earnings by the stock price 5 days before the announcement date. In this case, the stock prices are again stock-split adjusted. When a CEO announces an EPS point estimate, this variable is calculated as the difference between the point estimate and the actual earnings divided by the stock price 5 days before the announcement. Using this proxy, a CEO is classified as optimistic if this forecast error is a positive number. The second measure of optimism is a dummy that takes on the value 1 if the earnings forecast is exceeding the realized EPS. This measure uses the same intuition as was used in the paper of Otto (2014).
To be able to compare the quarterly forecast data with the annual forecast data, the quarterly data is adjusted by comparing the distributions of the two types of forecasts. Then the quarterly distribution is scaled to resemble the distribution of the annual forecasts. To be able to measure the relative deviation in risk perception of a CEO, which allows for the classification of a CEO with regard to the elements overconfidence according to the method explained in this section, the regressions include a variable that controls for the actual risk a company is exposed to. This is explained in further detail in the control variables section.
C. Control variables
To be able to research the influence of the two elements of overconfidence on the in- and output of innovation control variables are used, similar to the paper of Hirshleifer et al. (2012). The control variables included in the regressions are firm size, capital intensity, Market-to-Book, sales growth, return on assets, total book leverage and cash holdings. Firm size is measured by taking the natural logarithm of sales, capital intensity is calculated as total assets divided by sales. Market-to-Book is calculated by dividing the market value of equity plus the book value of assets minus common equity and minus deferred taxes by total assets. Return on assets, total book leverage and cash holdings are all normalized by total assets.
In addition to the general control variables, a control variable is used that specifically controls for the differing (absolute) amount of risk a firm is exposed to. R&D intensive firms are exposed to a relatively larger amount of risk. This risk can be expressed in stock price volatility. Chan, Lakonishok and Sougiannis (2001) show that the R&D intensity of a firm is positively associated with the volatility of the stock
price returns. This relatively larger amount of risk is incorporated in the forecast announcements of the CEO and therefore influences his forecast behavior. To measure the deviation in risk perception of a CEO relative to the actual firm risk (which allows for the classification of a CEO with regard to the elements of overconfidence), a variable must be included that controls for the company specific risk. This captures the noise of firms’ specific risk and allows the earnings forecast announcements in comparison with the realized earnings to be an unbiased measure of the elements of overconfidence. To control for this company specific risk, the stock price variance is included in the regressions. The stock price variance is calculated by taking the average of the daily stock price variance of the past 250 days.
All control variables are winsorized at the 2.5% level in both tails and lagged by one period.
D. Descriptive statistics
Table 1 presents the summary statistics for the final sample. Information on the main dependent variables is presented in Panel A, the information on the measures of overconfidence are presented in panel B and the information on the control variables is presented in panel C.
As shown in panel A, on average 4.8% of a firms’ assets are R&D investments. The average amount of raw citation counts is 88. The more realistic data on citations, which adjust the raw citations for the different propensity to cite in different industries and the difference in the propensity to cite over years, is respectively 22.4 and 0.8. The large difference in these averages can be contributed to this adjustment process. When constructing the variable Qcitation count, that controls for the difference in the propensity to cite over years, the raw number per patent is divided by the average number of citations in a year. A firm that has a relatively large amount of citations in a year where there were relatively few citations could end up with a relatively large Qcitation count number. This is shown in the, compared to the mean, modest median and the relatively large standard deviation. The variable TTcitation count, which adjusts the raw citation counts for the different propensity to cite in different industries, shows a lower average. This can be contributed to the fact that the relatively large difference in citations across industries is now eliminated. Which reduces the large spikes that were visible in the Qcitation count variable.
