Comparing the duration of supply chain link inattention for positive and negative earnings surprises

(1)

Comparing the duration of supply chain link inattention for positive

and negative earnings surprises.

ABSTRACT

In this thesis the duration of investor inattention to supply chain links is analyzed. This is done by comparing the duration of inattention to positive earnings surprises with negative earnings surprises. In line with the hypothesis a significant negative influence of negative earnings surprises on the duration of inattention is found. The Fama-French 3-factor model suggests 10.1% more under reaction for negative surprises, significant at the 1% level. The Carhart 4-factor model estimates a 7.79% higher under reaction for negative surprises, significant at the 5% level. Furthermore, the duration of inattention to Friday announcements is investigated. As expected, the Friday announcements seem to increase the duration of inattention, the market adjusted model suggests a 9.56% higher under reaction for Friday announcements, significant at the 10% level. Additionally, a significant (5% level) negative influence of institutional ownership on the duration of inattention is found, indicating that a 1% increase in institutional ownership decreases inattention to earnings surprises by 0.117%.

Name: Bas Geenen Student number: 6156533

Supervisor: dr. L. Zou Date: 29-06-2016

(2)

i

Statement of Originality

This document is written by Student S.W. Geenen 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)

ii

1. Introduction

Due to the increasing amount of attention that is paid to behavioral aspects of economics and the significant influence of behavioral factors on economic outcomes, a behavioral topic will be investigated in this thesis. While there are many economic theories that predict how economic phenomena are supposed to occur it has more recently become clear that behavioral factors, such as the limited attention hypothesis, play a significant role in the choices investors make. An example, on which will be elaborated further in the literature section, is the case where a portfolio based on the limited attention hypothesis, made a significant annual return of 18.6% per year (Cohen and Frazzini, 2008).

This thesis will investigate the duration of investor inattention to supply chain links and will focus on the difference between positive and negative earnings surprises. The inattention hypothesis has been introduced by Kahneman and Tversky in 1973. Their hypothesis states that due to limited attention, people are not able to process all available information. This hypothesis is not compatible with the efficient market hypothesis developed by Fama (1973). Many researchers have confirmed the limited attention hypothesis in different settings. For example, with respect to announcements Huberman and Regev (2001) find that the source in which the information is published plays a large role. Furthermore, Dellavigna and Pollet (2009) show that the weekday on which earnings per share information is released also plays a significant role, announcements on Friday seem to receive significantly less attention. Cohen and Frazzini (2008) introduce (subtle) customer-supplier links as a proxy for inattention. Since suppliers are obliged to publish the information of their largest customers on their balance sheets, this information is publicly available. They prove that the subtle customer-supplier links are often not being observed by investors by creating a portfolio in which they short sell the supplier stock after negative (customer) returns and buy the supplier stock after positive customer returns. They refer to the returns of this portfolio as the customer momentum portfolio.

Even though much research has been done into the effect of inattention, research regarding the duration of inattention is limited. Especially when we consider firm level data. In this thesis the customer-supplier links will be used as a proxy for inattention. The focus will be on the duration of the inattention for positive and negative earnings per share surprises. Whereas in the analysis by Cohen and Frazzini (2008) various shocks are analyzed, in this thesis the separate effect of EPS announcements will be analyzed.

(5)

2 Unlike in the study by Cohen and Frazzini the Customer stocks will be assigned to the ‘positive earnings surprise group’ or the ‘negative earnings surprise group’ based on the nature of the earnings surprise. Cohen and Frazzini (2008) divided the customer stocks into quintiles and arranged them based on the returns that followed after a certain shock, regardless of the nature of the shock. Taking the nature of the shock into account in this thesis should benefit the causal interpretation1

of the results since positive returns that follow after negative shocks are filtered out.

The separate positive and negative EPS announcement returns can be distinguished because the returns of the suppliers are analyzed separately. Therefore, the amount of attention paid to positive shocks can be compared to the attention paid to negative shocks. This is interesting because the study by Karlsson Loewenstein and Seppi (2009) suggests that (mostly private) investors will pay less attention to negative news than to positive news, a phenomenon appropriately called ‘the ostrich effect’. While the authors use a completely different proxy, namely the amount of logins of private investors into their Scandinavian brokerage accounts under varying market conditions, the ostrich effect may also be found by measuring the abnormal returns of supplier companies subsequent to negative customer earnings surprises.

After obtaining the abnormal returns for both positive and negative earnings surprises the duration of the inattention will be measured using the under reaction coefficient (URC). This measure, also addressed in the study by Cohen and Frazzini (2008) measures the size of the effect of a shock that has already been incorporated into the (supplier) stock price divided by the total effect over the entire event window. When this measure will be higher for positive shocks than for negative shocks it will indicate that investors have shorter duration of inattention to positive shocks. The study by Hou et al., (2009) suggests that reduced attention, for example because of the ostrich effect (in negative events), may lead to more under reaction. Therefore, negative earnings surprises are expected to have longer duration of inattention, hence more under reaction, than positive earnings surprises.

Besides the nature of the earnings surprise shocks, the influence of Friday announcements and investor type on the duration of inattention will be tested. It is expected that the reduced attention for Friday announcements will cause more under reaction and therefore lower URC’s. For the investor type hypothesis, it is expected that the reduced bias from professional investors will cause less under reaction and therefore higher URC’s.

1_{By stating that the causal interpretation of the results should be better it is not claimed to prove a causal}

(6)

3 This leads to the following research question:

Does the duration of investor inattention to earnings surprises depend on the nature of the earnings surprises, day of the week and investor type? Measuring inattention on the firm level through supply chain links.

The remainder of this thesis will be structured as follows: in section two the literature review will be presented, herein, the inattention hypothesis and the efficient market hypothesis will be addressed, together with relevant research with respect to announcements and supply chain links. Furthermore, the return models which are used in this thesis will also be discussed in this section. Section three sets out the methodology, in this section the precise steps that are taken to retrieve the final dataset will be discussed and the model that will be tested is explained. The fourth section will provide a closer look on the final data and descriptive statistics. Section five sets out the results and compares them to the existing literature and the theories provided in the second section. Robustness checks will be performed in the sixth section, among these will be two alternative models used for the returns and a model with robust standard errors. In the seventh section the conclusion is presented, and the study is finalized with the eight section, in which the limitations of the current research are discussed and suggestions for further research provided.

(7)

4

2. Literature Review

2.1.1 Inattention hypothesis

The framework for inattention has been laid by Kahneman and Tversky (1973). The authors find that through the limited cognitive resources of individuals, individuals will not be able to process all information at the same time. Therefore, individuals will most likely not be able to gather and understand all information. A study that focusses more on the inattention of investors is the study by Merton (1987). Merton shows that investors typically focus and only actively follow a small number of stocks. He suggests that gathering information about stocks requires resources, and that the cognitive resources, as suggested by Kahneman and Tversky (1973), are limited. Merton also finds that there should be no bias of investors buying attention-grabbing stocks since investors only monitor their ‘chosen’ stocks.

