The tendency of managers to use stock splits when they are afraid to miss the analyst´ consensus forecast

(1)

The tendency of managers to use stock splits when they are afraid to miss

the analyst´ consensus forecast.

Lisa Hoving 10177172 Abstract

In this paper it is found that the average forecast error for negative (positive) earnings surprises significantly increases (decreases) after stock splits. In addition, it is found that the average forecast error significantly decreases after a stock split for firms using positive discretionary accruals. Weak evidence is found that managers actually use stock splits when they are aware they cannot beat next quarters consensus forecast. Results suggest managers are more eager to use stock splits when they beat the forecast for four quarters. In addition to this, I find managers use significantly more discretionary accruals throughout the 4 quarters before a stock split. However, evidence with regard to the tendency of managers using stock splits after they have missed the analyst’ consensus forecast for four periods is still quite contradictive. Even I find most results to be significantly related to stock splits, the absolute effect is still relatively small.

Supervisor: Shivesh Changoer

University: University of Amsterdam

Studies : Bachelor Economics and Business specialization: Finance and Organisation

Time spent: 12 EC

Keywords: Forecast error, analyst’ consensus forecast, earnings management, negative earnings surprise.

Acknowledgements: I would like to thank Shivesh Changoer for guiding me through the whole process of writing a thesis

(2)

Table of contents:

Page

List of figures

Figure 1 (proposed probit model, using robust standard errors) 8 Figure 2 (Model for calculating Discretionary accruals by Ecker et al. (2013)) 9 List of tables

Table 1 (Descriptive statistics variables CRSP, I/B/E/S, and Compustat ) 7 Table 2 (Descriptive statistics for split=1 and DA>0) 12 Table 3 (Descriptive statistics for split=1 and DA≤0) 13 Table 4 (Descriptive statistics for split=1 and FCE≥0) 14 Table 5 (Descriptive statistics for split=1 and FCE<0) 15 Table 6 (Descriptive statistics for quarters with a stock split) 17 Table 7 (Descriptive statistics for quarters without a stock split) 18

Table 8 (Spearman correlations table) 19

Table 9 (Probit regression with control variables) 22 Table 10 (marginal effects to the probit regression, using dy/dx and means) 23

Table 11 (Variable description) 28

Table 12 (Dummy variables for meeting or beating the analyst’ consensus forecast) 30 Table 13 (Dummy variables indicating unmanaged earnings and earnings) 31 Table 14 (Dummy variables to indicate 10 major industries using SIC codes) 31 Table 15 (Dummy variable to indicate stock splits) 31 Table 16 (Means of all variables used in the regression) 31 I Introduction

2 Literature Review 2

(3)

Page 4 Additional Hypotheses 5 II Research design 1 Sample selection 6 2 Method 8 III Results

1 Replicating Cheong and Thomas (2011) 11

2. Descriptive statistics 15

3 Univariate Analysis. 16

4 Multivariate Analysis 20

IV Discussion and summary

1 Discussion and summary 25

V Bibliography 26

VI Appendices

Appendix A 28

Appendix B 30

(4)

I Introduction

Each quarter, analysts share their independent earnings expectations on different companies. Next, investors use these expectations to form their opinion on different kind of stocks. It is thus not all surprising that analyst´ earnings expectations influence stock prices one way or another. Managers who fail to beat the analyst´ consensus forecast can expect severe punishment by investors in the form of a sharp price decline, especially when the stock is labelled as a growth stock (Skinner & Sloan, 2002, p.289). Managers thus have the incentive to not miss those expectations. Managers try to beat analyst´ consensus forecast by trying to minimize the forecast error: real EPS (Earnings Per Share ) – analyst’ EPS consensus forecast. Being aware of (possibly too high) expectations, managers try to meet or beat these by using discretionary accruals. Discretionary accruals are the manageable part of earnings that can be used to either increase or decrease earnings. The downside to discretionary accruals is their reversibility: Using discretionary accruals to beat the analyst’ consensus forecast in one period can mean missing the forecast when discretionary accruals reverse in a later period. In this paper, another way for managers to minimize the forecast error is proposed: Using stock splits. Cheong and Thomas (2011) found that, in the US, the forecast error and its interquartile range declines with scale after stock splits (pp. 388-390). These results are remarkable as in all other cases the forecast error did not vary with scale (Cheong & Thomas, 2011, p.360). This could be evidence that managers can use stock splits to decrease the absolute forecast error. Drawing inspiration from these findings, I research whether managers use stock splits when they are aware they cannot beat next quarters analyst’ consensus forecast.

This papers finds evidence for the increase (decrease) of the average forecast error after a stock split, when the average forecast error at the time of the stock split is negative (positive). This indicates managers expecting to report negative (positive) earnings surprises

(5)

can increase (decrease) the forecast error by doing a stock split. In addition to this, I find an positive average forecast error for firms using positive discretionary accruals, and a

decreasing forecast error after the stock split. Overall, inconclusive evidence is found to the question if managers actually use stock splits when they are afraid to miss the analyst’

consensus forecast. However, in this paper evidence is found that managers are more eager to use stock splits when they have beaten the analyst’ consensus forecast for four periods. I explain this by stating managers who have beaten the forecast for the most periods, also built the most abnormal return (Bartov, Givoly, & Hayn, 2002, p. 202). In addition to this, I find managers use more discretionary accruals before a stock split. This could mean managers use stock splits to increase the forecast error before doing a stock split. Firms who have gained the most abnormal return also have the most to lose when they report a negative forecast error. Overall contradicting evidence is found for whether managers use more stock splits after missing the analyst’ consensus forecast for four periods.

The rest of the paper is structured as follows. In section 2 background information to this paper is provided. In section 3, hypotheses are explained. In section 4, additional

hypothesis to this study can be found. The research design is provided in section II. In section III, the results of the study is stated and explained. In section VI a discussion and summary of the findings can be found.

2 Literature Review

Managers can expect severe punishment by investors for not beating analyst’ consensus forecast, especially if their stock is labelled as a growth stock (Skinner & Sloan, 2002, p. 289). For value stocks, Skinner and Sloan (2002) find a maximum price decline around 5% for severe negative forecast errors (between -0.03 and -0.05) (p. 299). Growth stocks are less fortunate, a relatively large negative forecast error between -0.03 and -0.05 can already result in a price decline of -15% to -20% (Skinner & Sloan, 2002, p. 299). Given these, there are

(6)

also benefits to meeting the analyst’ consensus forecast. As found by Kasznik and McNichols (2002), firms who have consistently met the analyst´ consensus forecast reap a higher (and abnormal) return (p. 730). Skinner and Sloan (2002) have similar findings and report a small positive price increase of around 3% for firms meeting the analyst’ consensus forecast

(p.299). Given the negative price shocks to reporting negative earnings surprises and the benefits to meeting the analyst´ consensus forecast, managers are expected to be incentivized to report positive earnings surprises instead of negative ones. To prevent reporting a negative earnings surprise, a manager can try to manage earnings. A manager manages earnings when it uses discretionary accruals to increase (or decrease) earnings. Given the punishments to missing the analyst’ consensus forecast, it is not all surprising research finds an unusually high frequency of firms just meeting analyst’ expectations, meanwhile finding an unusually low frequency of firms missing expectations (Burgstahler & Dichev, 1997, p. 124. ; Bartov et al., 2002, p. 202; Brown, 2001, p.221).