Table 1 Descriptive statistics
This table shows the descriptive statistics for the dataset used in the regression analysis. The forecast data is retrieved from the Thomson Reuters’ I/B/E/S guidance database, the realized EPS data is retrieved from the I/B/E/S database, the accounting data for R&D and the control variables is extracted from the CRSP/COMPUSTAT merged database. Data on patent citations is retrieved from the database that was constructed for the paper of Kogan, Papanikolaou, Seru and Stoffman (2012). R&D is calculated as R&D scaled by assets. TTcitation count is the total industry adjusted number of citations to all the patents applied for during the year. To adjust for truncation bias the citation count for each patent is divided by the average citation count of all patents in the same industry in the same year. After that, the adjusted citation count of each patent belonging to a firm-year is added up to the firm-year level. Qcitation count is the total year adjusted number of citations to all the patents applied for during the year. To adjust for truncation bias the citation count for each patent is divided by the average citation count of all patents in the same year. After that, the adjusted citation count of each patent belonging to a firm-year is added up to the firm-year level. Range indicates the width of the range of the earnings per share forecast (EPS), scaled by the closing price of the stock five days before the announcement of the earnings forecast took place. The dummy variable optimistic takes on the value 1 if the lower bound or the point estimate of the earnings forecast (EPS) exceeds the actual earnings forecast. The variable forecast error indicates by how much the midpoint or point estimate of the earnings forecast (EPS) deviates from the actual earnings. The controls included in these regressions are the log of sales, capital intensity (calculated by dividing total assets by sales), Market-to-Book, sales growth, Return on Assets (calculated by dividing operating income after depreciation by total assets), Total debt (calculated as long-term debt plus current debt/total assets), cash (calculated as income before extraordinary items plus depreciation divided by total assets) and stock price variance (calculated by taking the average of the daily stock price variance of the past 250 days).
Mean Median Standard deviation N Panel A: Main Dependent Variables
R&D 0.048 0.025 0.059 54240
TTcitation count 0.794 0.202 1.380 16608
Qcitation count 22.435 5.223 35.692 16608
Cites 88.289 7.000 369.572 16608
Panel B: Main Independent Variables
Range 0.004 0.002 0.005 76887
Optimistic 0.237 0.000 0.426 76887
Forecast error 0.001 -0.001 0.018 76434
Panel C: Control Variables
Log(sales) 7.031 6.988 1.737 76887 Capital intensity 1.231 0.996 0.928 76887 Market to Book 2.083 1.721 1.146 76887 Sales growth -0.006 0.000 0.663 76887 ROA 0.102 0.100 0.081 76887 Book leverage 0.185 0.162 0.170 76887 Cash 0.094 0.096 0.078 76887
Stock Price Variance 0.230 0.150 0.288 70967
In this sample, 44% of the observations are quarterly earnings forecasts and 56% of the observations are annual earnings forecasts. The optimistic dummy in panel B shows that according to this method of classification, 23% of all forecasts can be
classified as optimistic, which indicates that on average most forecasts announced by the CEOs in the sample are met or exceeded. The forecast error, which shows by how much the earnings forecast is deviating from the actual earnings, shows a positive average of 0.0012 and a negative median of -0.0005 with a standard deviation of 0.01835. This can be contributed to the fact that forecasts announced by CEOs defined as optimistic on average deviates by a relatively large amount from the actual earnings in comparison with the amount the regular CEOs on average deviate from the actual earnings.
E. Empirical Specification
This section first explains the regression method used and then it shows the regression specification per hypothesis.
Similar to the regression used by Hirshleifer et al (2012), a panel data regression is performed. To be able to control for time-invariant industry specific factors with time-invariant effects that could lead to omitted variable bias in the regression with the dependent variables R&D, Qcitation count and TTcitation count, a fixed regression with industry fixed effects is used. In addition to this, omitted firms specific characteristics in the industry cluster could also influence the dependent variable and therefore interfere with the coefficients on the main independent variables of interest (proxies for the elements of overconfidence) when excluded from the regressions. To control for this kind of omitted variable bias, several control variables are included. Furthermore, to control for the influence time trends could have on the influence of the main independent variables (proxies for the elements of overconfidence) on the dependent variables used (R&D, Qcitation count and TTcitation count), time fixed effects are included. To control for the correlation of errors within firm clusters, standard errors are clustered at the firm level. This leads to the following regressions specified in the section below.
The intuition behind hypothesis 1 is based on the findings of Hirshleifer et al. (2012), who show that overconfidence leads to greater R&D investment and innovation. After the separation of overconfidence, the expectation is however that the overoptimism element of overconfidence will induce greater R&D investment (increase input), but overconfidence by itself does not induce greater innovation value (output). Overoptimism is defined as the overestimation of uncertain outcomes Hribar and Yang (2015), which could lead overoptimistic managers to overestimate the
positive contribution of R&D investment. This intuition leads to the following hypothesis:
Hypothesis 1: Overoptimism leads to more (wasteful) R&D spending and a lower value innovation.
To test hypothesis 1, the following regression is performed:
!&!!" = α! + !β!!"#$%$&%!"+ !!!"#$%&"'(%)"*+!"+ !!!"#$%"&'!"+ !!!"#$%&'(! + !!!!"#$!+!ε!"