The study by Hong and Stein (1999) uses the inattention hypothesis introduced by Kahneman and Tversky and applies it into an economic setting. In the model by Hong and Stein (1999) the translation of new information into stock prices is studied. The authors investigate a model that explains both the under reaction to attention grabbing events and the overreaction that follows after news has been picked up by traders. They explain the under reaction by inattention and the gradual (non-direct) interpretation of the information. The overreaction that follows is caused by momentum traders that have picked up the attention grabbing event and (over) exploit the new information.

In the study by Shapira and Venezia (2001) the variation in inattention among different groups of investors is studied. In their research they find that institutional investors have been trained to reduce their behavioral biases such as inattention and therefore suffer less from these biases. Private investors seem to display a larger degree of inattention and a more severe bias in their information processing. Barber and Odean (2008) confirm these findings. Furthermore, Barber and Odean (2008) focus on the tendency of investors to buy attention grabbing stocks (that are doing well) and less frequently sell attention grabbing stocks that are underperforming. Institutional investors suffer less from this tendency since they are able to use computer programs to distribute their attention and therefore reduce their inattention. Moreover, institutional investors will hold more stocks which they can possibly sell and will therefore have less asymmetric selling behavior than individual investors. Moreover, Courwin and Coughenour (2008) find that institutional investors also suffer from attention constraints. They confirm the findings that individual investors are primarily biased by their attention constraints. But they also find a significant impact of inattention on

(8)

5 professional investors in periods of increased activity on the market. They conclude that in periods of increased activity on the market professionals also focus on the most active stocks.

In the study by Karlsson Loewenstein and Seppi (2009) yet another behavioral bias is found, which the authors call selective attention. In their research they examine the amount of times investors log into their brokerage accounts, and find that when the markets are performing badly the amount of logins is significantly lower than when markets are performing well. They call the selective inattention when markets are underperforming the ostrich effect.

Hou et al., (2009) test whether the amount of attention paid, influences the under reaction that occurs after certain events. This can be expected because when less attention is being paid, the reaction that will follow (based on the event) will occur more slowly. Hence the under reaction driven earnings momentum should be larger in cases of low attention. Therefore, in their study Hou et al., (2009) use trading volume as a proxy for attention. High trading volume indicates that more attention is paid to the market, since investors need to pay attention to be able to trade. When this effect is combined with the ostrich effect, which is also suggested in the paper by Hou et al., (2009), we can expect the under reaction to maintain longer in cases of negative events; for example, for negative earnings surprises compared to positive earnings surprises.

2.1.2 Inattention to (earnings) announcements

In the study by Huberman and Regev (2001), an example is given of investor inattention to announcements. In their study the exact same information about a possible cure for cancer was published twice, firstly in ‘Nature’, after which no stock price reaction of the manufacturer followed, and later in the ‘New York Times’, after which an immediate (positive) stock price return followed. Apparently the attention of the average investor was limited to the big (news)papers such as the New York Times. Regarding inattention to announcements, not only the source the information is announced in, but also the day on which the information is announced plays a large role. DellaVigna and Pollet (2009) indicate that Earnings Per Share (EPS) announcements made on Fridays have a 15% reduced direct response and a 70% higher delayed response compared to EPS announcements made on other weekdays. Furthermore, Hirshleifer et al., (2009) show that the amount of announcements on the same day plays a role. Because of the limited attention hypothesis investors will have less attention for EPS announcements when an increasing amount of EPS announcements is made

(9)

6 on the same day. Therefore, the price and trading volume will underreact when many announcements are made on the same day and the post announcement drift will therefore be larger.

A measure often used in comparing earnings announcements to their expected values (regularly analyst forecasts), is the Standardized Unexpected Earnings (SUE) measure. This measure is also used to indicate the extent of the earnings surprise. For example, in the study by Chan, Jegadeesh and Lakonishok (1998) evidence is found that investors highly value the SUE statistic to decide whether to buy or sell stocks. They also find that the reaction to these EPS forecasts relative to their actual EPS values is slow, which supports the inattention hypothesis.

2.1.3 Inattention to supply chain links

In line with the limited attention hypothesis and limited attention to announcements described above, Menzly and Ozbas (2007) investigated the amount of attention paid to subtle links with regard to stock returns. They perform their analysis on the industry level and find that stocks which are in economically related customer and supplier industries cross predict each other’s returns. The authors suggest a trading strategy that exploits the industry links by buying the supplier stock with the highest customer returns and selling supplier stocks with the lowest customer returns earns significant annual abnormal returns of 7%. Shahrur et al., (2010) also investigate C-S links with respect to inattention using industry level linkages. In their paper the international evidence of return predictability due to inattention is studied. Again, it is shown that the customer shock is slowly incorporated into the suppliers share price, supporting the inattention hypothesis.

Peng and Xiong (2006) also indicate that because of the limited attention of (individual) investors the investors will pay more attention to market- and sector-wide information than to firm-specific links. The authors call this phenomenon category-learning behavior. Since in this paper the focus will lay on the amount of attention paid to firm-specific supply chain links, it is expected that investors will pay little attention to these firm-specific details. Peng and Xiong (2006) give the example of the dotcom bubble in which investors started to overvalue all companies whose names ended on .com, instead of looking up the firm specific information.

In the article by Cohen and Frazzini (2008) the (in)attention of investors to customer-supplier links is measured on the firm level. Cohen and Frazzini use individual stocks and the

(10)

7 company specific links, which makes their result robust to both inter- and intra-industry effects. By measuring the stock price increase (decrease) due to shocks to the customer the stock price the change of the supplier stock is studied. An example given by Cohen and Frazzini is the ‘Coastcast Callaway’ case. Callaway, a large manufacturer of Golf clubs is one of the biggest customers of Coastcast. When Callaway loses 30% of its market value due to a shock it takes two months before this shock is translated into Coastcast’s stock price. The authors construct a portfolio in which they buy the supplier stock after positive customer stock returns and short the supplier stocks after a negative customer stock returns, they refer to this portfolio as the customer momentum portfolio. To choose which supplier stocks need to be bought monthly customer returns are used. At the end of the month the portfolio of customers is balanced and divided into quintiles, the supplier stocks of the customers in the highest quintile are bought and the stocks in the lowest customer quintile are sold. Using this approach Cohen and Frazzini find that their strategy yields a significant monthly abnormal return of 1.45% using the Fama-French 3-factor model. Even after correcting for the fourth factor added in the Carhart (1997) model, the monthly returns remain significant and large (1.37%). This supports the limited attention hypothesis (and subsequent under reaction) described above and contradicts the efficient market hypothesis that will be described in the next section.

