The effect of managerial overconfidence on corporate investment forecasts

(1)

Corporate Investment Forecasts

June 30, 2017

Abstract

This paper examines the effect of managerial overconfidence on corporate investment plan-ning and deviations from these commitments. In previous literature it is found that overconfident CEOs overestimate their value-added to investment projects. This leads to excessive investing when they are not financially constraint. In this paper it is hypothesised that this excessive in-vesting lowers the accuracy of the investment forecast. The accuracy hypothesis is tested using a hand collected panel dataset on airlines’ investment commitments from 1994–2014. For each year, up to five years of investment forecasts are obtained from their respective 10-K statements. CEOs are classified as overconfident based on three measures. The results suggest that man-agerial overconfidence leads to lower investment plans but similar or higher actual investments. Lastly, robust support is found for the accuracy hypothesis. The findings are relevant to both the investment- and the behavioural literature.

Master Thesis

Name Stijn Smit

Student nr. 10559515

Supervisor prof. dr. T. Ladika Specialisation Corporate Finance

University University of Amsterdam, Amsterdam Business School Faculty Faculty of Economics and Business

(2)

ii

Statement of Originality

This document is written by Stijn Smit 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.

(3)

Acknowledgements

I would first like to thank my thesis supervisor dr. Tomislav Ladika for providing me with the data required to do the analysis. Furthermore, I want to thank him for all his advice and feedback. When-ever I had a question about my research or writing I could always approach him. He allowed this paper to be my own work, but steered me in the right direction whenever he thought I needed it. I could not have asked for a better supervisor. Thank you.

I would also like to thank Arthur van Eeden, Anne-Cor Halma, Thijs Dijkman, Laurens Hesse and Michael Steenkist for their support and being there to discuss all my ideas. They have made this final year of college a pleasure.

Finally, I must express my gratitude to my parents and to my girlfriend for providing me with their support and continuous encouragement throughout my years of study and through the process of researching and writing this thesis. This accomplishment would not have been possible without them. Thank you.

(4)

iv

1 |

Introduction

In finance it is often assumed that people are fully rational in their preferences and beliefs, this as-sumptions is known as the rationality constriction. It implies that markets are "efficient" (i.e., mar-kets are correctly priced), which is arguably not always true. A relatively new approach, known as be-havioural finance, has tried to link finance and psychology. Bebe-havioural finance seeks to improve re-alism by acknowledging deviations from this constriction. It does so by implementing non-standard preferences, non-standard beliefs, and inattention in its models. Examples of irrationality are loss aversion, anchoring and confirmation bias.1 Irrationality may lead to, among others, bubbles and crashes in the stock markets, speculative trading, overpricing in IPOs (e.g., Facebook), the failure of most mergers to deliver the promised benefits, and managerial overconfidence.

Overconfidence has gotten considerable attention in psychological literature. A famous exam-ple comes from Svenson (1981), who finds that 93% of his samexam-ple rated their driving skill as above average, compared to other people in the sample. This findings has been replicated in several other studies. For example, Cooper et al. (1988) find that 89% of entrepreneurs believe their chance of suc-cess lays between 70% and 100%; however, these entrepreneurs do not give other businesses just like theirs the same chances, in fact they estimate them much lower. Similarly, Pychyl et al. (2000) find that university students drastically underestimate the time needed to study for an exam.

Although well established in psychology, overconfidence is relatively new to finance literature. This is surprising since CEOs are more likely to be influenced by the ‘better than average’ effect (Moore & Kim, 2003). Moore & Kim (2003) conclude that lack of comparison of achieved objectives, which is rarely not the case at CEO-level, causes the BTA effect to be stronger. An example of an objective that is difficult to compare is large-scale investment. These investments are relatively com-plex and differ across firms, which makes them hard to compare; therefore, overestimation is nearly impossible to detect.

One reason why overconfidence is relatively new to finance is that ‘overconfidence’ is hard to quantify. However, Malmendier & Tate (2005) have done just that. They examined the portfolios of CEOs and concluded that rational CEOs would diversify the idiosyncratic risk they are facing as soon as possible. It is expected that when the CEO’s options become vested (and are in-the-money), a rational CEO exercises immediately. Different behaviour may indicate overly optimistic expectations. Using this (and several other) rationals, Malmendier & Tate (2005) were able to create a measure for overconfidence. Since then, previous statement can be tested.

However, research has mostly focused on the effect of overconfidence on mergers and

(6)

6 Introduction– MSc Thesis Stijn Smit

tions (M&A) (e.g. Ferris et al. (2013) and Brown & Sarma (2007)). This leaves areas like corporate investment under-researched. To fill the gap, this paper will focus on the effect of managerial over-confidence on corporate investment forecasts. Overover-confidence could show to be an important fac-tor because overconfident CEOs tend to systematically overestimate the net present value (NPV) of investment projects (Malmendier & Tate, 2005; Larwood & Whittaker, 1977). Meaning, if they are not financially constrained, they will over-invest. Theory predicts that the overestimation of NPV is credited to the self-esteem of the CEO; however, the exact relation is unclear. The suspicion is that overconfident CEOs overestimate their sales, which leads to over-investment (Larwood & Whittaker, 1977). This suspicion is expected to have a larger effect in the dataset used in this paper because it contains investment commitments made several years in advance (and thus depends on cash flow forecasts). When the actual sales are revealed they have to cancel some of the investment, which leads to differences between the actual and planned investments (i.e., forecast errors).

These theories lead to two hypotheses. The first focuses on the overestimation of the NPV. Since overconfident CEOs are likely to overestimate their NPV, it is hypothesised that they have larger in-vestment plans than their non-overconfident counterpart. The second hypothesis focuses on the accuracy of the forecast. When the actual sales are revealed, CEOs may discover that the commit-ments are difficult to finance. This could lead to cancellations, thus larger forecast errors.

To examine the hypotheses, this paper will use a unique hand collected dataset on airlines’ in-vestment commitments. Airlines are chosen because they are in an industry which tends to report their investment plans, or rather commitments, in their 10-K statement. They tend to report this data because their commitments (e.g., plane orders) are typically very large and must be ordered in advance. For each year, up to five years of investment forecasts are obtained from their respective 10-K statements. The effect of managerial overconfidence is tested on two measures of investment behaviour; the first tests the total investment plans a firm has in a given year. The second examines the actual level of investment. A final regression is done to test the absolute forecast error (i.e., the difference between the forecast and the actual investments).

The dependent variable ‘total investment plans’ is equal to the aggregate of the investment plans. An aggregate is used to utilise all the available data and to prevent a bias which could occur if certain forecast years are omitted. Then, actual capital expenditures are examined to provide suggestive evidence for the forecast error. Besides ‘overconfidence’, the main explanatory variables follow from investment literature; variables explaining invest should also explain investment plans.

The main results are as follows. In contrast to the first hypothesis, the results indicate that man-agerial overconfidence is negatively related to total investment plans. The finding predicts lower investment plans of approximately 300 million USD per year, and is robust to a test on financial

(7)

con-straints. Although overconfident CEOs are found to commit to less investments, they end up invest-ing the same as (or more than) non-overconfident CEOs. This suggests larger forecast errors, which is confirmed in the final regression; the combination of lower commitments but the same level of investment leads to significant forecast errors for overconfident CEOs.