This research proposes another way to decrease the absolute forecast error than earnings management: Doing a stock split. Cheong and Thomas (2011) find that the EPS forecast error does not naturally vary with scale (p. 360). However, an exception to this rule is found: After stock splits the EPS forecast error declines roughly by the magnitude of the stock split (Cheong & Thomas, 2011, pp. 388-390). Cheong and Thomas (2011) report lower

interquartile ranges and forecast error standard deviations after the stock split (pp. 388-390). Thus, if managers do a stock split when they are about to report a negative earnings surprise, the forecast error could increase. Investors are then expected to punish the firm less. Even though the forecast error is still negative, the decline in price would still be lower than when a firm would not use a stock split. This line of reasoning would suggest managers could use stock splits to signal negative earnings surprises. This is on the contrary to what former research suggest. Early research suggests shows that firms have the tendency to split their

(7)

stock after a period of growth (Fama, Fisher, Jensen, & Roll, 1969, p.20; Asquith, Healy, & Palepu, 1989, p.387). Fama et al. (1969) suggests stock splits are used to signal an increase of dividends to investors (p. 20). As firms can only keep up high dividends if past growth is sustainable, stock splits can be a signal of profitability. Louis and Robinson (2005)

corroborate on the findings by reporting evidence that managers use discretionary accruals and stock splits as a signalling device (p. 363). Louis and Robinson (2005) state: “At the split announcement, the market construes the pre-split abnormal accrual as a signal of managerial optimism rather than managerial opportunism.” (p. 363). Lakonishok and Lev (1987) also find (weak) evidence for the signaling motive previously described: Firms who split their stock had a somewhat higher growth in earnings and especially dividends after the stock split (p. 931). This research thus contradicts these findings by stating managers could use stock splits to signal a negative earnings surprise instead of a positive one. By looking at whether managers use stock splits when they are aware they can’t beat next quarters analyst’

consensus forecasts, this research thus aims to provide the reader with a totally new perspective on the usage of stock splits.

3. Hypothesis

Given the research question: Do managers use stock splits when they are aware they can’t beat next quarters analyst’ consensus forecasts?, several hypotheses were tested. The first hypothesis results from looking at the general results of Cheong and Thomas (2011). If a manager uses a stock split, the average forecast error and its interquartile range decline (Cheong & Thomas, 2011, pp. 388-390) . This could mean that managers who expect to report a negative earnings surprise could decrease the absolute forecast error (but not make it positive) by doing a stock split. As shown by Skinner and Sloan (2002), investors respond to a less negative earnings surprise by a less sharp price decline (p. 299). I therefore expect

(8)

managers to be incentivized to do a stock split when they expect to report a negative earnings surprise.

Hypothesis 1: Managers use stock splits when they are afraid to report a negative forecast error.

If a firm has been coping with a lot of negative earnings surprises, a stock split could be a way to kick start a clean slate. If a manager uses a stock split, the average forecast error and its interquartile range decline (Cheong & Thomas, 2011, pp. 388-390). This could mean that managers who expect to report a negative earnings surprise could decrease the absolute forecast error (but not make it positive) by doing a stock split. As shown by Skinner and Sloan (2002), investors respond to a less negative earnings surprise by a less sharp price decline (p. 299). Stock splits could thus give managers a way to report a higher forecast error and smoothly start beating the analyst’ consensus forecast again. This will however, only work for managers expecting to report positive surprises in het future. This line of reasoning is indirectly corroborated by Bartov et al. (2002) who states sporadic beaters have more abnormal return to gain from beating the forecast than habitual beaters (p. 190). Assuming that firms who missed the analyst´ consensus forecast for four periods are sporadic beaters, it would mean they have the most to gain from beating the analyst´ consensus forecast.

Hypothesis 2: Managers have the tendency to use stock splits after four periods of repeated missing analyst’ consensus forecast.

4. Additional hypotheses

The third hypothesis is made to support the first hypothesis. It builds on evidence found by Skinner and Sloan (2002) which states managers get punished by investors when they miss analyst consensus forecasts (p. 289). In addition to this, firms who have beaten the forecast for the most periods, also have build the most abnormal return (Bartov et al., 2002, p.202).

(9)

This could mean that firms who have beaten the analyst’ consensus forecast for four periods, also have the most abnormal return to lose. It could therefore be expected that (if the other hypotheses hold), rational managers beating the analyst’ consensus forecasts for the most periods are the most eager to do stock splits.

Hypothesis 3: Managers have the tendency to use stock splits after four periods of repeated beating analyst’ consensus forecast.

The usage of discretionary accruals can support hypothesis 4. After a manager uses

discretionary accruals, the firm will have to report a negative shock to its earnings in a later quarter when the discretionary accruals reverse. If a firm uses discretionary accruals to beat the analyst’ consensus forecast, it would be expected that a period of beating the analyst’ consensus forecast would follow up with a period of missing the analyst’ consensus forecast. Doing a stock split could then be a good idea as it could result in an increased forecast error. Hypothesis 4: Managers have the tendency to use more discretionary accruals before doing a stock split.

II Research design

1 Sample selection

Research by Cheong and Thomas (2011) focuses on US data between 1993-2006 (p. 263). Observations before 1993 were dropped, because of a methodology shift in calculating the actual reported EPS in the I/B/E/S database (Cheong & Thomas, 2011, p. 263). Stock splits are researched four quarters before and after stock splits. As this research embroiders on the results of Cheong and Thomas (2011) these stated decisions were incorporated into this research. For this research, data on 89.908 US firm quarters were used. The dataset consist of 2561 quarters where a firm split their stock and 87347 quarters where firms did not split their stock.