In this thesis the focus will be on the duration (time lag) after which the customer earnings surprise is translated into the supplier stock price. A fundamental difference in the approach followed in this thesis compared by the approach by Cohen and Frazzini is the way the returns are categorized. Where Cohen and Frazzini divide their sample into quintiles after the occurrence of a variety of customer shocks, this thesis takes into account the nature of the (earnings surprise) shock. In the study by Cohen and Frazzini it can therefore be the case that a positive shock yields significant negative results and the customer is assigned to the lowest quintile (short) portfolio, regardless of the fact that the shock was expected to yield positive abnormal returns. In this thesis the earnings surprises of the customers will be divided into the ‘positive surprise’ and ‘negative surprise’ group regardless of the returns that follow. This will be done by comparing the expected earnings per share with the actual earnings per share. This division makes the results more robust for effects that are simultaneously occurring and influence the customer stock price returns.

(11)

8

2.2 Efficient market hypotheses

The efficient market hypothesis by Fama (1970) predicts that prices immediately reflect all available information. Therefore, the ‘Callaway-Coastcast’ example displayed above should not be able to yield abnormal returns on a constant basis since prices should always incorporate all available information.

Fama (1970) has distinguished three forms of the efficient market hypothesis, which vary in the matter in which the market is efficient. The most efficient form is the strong efficient market hypothesis. For this form all information should be incorporated into asset prices. This form also assumes that all private/inside information is automatically incorporated into stock prices and that therefore not even insiders are able to make abnormal returns. The second form of the efficient market hypothesis assumes semi-strong efficient markets. In the semi-strong form, it is assumed that all information that can be accessed publicly is incorporated into stock prices. Acting upon such information should therefore not yield abnormal returns systematically. This form of market efficiency is most interesting for this thesis since the supply chain link information on which the inattention hypothesis is tested is publicly available in the financial statements of the supplier companies. Paying attention to the supply chain links of the customer with its supplier and exploiting this information should therefore not systematically yield abnormal returns. The third form of market efficiency is the weak form efficiency. In the weak form the focus is on the past prices, therefore strategies which use past prices to acquire systematic profits should not be successful.

2.3 Return models

In this thesis several return models will be used to compare the influence of the customer earnings surprises on the subsequent suppliers’ abnormal stock returns. Therefore, the return models will be explained shortly in this section, starting with the capital asset pricing model.

The capital asset pricing model (CAPM) was developed to explain stock returns while using market risk factors. Since the assumption was that all other risks could be neutralized by diversification, the market risk factor was the only factor to take into account (Mossin, 1966). However, there seemed to be other factors influencing the returns that were not included in the CAPM model since certain market phenomena could not be explained by the model. These were for example the book-to-market effect, found by Stattman (1980) and the momentum effect found by Jegadeesh and Titman (1993).

(12)

9 In the adjusted CAPM model used by Achaya and Pedersen (2005) the CAPM model is extended to control for liquidity factors that influence the stock returns.

The 3-factor model by Fama and French (1993) is one of the models Cohen and Frazzini (2008) use to calculate the abnormal returns for the customer and supplier returns and will also be used in this thesis. The 3-factor model identifies three stock market factors. Firstly, the overall market factor and furthermore characteristics related to firms’ size and the book to market equity of the firm. Since the excess market returns found using this model are close to zero, Fama and French find that the 3-Factor model does a good job at explaining the average returns of the stocks in their samples.

The last model used in this thesis, also addressed by Cohen and Frazzini (2008) is the Carhart 4-factor model. The model, described in the article by Carhart (1997) is an extension of the 3-factor model described above. It additionally controls for the (one-year) momentum effect that was found by Jegadeesh and Titman (1993). Carhart finds that the model reduces the average pricing errors significantly compared to the 3-factor model.

2.4 Hypotheses development

The studies by Karlsson et al., (2009) and Hou et al., (2009) suggest that the inattention, and the subsequent under reaction to negative events will be larger than for positive events. Therefore, it is expected that negative EPS announcements will suffer more from under reaction compared to positive EPS announcements. In line with the findings by Dellavigna and Pollet (2009) it is furthermore hypothesized that announcements on Friday will yield more under reaction because the amount of attention paid to announcements on Fridays is significantly lower compared to other week days. The third hypothesis considers the institutional ownership factor that has been addressed by Barber and Odean (2008) and Shapira and Venezia (2001). Based on their findings it can be expected that firms with larger institutional ownership shares will suffer less from inattention and therefore have less under reaction.

(13)

10

3. Methodology

In this section the methodology used in this thesis will be discussed. First, the method to calculate the abnormal returns will be discussed for the return models used in this thesis. When the retrieval of the returns is discussed the construction of the Under Reaction Coefficient (URC) will be illustrated. Then the regression model will be explained and the interpretation of the influence of the X’s on Y will be given and compared to the hypotheses. Lastly, the used control variables will be discussed.

3.1 Calculation cumulative abnormal returns (CAR’s)

For this thesis an event study methodology will be applied using data on firms that are available in the WRDS Customer Segment database. The events of the event study that is performed will be the earnings per share announcements of the customers of supplier firms. The date on which the (customer) earnings are announced will therefore be t=0 in the event study. Subsequently the 90-day supplier abnormal returns will be estimated using the abnormal return formula (1) displayed below.

(1) ARi,t = Ri,t – ERi,t

• Where: ARi,t are the abnormal returns of firm i on day t

• Ri,t are the actual returns of share i on day t and

• ERi,t, are the normal or expected returns of share i on day t

• Date 0 is the announcement date of the EPS announcement

The abnormal returns will be cumulated (CAR’s) and calculated for each additional day until 90 days. Hence the CAR for each day after t=0 will be available. Therefore, it is implicitly assumed that the effect of the EPS announcement will not last longer than 90 days. The largest event window has been chosen because the earnings announcements are done quarterly, this way there will be less bias from the subsequent earnings announcements. This methodology will be followed for both the supplier and the customer stock returns. By doing so the size of the effect translated into the supplier stock returns can be estimated.

The abnormal returns described above are computed for four different return models, whose theory was briefly introduced in the literature section. The returns for all four models are computed in the event study tool provided by WRDS, called ‘event study by WRDS’.2

2_{Additional information about requirements and assumptions in the models used can be found on the WRDS}

(14)

11 The output of the stock returns (and the CAR’s) per firm are then merged back into the dataset based on the cusip identifiers and announcement dates.