The remainder of this paper is organised as follows. Section 2will give a summary of previous literature, both on the origin of the overconfidence measure and its applications. The section closes with the hypotheses examined in this paper. Then,Section 3discusses the data required, how it is combined, and the methodology used to answer the hypotheses. Furthermore, it displays several summary statistics. The paper continues withSection 4, which presents the results found, and their implications. ThenSection 5 tests the robustness of the empirical results by performing tests on critical assumptions. Lastly,Section 6summarises the paper and discusses the concluding remarks.

2 |

Literature Review

2.1 Investment Theories

Over the years several investment theories have tried to explain firm investment behaviour. The most notable was Tobin’s Q-theory, where Q is defined as the market value of a firm divided by the replacement cost of capital (Von Furstenberg et al., 1977). If Q is sufficiently high (i.e., above 1), it is logical for firms to invest since firms with a Q > 1 generate more value using capital than other investors or firms.2Similarly, firms with Q < 1 are wasting some capital and are better off downsizing. Research on Q-theory has been elaborate. The focus was on explaining why this variable, stem-ming from investment models, had such low statistical power. For example, Erickson & Whited (2000) show that Tobin’s marginal Q is a good predictor when the measurement error is removed. They do so by using “measurement error-consistent generalised methods of moments estimators”. Erickson & Whited (2000) also name other important variables which are proven to have statistical power: sales and measures of internal funds (e.g. cash flows).

However, a high Tobin’s does not guarantee high levels of investment. Almeida et al. (2004) argue that financially constrained firms are unable to undertake all of their positive NPV projects. This cre-ates both a cost and a benefit to holding cash. The benefit is the ability to finance future investment, but is simultaneously costly because it potentially requires not pursuing positive NPV projects today. Being financially constraint forces firms to have an optimal cash policy (Almeida et al., 2004).

A more recent contribution to the literature comes from Peters & Taylor (2017) who point out a flaw in previous research. Namely, that previous measures of investment fail to include intangible

(8)

8 Literature Review – MSc Thesis Stijn Smit

assets, which typically amount to approximately 30-40% of firms’ capital. Again, when this is ac-counted for, the regression R2increases; suggesting flaws in prior research. Although future research should account for this, the dataset used in this thesis does not because approximately 99% of airline investment is physical, rendering the intangible part insignificant; planes are the priority investment. A noticeable gap in empirical literature is firm investment planning behaviour and deviations from those plans, previous literature tends to focus on explaining investment with data from actual investments. Although plans are seemingly forgotten, they may hold new insights. This thesis will try to fill the knowledge gap by examining both the aggregate investment plans and the forecast errors.

2.2 Overconfidence

2.2.1 Psychological evidence

Although new to finance, psychological literature has given considerable attention to overconfidence. A famous paper is from Svenson (1981), who examines the tendency of people to consider themselves ‘above average’. Surprisingly, he finds that a large majority believe they are above average car drivers, compared to the other people in the sample.

Other papers have found similar results. Cooper et al. (1988) ask entrepreneurs about their per-ceived chance of success. 89% of entrepreneurs attached a probability between 70% and 100% to their success , 33% went as far as attaching a probability of 0% to failure. When asked how similar business would thrive, the perceived odds were significantly lower, 68% believed their chances to be better than their competitor’s. Cooper et al. (1988) calls this phenomenon ‘entrepreneurial euphoria’, in this paper it is labelled as ‘overconfidence’.

An important consequence of the ‘better than average’ effect (from here on, BTA effect) found by Svenson (1981) and Cooper et al. (1988) is that the individual only attributes success to himself, and assumes it is bad luck when there is underperformance (Feather & Simon, 1971; Miller et al., 1975). The BTA effect is strongest among top students and individuals at the top (like CEOs) because they are unable to make fair comparison (Kruger, 1999; Camerer & Lovallo, 1999). For example, CEOs should compare themselves to other CEOs and not to subordinate managers, otherwise they may conclude they are better than average.

Furthermore, Moore & Kim (2003) show that the BTA effect tends to be stronger if the goals are abstract; abstract goals are more difficult to compare. Choices on mergers or investment projects are complex, and often unique. Therefore, CEOs are rarely able to truly compare themselves to other CEOs. This results in a strong BTA effect, causing managers to have tendency for overconfidence. For example, Hirshleifer et al. (2012) estimate that 61% of CEOs in their sample show signs of overconfi-dence.

(9)

Overconfidence behaviour among CEOs may lead to overly optimistic planning for the future (Larwood & Whittaker, 1977). This hypothesis has been examined with management students and corporate presidents. Larwood & Whittaker (1977) tested several hypotheses. Their most important finding is that CEOs overestimate the performance of their firm. However, the effects of overconfi-dence and overly optimistic planning are under-exposed in the literature following Larwood & Whit-taker (1977). Therefore, this paper will try to examine the effects more closely.

2.2.2 The evolution of the overconfidence variable

Finance literature has an increasing interest in behavioural factors. The focus lies on three irrational-ities: non-standard preferences, non-standard beliefs, and inattention. Recently, the interest of lit-erature turns to managerial overconfidence and its effects. This interest stems from the paper of Malmendier & Tate (2005). They were the first to (properly) quantify overconfidence, opening the doors for future research. Their paper builds on the assumptions of portfolio theory. According to Malmendier & Tate, overconfidence among CEOs ‘reveals’ itself in the under-diversification of the CEO’s investment portfolio; most likely due to overestimation of their companies’ returns. Based on this logic, some portfolio aspects would classify a manager as overconfident. This classification is known as the ‘revealed beliefs’-estimator. Specifically, in their paper, a CEO is classified as overconfi-dent when they hold options beyond reasonable levels. What is considered as ‘unreasonable’ follows from Hall & Murphy (2002).

In the paper of Malmendier & Tate (2005) three definitions of overconfidence are used. The first classifies a CEO as overconfident if their vested options are at least twice 67% in-the-money in the first year the options are exercisable (i.e., are fully vested). This percentage represents a risk-aversion of 3 in a constant relative risk-aversion utility specification, described in Hall & Murphy (2002). The second definition looks at the expiration date of the options. A CEO is flagged (i.e., marked as over-confident) when he holds an option until the last year of its duration. Their final measure uses a subsample where CEO tenure within the company is at least 10 years, and looks at stocks rather than options. The last measure flags a CEO if he purchases more stock than he sells (i.e. being a ‘Net Buyer’) in a given year, for the majority of the years of the sample.

Malmendier & Tate (2008) expand their idea on measuring overconfidence with a second mea-sure: the perception of outsiders (i.e., how the press portrays the CEO). To do so, press data is hand collected from large magazines and newspapers like The New York Times and The Economist. They report the number of articles regarding the CEO, which mentions positive words (like ‘confident’ or ‘confidence’) or negative words (like ‘conservative’ or ‘cautious’). A CEO is then classified as over-confident if he is more often described as over-confident than cautious. Although this method tends to

(10)

10 Literature Review – MSc Thesis Stijn Smit

be time-consuming to build, it can be constructed without extensive option data. Therefore, it lends itself for companies outside the US, who often report far less detailed data.

The first measure, as provided in Malmendier & Tate (2005), requires elaborate data on the option portfolio of the CEO, which is not publicly available. Therefore, Hribar & Yang (2015) use a different measure. They have similar reasoning as in Malmendier & Tate (2005), however, choose a slightly different approach. Their primary measure of overconfidence marks a CEO as overconfident when his vested options are, on average, 67% in the money at least twice during the sample period. This measure can be constructed with data available in ExecuComp. Therefore, this version of the over-confidence variable will be replicated in this paper. The measure is extensively defined and discussed inSection 3.