The dependent variable in this regression is R&D (as stated in the equation above) or the proxies for innovation value (TTcitation count and Qcitation count). The control variables are the controls specified in the control variable section (firm size, capital intensity, Market-to-Book, sales growth, return on assets, total book leverage, cash holdings and stock price variance).
The intuition behind hypothesis 2 builds on the theory of Campbell et al. (2011), which shows an optimum degree of, what they define as, optimism to offset risk aversion. The idea is that miscalibrated CEOs underestimate the variance in future returns, which induces them to invest in risky R&D investments. A medium amount of miscalibration will induce these CEOs to choose the first-best R&D investment level, which will lead to a greater value of innovation.
Hypothesis 2: Only a medium amount of miscalibration leads to a greater value of innovation
To test hypothesis 2 the following regression is performed:
!!"#$%$#&'!!"#$%!" = α! + β!!!"#$%&"'(%)"*+!" + γ!!"#$%$&%!" + δ!!"#$%"&'!" + !!!!"#$%&'(! + !!!!"#$!+ ε!"
The dependent variable is TTcitation count (as stated in the equation above), and the alternative measure of innovation value (Qcitation count). The control variables are again the variables specified in the control variable section (firm size, capital
intensity, Market-to-Book, sales growth, return on assets, total book leverage, cash holdings and stock price variance).
IV. Regression results
"
To test the first hypothesis that hypothesizes that overoptimism leads to more R&D spending and lower innovation and the second hypothesis that hypothesizes that only a medium amount of miscalibration leads to greater innovation, several regressions have to be run. In the sections below, these regressions are first categorized by their main dependent variable (R&D scaled by assets) and the measures of innovation (TTcitation count and Qcitation count) and described separately. The last section of this paragraph uses these results to comment on the hypotheses.
A. R&D regression
To test what the relation is between the two separate elements of overconfidence (miscalibration and overoptimism) and the spending on the input of innovation, according to the regression specification, a few regressions are performed that use R&D scaled by assets as the dependent variable and the proxies for miscalibration and optimism as the main independent variables. These regressions are shown in Table 2, which shows the main regressions regarding overconfidence and R&D.
Model (1) shows a positive and significant coefficient for the variable Range. The degree of miscalibration is displayed in the width of the forecast the CEO issues: the
smaller the width, the higher the degree of miscalibration. The positive coefficient in
model (1) shows that an increasing amount of CEO miscalibration causes R&D expenditures to decrease. If the miscalibration of a CEO increases with one standard deviation (table 1: 0.0049), which means the forecast range decreases with one standard deviation, R&D/Assets decreases with 0.51%. To contextualize this, the mean R&D/Assets shown in Table 1 are 4.81%. An increase in miscalibration with one standard deviation decreases the amount of R&D/Assets on average by 10.6%.
Model (2) and (3) show negative and significant coefficient for the variable forecast error and for the dummy optimistic. These coefficients both show that the optimism element of overconfidence causes R&D expenditures to decrease. The optimism dummy shows that being overoptimistic leads to a decrease of R&D/Assets by 0.3%. Scaling this to the mean R&D/Assets again shows that
Table 2
R&D expenditures and overconfidence
The table shows the results of the regression of R&D expenditures on miscalibration and optimism. The dependent variable, R&D, is the ratio of R&D expenditures to total assets. The independent variable range indicates the width of the range of the earnings per share forecast (EPS), scaled by the closing price of the stock five days before the announcement of the earnings forecast took place. The independent dummy variable optimistic takes on the value 1 if the lower bound or the point estimate of the earnings forecast (EPS) exceeds the actual earnings forecast. The independent variable forecast error indicates by how much the midpoint or point estimate of the earnings forecast (EPS) deviates from the actual earnings. The controls included in these regressions are the log of sales, capital intensity (calculated by dividing total assets by sales), Market-to-Book, sales growth, Return on Assets (calculated by dividing operating income after depreciation by total assets), Total debt (calculated as long-term debt plus current debt/total assets), cash (calculated as income before extraordinary items plus depreciation divided by total assets) and stock price variance (calculated by taking the average of the daily stock price variance of the past 250 days). All of these independent control variables are lagged by 1 year. All regressions include year and industry fixed effects, defined based on the two-digit sic codes. Standard errors are corrected for clustering of observations at the firm level.