3.2 Primary independent variable: Earnings surprise type

After the returns for the various models have been computed for both the suppliers and customers, the dummy variable for positive and negative earnings surprises is computed using the actual and forecasted EPS data. When the expectations exceed the actual value of the EPS the dummy is assigned a negative surprise and vice versa.

The current approach differs from the approach used in the study by Cohen and Frazzini (2008). They use various types of shocks (not only EPS announcement) and assign customers to different quintiles based on the return that follows after the earnings surprise. In this study the nature of the earnings surprise (positive or negative) is taken into account by assigning the suppliers returns to the positive or negative surprise group based on the nature of the surprise. This should contribute to the causal interpretation of the results.

3.3 Construction of URC’s

To estimate the duration of the inattention the under reaction coefficient (URC) measure is used, this measure is also used in the paper by Cohen and Frazzini (2008). Cohen and Frazzini use this measure to illustrate the under reaction of investors for different dates in the event window. In this thesis the URC measure will be used as the dependent variable. Therefore, differences with regard to the duration of the inattention can be interpreted. The supplier URC’s will be analyzed for positive and negative EPS surprises, hence the effect of positive/negative (customer) EPS earnings surprises on the duration of inattention can be analyzed. The URC measures the amount of the CAR’s (between t = 0 - 90) which has been realized divided by the total CAR for the chosen event window. When the entire effect of the inattention has been realized the measure will equal 1. When this is not yet the case the measure will be between 0 and 1. When the measure is higher than 1 we can speak of overreaction. This measure will be computed for each firm and for each day in the event window.

The formula for the (firm specific) URC’s is the following:

(2) 𝑼𝑹𝑪𝒊,𝒕= _𝐂𝐀𝐑𝐂𝐀𝐑𝒊,𝒕

𝒊,𝟗𝟎

• Where: CAR1,2 is CAR for firm i at (event)time t

(15)

12

3.4 The model: testing the duration of inattention for earnings surprise types

Using the firm specific URC’s, a panel data regression will be performed with the URC’s as the dependent variables. This regression will be performed for all four return models. The Fama-French and Carhart model will be used for the main results and the market and market adjusted model will be added for robustness. The panel data regression on the firm specific URC’s will also be performed for the customer firms to be able to compare the customer duration of inattention with the supplier’s values in the robustness section.

The regression model (3) is provided below:

(3) 𝑼𝑹𝑪𝒊,𝒕 = 𝜷𝟎+ 𝜷𝟏𝑵𝒆𝒈𝑺𝒖𝒓𝒑 + 𝜷𝟐𝑰𝑶𝒔𝒉𝒂𝒓𝒆 + 𝜷𝟑𝑭𝒓𝒊𝒅𝒂𝒚 + 𝜷𝟒𝑪𝒓𝒊𝒔𝒊𝒔 +

𝜷_𝟓𝐥𝐧 (𝒂) + 𝜷_𝟔𝑹𝒆𝒍𝒔𝒂𝒍𝒆𝒔 + 𝜷_𝟕𝑳𝒆𝒗 + 𝜷_𝟖𝑺𝑼𝑬 + 𝜷_𝟗𝑺𝑰𝑪 + +𝝐_𝒊,𝒕

This regression will show how the under reaction coefficient is influenced by the independent variables. When the URC’s are higher, the duration of the inattention effect will be smaller3

. Therefore, positive coefficients will indicate that the duration of the inattention is decreased by the coefficient. From the studies by Hou et al., (2007) and Karlsson et al., (2009) it can be expected that the 𝛽X, the influence of negative surprises on the URC’s will be

negative. Hence for negative surprises the under reaction will be larger and therefore the duration of the inattention will be greater for these surprises.

To correct for possible endogeneity a correlation test of the regressors will be performed, when the correlation is high the independent variables will be excluded from the analysis. Furthermore, all the regressions will be performed using a firm fixed effects model. Using firm fixed effects controls for omitted variables that differ between firms but are stable over the time period that is analyzed (Stock and Watson, 2012). Whether the use of firm fixed effects is justified will be discussed in the robustness section.

3.5 Secondary independent variables

The day of the week dummy (Friday) is added to test the influence of Friday surprises on the duration of the inattention. By only including Friday the effect of Friday announcements to announcements on other days of the week is shown. The study by Dellavigna and Pollet (2009) found that the day of the week has a significant influence on the amount of attention

3_{Please not the somewhat contradictory interpretation of the URC measure. When more attention is paid a}

(16)

13 paid to shocks. Since Fridays receive significantly less attention than other days, it can be expected that the URC’s will likely be smaller on Fridays compared to other weekdays. This variable will therefore provide the results for the second hypothesis.

The IOshare variable measures the percentage of shares held by the institutional investors. The studies by Odean (2008) and Shapira and Venezia (2001) suggest that high institutional ownership reduces the inattention bias. The URC coefficient is therefore expected to be positive, since a higher value for the URC’s means the duration of the under reaction is shorter compared to firms with lower institutional ownership shares.

3.6 Control variables

The crisis dummy is added to the model because in the study by Lumsdaine (2011) it is found that the amount of attention paid to stocks may have changed in the build-up of the financial crisis. Therefore, the dummy is specified to be 1 after the beginning of the financial crisis, for which the fall of the Leeman Brothers is taken as the start date (15-09-2008). It can be expected that this coefficient will be positive since the financial crisis has increased the amount of attention paid to stock markets (Lumsdaine, 2011). The second control variable used in the model is the (supplier) firm size variable, which is measured by the natural logarithm of the assets of the supplier firm and depicted as ‘ln(assets)’ in the equation. From the study by Hirshleifer et al., (2009) it can be expected that larger firms will draw more attention and therefore have less under reaction. The URC’s can therefore be expected to be closer to one. The coefficient 𝛽Y will therefore be expected to be positive. Thirdly the relative

sales measure will be used as a control variable, this is also done in the study by Cohen and Frazzini (2008). This variable is included because when the relative amount of sales to a customer is higher, the shock is more likely to affect the supplier. This variable is constructed by dividing the sales to the specific customer by the total supplier sales for the specific period. The leverage control variable is included in the model because this is also done in the study by Dellavigna and Pollet (2009), which also investigates the inattention hypothesis with respect to stock returns, but focusses on the Friday announcements. Furthermore, the size of the earnings surprise will be included into the model as a control variable. In the study by Latane and Jones (1977) it is found that higher values of the SUE lead to larger reactions, hence less inattention in the stock price. Lastly, a control variable that takes the supplier industry into account will also be added to the model since it can be expected that the amount of attention paid to EPS shocks may differ between industries, the industry specific control

(17)

14 variable is also used in the study by Dellavigna and Pollet (2007). The first digit of the SIC code will therefore be used to control for the main industries of the suppliers in the sample.