Hribar & Yang (2015) were, however, not the first to implement this measure. Campbell et al. (2011) test whether over- or under optimistic CEOs should be facing forced turnover using this partic-ular measure. Furthermore, they validate that using aggregate data, as provided by ExecuComp, can produce empirically useful measures. First, they examine whether the main results of Malmendier & Tate (2005) hold using the Execucomp based measure. In correspondence with Malmendier & Tate (2005), they find that overconfident CEOs have significantly greater investment-cashflow sensitivity. Therefore, the Execucomp based measure is believed to hold similar empirical strength. For the sec-ond validation, they use the press-based measure. Both measures suggest that the Execucomp based measure can produce empirically useful measures of CEO overconfidence.

2.2.3 The findings of overconfidence literature

Malmendier & Tate have arguably contributed the most regarding overconfidence. First by quantify-ing overconfidence and later by applyquantify-ing their measures to several subjects. In their original paper, Malmendier & Tate (2005) test for an increased sensitivity of investment to cashflow when CEOs are overconfident.

This prediction is based on the assumption that overconfident CEOs overestimate the returns of investment projects they select. If they have sufficient cash to fund the investment they are expected to over-invest. However, if there is insufficient funding they are less likely to seek funding on the capital markets because they believe the stock market undervalues the company. This combined should lead to increased sensitivity of investment to cash flow. The main finding from this paper is that investments of companies run by an overconfident CEO indeed are more sensitive to cash flow, especially if firms are equity dependent.

In Malmendier & Tate (2008) the effect of overconfidence on takeovers is examined. Their hy-pothesis is that overconfident managers will overestimate the value they can create in a takeover.

(11)

This creates an eagerness to acquire other firms. Overconfidence, therefore, could (to an extent) explain merger activity. Their results validate this by using both the press-based measure and the original measure: late-exercise of options. Both have a positive effect on the probability of engaging in a takeover bid.

Furthermore, they test whether these deals have lower average quality; most likely due to over-bidding. This should be priced by the market: the market’s reaction to bids from overconfident CEOs should be lower than for non-overconfident CEOs. They confirm this hypothesis by looking at an-nouncement returns. Merger bids from ‘long holders’ are significantly lower.

More recently, the research in this field has been extended by Ferris et al. (2013). They continue the work of Malmendier & Tate (2008) by looking at the effect of overconfidence on international takeover activity. According to Ferris et al. (2013) the cultural differences between countries can dis-perse the overconfidence among CEOs. Unfortunately, there is no option holding data available of international CEOs. Therefore, they are limited to the second proposed measure: the perception of outsiders. They conclude that overconfidence affects not only the number of offers made by CEOs, but also that overconfident CEOs tend to have more diversifying acquisitions, and more often use cash to finance the acquisition.

These findings raise the question why firms hire overconfident CEOs. Hirshleifer et al. (2012), therefore, ask whether overconfident CEOs are better innovators. Overconfident CEOs are expected to prefer companies which are risky and challenging because they believe they can make the most difference there. Hirshleifer et al. (2012) argue that innovation tends to have these characteristics; therefore, overconfident CEOs should be more present in innovative companies. Their findings show that overconfidence CEOs indeed are associated with riskier projects and they invest more in inno-vation. Additionally, they find that the investment in innovation is more productive than for non-overconfident CEOs. However, this result holds only in innovative industries.

In contrast to earlier mentioned papers, Goel & Thakor (2008) argue that overconfident behaviour is not necessarily irrational. Namely because firms hire based on the CEO’s perceived ability, which depends on the investment choices made. Since CEOs are aware of the selection criteria, they are found to take on more project risk than they would without this criteria. Overconfident behaviour is under this assumption actually rational behaviour because higher levels of risk lead to a higher probability of being promoted/appointed as CEO (assuming a successful project). Overconfidence is thus not by definition irrational. Although ‘excessive’ overconfidence leads to over-investment and destroys firm value, Goel & Thakor (2008) further find that moderate levels of overconfidence actually is in the shareholders’ interests since non-overconfident CEOs tend to under-invest.

(12)

Over-12 Literature Review – MSc Thesis Stijn Smit

confident CEOs are able to improve decision implementation (Russo & Schoemaker, 1992), are able to encourage agents to take sufficient levels of risk (Goel & Thakor, 2008), and they stimulate en-trepreneurship (Bernardo & Welch, 2001). Lastly, Hirshleifer et al. (2012) show that they are more productive in innovation industries, leading to higher innovation output, even after controlling the R&D expenses.

Above mentioned papers are pioneering in the CEO overconfidence effect. However, since this area is relatively new, there is still much to research. This paper will add to the literature by examining the effect of CEO overconfidence on investment-plans. Until now, research has focused on actual investments, but overlooked investment planning behaviour.

2.3 Hypotheses Development

The research of Moore & Kim (2003), Kruger (1999), and Camerer & Lovallo (1999) predicts that CEOs, compared to an average person, are more likely to suffer from the BTA effect or, in other words, over-confidence. Malmendier & Tate (2005) have quantified overconfidence which led to a trend in litera-ture focusing on the effects of overconfidence on various aspects of business.

Furthermore, they find behaviour which suggests that overconfident CEOs believe their firm is undervalued. This makes overconfident CEOs reluctant to issue equity, which creates a dependency on internal cash flow. They continue their research on overconfidence by examining the effects on mergers. Their prediction is that overconfident CEOs overestimate the value they can add in a merger. This leads to over-bidding and increased merger activity (Malmendier & Tate, 2008).

The first hypothesis builds on this logic. Malmendier & Tate (2005) have shown that, due to overestimation of investment project returns, overconfident CEOs will over-invest when they have sufficient funds. Since investments are required to be made in advance in the airline industry and overconfidence leads to over-investment, the normalised investment plans of overconfident CEOs should be higher as well, holding constant the investment opportunities. The effect is expected to be positive because plans are dependent on the forecast ability of the CEO, and overconfident CEOs overestimate their ability to create value. This could lead to overestimation of future cash flows, which may affect investment plans. Therefore, the first hypothesis is as follows:

Hypothesis 1 Overconfident CEOs have higher investment commitments relative to non-overconfident CEOs.

The second prediction builds on the first. While the first prediction simply looks at planned in-vestments, the second looks at how well the plan is executed (i.e., forecast accuracy). It builds on the work of Larwood & Whittaker (1977), who show that managerial overconfidence may lead to overly

(13)

optimistic planning (e.g., due to overestimation of sales). It is expected that the (overly) optimistic expectations for sales are a driver for investment plan deviations; mostly because this implies that they will overestimate the funds they have available for investing, leading to cancellations or down-ward adjustments of investment when the actual sales become known. These adjustments create gaps between actual- and planned investment, and are expected to occur in larger amounts for CEOs marked as overconfident. The second hypothesis, therefore, is as follows:

Hypothesis 2 CEO overconfidence leads to larger absolute differences between actual- and planned investments compared to non-overconfident CEOs.

3 |

Methodology and Data

3.1 Data and Summary Statistics

To examine the hypotheses, this paper will use a unique hand collected dataset on airlines’ invest-ment commitinvest-ments. Airlines are chosen because they are among the few who tend to report their investment plans, or rather commitments, in their 10-K statement. They tend to report this data be-cause their commitments (e.g., plane orders) are typically very large and must be ordered in advance due to the construction time required.