(10)

From the CRSP database, data on share code (SHRCD), distribution code (DISTCD), return on shares (RET), and the Standard Industrial Classification (SICCD) was obtained. Being only interested in return on common stock, all data where SHRCD did not equal 1 (1 indicating the observation was on common stock) was deleted. DISTCD is a variable that indicates a stock split when it is in the 5000-5999 range. The return on shares is the return on a stock including dividends for one month. The standard industrial classification code

indicates in which industry a firm lies. Using the I/B/E/S database, data on both median

Table 1

(Descriptive statistics variables obtained in the CRSP, I/B/E/S, and Compustat databases)

Variable N Mean 1st Quartilea Medianb 3rd Quartilec Standard Deviation Min Max ACTUAL 89908 0.241 0.080 0.200 0.370 0.336 -14.450 3.740 ATQ 89908 3257.247 358.735 969.466 2781.305 6455.385 11.727 136522 CHEQ 89908 286.877 20.735 68.808 204.648 877.491 -1.699 21761 DLCQ 89908 145.470 0.016 6.392 64.405 501.277 0.000 11441 DPQ 89908 37.850 3.708 10.908 34.034 80.367 0.000 1648 IBQ 89908 37.463 2.569 12.081 39.334 78.684 -98.100 350 LCTQ 89908 745.238 60.874 179.148 607.199 1668.521 0.551 32457 MEDEST 89908 0.233 0.080 0.200 0.360 0.315 -4.6 3.53 NUMEST 89908 9.525 5 8 12 5.718 3 44 PPENTQ 89908 1252.564 59.856 246.642 887.078 2949.237 0 31975 RECTQ 89908 393.275 35.250 117.562 353.365 969.255 -41 31622 RET 89908 0.016 -0.059 0.011 0.082 0.150 -0.830 9.374 SALEQ 89908 784.983 80.215 234.287 715.987 1639.046 -0.095 29834

a) The first quartile is defined as the 25th percentile. b) The median is defined as the 50th percentile. c) The third quartile is defined as the 75th percentile.

analyst´ consensus forecast (MEDEST), the number of estimates (NUMEST) and actual EPS (ACTUAL) was obtained. To allow for meaningful comparison, all firm quarters on which earnings announcement is before analyst forecast were deleted. In addition, all firm quarters having less than 2 forecasts were deleted. In the Compustat database assets total (ATQ), income before extraordinary items (IBQ), cash and short term investments (CHEQ), current liabilities total (LCTQ), debt in current liabilities (DLCQ), depreciation and amortization total (DPQ), sales/turnover(net) (SALEQ) , receivables total (RECTQ), and property plant and

(11)

equipment total net (PPENTQ) were obtained. Descriptive statistics on all the obtained variables can be found in table 1. I note additional information on the meaning of the data can be found in appendix A.

2 Method

To test the hypotheses the following regression was run: Figure 1

(proposed probit model, using robust standard errors)

P(SPLIT=1)= Φ(β25RET12months+β27industry1+β28industry2+β29industry3

+β30industry4 +β31industry5+β32industry6+β33industry7+β34industry8

+β35industry9+ β5BBBB1+β6BBBM2+β7BBMM3+β8BMMM4

+β9MMMM5 +β10MBBB6+β11MMBB7+β12MMMB8+β13BMBB9

+β14MBBM10+β15MMBM11+β16BBMB12+β17BMBM13

+β18MBMB14 +β19BMMB15+β21DAqmin1+DAqmin2

+β23DAqmin3 +β24DAqmin4+β26IBQincrease

+β1unman1+β2unman2+β4unman4+)

SPLIT is a dummy variable created to indicate quarters where a stock split occurred (SPLIT). To accomplish this, the distribution code (DISTCD) was used. DISTCD indicates a stock split if it is in the 5000-5999 range. Therefore split equals to 1 if DISTCD is in the 5000-5999 range for at least one of the three months in a specific quarter. RET12months is the yearly return of a stock. It is created by lagging the monthly return (RET) for 12 months and multiplying them to create yearly returns (RET12months). The 10 industry dummies

(industry1-industry10) indicate the ten major industries a company can be in. To accomplish this, SIC codes were truncated to the first digit. BBBB1-MBMM16 are dummy variables indicating all 16 ways in which a firm can beat or miss the analyst’ consensus forecast in 4 quarters. It is assumed that a firm has beaten the quarterly forecast if the forecast error is positive and missed if the forecast error is negative. The forecast error (FCE) was defined as the actual EPS (ACTUAL) minus the median estimate (MEDEST). As it is beyond the scope of this research, it was decided not to make a distinction between meeting the forecast error (FCE is close to 0) and beating it. IBQincrease is the increase in earnings for four quarters. It

(12)

is obtained by lagging earnings (IBQ) for four quarters and calculating the ratio of its

increase. DAqmin1, DAqmin2, DAqmin3 and DAqmin4 are discretionary accruals lagged for 1,2,3, and 4 quarters respectively. Discretionary accruals are calculated using the model proposed by Ecker, Francis, Olsson and Schipper (2013)1. To calculate discretionary accruals, Assets Total (ATQ), income before extraordinary items (IBQ), cash and short term

investments (CHEQ), current liabilities total (LCTQ), debt in current liabilities (DLCQ), depreciation and amortization total (DPQ), sales/turnover(net) (SALEQ) , receivables total (RECTQ), and property plant and equipment total net (PPENTQ) were used. To prevent multicollinearity, only lagged discretionary accruals are added to the regression

(discretionary accruals are used to calculate unman1-unman4). Unman1, unman2, unman3 and unman4 are dummy variables created to indicate all four possibilities in which the positivity or negativity of both earnings (IBQ) and unmanaged earnings (the portion of earnings that cannot be actively managed) could occur. Unmanaged earnings was calculated by deducting discretionary accruals (DA) from earnings (IBQ). Unman1 equals 1 if earnings (IBQ) are positive, but unmanaged earnings are negative. Unman2 equals to 1 if both earnings (IBQ) and unmanaged earnings (unman) are negative. Unman3 equals to 1 if both earnings (IBQ) and unmanaged earnings (unman) are positive. Unman4 equals to 1 if earnings (IBQ) are negative but unmanaged earnings (unman) are positive. To take account of the dummy variable trap, unman3, MBMM16 and industry 8 were deleted from the regression. These variables were chosen, as testing their relationship with stock splits is not of importance for

1

The used model by Ecker et al. (2013) is based on a model proposed by Jones (1991), but has a significant advantage over the Jones (1991) model. Estimation samples in the model proposed by Ecker et al. (2013) are based on firm size (lagged total assets) instead of industry group, resulting in less dropped data (p. 191).

Figure 2

(13)

the stated hypotheses. All data used in the regression is Winzorized to ensure outliers do not wrongly influence the results of this research. I note additional information on data definitions can be found in appendix A. Information on how dummy variables are defined can be found in appendix B. Descriptive statistics on the variables used in the regression can be found in table 2 and 3 in the results section.

As it is not directly measureable whether managers are afraid to report negative earnings surprises, a variable is set up to indicates this. A dummy variable that equals 1 if unmanaged earnings are negative, but earnings are positive (unman1) is chosen. The reason behind this is as follows: If unman1 is equal to 1, it can be said that managers are trying to artificially increase earnings over the threshold of zero to cover up negative earnings

surprises. One reason for this could be that managers are afraid they might miss the analyst’ consensus forecast. If this phenomenon consistently occurs together with stock splits, it could be said that managers use stock splits when they are afraid to report negative earnings

surprises. In this research, I thus expect to find a positive significant relationship between unman1 and stock splits. However, this measure is far from perfect. For example, firms that report positive earnings and unmanaged earnings can still miss forecast due to lower reported earnings than expected. Unman1 will therefore only indicate one distinct case of managers being afraid to miss the analyst’ consensus forecast.