4. Data and Descriptive statistics

The most important data restriction for this thesis comes from the Customer Segment database from Compustat. In this database the firm specific Customers supplier links are available, the proxy used for inattention in this thesis. Since these links are crucial for the current analysis the Customer Segment database will be taken as a starting point for data collection. The analyzed period is January 2002 until December 2015, this period includes the financial crisis, so that a possible effect of the crisis on the duration of inattention can also be estimated.

4.1 Sample Selection

Due to regulation SFAS NO. 131 suppliers need to report all customers that account for more than 10% of their sales, therefore these suppliers report this information on their financial statements. This information about the supply chain links is available in the Compustat Customer Segment database. This database unfortunately does not provide any identification numbers for the customer firms. Therefore, this dataset is merged on customer name with the Crisp Compustat Merged database. In order to match as many customers as possible, a do-file is used to generalize customer names4_{. The adjustment increases the amount of matches} because many customers have for example ‘limited’ in the Customer Segment database and ‘ltd’ in the CCM database. The choice to merge on name with the CCM database and not the Compustat Quaterly database is made because merging with the CCM database provides almost twice as many matches. After having merged on name with the CCM database 10,753 customer supplier link observations remain in the sample.

The database with the customer supplier links is now merged with the Institutional Brokers' Estimate System (I/B/E/S) database that provides the earnings forecasts, actual EPS values, and the announcement dates that are required to calculate the (customer specific) earnings surprises. The dummy variable for negative surprises is created and is assigned the value 1 when the expected earnings are higher than the actual earnings. For the expected earnings the mean value of the quarterly predictions is used. After the merge with I/B/E/S 9,553 customer supplier link observations remain in the sample.

(18)

15 Since the supplier information provided in the Customer Segment database is limited, a merge is performed with the CCM database to retrieve additional information about control variables such as total assets for the supplier firms. Furthermore, a threshold is set with respect to the minimum percentage of sales to the customers. This threshold is set at 5%. When the customer accounts for less than 5% of the supplier sales the shock is assumed not to influence the supplier stock returns5

. The Standardized Unexpected Earnings measure (SUE) is also set at 5%, this way very small earnings surprises are excluded from the sample6_{. After} the merge and applying the thresholds, 4,397 customer supplier link observations remain in the sample.

Subsequently the dataset is merged with the 13f database from Thomson One7 . This dataset is used to retrieve information about institutional ownership in the supplier firms. Information on the total shares held by the biggest institutional traders is also available in this database and therefore a ratio of the total institutionally held shares divided by total shares is added to the dataset.

The supplier stock returns are then retrieved using the ‘Event study by WRDS’ tool on the WRDS website8

. The daily stock returns and abnormal returns are retrieved using this tool for the four return models discussed in the literature review section. The returns are merged based on the announcement dates and the supplier cusips. The dataset now has 91 return observations for t=0 and the 90 days after the customer earnings surprise date, and contains 4 different return variables per observation. The customer returns are retrieved using the same methodology and merged into the dataset. After this merge 268,164 returns are available for 2,9479

customer supplier link observations.

5_{This threshold excludes 1620 observations out of the dataset.} 6_{This threshold excludes 3565 observations out of the dataset.} 7_{Access to the Thomson One 13f database was granted by T. Jochem.}

8_{The specification of the ‘Event Study by WRDS’ can be found on:}

https://wrds-web.wharton.upenn.edu/wrds/ds/wrdseventstudy/index.cfm?navId=355

(19)

16 A snapshot from selected variables in the dataset is displayed below in table 1.

Table I: Snapshot of dataset

Notes: cusipc represents the 8dgt. cusip code for the customer firm. Cusips8 represents the 8dgt cusip code for the linked supplier firm. Anndats is the date on which the actual earnings were announced and is the start of the event window. SUE is a measure of the size of the earnings surprise defined in the literature section. Atq is the total assets of the supplier firm for the relevant quarter. Ltq is the total liabilities for the supplier firm for the relevant quarter. Etime is the event time, increasing with 1 every day after the start of the event window. The return measure represents the daily stock return. The 4 abnormal return measures measure the daily abnormal return per model. Instshare measures the percentage of shares held by institutional investors as a fraction of total shares.

cusipc cusips8 anndat SUE Negs atq ltq etime ret ar3 ar4 arma arM instsh

17004010 36866T10 31-Jan-06 1.12 0 510.23 62.863 0 -0.005 -0.007 -0.007 -0.003 -0.002 0.908

17004010 36866T10 31-Jan-06 1.12 0 510.23 62.863 1 -0.024 -0.027 -0.027 -0.026 -0.026 0.908

17004010 36866T10 31-Jan-06 1.12 0 510.23 62.863 2 -0.020 -0.008 -0.008 -0.011 -0.006 0.908

In table I a selection of the available variables is depicted. From left to right the Customer identifier and its respective supplier are displayed. The ‘anndats’ variables provide the date on which the actual earnings where announced and functions as the start of the event window (t=0). The SUE measures how large the deviation from the projected earnings is and the ‘etime’ variable counts the days of the event window (until 90 days). The ‘ret’ column provides the returns for the supplier for each day and the four columns to the right give the abnormal returns for the four return models discussed in this thesis. Lastly the ‘instshare’ variable measures the amount of shares held by institutional investors as a percentage of total shares.

The under reaction coefficients are constructed by following the described methodology in the previous section. The URC’s are then winsorized because the ratio of the CAR’s takes on extreme values when the numerator is very small. These extreme values, created by the fracture in the dependent variable, create outliers. Just as in the study by Jegadeesh et al. (2006) winsorizing is done at the 2.5% level. A description of the values is given in the next section.

4.2 Descriptive statistics

In this section the dependent and independent variables will be described. Firstly, the variables in the main regression model will be described. Subsequently the sub-samples of negative and positive earnings surprises will be tabulated and compared.

(20)

17

4.2.1 Main description

In table II the descriptive statistics for the main dependent and independent variables are provided. These statistics include all observations, and therefore give a summary of both the positive and the negative earnings surprises. The dummies for the supplier industry and Friday are excluded from the descriptive statistics. For the first three models the average CAR’s are negative, in the market adjusted model the average CAR becomes positive. For all return models the values lie between -5.50% and 3.2%. The market adjusted model provides the lowest range in CAR’s since the values are between -1.77 and 2.42. The 3-factor model provides the broadest range, since the returns differ between -5.42 and 2.9310

.