Dept. variable: R&D/Assets
(1) (2) (3) (4) Range 1.039*** 1.043*** (5.8) (5.8) Forecast error -0.078** -0.072** (-2.5) (-2.4) Optimistic -0.003** (-2.3) Log(sales) -0.004*** -0.004*** -0.004*** -0.004*** (-4.0) (-4.5) (-4.6) (-4.0) capital intensity -0.003** -0.004*** -0.004*** -0.003** (-2.4) (-2.9) (-2.9) (-2.5) Market to Book 0.016*** 0.015*** 0.015*** 0.016*** (12.0) (11.5) (11.6) (11.9) Sales growth 0.002*** 0.002*** 0.002*** 0.002*** (2.9) (3.5) (3.4) (3.1) ROA -0.151*** -0.157*** -0.161*** -0.148*** (-6.8) (-7.0) (-7.1) (-6.7) Book leverage -0.041*** -0.040*** -0.039*** -0.042*** (-6.1) (-6.0) (-5.8) (-6.2) Cash -0.079*** -0.088*** -0.084*** -0.084*** (-3.8) (-4.2) (-3.9) (-4.0)
Stock Price Variance 0.006* 0.009*** 0.008** 0.006*
(1.7) (2.6) (2.5) (1.9)
N 49886 49559 49886 49559
R2 adj. 0.520 0.516 0.515 0.522
t statistics in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01"
being overoptimistic decreases the amount of R&D/Assets on average with about 6.2%. An increase in the forecast error of one standard deviation (table 1: 0.0183) shows an impact of comparable magnitude on R&D/Assets. A one standard deviation increase in forecast error causes the amount of R&D/Assets to decrease by about 0.14%. This shows that being optimistic increases the amount of R&D/Assets on average by 3%.
Model (4) shows that controlling for optimism, the significant negative relationship between miscalibration and R&D/Assets still holds and is of the same magnitude as shown in Model (1). Controlling for miscalibration, the significant negative relationship between optimism and R&D/Assets also holds and the coefficient on optimism is of similar magnitude as the coefficient shown in Model (3).
As was found in Hirshleifer et al. (2012) the results in Table 2, consistently indicate that relatively small firms, with a high Market-to-book figure, poor operating performance, low leverage and a large amount of cash show higher R&D expenditures. In addition to this, in line with the findings of Chan et al., the results in Table 2 also consistently indicate that more volatile firms show higher R&D.
B. Citation count regressions
To test what the influence is of the two separate elements of overconfidence on the output of innovation, a few regressions are performed that use two kinds of measures of citations counts as the main dependent variable and the proxies for miscalibration and optimism as the main independent variables. Because citation counts are an indication of the value of innovation, these regressions allow me to check the influence of the two elements of overconfidence on the value of innovation. The results of these regressions are tabulated in Table 3.
As mentioned before, the two kinds of measures of citation counts, Qcitation count and TTcitation count, are the raw citation counts adjusted for truncation bias and the changing propensity to cite. Both measures of citation counts show a positive and significant coefficient for the variable range. As was mentioned in the section regarding the R&D regressions, this means that the larger the degree of miscalibration, the smaller the citation counts. To put this in context again the same method as in the previous section is used. The main of TTcitation count (Qcitation count) in Table 1 is 0.79 (22.43), Models (1) and (6) of table 3 show that if the miscalibration of a CEO increases with one standard deviation (table 1: 0.0049), the TTcitation count (Qcitation count) on average decreases with 11.3% (5.6%). Models (5) and (10) show that when controlling for optimism, this significant negative influence of miscalibration still holds and is of similar magnitude. In this case, a one standard deviation increase in miscalibration on average decreases Qcitation count
Table 2
Patent citations and overconfidence
The table shows the results of the regression of TTcitation count and Qcitation count on miscalibration and optimism. R&D expenditures are calculated as R&D scaled by assets. TTcitation count is the total industry adjusted number of citations to all the patents applied for during the year. To adjust for truncation bias the citation count for each patent is divided by the average citation count of all patents in the same industry in the same year. After that, the adjusted citation count of each patent belonging to a firm-year is added up to the firm-year level. Qcitation count is the total year adjusted number of citations to all the patents applied for during the year. To adjust for truncation bias the citation count for each patent is divided by the average citation count of all patents in the same year. After that, the adjusted citation count of each patent belonging to a year is added up to the firm-year level. The independent variable range indicates the width of the range of the earnings per share forecast (EPS), scaled by the closing price of the stock five days before the announcement of the earnings forecast took place. The independent dummy variable optimistic takes on the value 1 if the lower bound or the point estimate of the earnings forecast (EPS) exceeds the actual earnings forecast. The independent variable forecast error indicates by how much the midpoint or point estimate of the earnings forecast (EPS) deviates from the actual earnings. The controls included in these regressions are the log of sales, capital intensity (calculated by dividing total assets by sales), Market-to-Book, sales growth, Return on Assets (calculated by dividing operating income after depreciation by total assets), Total debt (calculated as long-term debt plus current debt/total assets), cash (calculated as income before extraordinary items plus depreciation divided by total assets) and stock price variance (calculated by taking the average of the daily stock price variance of the past 250 days). All of these independent control variables are lagged by 1 year. All regressions include year and industry fixed effects, defined based on the two-digit sic codes. Standard errors are corrected for clustering of observations at the firm level.