The full dataset contains 212 airline-year observations of 18 individual airlines. The sample is similar to Rampini et al. (2014), who managed to form a sample of 23 airlines. This sample is smaller because airlines are not obligated to report their commitments. For an overview of the airlines in the sample seeTable I.

The airline dataset contains data from 1994 to 2014 and includes commitments for up to 5 years and larger. Airlines typically report the first five years in detail and aggregate the final years (i.e., I_{t +1},

I_{t +2}, I_{t +3}, I_{t +4}, I_{t +5}, and I_{t >5}). These commitments are not rigid but may change due to renegoti-ations and overall price changes. Also, cancellrenegoti-ations may occur. This creates differences which are interesting to examine. The commitments are, therefore, crucial for this paper since they allow for a clear examination of investment plans compared to the actual investments.

The main objective of this paper is to examine the planning behaviour of overconfident CEOs and to test whether the differences between actual and planned investments can be partially explained by the behavioural bias. Three distinct, yet related, measures of overconfidence are used. The rationale, and how they are measured is explained inSubsection 3.2, other variables are discussed below.

To complete the dataset, the airline database is merged with several other databases. First, the airline dataset is supplemented with accounting data from the Compustat database. Following Mal-mendier & Tate (2005), cash flow is measured as earnings before extraordinary items plus

(14)

deprecia-14 Methodology and Data – MSc Thesis Stijn Smit

tion, and investment as capital expenditures. Both these variables are normalised by Property, Plant & Equipment (from hereon, capital). Tobin’s Q is measured as the ratio of market value of assets to book value of assets.3 Where market value of assets follows from the book value of assets plus the market value of equity (stock price at the end of the fiscal year times common shares outstanding) minus the book value of equity (i.e., share holders equity). Book value of assets is the item ‘total assets’ from Compustat.

Second, the Execucomp database provides the data on the realisable value of the vested options, and how many options the CEO owns. These variables, combined with the stock price at the end of the fiscal year from the Center for Research in Stock Prices (CRSP) database, allows for an estimation of how much the options are, on average, in-the-money, or have ‘moneyness’. This variable is crucial in calculating the first overconfidence measure, Hol d er 67. The stock-based measure of overconfi-dence, Net Bu yer , uses data on the CEO’s stock holdings. This is based solely on differences within the shares owned, excluding options, variable from Execucomp.

Lastly, the Long Hol d er variable requires data on option exercise dates from the Thomson Reu-ters Insiders database. This database contains more elaborate data on options held by the CEO, and is the closest match to the one used in Malmendier & Tate (2005). The main variables it provides are the date the options become vested, the transaction date (i.e. the date the option is exercised), and the expiration date. This dataset is merged with CRSP stock data to create the Long Hol d er variable. The Execucomp database further provides data on several control variables. CEO tenure is calcu-lated as the number of years a CEO is listed in the Execucomp database, the CEO’s ownership in the company is calculated as the sum of the CEO’s shares and his options (both vested and unvested) as a fraction of the common shares outstanding.

The CRSP database provides monthly data on the individual airline’s stock prices, which is used to calculate the annual volatility and the annual buy-and-hold returns. The final control variable is the return on oil prices, where the prices are collected from investing.com. Oil prices are included because of the dependency the airline industry has on oil, which may affect their costs and, therefore, influences the funds available for investment.

Relevant firm statistics are shown inTable II. Panel A shows that the airlines are, with an average (median) dollar value of 10,786 (4,776) million, quite large. This size shows in their average (median) yearly capital expenditures of 833 (566) million USD, which roughly equals 15% of capital owned, or 82% of EBITDA. Panel B provides insight in the companies led by managers showing signs of over-confidence. It is important to note that those companies (on average) seem to have better investment opportunities. Companies led by overconfident CEOs have higher Tobin’s Q and higher cash flows.

3_{This measure has recently been under debate (e.g., Peters & Taylor (2017)), however, since airlines’ investments are}

(15)

Investment theory predicts that those companies should invest more, providing suggestive evidence for the first hypothesis.

CEO-data is summarised inTable III. It shows an average (median) CEO age of 54.95 (55.00). From these statistics it is not evident that overconfident CEOs are either much younger or older than non-overconfident CEOs. However, overconfident CEOs do appear to have more experience. This may be explained by the ‘better than average’ effect as described in Moore & Kim (2003). The longer CEOs work isolated from any means of comparison, the more overconfident they might become. This finding is also inline with Gervais & Odean (2001), who suggest that overconfidence can be ‘learned’ when CEOs attribute good performance to themselves and bad performance to chance; assuming that a longer tenure contains more successes, this implies that tenure might lead to overconfidence. As discussed inSubsection 3.2, tenure is expected to have a negative effect on the forecast errors. In contrast to the second hypothesis, this suggests that overconfident CEOs might have lower forecast errors.

Lastly, Panel A shows that 45% of the full sample is overconfident based on the option-based method, 31% based on the stock-based method, and only 18% based on the expiration-based method. This is comparable to previous findings of overconfidence like presented in Hribar & Yang (2015) and Campbell et al. (2011). It is worth noting that the far majority (66%) of overconfident CEOs hold their options beyond rational in-the-money thresholds. Overconfident-purchasing and holding options until expiration is less common behaviour.

(16)

16 Methodology and Data – MSc Thesis Stijn Smit

Table I Sample of Airlines

First Last Average

Airline Year Year Assets

Airtran Holdings Inc. 2003 2010 1,787

Alaska Air Group Inc. 1994 2014 4,016

Allegiant Travel Inc. 2007 2014 688

American Airlines Group Inc. 1994 2011 25,774

Atlas Air Worldwide Holding Inc. 2006 2012 1,908

Delta Air Lines Inc. 1994 2014 35,228

FedEx Corp. 1998 2013 23,848

Frontier Airlines Holdings 2006 2007 1,146

Hawaiian Holdings Corp. 2007 2014 1,502

Jetblue Airways Corp. 2002 2014 5,824

Mesa Air Group Inc. 1999 2008 1,002

Midwest Air Group Inc. 1998 2006 335

Northwest Arlines Corp. 1996 2007 14,091

Republic Airways Holdings Inc. 2007 2014 3,641

Skywest Inc. 1995 2014 3,067

Southwest Airlines 1994 2014 12,848

United Continental Holdings Inc. 1994 2012 24,805

US Airways Group Inc. 1994 2004 1,545

Note:Table Ilists all 18 airlines used in the sample. First (last) year is the first (last) year the airline appears in the sample. Average assets is the average asset size during the sample period in millions of US dollars. However, do note that some years are missing from the dataset due to data availability.

(17)

Table II

Firm Summary Statistics

Panel A: Full Sample (N = 212)

Obs. Average Median Std. Dev.