When the URC’s are considered, the mean values lie between 0.45 and 0.49. This means that for the four models considered on an average day, the effect of the earnings surprise that has already been incorporated into the stock prices lies between 45 and 49 percent. In table II the winsorized values of the URC’s are used since the unwinsorized values led to extreme values. Especially when the numerators where very close to zero and the denominators where relatively large outliers were created.11

The IOshare variable has a mean of 0.62, indicating that for the average supplier 62% of its shares are held by institutional investors. What stands out is that the maximum of IOshare is higher than 1. These observations have been excluded from the final sample. The percentage sales variable is also taken into account; it can be seen that the average customer accounts for 20.57% of the sales of the supplier. Since the threshold for the minimum amount of sales to a customer is set at 5% the minimum value is 0.05. For the SUE measure, that estimates the impact of a surprise, the threshold is set at 0.05 as well.

10_{Please note that the descriptive statistics for the customer CAR’s and URC’s can be found in appendix III. As}

expected the values for the customer URC’s are higher than the supplier URC’s. This can be explained by the fact that investors will pay more attention to earnings announcement surprises of the customers directly than through the supply chain links.

11_{In appendix II the unwinsorized values of the URC’s are provided. Winsorizing is done at the p(0.025) level.}

To increase the robustness of the analysis a lower winsorizing level of the dependent variable will be discussed in the robustness section.

(21)

18

Table II: Descriptive Statistics

Notes: Descriptive statistics of the main dependent, independent and CAR’s in the sample. For both the CAR’s and the URC’s the four return models are described.

Dependent vars mean p50 sd min max count

URC3f .4920527 .5038824 1.288578 -3.653322 4.703781 268,164 URC4f .4723498 .4985035 1.163562 -3.245937 3.987447 268,164 URCM .4564527 .5001627 1.278208 -3.91196 4.225025 268,164 URCMA .4948552 .5151011 1.370068 -3.869138 4.765556 268,164 Independent vars IOshare .6236953 .6837407 .3145714 2.27e-06 1.793189 251,816 Ln(a) 6.369548 6.326611 1.95214 .0516432 12.26881 268,073 Perc. sales .2057767 .155 .1599744 .05 1 268,073 SUE 1.115926 .3025 5.261181 .05 129.96 247,204 Leverage .4691836 .4407663 .3207699 0 6.72462 268,073 CAR’s CAR3F -.0010529 -.0019832 .2513381 -5.423238 2.922209 268,164 CAR4F -.0035664 -.0028912 .2467272 -5.509147 2.513082 268,164 CARM -.0014747 -.0025283 .2523608 -5.081114 3.136619 268,164 CARMA .0151683 .0054191 .2146181 -1.777187 2.416705 268,164 Observations 268,164

In appendix IV the correlation coefficients of the variables are provided. Analysis of the correlation shows that the correlations are low. No correlations of higher than 0.2 are found, except for the correlation of IOshare with the logarithm of assets. The value of the coefficient is 0.43 and can be considered low. Therefore, no bias from multicollinearity is expected (Stock & Watson, 2012).

4.2.2 Comparison of positive and negative surprises

In this section the differences in descriptive statistics between the two sub-samples; the positive and the negative earnings surprises will be discussed.

The CAR’s are primarily negative for the return models in case of negative shocks and positive in case of positive shocks. The average CAR’s range between 0.82% and 1.91% for positive surprises and between -0.88% and 1.3% for the negative surprise sub-sample. This was also expected since negative (customer) shocks are expected to cause negative (customer and) supplier CAR’s through the customer supplier links. What stands out is that the market adjusted model finds a positive mean CAR for negative events. Despite the positive value for the market adjusted model in the negative panel the values are still lower compared to the mean of the market adjusted model in the positive panel (0.019%) .

For the URC’s the values differ between 40-51% across the sub-samples and models indicate that on average, on an average day in the event window, 40-51% of the effect of the

(22)

19 earnings surprise is incorporated into the stock prices. It can furthermore be seen that the values are higher for the positive surprise sub-sample (47%-51%), compared to the negative sub-sample with average values between 40%-47%, suggesting that the average amount of under reaction is larger in case of negative surprises. This also corresponds with the expectations since positive surprises are expected to draw more attention and subsequently have URC’s closer to one. For the negative surprise sub-sample more observations are available; 188,326 of the total 268,164 observations are negative. Despite the larger negative sub-sample, the samples is expected to be comparable since the positive subsample is sufficiently large with nearly 80,000 observations.

Table III: Descriptive Statistics Positive surprises

Notes: A comparison of the positive surprise and the negative surprise subsample has been set out in tables III and IV. What stands out is that the market adjusted model provides a positive mean in the negative surprise panel.

Dependent vars. Mean p50 sd min max count

URC3f .5167639 .5129648 1.304142 -3.653322 4.703781 79,838 URC4f .4937649 .5066033 1.173817 -3.245937 3.987447 79,838 URCM .4762652 .507999 1.292484 -3.91196 4.225025 79,838 URCMA .5040529 .5225413 1.379363 -3.869138 4.765556 79,838 CAR’s CAR3F .0150508 .0033978 .2602005 -2.008096 2.272694 79,838 CAR4F .0082855 .0002557 .2540484 -1.952764 2.16611 79,838 CARM .0158617 .0039952 .2622605 -2.061372 2.297699 79,838 CARMA .0191241 .007454 .2296349 -1.742554 2.045527 79,838 Observations 79,838

Table IV: Descriptive Statistics Negative Surprises

Dependent vars. Mean p50 sd min max count

URC3F .4337626 .4806181 1.249171 -3.653322 4.703781 188,326 URC4F .4218346 .4785222 1.137419 -3.245937 3.987447 188,326 URCM .409718 .4806752 1.242642 -3.91196 4.225025 188,326 URCMA .4731592 .4976763 1.347648 -3.869138 4.765556 188,326 CAR’s CAR3F -.0078798 -.0045474 .2471695 -5.423238 2.922209 188,326 CAR4F -.0085908 -.0041943 .2433835 -5.509147 2.513082 188,326 CARM -.0088242 -.0054058 .2476794 -5.081114 3.136619 188,326 CARMA .0134913 .0045757 .2079028 -1.777187 2.416705 188,326 Observations 188,326

4.2.3 Illustration of the URC’s

A graphical illustration of the under reaction coefficients is provided in graph I below. A distinction between the two subsamples, the positive and the negative earnings surprises is shown by the two distinct lines. The upper, blue line represents the URC’s of the positive

(23)

20 surprises. These are, as also illustrated in the descriptive statistics above, higher for the average day in the event window (on the x-axis) than the negative surprises. The negative surprises are represented by the bottom red line. Until day 10 in the event window the URC’s between the subsamples stay close to one another. From day 10 until the end of the event window the positive surprises seem to suffer less from under reaction than the negative surprises since the value for the positive surprises is closer to one. In the following section the regression will be performed to evaluate whether this graphical difference is significant.

Graph I: An illustration of the URC measure for positive and negative surprises

Notes: The positive URC’s, depicted in blue and the negative URC’s depicted in red. On the x-axis the event time (in days) relative to the start announcement date of the actual EPS. To be able to plot the URC’s provided in this graph the average URC’s have been calculated for both the positive and the negative surprise group, this has been done for each day in the event window.