To test whether managers have the tendency to use stock splits after four periods of repeated missing analyst’ consensus forecast MMMM5 is used. This variable indicates the case where managers have missed the analyst’ consensus forecast for four periods. As I expect that a manager will be more eager do a stock split if she has missed the analyst’

consensus forecast for four periods, I expect to find a positive significant relationship between MMMM5 and stock splits.

(14)

the forecast for four periods) and lagged discretionary accruals were used. As I expect managers to be more tempted to use a stock split after beating the analyst consensus forecast for four periods, I expect a to find a significant positive relationship BBBB1 and stock splits. As I expect managers to use more discretionary accruals in the quarters before stock splits, I expect the chance of a stock split increases the more managers use discretionary accruals. Thus, it is expected to find a positive significant relationship between Daqmin1, DAqmin2, DAqmin3, DAqmin4 and stock splits.

In order to try to control for omitted variable bias, control variables were added to the regression. The 12 month return (RET12months), increase in earnings (IBQincrease), major industries (industry1-10) and dummies indicating the 16 possibilities in which a firm can beat or miss the analyst’ consensus forecast for four periods were added as control variables (except MMMM5 and BBBB1, these are used to test hypothesis 2 and 3). RET12months and IBQincrease was added as researchlike fama et al. (1969) suggest firms experience growth before a stock split (p.21). Industry1-10 was added to take account of industry differences that might arise between different industries. Since I find no evidence for the insignificance of any of the other dummy variables (BBMM2- BMMB15 and unman2-unman4), the decision is made to be safe and add them all to the regression.

III Results

1 Replicating Results of Cheong and Thomas (2011)

In the first part of this research some results by Cheong and Thomas (2011) were recreated. Cheong and Thomas (2011) find that the EPS forecast error does not naturally vary with scale (p. 360). However, after a stock split, the mean forecast error declines (Cheong & Thomas, 2011, pp. 388-390). Not only does the mean forecast error decline, they also find the

interquartile ranges to decline after stock splits (Cheong & Thomas, 2011, pp. 388-390). To get a broader understanding of what happens to negative earnings surprises after a stock split,

(15)

negative and positive earnings surprises at the time of the stock split were separated. Positive and negative discretionary accruals at the time of the stock split were also separated. This is to get a broader understanding of what happens to the forecast error of firms using discretionary accruals. FCE was lagged and forwarded for four periods for all cases, and its descriptives can be found in table 2, 3, 4, and 5.

Table 2

(Descriptive statistics of the forecast error lagged and forwarded for four periods. Quarters with a stock split and positive discretionary accruals. SPLIT=1& DA>0)

Variable Mean p-valuea Median Interquartile rangeb Standard deviation Number of observations FCEqmin4c 0.0115 0.7771 0.010 0.020 0.0340 1394 FCEqmin3c 0.0134 0.3348 0.010 0.020 0.0418 1394 FCEqmin2c 0.0160* 0.0149 0.010 0.020 0.0480 1394 FCEqmin1c 0.0159* 0.0127 0.010 0.020 0.0435 1394 FCE 0.0119 0.010 0.020 0.0406 1394 FCEqplus1d 0.0049** 0.0000 0.000 0.020 0.0375 1394 FCEqplus2d -0.0005** 0.0000 0.000 0.010 0.0736 1394 FCEqplus3d -0.0002** 0.0000 0.000 0.010 0.0709 1394 FCEqplus4d 0.0009** 0.0000 0.000 0.010 0.0616 1394

a) The p-value is the result of a t-test test for equal variances for FCE and the researched variable. The p-value is two tailed

b) The interquartile range is defined as the third quartile (75%) minus the first quartile (25%).

c) qmin1 stands for the specific variable lagged for one period etc. d) qplus1 stands for the specific variable forwarded for one period etc.

* Corresponds to a p-value smaller than 0.05. The difference between means is thus significantly different from 0 for an alpha of 5%

**Corresponds to a p-value smaller than 0.01. The difference between means is thus significantly different from 0 for an alpha of 1%.

Table 2 shows managers who use discretionary accruals at the time of a stock split will, on average, have a positive forecast error. This forecast error decreases significantly after a stock split and gets close to zero. The average decline after a stock split if a firm uses discretionary accruals is 0.011. The median, and interquartile range decrease only lightly. These results suggest doing a stock split when using positive discretionary accruals would , on average, result in reporting a lower forecast error than before the stock split. As described by Skinner and Sloan (2002), this could (but does not have to) result in a lower price increase. Table 3

(16)

shows that the forecast error stays approximately the same for firms using no or negative discretionary accruals. No forecast error after the stock split is significantly different from the forecast error at the time of the stock split. In this dataset, the forecast error is already

incredibly low at the time of the stock split, significantly lower than the quarters before the split. A reason for this could be the time of the stock split. If the stock split occurs at the beginning of the quarter, the decrease in forecast error could already be incorporated in the

Table 3

(Descriptive statistics of the forecast error lagged and forwarded for four periods. For quarters with a stock split and zero or negative discretionary accruals. split=1 DA<=0)

Variable Mean p-valuea Medianb Interquartile rangec Standard deviation Number of observations FCEqmin4 0.0105** 0.0013 0.0100 0.0200 0.0385 1167 FCEqmin3 0.0131** 0.0000 0.0100 0.0200 0.0392 1167 FCEqmin2 0.0126** 0.0014 0.0100 0.0200 0.0694 1167 FCEqmin1 0.0120** 0.0001 0.0100 0.0200 0.0389 1167 FCE 0.0047 0.0100 0.0200 0.0476 1167 FCEqplus1 0.0066 0.3767 0.0100 0.0200 0.0582 1167 FCEqplus2 0.0052 0.7957 0.0000 0.0200 0.0435 1167 FCEqplus3 0.0078 0.1245 0.0000 0.0200 0.0491 1167 FCEqplus4 0.0037 0.6828 0.0000 0.0200 0.0680 1167

average forecast error for the quarter of the stock split. This would mean the decrease in forecast error could decrease significantly, but not compared to the quarter of the stock split. In this dataset however, only the median seems to decrease slightly. These results suggest that doing a stock split will have no effect on the forecast error if a manager uses zero or negative discretionary accruals. In table 4 it can be seen that the forecast error significantly decreases after a stock split for firms with a zero or positive forecast error. In addition to this, the

(17)

median also seems to decrease lightly. This suggest firms who have a positive or zero forecast error at the time of the stock split better not split their stock to increase the forecast error. However, the average decrease in forecast error by 0.013 is quite small. However, as described by Skinner and Sloan (2002), such a decrease in forecast error will still result in a small loss of the value of the stock (p.299). The magnitude depending on whether the stock is a growth or value stock (Skinner & Sloan, 2002, p. 299). However, if a firm has good reasons to split their stock, there is a good chance the positive effects will overrule the relatively small negative effects just described.