Dept. Variable: TTcitation count Qcitation count
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Range 18.281*** 18.015*** 255.300** 246.671* (3.1) (3.0) (2.0) (1.9) Forecast error 0.274 0.497 -13.615 -10.565 (0.2) (0.4) (-0.6) (-0.4) Medium miscalibration -0.116** -0.865 (-2.1) (-0.7) Optimistic -0.027 -1.493 (-0.6) (-1.5) Log(sales) 0.502*** 0.496*** 0.500*** 0.495*** 0.503*** 13.276*** 13.179*** 13.212*** 13.156*** 13.269*** (9.1) (9.1) (9.1) (9.1) (9.1) (14.6) (14.6) (14.6) (14.7) (14.6) capital intensity 0.145*** 0.139*** 0.142*** 0.137*** 0.146*** 5.028*** 4.904*** 4.955*** 4.898*** 5.007*** (3.8) (3.6) (3.7) (3.6) (3.8) (5.3) (5.2) (5.2) (5.2) (5.3) Market to Book 0.170*** 0.161*** 0.168*** 0.160*** 0.170*** 4.717*** 4.571*** 4.636*** 4.551*** 4.700*** (4.3) (4.1) (4.2) (4.1) (4.2) (4.9) (4.8) (4.9) (4.8) (4.9) Sales growth -0.002 0.007 0.002 0.007 -0.002 0.266 0.418 0.355 0.423 0.292 (-0.1) (0.3) (0.1) (0.3) (-0.1) (0.5) (0.7) (0.6) (0.7) (0.5) ROA -3.445*** -3.576*** -3.492*** -3.595*** -3.424*** -79.963*** -81.848*** -81.309*** -81.881*** -79.768*** (-4.1) (-4.2) (-4.1) (-4.2) (-4.1) (-4.4) (-4.5) (-4.5) (-4.5) (-4.3) Book leverage -0.869*** -0.865*** -0.860*** -0.856*** -0.876*** -24.243*** -24.327*** -24.087*** -24.099*** -24.480*** (-3.3) (-3.3) (-3.3) (-3.2) (-3.3) (-3.8) (-3.8) (-3.8) (-3.8) (-3.9) Cash 1.865** 1.799** 1.821** 1.808** 1.851** 40.269*** 38.791** 39.639*** 39.032** 39.502*** (2.3) (2.2) (2.3) (2.3) (2.3) (2.7) (2.5) (2.6) (2.6) (2.6)
Stock Price Variance 0.332*** 0.378*** 0.361*** 0.382*** 0.329*** 8.355** 8.888*** 8.898*** 9.030*** 8.227**
(3.0) (3.3) (3.2) (3.4) (3.0) (2.5) (2.6) (2.6) (2.6) (2.4)
N 16427 16398 16427 16427 16398 16427 16398 16427 16427 16398
R2 adj. 0.282 0.280 0.281 0.280 0.283 0.415 0.415 0.414 0.415 0.415
by about 5.4% and TTcitation decreases on average by about 11.2%. Models (3) and (8) specifically show that a medium amount of miscalibration scaled to the average TTcitation count (Qcitation count) causes TTcitation count (Qcitation count) on average to decrease with 14.6% (3.9%).
The regressions on citation counts and optimism do not show a consistent and significant influence of optimism on innovation value. Model (2) shows an insignificant positive influence of optimism on TTcitation count. In contrast, model (7) shows an insignificant negative influence of optimism on Qcitation count. The optimism dummies in models (4) and (9) do show a consistent negative influence of overoptimism on the two measures of citation counts. However, these coefficients are also insignificant.