5. Results

5.1 Expected results

In this section the influence of negative and positive earnings surprises on the duration of the inattention is discussed. This will be done for four return models which are discussed in the methodology and literature review sections. The first two regression models that are run, are the Fama 3-factor model and the Carhart 4-factor model. These models are also used in the study by Cohen and Frazzini (2008) and are therefore most comparable with their results12

.

12_{Please note that Cohen and Frazzini (2008) do use the URC’s based on the Fama 3-factor and the Carhart}

4-factor model for illustrative purposes but that they do not use the URC’s as dependent variables in the regression

0 0,2 0,4 0,6 0,8 1 1,2 0 10 20 30 40 50 60 70 80 90

(24)

21 The main coefficient of interest will be the coefficient of the negative surprise dummy. This dummy is given the value ‘1’ when the expected returns are higher than the actual returns and vice versa. When the coefficient of this dummy is negative (and significant) this will support the hypothesis that the under reaction is larger when the earnings surprises are negative. The inattention hypothesis by Kahneman and Tversky (1973), and more specifically the reduced attention to negative events found by Karsson et al., (2009) predicts that less attention will be paid to negative surprises. The study by Hou et al., (2009) suggests that this lower attention will lead to more under reaction and therefore a lower value of the under reaction coefficient. Therefore, the coefficient of the dummy is hypothesized to be negative. The values for the customer URC’s are provided in appendix III, as can be expected the values for the customer URC’s are lower than the supplier URC’s since more attention is being paid to the customers directly then through the supply chain links13

.

5.2 Fama-3-Factor model

The first model that is discussed is the model with the Fama 3-factor model used for the Cumulative Abnormal Returns. These CAR’s are subsequently used to compute the under reaction coefficients, the dependent variable of the first model. The first column of table V shows that for the main independent variable ‘Negative Surprise’ a significant negative coefficient of -.101 is found. This result confirms the hypothesis that negative shocks suffer more from under reaction compared to positive shocks. The interpretation of the coefficient ‘Negative surprise’ is therefore that a negative surprise would lead to a 10.1% lower under reaction coefficient than a positive surprise. This finding is in line with the expectations based on the studies by Karlsson et al., (2009) and Hou et al., (2009). The coefficient is significant at the 1% significance level and can therefore be expected to be reliable. The adjusted R² is 0.13, this indicates that the model used explains about 13% of the variance in the data. The adjusted R² adjusts for the amount of variables that are added to the model. This is required since the normal R² will always increase when additional variables are added (Stock & Watson, 2012).

The variable ‘institutional ownership share’ has an insignificant negative coefficient. Based on the studies of Courwin and Coughenour (2008), and Barber and Odean (2008),

analysis. In this study the URC’s are used as dependent variables to analyze the duration of the inattention between different earnings surprises, investor types and days of the week.

13_{Even though all average customer URC’s are higher than the average supplier URC’s (0.54-0.57 vs. 0.45-0.49)}

the differences are small. Therefore, is seems that relatively little attention is being paid to the customer URC’s directly.

(25)

22 professionals were expected to suffer less from inattention and the coefficient was therefore expected to be positive. However, due to the insignificance of the coefficient, this hypothesis cannot be rejected.

As expected in the study of Dellavigna and Pollet (2009) the amount of attention paid to announcements seems to be smaller on Fridays compared to other weekdays. The coefficient suggests that the subsequent under reaction will be greater for Friday announcements and will therefore decrease the URC. Even though the coefficient of the Friday dummy is negative (-0.00328), suggesting a longer duration of inattention to earnings surprises on Fridays, the coefficient is highly insignificant. Based on these results the hypothesis based on the findings of Dellavigna and Pollet (2009) cannot be confirmed.

The coefficient of the Crisis control dummy was expected to be positive since the amount of attention paid to stock markets as a whole has increased after the crisis (Lumsdaine, 2011). The coefficient is indeed positive (0.0175) but insignificant. Therefore, it cannot be concluded that the financial crisis has increased the amount of attention paid, and reduced the subsequent under reaction to earnings surprises.

5.3 Carhart 4-Factor model

The second regression model, displayed in the second column of table V, uses the Carhart 4-factor model to determine the CAR’s that are used in the URC’s. After having corrected for the momentum factor, a smaller but still negative and significant coefficient is found for the ‘Negative Surprise’ variable. The coefficient is significant at the 5% level, compared to the significance at the 1% level in the first regression. The size of the coefficient has become smaller. Based on the Carhart model a negative surprise would lead to a 7.69% higher14

under reaction coefficient than a positive surprise. With these findings the hypothesis that negative earnings surprises receive less attention and cause more under reaction, based on the studies by Karlsson et al., (2009) and Hou et al., (2009), can still be supported. The adjusted R² of the model has increased until 0.15, indicating that the Carhart model explains roughly 2% more of the variance in the data than the 3-Factor model and therefore better fits this study’s data.

While the Crisis dummy has remained positive and insignificant, in line with the findings in the 3-factor model, the institutional ownership share variable has become negative and significant at the 5% level. Contrary to the findings in the first model these findings are in line with the hypothesis that larger institutional ownership reduces inattention. Since

(26)

23 professional investors are expected to suffer less from inattention based on the studies by Barber and Odean (2008), and Courwin and Coughenour (2008) it was also expected that the duration of the inattention would be reduced more quickly. The positive coefficient of 0.117 shows that a one percent increase in institutional ownership share increases the under reaction coefficient by 0.117%15

.

The Percentage of (supplier) sales variable has remained negative in the Carhart model. Interestingly, the coefficient has become more significant and now has a p-value of 0.188 vs. a p-value of 0.845 in the 3-factor model. The negative coefficient suggests that a larger share of sales to the customer would decrease the URC. The smaller URC would then result in a larger under reaction. This is not in line with the prediction that larger customers would draw more attention and therefore suffer less from under reaction. Due to the insignificance of the coefficient this hypothesis can however not be rejected. Furthermore, the percentage of sales variable is a control variable and can therefore not be given causal interpretation.

5.4 Market Model

In the market model the size of the coefficient for negative surprises has increased to -.055 compared to -0.0769 in the 4-factor model. Furthermore, the coefficient has a p-value of 0.119, and is therefore no longer significant. The model is therefore not able to fit data as good as the previously discussed models. With respect to the adjusted R² it can be confirmed that the model has less explanatory power compared to the second model. Compared to the first model the adjusted R² has the same value of 0.13. Since the 3-factor model does yield significant returns, the book to market factor and the firm size factor that are included in the 3-factor model seem to improve the significance.