Assets 212 10,786 4,776 11,935 Log (Assets) 212 8.50 8.47 1.41 Sales 212 9,548 3,403 10,948 Log (Sales) 212 8.39 8.13 1.35 Capital 212 9,698 4,454 10,301 Capital expenditures 212 833 566 888 Normalised Capital Expenditures 185 0.15 0.11 0.15 EBITDA 212 1,022 399 1,373 Cash flow 212 701 233 2246

Normalised Cash flow 185 0.09 0.09 0.13

Tobin’s Q 212 1.29 1.18 0.40

Stock return (%) 170 4.50 -0.41 48.42

Return volatility (%) 171 14.18 12.69 6.62

Panel B: Overconfident Sample (N = 146)

Assets 146 10,048 5,552 10,897 Log (Assets) 146 8.44 8.62 1.42 Sales 146 9,676 3,638 11,165 Log (Sales) 146 8.39 8.20 1.37 Capital 146 10,121 5,194 10,894 Capital expenditures 146 913 620 965 Normalised Capital Expenditures 137 0.15 0.12 0.14 EBITDA 146 1,129 490 1,433 Cash flow 146 810 304 1,395

Normalised Cash flow 137 0.10 0.10 0.08

Tobin’s Q 146 1.34 1.21 0.43

Stock return (%) 118 2.94 -3.43 48.11

Return volatility (%) 118 14.00 12.71 6.69

Note: The data shown is either from Compustat or a combination of Compustat and CRSP. Panel A shows

the statistics for the entire sample, Panel B shows the statistics for the CEOs marked as overconfident based on any of the overconfident variables. All variables are in millions of dollars, unless otherwise indicated. Asset s equals the book value of total assets, C api t al is defined as ‘Property, Plant & Equip-ment’, C ash F l ow is calculated as the earnings before extraordinary items plus depreciation, Tobi n0s Q is the ratio of market value of assets to book value of assets, St ock Ret ur n is one-year buy-and-hold return based on monthly stock prices, and Ret ur n V ol at i l i t y is the annual standard deviation of the stock’s returns. Capital expenditures and cash flow are normalised by capital at the beginning of the year. To address severe outliers, all normalised variables are winsorized at the 0.5% and 99.5% level.

(18)

Table III

CEO Summary Statistics

Panel A: Full Sample (N = 212)

Age 199 54.95 55 6.28

CEO Tenure 212 5.54 4.50 4.44

CEO & President 212 0.12 0 0.33

Moneyness 212 0.64 0.16 1.42 Ownership (%) 201 3.30 1.84 4.28 Vested Options (%) 211 0.79 0.48 0.87 Holder67 212 0.45 0 0.50 Net Buyer 212 0.31 0 0.46 Long Holder 212 0.18 0 0.38

Panel B: Overconfident Sample (N = 146)

Age 140 56.15 56 5.72

CEO Tenure 146 6.49 5 4.79

CEO & President 146 0.10 0 0.30

Moneyness 146 0.91 0.36 1.62 Ownership (%) 140 3.55 2.14 4.39 Vested Options (%) 146 0.83 0.50 0.88 Holder67 146 0.66 1 0.48 Net Buyer 146 0.45 0 0.50 Long Holder 146 0.26 0 0.44

Note:Table IIIshows summary statistics on all CEOs in the sample (N = 42). Panel A shows the statistics for the entire sample, Panel B shows the statistics for the CEOs marked as overconfident based on any of the overconfident variables. Hol d er 67 is a dummy equal to 1 if the CEO holds fully vested options which are at least 67% in-the-money at least twice during the sample period, starting the first year he shows this behaviour. Net Bu yer is a dummy equal to 1 if the CEO purchases more stock than he sells for more than 60% of his tenure at the firm, starting the first year he shows this behaviour. Long Hol d er is a dummy equal to 1 if the CEO ever holds in-the-money options until expiration. Mone yness is a fraction equal to the percentage difference between the stock price at the end of the fiscal year and the implied strike price. Tenur e is defined as the number of years a CEO is listed in the Execucomp database. Ow ner shi p is the sum of the CEO’s shares and his options (both vested and unvested) as a fraction of the common shares outstanding. V est ed Op t i ons is the fraction of vested options to common shares outstanding. C EO & P r esi d ent is a dummy equal to one if the CEO holds the title ‘Chairman, Chief Executive Officer, and President’.

(19)

3.2 Methodology

3.2.1 Measuring overconfidence

The rational of Malmendier & Tate (2005) for measuring overconfidence is as follows. CEOs often receive large compensation packages, including options, to align their interests with those of the shareholders. To ensure the effect of the packages, the actions of the CEO are restricted by the re-spective company. For example, options typically have a vesting period in which management is not allowed to exercise or sell the options. Also, management is often restricted from selling company stock short, which would allow for perfect hedging. Therefore, the alignment of interest comes (to the management) at a cost of high exposure to their firm’s idiosyncratic risk. As shown by Hall & Murphy (2002), these factors imply that a rational CEO should exercise the options as soon as they are both vested and in-the-money. Overconfident CEOs, however, overestimate the value they add in investment project and thus the company’s future returns as well. To profit from the expected value increase, managerial overconfidence causes management to postpone option exercise or to take on additional risk by purchasing stock.

This rationale translates into a quantitative measure via the option exercise behaviour, which is implicit in aggregate option data. As described inSubsection 2.2, there are several ways of measuring overconfidence. Malmendier & Tate (2005) were able to use three distinct, yet related, definitions of overconfidence. However, due to data availability, not all measures are easily replicated. Therefore, this paper will follow Hribar & Yang (2015) and Campbell et al. (2011), who mark CEOs as overconfi-dent based on aggregate option values.

Specifically, an aggregate implied strike price is calculated using the realisable value of the vested options, the quantity of the vested options, and the actual stock price at the end of the fiscal year. Di-viding the realisable value by the quantity, yields an average realisable value per option. Subtracting this from the stock price at the end of the fiscal year yield the aggregate implied strike price. Then the average ‘moneyness’ follows from the percentage difference between the stock price and the implied strike price.

The first measure for overconfidence used in this paper, Hol d er 67, is a dummy equal to 1 if a CEO holds exercisable options with an average moneyness of 67% at least twice during the sample. Because it is likely that unbeneficial overconfident behaviour gets punished overtime by either the board or investors, and therefore likely to decline, the CEO is marked as overconfident for only the first three years of the period he shows this behaviour. This allows for better identification of the overconfidence-effect.

(20)

on the Net Bu yer variable described in Malmendier & Tate (2005). This measure examines the stock purchasing behaviour of the respective CEO. A CEO is flagged (i.e., marked as overconfident) when he purchases more stock in a given year than he sells. The logic is similar to the Hol d er 67 measure, if the CEO was rational, he should be diversifying his risk. Instead, he takes on more idiosyncratic risk, which is a sign of overconfidence. To construct this measure Malmendier & Tate (2005) look at CEOs who are at least 10 years in the dataset. If he is a net buyer (i.e., buys more stock than he sells) in their first five years they mark him as overconfident. Due to a lack of data availability, this paper will mark a CEO as overconfident based on the stock-measure when the CEO shows net-buying behaviour in at least 60% of his tenure at the company.

The third measure is based on the Long Hol d er variable used in Malmendier & Tate (2008). This measure is also based on the option exercise behaviour of the CEO, but instead of moneyness it is based on the period the option is held by the CEO. If the CEO holds the option until expiration he is facing idiosyncratic risk for a period beyond a rational threshold. Therefore, he is assumed to be overconfident. This paper will mark a CEO as a ‘long holder’ if he exercises the option in the expiration year.

Based on these measures it is found that, of the 42 managers included in the sample, 15 managers (≈ 36%) show signs of managerial overconfidence based on the Hol der 67 measure, 13 (≈ 31%) based on the Net Bu yer measure, and 5 (≈ 12%) based on the Long Hol der measure.