Table 4

(Descriptive statistics of the forecast error lagged and forwarded for four periods. For quarters with a stock split and a zero or positive forecast error. split=1 FCE>=0)

Variable Mean p-valuea Median Interquartileb range Standard deviation Number of observations FCEqmin4c 0.0110** 0.0000 0.0100 0.0200 0.0332 2195 FCEqmin3c 0.0125** 0.0000 0.0100 0.0200 0.0347 2195 FCEqmin2c 0.0158 0.0586 0.0100 0.0200 0.0408 2195 FCEqmin1 0.0154* 0.0155 0.0100 0.0200 0.0388 2195 FCE 0.0179 0.0100 0.0200 0.0291 2195 FCEqplus1d 0.0096** 0.0000 0.0100 0.0200 0.0312 2195 FCEqplus2d 0.0034** 0.0000 0.0000 0.0200 0.0624 2195 FCEqplus3d 0.0051** 0.0000 0.0000 0.0100 0.0622 2195 FCEqplus4d 0.0019** 0.0000 0.0000 0.0100 0.0671 2195

Table 5 shows that, in my dataset, the forecast error significantly increases after a stock split for firms having a negative forecast error at the time of the stock split. Reported forecast errors are on average still negative for the first 3 quarters, but are almost 7 times as low as the forecast error on t=0. The average change in forecast error of about + 0.039 over the four

(18)

quarters seems small, but can be quite substantial for firms. For value stocks, Skinner and Sloan (2002) find a maximum price decline around 5% for severe negative forecast errors (between -0.03 and -0.05) (p. 299). For growth stocks, a large negative forecast error between -0.03 and -0.05 can already result in a price decline of -15% to -20% (Skinner & Sloan, 2002, p. 299). Next to an increase in average forecast error, I also find an increase in the median and small decrease in interquartile ranges and standard deviations.

Table 5

(Descriptive statistics of the forecast error lagged and forwarded for four periods. For quarters with a stock split and a negative forecast error. split=1 FCE<0)

Variable Mean p-valuea Median Interquartile rangeb Standard deviation Number of observations FCEqmin4c 0.0115027** 0.0000 0.01 0.02 0.0502507 366 FCEqmin3c 0.0175683** 0.0000 0.01 0.03 0.0658087 366 FCEqmin2c 0.0061475** 0.0000 0 0.02 0.1187384 366 FCEqmin1c 0.0064754** 0.0000 0 0.03 0.0543846 366 FCE -0.0468033 -0.02 0.04 0.0703564 366 FCEqplus1d -0.0179781** 0.0000 0 0.03 0.0982352 366 FCEqplus2d -0.0060383** 0.0000 0 0.03 0.0572989 366 FCEqplus3d -0.0068579** 0.0000 0 0.03 0.060198 366 FCEqplus4d 0.0003388** 0.0000 0 0.03 0.0476517 366

2. Descriptive statistics

In table 2 and 3 the descriptive statistics of all variables used in the regression can be found. Table 6 depicts the quarters with a stock split, and table 7 depicts the quarters without. In the dataset, 2.85% of all quarters had a stock split. As there are zero observations for unman1 for stock split quarters, the mean and standard deviation are zero in table 6. In total, for only

(19)

0.4666% of the observations it happens that unmanaged earnings is negative meanwhile earnings are positive. As unman1 only covers a distinct case in which managers are afraid they might miss the analyst’ consensus forecast. This distinct case where unman1 equals to 1 just does appear not happen enough for it to be testable using a t-test. It is found that 1.457% of all firms has missed the analyst consensus forecast for quarters with a stock split

(MMMM5). 2.0% Of all firms has missed the analyst´ consensus forecast for four periods for quarters without a stock split (MMMM5). Asignificant difference between the two means is found. Finding that the percentage of firms missing the forecast for four periods is lower for stock split quarters than non stock split quarters could indicate that, on the contrary to

hypothesis 2, managers are not more tempted to do a stock split when they missed the analyst´ consensus forecast for four periods. I find 79,9% and 48,7% of firms has beaten the analyst´ consensus forecast for quarters with and without a stock split respectively. The difference between those means is significant 2. Finding a significant higher proportion for firms with a stock split could mean that managers are more tempted to do a stock split if they have beaten the analyst´ consensus forecast for four periods. The fact that lagged discretionary accruals are significantly higher for stock splitting quarters in all cases gives us evidence that

managers might have tried to beat the analyst´ consensus forecast with discretionary accruals before doing a stock split. It appears to be that the mean of rest of the 16 dummies all are significantly different from 0 to take account of the extremely high increase by BBBB1.

2 Univariate Analysis

In table 8 correlations of all variables tested in the hypotheses can be found. The whole correlation table can be provided to the reader by request. A small insignificant correlation is found between stock splits and unman1. Against expectations stated by hypothesis 2

2_{the amount of firms beating the forecast for four periods is extremely high compared to the} other dummies. This finding is consistent with the findings of Burstahler and Dichev (1997) and Bartov et al. (2002) who also find a disproportionate amount for firms just meeting or beating the analyst’ consensus forecast.

(20)

Table 6

(Descriptive statistics for quarters with a stock split)

Variable N Mean P-valuea 1st Quartile Median 3rd Quartile Standard Deviation Min Max Dependent variable SPLITb 2561 1 - 1 1 1 0 1 1 Tested variables unman1 2561 0.000 0.285 0 0 0 0.000 0 0 BBBB1 2561 0.779** 0.000 1 1 1 0.500 0 1 MMMM5 2561 0.014* 0.032 0 0 0 0.140 0 1 DAqmin1c 2561 0.007** 0.000 -0.013 0.005 0.029 0.046 -0.211 0.204 DAqmin2c 2561 0.007** 0.000 -0.016 0.004 0.027 0.048 -0.211 0.204 DAqmin3c 2561 0.004* 0.018 -0.017 0.001 0.023 0.045 -0.211 0.204 DAqmin4c 2561 0.004* 0.017 -0.018 0.003 0.025 0.046 -0.211 0.204 Control Variables unman2 2561 0.058** 0.000 0 0 0 0.234 0 1 unman3 2561 0.942** 0.000 1 1 1 0.234 0 1 unman4 2561 0.000 0.675 0 0 0 0.000 0 0 BBBM2 2561 0.043** 0.000 0 0 0 0.254 0 1 BBMM3 2561 0.007** 0.000 0 0 0 0.175 0 1 BMMM4 2561 0.002** 0.000 0 0 0 0.139 0 1 MBBB6 2561 0.044** 0.000 0 0 0 0.263 0 1 MMBB7 2561 0.010** 0.000 0 0 0 0.176 0 1 MMMB8 2561 0.004** 0.000 0 0 0 0.138 0 1 BMBB9 2561 0.029** 0.000 0 0 0 0.245 0 1 MBBM10 2561 0.011** 0.000 0 0 0 0.144 0 1 MMBM11 2561 0.006** 0.000 0 0 0 0.120 0 1 BBMB12 2561 0.024** 0.000 0 0 0 0.242 0 1 BMBM13 2561 0.007** 0.000 0 0 0 0.144 0 1 MBMB14 2561 0.009** 0.000 0 0 0 0.146 0 1 BMMB15 2561 0.007** 0.000 0 0 0 0.161 0 1 MBMM16 2561 0.004** 0.000 0 0 0 0.120 0 1 RET12month s 2561 1.794** 0.000 1.253 1.545 2.031 0.854 0.129 4.298 IBQincrease 2561 0.649** 0.000 0.082 0.359 0.778 3.434 -24.304 21.697 industry1 2561 0.044** 0.000 0 0 1 0.205 0 1 industry2 2561 0.132** 0.000 0 0 1 0.338 0 1 industry3 2561 0.330 0.088 0 0 1 0.470 0 1 industry4 2561 0.080** 0.001 0 0 1 0.271 0 1 industry5 2561 0.170** 0.000 0 0 1 0.376 0 1 industry6 2561 0.032* 0.019 0 0 1 0.176 0 1 industry7 2561 0.134 0.620 0 0 1 0.341 0 1 industry8 2561 0.079** 0.000 0 0 1 0.270 0 1 industry9 2561 0.000 0.570 0 0 0 0.000 0 0 Industry10 2561 0.000 - 0 0 0 0.000 0 0