C. Hypothesis 1
Combining the above-mentioned findings on the influence of overoptimism on R&D spending and innovation, the following things can be inferred with regard to hypothesis 1. When controlling for miscalibration, the results indicate that overoptimism leads to less R&D spending, which contradicts the first part of the hypothesis. The second part of hypothesis 1 hypothesizes that when controlling for miscalibration, overoptimism leads to a lower value of innovation measured by citation counts. The results seem to indicate that this is the case. However, these results are not significant, so no conclusions can be drawn from these results. Therefore, the results do not confirm the rationale used in hypothesis 1 that after splitting up overconfidence in overoptimism and miscalibration, the overoptimistic element causes overconfident CEOs to increase R&D investment (input) but by itself does not induce greater innovation value (output).
D. Hypothesis 2
Combining the above-mentioned results on the influence of miscalibration on the value of innovation, the following things can be inferred with regard to hypothesis 2. The results do not confirm the rationale that a medium amount of miscalibration is needed to offset the risk aversion of a CEO, which induces him to choose the first-best investment level. In addition to this the results seem to indicate an opposite relation between miscalibration and the value of innovation: the lower the degree of miscalibration, the higher the value of innovation measured by the citation counts.
A possible explanation for the above-mentioned results that contradict the drafted hypotheses could be that founders of R&D intensive firms show the biased beliefs that could induce R&D spending and innovation value as was hypothesized. However, when an innovative firm goes public, it could be that the founders are replaced by more cautious CEOs that carefully build forward on the innovation that caused the company to grow. This theory is further explained in the conclusion and
discussion section.
V. Robustness tests
#
As mentioned before, the regression results described in the results section contradict the hypotheses that are building on previous research. This could be the case because this thesis separates overconfidence in two elements and measures these elements in a way that has not been done before in this particular context.
It could be that the method used for measuring optimism and miscalibration is not a correct way to measure these two elements of overconfidence. Possibly assessing every forecast entry individually and classifying this forecast entry on the scale of miscalibration and overconfidence is not the right classification method. It could be that the fluctuation in the announced forecasts is interfering with the results. To control for this possible error in measurement method of optimism and miscalibration, alternative measures are used to check if the relations as reported in the results section still hold up. The sections titled rolling average shows a method to control for this possible measurement error. The section titled average per CEO describes in addition to controlling for this possible measurement error of optimism and miscalibration, another type of measurement error that could come from the combined use of quarterly and annual forecast data.
The section titled timing of announcement controls for the influence that timing could have on the forecast behavior of a CEO.
A. Rolling average
One of the methods to eliminate the possible fluctuations in individual forecast announcements is to take a rolling average using lagged years. For this method to work, first the forecast data is averaged for the firm and then this average is used to
calculate the three-year rolling average using two lags. The results of these regressions are tabulated in appendix I table 4.
Using this method, the significant negative influence of miscalibration on R&D expenditures and the two measures of citation counts holds. In the regressions with R&D expenditures as the dependent variable, the negative and significant influence of optimism on R&D as tabulated in Table 2 also holds. The coefficients in Table 4 of appendix I are in absolute terms somewhat bigger compared to the coefficients in the main regression tables 2 and 3.
Using the same approach to show the contextual magnitude of these coefficients, model (4) of table 4 shows that when controlling for miscalibration (optimism), a one standard deviation increase in optimism (miscalibration) causes on average a 5.5% (6.6%) decrease in the amount of R&D/Assets compared to the 2.7% (10.7%) in the main regression results.
In the regressions with the industry adjusted citation count (TTcitation count) as the dependent variable, the coefficients on the optimism measures are now positive but insignificant again. In the regressions with the yearly adjusted citation count (Qcitation count), all of the coefficients on the optimism measures are also positive and insignificant, only the coefficient on the continuous optimism measure in model (12) of table 4 is negative but again insignificant. For comparison, a one standard deviation increase in miscalibration controlling for optimism causes on average a 20.9% (9.2%) decrease in TTcitation count (Qcitation count) compared to the 11.2% (5.4%) decrease in the main results section.
Using the rolling average method to control for fluctuations in individual forecast data shows that the results found in the main regression tables regarding R&D expenditures and innovation value hold up.