3.2.2 Analysis

This paper focuses on explaining differences between investment plans and actual investments and tests whether the behavioural bias of ‘overconfidence’ plays a role. It does so by first testing whether overconfident CEOs have larger than average investment plans, which is a possible explanation for larger differences. All measures of overconfidence (O), either Hol d er 67, Net Bu yer or Long Hol d er , are regressed separately to improve the robustness of the results.

Since total investment plans are likely to be explained by similar variables as regular capital ex-penditures, recent investment literature is utilised to find explanatory variables. Furthermore, Mal-mendier & Tate (2005) offers guidance for control variables as well. Specifically, the regression will be extended with Tobin’s Q, size approximations, data on stock returns, annual volatility, and cash flow data. To match the period related to the volatility, a yearly buy-and-hold return is calculated over the same period (i.e., an annual return).

Included control variables are ownership, tenure, oil prices, the number of vested options, sales, and information on the actual role of the CEO in the company. Ownership is the stake the CEO has in the (shares and options) company, as a fraction of common shares outstanding. Tenure is defined as

(21)

the number of years the CEO is listed in Execucomp and serves as a proxy for experience. Oil prices are included because the sample consist solely of airlines, which tend to be dependent on oil in their day-to-day operations. Vested options are all exercisable options as a fraction of common shares outstanding. Lastly, a dummy equal to one if the CEO holds the title ‘Chairman, Chief Executive Officer, and President’ is included.

The first prediction, discussed inSubsection 2.3, is related to the general planning behaviour of CEOs. To test whether managerial overconfidence is related to higher levels of investment planning, regressions on the planning variables are executed. Specifically, the aggregate investment plan (i.e., the sum of investment plans, T I Pt) will be the dependent variable. This variable is chosen over,

for example, the one-year forecast to better account for the production time needed. It may be the case that the one-year forecast of an overconfident CEO is low, but he has recently committed to a large order. Unlike the one-year forecast, T I Pt will contain the large order, and those in between.

Therefore, it is believed to hold more explanatory power. The following specification summarises the regression:

T I Pi ,t= β0+ β1Oi+ β2Qi ,t+ β3Ci ,t+ β4Ri ,t −1+ β5σi ,t −1+ β6Ei ,t+ β7(Ci ,t×Qi ,t) + X_{i ,t}0 B9+ ²i ,t, (1)

where T I Pi ,t is the total investment plan (i.e.,P5_k=1I_{t +k}) at year t of firm i , Oi is the overconfidence

measure, Qi ,t is Tobin’s Q, Ci ,t is normalised cash flow, Ri ,t −1is the lagged total buy-and-hold return

in a given year,σ_{i ,t −1}is the lagged volatility of the monthly stock returns in a given year, Ei ,t is

nor-malised EBITDA, X is the set of controls including stock ownership (as a percentage of total common shares outstanding), the total number of vested options (as a fraction of total shares outstanding), CEO tenure, firm size (the natural log of total assets at the beginning of the year), and the total buy-and-hold return on oil in a given year. An interaction between Tobin’s Q and cash flow is included because a firm’s ability to raise capital should have a relation with its investment opportunities (i.e.,

Q). Therefore, firms with a high Q should be less dependent on their internal cash flows to finance

their capital expenditures; a negative sign is expected (Gugler et al., 2004). Furthermore, year fixed effects are included to account for trends. The regression specified inEquation (1)will be a panel data regression using random effects. Firm fixed effects are intentionally not included because CEO overconfidence is likely to vary little across time, thus trying to identify the coefficient off of variation with the firm is not useful.

One common objection to behavioural theories is that people with “wrong” beliefs will learn over time from their mistakes (or be eliminated), hence such distortions will disappear. Therefore, it is assumed that investment decisions early in the CEO’s tenure are affected more by overconfidence then the decisions made when he is more experienced. To isolate the true overconfidence effect the

(22)

regression is run on the first 3 years the CEO shows overconfidence behaviour and the years before. Building on the results from this regression, the second regression tests whether overconfident CEOs have larger forecast errors. However, since the data is rather noisy two separate regression are run to verify implicit results. First, the same regression as specified inEquation (1)is run on normalised capital expenditures. This allows for testing of the actual investment behaviour of confident CEOs compared to non-overconfident CEOs. From the results it is implicit whether over-confidence leads to larger errors. For example, ifEquation (1)suggests that overconfident CEOs plan to invest more than non-overconfident CEOs andEquation (2)suggests they end up investing less, then it is implicit that the forecasts errors are larger.

C APE Xi ,t= β0+β1Oi+β2Qi ,t+β3Ci ,t+β4Ri ,t −1+β5σi ,t −1+β6Ei ,t+β7(Ci ,t×Qi ,t)+Xi ,t0 B9+µi ,t, (2)

To further test this implicit finding, a regression on investment differences is run. Existing liter-ature offers little guidance on relevant control variables for differences between actual and planned investments. However, the differences are created by increases/decreases of investments relative to the investment plans. And since it is well established what influences investment, it is assumed that relevant explanatory variables for investment also hold for differences between actual and planned investment. For example, if Tobins Q is higher than expected in year t then the firm is likely to invest more than planned. This, ceteris paribus, leads to larger forecast errors. This element of ‘unexpected-ness’ is unobservable, however, regular Q is likely to still hold sufficient explanatory power. Following this logic, the regression specification is as follows:

T I Di ,t= β0+ β1Oi+ β2Qi ,t+ β3Ci ,t+ β4Ri ,t −1+ β5σi ,t −1+ β6Ei ,t+ β7(Ci ,t×Qi ,t) + Xi ,t0 B9+ ξi ,t, (3)

where T I Di ,t is the total investment difference between actual capital expenditure at time t and

all relevant forecasts in the previous year of firm i (i.e., the sum of differences between the actual investments and the one-year forecast at time t −1, the two-year forecast at time t −2, etc) normalised by capital. T I Di ,tis chosen as dependent variable since it is likely to be less affected by the noisy data

than, for example, the one-year forecast. Specifically, it is assumed that the pluses from the one-year forecast are offset by minuses from the two-year forecast. This leads to lower extremes and thus a less noisy variable. This specification is able to answer the second hypothesis. The predicted value forβ1inEquation (3)will show whether differences indeed become larger when there are signs of managerial overconfidence.

Above mentioned methodology has as foremost strength that it uses panel data regression. This allows it to control for unobservable macro-variables, reducing omitted variable bias (OVB). Note

(23)

that it does not account for omitted variables that vary across firms in a single year, thus the re-gressions might still suffer from OVB. One potential threat to external validity is sample selection bias. The results discussed inSection 4may not be relevant to all industries, as they are based on data from the airline industry alone. Furthermore, because only airlines are used, the sample size is rather low. Therefore, the estimates have higher standard errors, which lowers the likelihood of finding significant results.

However, due to data availability, it is impossible to run similar analyses on, e.g., the firms in the S&P500. Simply, because those firms do not report investment planning data in their annual reports. To address this issue,Section 5tests several robustness concerns.