(21)

a) The p-value comes from a t-test, testing the equality of the means for split=1 and split=0. The t-test is two sided.

b) No p-value for split exist, as that would mean testing whether the mean of split=1 is significantly different from split=0 for split.

c) qmin1 stands for the specific variable lagged for one period, qmin2 stands for the specific variable lagged for two periods etc.

Table 7

(Descriptive statistics for quarters without a stock split)

Variable N Mean P-valuea 1st Quartile Median 3rd Quartile Standard Deviation Min Max Dependent variable SPLITb 87347 0 - 1 0 0 0 0 0 Tested variables unman1 87347 0.000 0.285 0 0 0 0.021 0 1 BBBB1 87347 0.487** 0.000 0 0 1 0.500 0 1 MMMM5 87347 0.020* 0.032 0 0 0 0.140 0 1 DAqmin1c 87347 0.002** 0.000 -0.016 0.001 0.019 0.040 -0.211 0.204 DAqmin2c 87347 0.002** 0.000 -0.017 0.001 0.020 0.040 -0.211 .204 DAqmin3c 87347 0.002* 0.018 -0.017 0.001 0.020 0.041 -0.211 0.204 DAqmin4c 87347 0.002* 0.017 -0.017 0.001 0.020 0.041 -0.211 0.204 Control variables Unman2 87347 0.167** 0.000 0 0 0 0.1397 0 1 Unman3 87347 0.832** 0.000 1 1 1 0.139 0 1 Unman4 87347 0.000 0.675 0 0 0 0.008 0 1 BBBM2 87347 0.069** 0.000 0 0 0 0.254 0 1 BBMM3 87347 0.031** 0.000 0 0 0 0.175 0 1 BMMM4 87347 0.020** 0.000 0 0 0 0.139 0 1 MBBB6 87347 0.074** 0.000 0 0 0 0.263 0 1 MMBB7 87347 0.032** 0.000 0 0 0 0.176 0 1 MMMB8 87347 0.019** 0.000 0 0 0 0.137 0 1 BMBB9 87347 0.064** 0.000 0 0 0 0.246 0 1 MBBM10 87347 0.022** 0.000 0 0 0 0.144 0 1 MMBM11 87347 0.015** 0.000 0 0 0 0.120 0 1 BBMB12 87347 0.062** 0.000 0 0 0 0.241 0 1 BMBM13 87347 0.021** 0.000 0 0 0 0.144 0 1 MBMB14 87347 0.022** 0.000 0 0 0 0.146 0 1 BMMB15 87347 0.027** 0.000 0 0 0 0.161 0 1 MBMM16 87347 0.015** 0.000 0 0 0 0.120 0 1 RET12month s 87347 1.175** 0.000 0.846 1.099 1.376 0.561 0.129 4.299

(22)

IBQincrease 87347 -0.011** 0.000 -0.437 0.078 0.417 3.496 -24.303 21.696 industry1 87347 0.064** 0.000 0 0 0 0.244 0 1 industry2 87347 0.172** 0.000 0 0 0 0.377 0 1 industry3 87347 0.314 0.088 0 0 1 0.464 0 1 industry4 87347 0.100** 0.001 0 0 0 0.300 0 1 industry5 87347 0.134** 0.000 0 0 0 0.340 0 1 industry6 87347 0.025* 0.019 0 0 0 0.155 0 1 industry7 87347 0.137 0.620 0 0 0 0.344 0 1 industry8 87347 0.054** 0.000 0 0 0 0.226 0 1 industry9 87347 0.000 0.570 0 0 0 0.011 0 1 Industry10 87347 0 - 0 0 0 0.000 0 0

a) The p-value comes from a t-test, testing the equality of the means for split=1 and split=0. The t-test is two sided.

b) No p-value for split exist, as that would mean testing whether the mean of split=1 is significantly different from split=0 for split.

c) qmin1 stands for the specific variable lagged for one period, qmin2 stands for the specific variable lagged for two periods etc.

Table 8

(Spearman correlation table)

Split Unman1 MMMM5 BBBB1 DAqmin1 DAqmin2 DAqmin3 DAqmin4

Split -0.0036 -0.0071* 0.0969** 0.0249** 0.0174** 0.0059 0.0081* Unman1 -0.0036 -0.0030 -0.0057 0.0040 -0.0012 0.0000 0.0119** MMMM5 -0.0071* -0.0030 -0.1412** -0.0236** -0.0197** -0.0036 -0.0038* BBBB1 0.0969** -0.0057 -0.1412** 0.0507** 0.0373** 0.0249** 0.0079* DAqmin1 0.0249** 0.0040 -0.0236** 0.0507** -0.0310** 0.0394** 0.0095** DAqmin2 0.0174** -0.0012 -0.0197** 0.0373** -0.0310** -0.0050 0.0402** DAqmin3 0.0059 0.0000 -0.0036 0.0249** 0.0394** -0.0050 0.0119** DAqmin4 0.0081* 0.0119** -0.0038* 0.0079* 0.0095** 0.0402** 0.0119**

* Corresponds to a p-value smaller than 0.05. The correlation is thus significantly different from 0 for an alpha of 5%

**Corresponds to a p-value smaller than 0.01. Correlation is thus significantly different from 0 for an alpha of 1%.

a negative significant correlation between missing the analyst’ consensus forecast for four periods and stock splits is found. As predicted by hypothesis 3, a positive significant

correlation between beating the analyst consensus forecast for four periods and stock splits is found. As predicted by hypothesis 4, a (in most cases) significant positive correlation between discretionary accruals and stock splits exists in this dataset. However, most correlations are

(23)

small (below |0.01|). Therefore, a regression should give us a better idea how the chance of a stock split is explained by the researched variables and by how much.

3 Multivariate Analysis

Table 9 depicts the results of the robust probit regression. In my data set there are zero cases where a stock split occurred simultaneously with unman1 (earnings being positive and unmanaged earnings being negative). In addition to this, there are also zero cases where a stock split occurred simultaneously with unman 4 (earnings being negative and unmanaged earnings being positive). Therefore, the variables indicating this, namely unman1 and unman4, could not be used in the regression. The same counts for the control variables industry1 and industry10.