B. Average per CEO
Another way to eliminate the fluctuations in individual forecasts is based on the method Otto (2014) uses in his paper to classify a CEO as optimistic using forecast data. Using this new method of classification, for every CEO that could be identified, the measurements for overoptimism and miscalibration are first averaged over a year. Then these measurements are averaged over the total tenure of the CEO. Instead of placing every single forecast of a CEO on a scale of miscalibration and optimism and in the process determining the degree of overconfidence and miscalibration of a CEO,
using this method a CEO is now personally placed on a scale of miscalibration and optimism by assessing his overall forecast behavior during his tenure.
R&D regressions
Table 5 in appendix II shows that the negative and significant influence of miscalibration on R&D holds when averaging miscalibration over CEO tenure. Model (1) in Table 5 of appendix II shows that the coefficient for miscalibration is now somewhat larger in magnitude in absolute terms compared to the coefficient for the basic method of miscalibration classification used in the main regression of Table 2 model (1). Scaling this to the mean R&D/Assets again, one standard deviation increase in miscalibration now causes R&D to decrease with about 9.2%, which is similar compared to 10.6% in the regression of Table 2, model (1), that shows the main regressions on overconfidence and R&D expenditures.
In addition to this, appendix II table 5 also shows that the negative and significant relation between optimism and R&D holds when averaging miscalibration over CEO tenure. Model (2) in appendix II table 5 shows that one standard deviation change in forecast error on average causes a 4.8% decrease in R&D/Assets, compared to 3% in Model (2) of main regression table 2. Model (3) of appendix II table 5 shows that being overoptimistic decreases the amount of R&D/Assets on average by about 18.7% compared to 6.2% in Model (3) of main regression table 2.
Controlling for optimism (miscalibration), the same comparison can be made, the negative and significant relation between miscalibration (optimism) and R&D expenditures in Model (4) (appendix II table 5) hold up and are of comparable magnitude in comparison with the magnitude in Model (1) (Model (2)). This shows that when evening out the yearly fluctuations in forecast data, miscalibration and overoptimism still have a negative influence on R&D expenditures.
Citation count regressions
Similar to the regressions regarding R&D expenditures, appendix II Table 5 also shows that the same relations between the two elements of overconfidence and citation counts hold and are in absolute terms of greater magnitude compared to the results tabulated in Table 3 that show the main regression results with regard to the innovation value. The optimism coefficients are now still insignificant with the same sign. However, the regressions with Qcitation count as the dependent variable do not
show significant coefficients anymore with respect to the two elements of overconfidence. In addition to this, the forecast error coefficient, which is a proxy for optimism, that was positive in model (2) of main regression table 3 is now also negative, but still insignificant.
Models (5) and (8) of Table 5 in appendix II show that one standard deviation increase in miscalibration now causes TTcitation count to decrease with respectively 17.3% and 17.4% which is similar compared to 11.3% in Model(1) of Table 3 and 11.2% in Model (4) of Table 3. This shows that when evening out the yearly fluctuations, miscalibration still has a negative influence on citation counts.
Separating annual and quarterly forecasts
In all the regression that are performed up until now, after adjusting for the difference in distribution between quarterly and annual forecasts, both types of forecasts are used to run the regressions. However, it could be that the method of adjustment is not correct, which could cause a possible uncorrected difference of quarterly forecast behavior of a CEO in comparison with annual forecast behavior to interfere with the way a CEO is classified with regard to miscalibration and optimism. To check if these results still hold when quarterly and annually data is used separately in a regression, all of the regressions in this section are run again both for average CEO overconfidence using only quarterly forecasts and for average CEO overconfidence using only annual data.
The results are tabulated in Table 6 and 7 of appendix II and consistently show the negative influence of miscalibration on R&D expenditures and the industry adjusted citation count (TTcitation count). However, in case of the regressions with R&D expenditures as the dependent variable the annual coefficients on miscalibration are not significant. The magnitude of the coefficients is greater in case of the quarterly data. To put the magnitude of the quarterly miscalibration coefficients in context again, the same approach as before is used. Controlling for optimism (miscalibration), a one standard deviation increase in miscalibration (optimism) causes the amount of R&D/Assets on average to decrease with 7.8% (5.0%). The influence of optimism (miscalibration) is a bit larger (smaller) compared to the decrease in the amount of R&D/Assets described in the main results section. (In this section miscalibration (optimism) causes a 10.6% (2.7%) decrease.)