3.2.3 Predictions

The expectations for the overconfidence measures are discussed in the hypothesis subsection in Section 2. The predictions for the control and explanatory variables are overall based on previous research. Tobin’s Q and cash flow are two proven explanatory variables for investment and are ex-pected to be significantly positive (Erickson & Whited, 2000; Malmendier & Tate, 2008). The causality is that if cash flow increases, more funds are available for investing, meaning that investment is likely to increase as well. EBITDA is another measure for internal funds and is expected to behave similarly as cash flow since EBITDA is often regarded as an estimate for ‘free cash flow’. Similarly, Tobin’s Q is proven to positively affect investment (Von Furstenberg et al., 1977). When the market value of a company’s assets is higher than the book value of its assets, the company is said to create value. Therefore, it should invest. In this paper it is expected that this prediction holds for both investment plans and regular capital expenditures.

Stock returns and return volatility are included because they are expected to affect investment plans for overconfident CEOs specifically. When previous year stock returns are high, the CEO is likely to accredit this to his own performance (Feather & Simon, 1971; Miller et al., 1975). After a period of success the investment plans are predicted to increase, especially for overconfident CEOs.

The effect of return volatility is less clear. On the one hand, volatility may influence the price of the CEO’s option package. One theory is that, if the returns are highly fluctuating, the CEO will at-tempt to boost the stock price (even further) to ensure the stock price does not fall below the strike price of his option; even with large price movements. Stock prices can be increased with (unex-pected) value creating investments. However, volatility also affects the level of capital a company can raise on the equity market with certainty. This could withhold the company from issuing equity, leading to lower levels of cash, thus lower available funds for investment. This argument is debat-able since pecking order theory predicts that, due to information asymmetry, companies prefer to

(24)

raise debt before accessing the equity markets. Overall, there seems to be no clear direction for the estimate.

Corporate governance theories suggest that providing the CEO with ownership should align his interests with those of the shareholder’s. This suggest that a properly motivated CEO will not show excessive investment behaviour. Especially, when assuming deviations from the forecast negatively affect the stock price. Following this logic, ownership should have a negative sign for both total in-vestment plans and deviations. However, overconfident CEOs overestimate the NPV of their invest-ment projects; therefore, additional ownership may further increase investing behaviour. Again, the direction of the ownership variable is unclear.

Although one may argue that young CEOs are more eager to prove themselves and will, therefore, commit to more investment, the sign of experience (i.e., CEO tenure) is unclear. Mostly, because Gervais & Odean (2001) show that overconfidence may actually be ‘learned’ over time. However, as-suming people learn from their mistakes, experience should lower excessive investing and make the forecasts more accurate. This suggests a negative relation between CEO tenure and the dependent variables. A similar argument can be made for CEO age.

Lastly, when oil returns are positive the price of oil has increased. Dependent on the hedging strategy of the airline, the costs will already have increased significantly, or will in the future (ceteris paribus). As a result, less cash is available for investing. Thus, positive oil price returns should lower total investment plans; a negative sign is expected.

3.3 Data limitations

One important note to this data is that it seems to suffer from (a form of ) constant measurement error. Specifically, the forecasts tend to fluctuate heavily from year to year. For example, the airline ‘Airtran Holdings Inc.’ has planned to invest 14.9 million USD in 2006, but actually invested 105 mil-lion in 2006, roughly seven times higher. No significant event had occurred in that period to account for the difference (like a change of control) and the 10-K statement does not offer an explanation either. The investment commitments have similar irregularities. For example, in 2012 the airline ‘Republic Airways Holdings Inc.’ had commitments to investment 0.5 million in 2014, but in 2013 the commitment increase for no apparent reason to 710.5 million. In 2014 they had actual invest-ments of 569.2 million, far below the commitinvest-ments reported in 2013 and far above those reported in 2012. Although one would expect that short-term forecasts are relatively accurate, these values al-ready show large differences. This casts doubt upon either the reporting quality of the commitments or the management’s forecasting ability, and is an important limitation of this research.

(25)

4 |

Results

To illustrate the overconfidence effect on both the commitments and the forecast error, the sample is divided into 3 subsamples, each representing a specific overconfidence measure. Figure Iplots those subsamples. Additionally, the complete sample is plotted to allow for relative evaluation of the subsamples. Each ‘dot’ is obtained as the equally-weighted average across the sample it represents. The graph provides contradictory evidence to what is hypothesised inSection 2, overconfident CEOs appear to have lower investment plans at any given point, with the exception of Hol d er 67.

Similarly,Figure II graphs the same samples but illustrates the average difference between the actual capital expenditures and the respective lagged forecast (i.e., the forecast error). The data over-all illustrates that the longer the forecast, the larger the forecast error, which is to be expected. Again, the overconfidence measures seem to contradict the hypothesis: overconfident predictions tend to have below average forecast errors. Further analysis will have to verify the illustrated results.

Figure I

Aggregate investment forecasts

Figure II Forecast error

Note:Figure Ishows the average investment forecasts aggregated over forecast years.Figure IIshows the average forecast error. Both are separated by overconfidence measure.

4.1 CEO planning behaviour

The dependent variable inTable IVis the sum of investment forecasts up to 5 years in a given year, normalised by capital at the beginning of the fiscal year. To test whether overconfident CEOs plan more optimistically, it is regressed on an overconfidence measure together with variables which have a proven impact on levels of investment. For robustness, two regression theories are applied: pooled OLS, and panel data with random effects.

Each regression series starts with a baseline regression. Columns (1) and (6) are those baseline regressions (i.e., do not include the overconfidence measure). Comparing estimates with those after the inclusion allows for better identification of the effect of including the overconfidence measures.

(26)

26 Results – MSc Thesis Stijn Smit

Columns (5) and (10) include additional control variables for the overconfidence measure with the strongest effect in the three columns prior. This to indicate that the results found are not driven by omitted variable bias. In general these regressions indicate what previous literature expects: Tobin’s

Q and cash flow increase the level of investment. However, the size of the airline does not seem to

behave as expected. The level of total assets does seems to lower the investment commitments. The variables Tobin’s Q and cash flow are robust to the inclusion of year fixed effects.

Unlike Tobin’s Q and cash flow, the overconfidence variables do not behave as hypothesised. Re-sults seem to indicate that overconfidence lowers the fraction total planned investment to capital by 0.53 (0.33) for the Hol d er 67 (Long Hol d er ) measure in the pooled OLS specification. With a median capital level of 4.45 billion USD this represents a difference in investment plans of 1.9 and 1.5 billion USD, respectively. However, due to the small sample size these findings are rather noisy. The true value ofβ1 is with 95% probability between -0.132 and -0.928 (-0.042 and -0.618), which suggests lower investment commitments between 436.1 and 4129.6 million USD (186.8 and 2750.1 million USD) for the Hol d er 67 (Long Hol d er ) measure. These findings are both economically sig-nificant and statistically sigsig-nificant at 5%. Note that these findings are an aggregate, thus have to be divided by 5 to get to the yearly average deviation. The findings are robust to controlling for financial constraints, discussed inSection 5. Unlike the option-based measures, the stock-based measure is statistically insignificant in both the pooled OLS and the random effects regression. This could be ex-plained by a difference in data quality compared to earlier research. Furthermore, neither the stock returns nor their volatility is not found to impact total investment plans significantly.

In both regression specifications the measure Hol d er 67 has the strongest overconfidence effect and is, therefore, further tested. To test whether the result is driven by omitted variable bias, the parsimonious models are extend with additional control variables. The results are robust to this extension. The strength of the effect decreases (increases) to 0.44 (0.76) in the pooled OLS (random effects) specification. The estimates are, however, not significantly different from the earlier findings (i.e., are within their respective confidence intervals).