Consistent with hypothesis 2, I find MMMM5 has a positive coefficient 0.3677 and a very small P-value of 0.006 (missing the analyst’ consensus forecast for four periods before a stock split). These results would suggest that managers are more likely to do a stock split after a missing analyst’ consensus forecasts for four periods. Finding that the chance of a stock split increases if a firm misses the analyst´ consensus forecast is relatively new. I could not find any literature directly relating to this problem. My explanation to this finding is that, it might be that doing a stock split can be a good way to increase the negative forecast error and get a steady step up to positive earnings surprises. These results are indirectly backed up by Bartov et al. (2002) who states sporadic beaters have more abnormal return to gain from beating the forecast for one quarter (p. 190). If it is assumed that firms who missed the analyst´ consensus forecast for four periods are sporadic beaters, then they would

automatically have the most to gain from beating the analyst´ consensus forecast. However, the overall results are quite contradictive. A positive significant coefficient is found in the regression, but a negative significant correlation between split and MMMM5 is found. In addition to this, the proportion of MMMM5 is lower for quarters with a stock split than

(24)

without. These results should therefore be taken lightly.

Consistent with hypothesis 3, BBBB1 has a positive coefficient of 0.4135 and a p-value smaller than 0.000. This indicates that the probability of a stock split could increase when the analyst’ consensus forecast is beaten for four periods. As previous research like Bartov et al. (2002) shows there is an abnormal return created for beating the analyst´ consensus forecast (p.202). As firms who have beaten the most analyst´ consensus forecasts have the most abnormal return to lose, I expect they will be the most eager to do a stock split and report a less negative result. This will also result in a less severe price decline as

described by Skinner and Sloan (2002). In addition to this I find, in line with hypothesis 4, positive coefficients significantly different from 0 for DAqmin1, DAqmin2, DAqmin3, and DAqmin4. This indicates that the more managers use positive discretionary accruals the more likely they are to engage in stock splits. Finding significant coefficients for DAqmin1,

DAqmin2, DAqmin3 and DAqmin4 could be an indication that mangers first tried to beat the analyst’ consensus forecast by using discretionary accruals. As discretionary accruals are reversible, it would be expected that a firm using discretionary accruals would first beat the forecasts and then do a stock split to minimize the price decline when the discretionary accruals reverse. However, these results can also be evidence of the signalling theory. If managers use discretionary accruals and stock splits to signal profitability, then you would also expect to find a significant positive relationship between stock splits and discretionary accruals (Louis & Robinson, 2005, p. 363).

Even though I find inconclusive results for the variable that was supposed to test whether managers are afraid to miss the analyst’ consensus forecast, finding significant positive coefficients for DAqmin1-min4 could also be an indication that managers tend to be afraid to miss the analyst’ consensus forecast before deciding to use a stock split (hypothesis 1). As former research proposes mangers use discretionary accruals to beat the analyst

(25)

Table 9

(Probit regression with control variables)

Variable Sign Coefficient Robust standard error P-valuea Intercept - -2.7231** 0.1197 0.000 Tested variables BBBB1 + 0.4135** 0.1150 0.000 MMMM5 + 0.3677** 0.1341 0.006 DAqmin1 + 0.5851* 0.2306 0.011 DAqmin2 + 0.8622** 0.2359 0.000 DAqmin3 + 0.4780* 0.2268 0.035 DAqmin4 + 0.6510** 0.2285 0.004 Control variables unman2 - -0.3237** 0.0368 0.000 BBBM2 + 0.0636 0.1213 0.600 BBMM3 - -0.2546 0.1441 0.077 BMMM4 - -0.4161* 0.1851 0.025 MBBB6 + 0.1838 0.1206 0.127 MMBB7 + 0.0002 0.1374 0.999 MMMB8 - -0.1133 0.1659 0.495 BMBB9 + 0.0185 0.1235 0.881 MBBM10 + 0.1418 0.1382 0.305 MMBM11 + 0.1150 0.1501 0.443 BBMB12 - -0.0648 0.1250 0.604 BMBM13 - -0.0509 0.1469 0.729 MBMB14 + 0.0651 0.1413 0.645 BMMB15 - -0.1035 0.1428 0.469 RET12months + 0.4809** 0.0111 0.000 IBQincrease + 0.0081** 0.0025 0.001 industry1 - -0.2106** 0.0538 0.000 industry2 - -0.1272** 0.0416 0.002 industry3 - -0.1261** 0.0380 0.001 industry4 - -0.1009* 0.0462 0.029 industry5 - -0.0350 0.0410 0.394 industry6 - -0.0080 0.0621 0.897 industry7 - -0.1406** 0.0427 0.001 Log likelihood -10,280.219 Chi squared 2,811.20 p-value 0.0000 Pseudo Rsquare 0.1164 Number of observations 89,828

a) P-value is two tailed

* corresponds to a p-value smaller than 0.05. The coefficient is thus significant for an alpha of 5%

**Corresponds to a p-value smaller than 0.01. The coefficient is thus significant for an alpha of 1%.

(26)

Table 10

(marginal effects to the probit regression, using dy/dx and means)

Variable margina Standard error 95% confidence interval Tested variables BBBB1 0.0184 0.0051 0.0084 0.0284 MMMM5 0.0163 0.0060 0.0047 0.0280 DAqmin1 0.0260 0.0103 0.0059 0.0461 Daqmin2 0.0384 0.0105 0.0178 0.0590 DAqmin3 0.0213 0.0101 0.0015 0.0411 DAqmin4 0.0290 0.0102 0.0090 0.0490 Control variables unman2 -0.0144 0.0016 -0.0175 -0.0113 BBBM2 0.0028 0.0054 -0.0077 0.0134 BBMM3 -0.0113 0.0064 -0.0239 0.0012 BMMM4 -0.0185 0.0082 -0.0346 -0.0024 MBBB6 0.0081 0.0054 -0.0023 0.0190 MMBB7 0.0000 0.0061 -0.0120 0.0120 MMMB8 -0.0050 0.0074 -0.0195 0.0094 BMBB9 0.0008 0.0055 -0.0099 0.0116 MBBM10 0.0063 0.0061 -0.0057 0.0184 MMBM11 0.0051 0.0067 -0.0080 0.0182 BBMB12 -0.0029 0.0056 -0.0137 0.0080 BMBM13 -0.0023 0.0065 -0.0151 0.0105 MBMB14 0.0029 0.0063 -0.0094 0.0152 BMMB15 -0.0046 0.0064 -0.0170 0.0079 RET12months 0.0214 0.0006 0.0202 0.0226 ibqincrease 0.0004 0.0001 0.0001 0.0006 industry1 -0.0094 0.0024 -0.0141 -0.0047 industry2 -0.0057 0.0019 -0.0093 -0.0020 industry3 -0.0056 0.0017 -0.0089 -0.0023 industry4 -0.0045 0.0021 -0.0085 -0.0005 industry5 -0.0016 0.0018 -0.0051 0.0020 industry6 -0.0004 0.0028 -0.0058 0.0051 industry7 -0.0063 0.0019 -0.0100 -0.0025

a) margins are calculated by taking the derivative of the probit model. The means of all variables, except for the one the margin is calculated for, are filled into this equation to obtain the margin. The margin thus equals dsplit/dvariable=margin.

consensus forecast (Burgstahler & Dichev, 1997,p. 124. ; Bartov et al., 2002, p. 202; Brown, 2001, p.221). If the chance of a stock split increases when managers use discretionary

(27)

accruals, then it can be an indication that managers were afraid to miss the analyst consensus forecast before the stock split. Given the found positive effects of a stock split on a negative forecast error, it could be a signal that managers are using stock splits to increase the analyst consensus forecast. However, to find a definite relationship between the two, we should further research this phenomenon.