The other coefficients in the market model are insignificant and show no difference compared to the previously discussed models.

(27)

24

Table V: Regressions

Notes: Table V provides the results of the main regression model that tests the duration of the investor inattention for the four return models that are considered in this thesis. As indicated by the stars next to the coefficients the results for the main independent variable ‘Negative Surprise’ are significant in the 3-and 4-factor models. The institutional ownership variable is significant only in the 4-factor model and the Friday dummy is significant in the market adjusted model.

(1) (2) (3) (4)

URC 3-Factor URC 4-Factor URC Market URC Market-adj. Negative Surprise -0.101*** (-2.87) -0.0769** (-2.38) -0.0547 (-1.56) 0.0107 (0.28) Crisis 0.0175 (0.51) 0.00375 (0.12) 0.0159 (0.46) -0.0200 (-0.55) Inst. Own. Share -0.0196

(-0.34) 0.117** (2.27) 0.0934 (1.62) -0.0916 (-1.38) Log(total assets) 0.0106 (1.06) -0.00240 (-0.25) -0.0108 (-1.05) 0.0118 (1.13) Percentage of sales -0.0207 (-0.20) -0.128 (-1.32) -0.0387 (-0.36) 0.129 (1.13) SUE -0.00320 (-0.97) -0.00223 (-0.76) -0.00701 (-1.63) -0.000872 (-0.29) Friday -0.00328 (-0.05) -0.00387 (-0.07) -0.00940 (-0.13) -0.0956* (-1.71) Industry -0.00438 (-0.45) -0.00693 (-0.81) -0.00205 (-0.22) -0.00946 (-0.93) Leverage -0.0423 (-1.14) -0.00868 (-0.25) -0.00716 (-0.21) 0.0573 (1.42) Constant 0.499*** (6.10) 0.490*** (6.67) 0.502*** (6.33) 0.480*** (5.58)

Firm fixed effects Yes Yes Yes Yes

Observations 231,493 231,493 231,493 231,493

Adjusted R² 0.13 0.15 0.13 0.14

t statistics in parentheses * p<0.10, ** p<0.05, *** p<0.01

5.5 Market adjusted model

In the liquidity adjusted market model the negative surprise coefficient is positive. Even though the coefficient is not significant, the positive coefficient is contrary to the expectations based on the studies by Karlsson et al., (2009) and Hou et al., (2009) that had predicted that negative surprises would draw less attention and would therefore have lower URC’s. Based on the high p-value of 0.779 this hypothesis cannot be rejected. Despite the insignificance of the main independent variable the model seems to explain more variance in the data than the market model since the adjusted R² has (slightly) increased until 0.14 compared to an adjusted R² of 0.13 in the market model.

An interesting coefficient in the market adjusted model is the Friday dummy. This dummy is still negative, like in the previous models, but is now significant at the 10% level.

(28)

25 The hypothesis predicting that the amount of attention paid to earnings surprises on Fridays compared to other weekdays is lower was based on the study by Dellavigna and Pollet (2009). The lower amount of attention paid on Fridays was subsequently expected to cause more under reaction, hence lower URC’s. Since the significant negative coefficient reduces the URC’s this hypothesis can be confirmed. The coefficient of -0.096 suggests that announcements on Fridays have on average a 9.6% lower URC than announcements on other weekdays.

6. Robustness checks

In this section the additional results and the robustness checks are provided. Please note that the results provided in table V in the results section already include several robustness checks. For example, the use of firm level stock prices and company-specific customer-supplier links make the results robust to inter- and intra-industry effects (Cohen & Frazzini, 2008). Furthermore, the use and comparison of the four return models provides a robustness check. Since the results for the main independent variable are comparable in three of the four models this increases the robustness of the results. Another factor that improves the robustness of the results is the assignment to the ‘Negative Surprise dummy’ based on the nature of the earnings surprise and not on the subsequent abnormal returns. Furthermore, the use of daily stock returns in this thesis should add to the precision of the results since more observations are available. In the remainder of this section additional tests are run to critically asses the robustness of the results.

6.1 Robust standard errors

In the results section, robust standard errors are used for the estimations. These standard errors correct for possible heteroscedasticity of the error terms. In appendix VI the results without robust standard errors are provided for the four models tested. Correcting for heteroscedasticity greatly differs the results between the robust and non-robust models. For the variable SUE, that measures the size of the surprise, the results are significant in the non-robust model but not in the non-robust model. For the Friday dummy the opposite holds. This variable is not significant in the non-robust model provided in the appendix but is significant at the 10% level in the robust model. This can be explained by the fact that robust standard errors are (usually) larger and can (sometimes) be smaller than normal standard errors depending on the correlations within the model (Auld, 2013). Since the significance levels

(29)

26 differ between the models it cannot be assumed that the error terms are homoscedastic. Therefore, the use of robust standard errors provides more reliable results.

6.2 Lower winsorizing threshold

Appling a lower level of winsorizing to the dependent variables will provide a sample that has been less modified. Therefore, the robustness will increase when the results stay comparable at lower winsorizing thresholds. Barnett and Lewis (1998) explain the influence of winsorizing the data. By winsorizing the data, the outliers, both positive and negative, are relocated to the closest observation that falls outside the threshold on which the winsorizing is set. This way the extreme values are reduced. To contribute to the robustness of the results a lower winsorizing threshold (1%) will be evaluated and compared with the results of the winsorizing level used in the main model (2.5%). The results, provided in appendix VII, show that the lower level of winsorizing hardly changes the significance and size of the main independent variable. Other independent variables do show significant differences with respect to the lower winsorizing threshold used. For example, the IOshare variable now shows a significant negative coefficient in the market adjusted model while a non-significant negative coefficient was found in the 2.5% winsorizing model. Furthermore, the ‘ln(assets)’ variable has become significant in the market model after winsorizing at the 1% level. The lower level of winsorizing offers comparable results and provides more significant coefficients. Therefore, the choice for the higher winsorizing threshold of 2.5% may have adjusted the dataset more than strictly necessary.

6.3 Hausman test

The results of the Hausman (1978) test are provided in appendix VIII. The Hausman test compares the results of the panel data model with fixed effects with the random effects model and specifies whether the fixed effects are needed to improve the estimation. Since the coefficient of the Chi-2 distribution is highly significant (p=0.000) we can conclude that the fixed effects that are added to the model are required. When the test would not have been significant this would have been an indication to use random effects (Stock and Watson, 2012).

Comparing the duration of supply chain link inattention for positive and negative earnings surprises

Comparing the duration of supply chain link inattention for positive

and negative earnings surprises.

Table of Contents

1.

Introduction

2.

Literature Review

3.

Methodology

4.

Data and Descriptive statistics

5.

Results

6.

Robustness checks