Notably, the results are in conflict with the first hypothesis. These results indicate that overcon-fident CEOs plan to invest less than non-overconovercon-fident CEOs. These results are surprising since it was argued in Malmendier & Tate (2005) that managerial overconfidence leads to over-investment when there are sufficient funds available. As demonstrated inTable IItheir cash flow is above av-erage, thus a positive sign was expected. Therefore, reasons not discussed in this paper might play a role. Possibly, overconfidence behaviour manifests itself in other areas than investment. For ex-ample, in overestimation of the life time of machinery (i.e., aeroplanes), leading to postponement of investment.

(27)

Summarising, the overconfident measures suggest lower levels of investment plans. This may be because their internal funds are insufficient to fuel their capital expenditure needs, combined with a reluctance to raise capital on the equity- and bond markets, or they overestimate the life time of the aeroplanes. This prediction is further tested inSection 5. Whatever the explanation, based on these results, the first hypothesis is disproven: CEO overconfidence does not lead to higher levels of total investment. It is, however, interesting to see which CEO type makes the more accurate forecast. Therefore, the following section will examine actual investments and deviations from the investment plans.

(28)

28 R esul ts – MS c Thesis S tijn S mi t Table IV

Do overconfident CEOs plan more optimistically?

Dependent Variable = Total planned investment

Pooled OLS Random Effects

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Holder67 -0.53** -0.44** -0.55* -0.76** (0.203) (0.206) (0.302) (0.291) Net Buyer 0.11 0.26 (0.153) (0.246) Longholder -0.33** -0.36 (0.147) (0.312) Tobin’s Q 0.82* 1.34* 1.19* 1.04* 1.76*** 0.92 1.26 1.37* 1.19 2.24** (0.468) (0.737) (0.642) (0.545) (0.589) (0.549) (0.928) (0.772) (0.689) (0.819)

Normalised Cash Flow 9.24*** 11.63*** 10.42*** 10.33*** 19.25*** 10.88*** 12.28** 12.06*** 12.31*** 23.89***

(2.850) (3.462) (3.424) (3.201) (4.218) (3.216) (4.236) (3.396) (3.940) (4.864)

Normalised EBITDA 4.44*** 4.01*** 3.98*** 4.29*** 5.08*** 5.14*** 4.83*** 4.13*** 4.92*** 4.36***

(0.921) (1.155) (1.086) (0.937) (1.336) (1.263) (1.191) (1.222) (1.280) (1.326)

Tobin’s Q × Cash Flow -7.97*** -9.77*** -9.09*** -8.87*** -16.49*** -9.32*** -10.41*** -10.29*** -10.42*** -20.42***

(2.470) (2.953) (2.988) (2.759) (3.692) (2.697) (3.516) (2.879) (3.249) (4.059) Log(Size) -0.23*** -0.29*** -0.28*** -0.22*** -0.14 -0.21** -0.27*** -0.27** -0.20* -0.16 (0.050) (0.065) (0.071) (0.052) (0.280) (0.092) (0.087) (0.106) (0.101) (0.475) Stock Return -0.21 -0.29 -0.20 -0.22 -0.23 -0.31 -0.42 -0.30 -0.32 -0.34* (0.175) (0.185) (0.197) (0.175) (0.172) (0.229) (0.244) (0.237) (0.219) (0.171) Return Volatility (%) 1.25 0.87 1.89* 1.67 0.28 1.93 1.34 3.81* 2.55 0.23 (0.914) (1.092) (1.026) (1.018) (0.900) (1.815) (2.664) (2.074) (1.960) (2.436)

Additional Control No No No No Yes No No No No Yes

Year Fixed Effects No No No No No Yes Yes Yes Yes Yes

Observations 152 109 128 152 104 152 109 128 152 104

Adjusted R-squared 0.3018 0.3474 0.3042 0.3165 0.5057 0.2850 0.3567 0.3121 0.3039 0.5660

Note: The dependent variable is the sum of investment plans up to 5 years (i.e.,P5

k=1It +k) normalised by Property, Plant & Equipment at the beginning of the year. Hol d er 67 is a dummy equal to 1 if the

CEO holds fully vested options which are at least 67% in-the-money at least twice during the sample period, starting the first year he shows this behaviour. Net Bu yer is a dummy equal to 1 if the CEO purchases more stock than he sells for more than 60% of his tenure at the firm, starting the first year he shows this behaviour. Long Hol d er is a dummy equal to 1 if the CEO ever holds in-the-money options until expiration. St ock Ret ur n is one-year monthly buy-and-hold return in decimals. Ret ur n V ol at i l i t y is the lagged annual standard deviation of the stock’s returns in percent. The additional controls are Ag e, Tenur e, C EO & P r esi d ent , Ow ner shi p, V est ed Op t i ons, Oi l Ret ur ns and Sal es. Clustered standard errors are given in parenthesis. *, ** and *** indicate significance at 10%, 5% and 1%, respectively.

(29)

4.2 Objectivity of investment plans

As discussed above, overconfident CEOs are found to have lower investment plans. This subsection will verify whether these lower investment plans also lead to lower forecast errors. It is hypothesised that overconfident CEOs overestimate their future sales leading to over-investment and order can-cellations when actual sales are revealed. These cancan-cellations should be a large driver of investment differences and affect differences through the overconfidence measure.

First, actual capital expenditures are examined. Combining the results from the previous section with results on actual investment should provide suggestive evidence for the overconfidence effect on deviations from the committed investments. For example, if it is found that overconfident CEOs have larger capital expenditures than non-overconfident CEOs then – combined with their lower tar-gets – this suggest larger deviations. The results of these regressions are summarised inTable V.

Second, the suggestive evidence is verified with a regression on the cumulative difference (i.e., the sum of differences between the actual investments and the one-year forecast at time t − 1, the two-year forecast at time t − 2, etc.). The dependent variable is the normalised aggregate difference to account for data inconsistencies; large positive estimates are offset by large negative estimates when they are aggregated. The results are presented inTable VI.

The results on actual investments are less consistent than those discussed earlier. The Hol d er 67 measure suggest that overconfidence increases the ratio of capital expenditures to capital by either 0.06%-point or 0.04%-point, dependent on the regression technique. For these estimates the 95% confidence interval for the effect is either between 84.5 and 449.6 million USD or 3.6 and 352.4 mil-lion USD, based on a median capital level of 4.45 bilmil-lion USD. Compared to the median capital ex-penditures of 566 million USD this represent an increase of at least 15%. This is believed to be eco-nomically significant. Furthermore, the findings are statistically significant at the 1%- and 5%-level, respectively. The estimate slightly drops when additional controls are included. Overall, it provides strong suggestive evidence that the forecast error will be larger for overconfident CEOs compared to non-overconfident CEOs.

However, the other measures for overconfidence are statistically insignificant in both regression forms. This suggests that overconfident CEOs are, statistically speaking, investing the same (per dol-lar of capital) as non-overconfident CEOs. This raises doubt on the earlier findings, but is not unlikely. In Malmendier & Tate (2005) the measures were also inconsistent in estimating the effect, only the interactions with cash flow were. These are not included because this paper does not intend to test for cash flow sensitivity, but rather the actual effect.

Unlike previous literature, the relation between Tobin’s Q, cash flow and actual capital expendi-tures is barely significant. This may be a result of the small sample size, which leads to relatively high