4. Marginal effects

In table 103 the marginal effects of the coefficients using the probit regression are shown. The means used to calculate the margins can be found in appendix C. Most marginal effects are, even though significant, relatively small. If a firm has beaten the forecast for four periods, the chance of a stock split increases by 1.84%. If a firm has missed the forecast for four periods, the chance of a stock split increases by 1.63%. If the discretionary accruals lagged by 1 period increases by 0.01, the chance of a stock split increases by 0.0260% (the average discretionary accruals are relatively small, the chance of them increasing by more than 0.01 is very little). The marginal effects for discretionary accruals lagged for 1,2,3, and 4 periods are relatively the same. This could be an indication that, even though most of the researched variables are found to be significant, their true effects on the chance of a stock split might be relatively small.

3

The reader has to note interpreting marginal effect calculated by using means can be quite contradictive. By choosing to calculate the marginal effects I force linearity on a model that is not linear by nature. All dummy variables will equal its averages, which is impossible in reality. Marginal effects will change throughout the model for all variables if other numbers then the means are chosen. Thus, the marginal effects shown are meant to be illustrative and are never meant to be interpreted as fixed .

(28)

IV Discussion and summary

1 Discussion and Summary

Research by Cheong and Thomas (2011) indirectly suggests the forecast error for firms reporting negative earnings surprises could increase the forecast error by doing a stock split. Given the extreme price declines small negative forecast errors can cause as shown by Skinner and Sloan (2002), managers could possibly prevent great price declines by doing a stock split. In this paper I research whether managers use stock splits when they are aware they can’t beat next quarters analyst’ consensus forecasts. In order to try to answer this research question, I use univariate and multivariate analysis. I find firms can increase the forecast error significantly by doing a stock split. However, the if a firm reports a positive forecast error at the time of the split, the forecast error will significantly decrease. I find the forecast error decreases (stays the same) after stock splits if managers use positive (zero or negative) discretionary accruals. In addition to this, evidence is found that managers will be more eager to do a stock split after a period of beating the analyst’ consensus forecast (hypothesis 3). I also find evidence that managers first try to beat the analyst’ consensus forecast by using discretionary accruals before a stock split. Sadly enough, the variable that was going to test whether managers were afraid to report negative earnings surprises could not be tested due to a lack of observations. As there were zero observations where a stock split occurred and unmanaged earnings were negative meanwhile earnings were positive, it is impossible to test for unman1 using a regression. However, I find a positive relationship between the chance of a stock split and discretionary accruals. I suggest that the fact that managers use more discretionary accruals before a stock split can be an indication that managers were afraid to miss the analyst´ consensus forecast before a stock split. This together with the found positive relationship between stock splits and beating the analyst

(29)

consensus forecast could be an indication that managers indeed use stock splits when they are afraid to report negative earnings surprises. However, as results are not yet conclusive, this relationship should be further investigated. Contradicting evidence is found that managers are more eager to use stock splits after missing the analyst’ consensus forecast for four periods (hypothesis 2). Descriptive statistics and correlations seem to indicate a negative relationship between stock splits and missing the analyst’ consensus forecast, but the used probit

regression seems to suggest the opposite. As finding a positive relationship is relatively new and the given contradicting evidence, more research is needed on this subject. There is also more research needed on the topic of decreasing/increasing forecast errors after stock splits. just because it is found that the forecast error increases (decreases) at the time of a stock split if the forecast error is negative (positive), does not imply a causality is found. The causality between these variables should therefore be further investigated.

One of the limitations to this study is omitted variable bias: If the possibility of a stock split is not regressed on all independent variables the coefficients become biased. However, including many variables will also increase the chance of making the wrong conclusion that a variable is significantly related to stock split, meanwhile it is not (type I error). Another limitation to this study is the choice of not making a distinction between meeting and beating the analyst’ consensus forecast. As described by Skinner and Sloan (2002), the increase in value of stocks can be quite different from meeting or beating the analyst’ consensus forecast. Beating the analyst’ consensus forecast or meeting it could therefore incentivize managers quite differently (even though they both result in an increase in value), but are taken together in my research.

Bibliography

(30)

387-403.

Bartov, E., Givoly, D., & Hayn, C. (2002). The rewards to meeting or beating earnings expectations. Journal of Accounting and Economics, 33(2), 173-204.

Brown, L. D. (2001). A temporal analysis of earnings surprises: Profits versus losses. Journal

of Accounting Research, 39(2), 221-241.

Burgstahler, D., & Dichev, I. (1997). Earnings management to avoid earnings decreases and losses. Journal of Accounting and Economics, 24(1), 99-126.

Cheong, F. S., & Thomas, J. (2011). Why do EPS forecast error and dispersion not vary with scale? Implications for analyst and managerial behavior. Journal of Accounting

Research, 49(2), 359-401.

Ecker, F., Francis, J., Olsson, P., & Schipper, K. (2013). Estimation sample selection for discretionary accruals models. Journal of Accounting and Economics, 56(2), 190-211. Fama, E. F., Fisher, L., Jensen, M. C., & Roll, R. (1969). The adjustment of stock prices to

new information. International Economic Review, 10(1), 1-21.

Jones, J. J. (1991). Earnings management during import relief investigations. Journal of

accounting research, 193-228.

Kasznik, R., & McNichols, M. F. (2002). Does meeting earnings expectations matter? Evidence from analyst forecast revisions and share prices. Journal of Accounting

Research, 40(3), 727-759.

Lakonishok, J., & Lev, B. (1987). Stock splits and stock dividends: Why, who, and when. The

Journal of Finance, 42(4), 913-932.

Louis, H., & Robinson, D. (2005). Do managers credibly use accruals to signal private information? Evidence from the pricing of discretionary accruals around stock splits. Journal of Accounting and Economics, 39(2), 361-380.

(31)

returns or don't let an earnings torpedo sink your portfolio. Review of Accounting Studies, 7(2-3), 289-312. VI Appendices Appendix A Table 11 (Variable description)

Label Description Unit Source

ACTUAL Actual value earnings from the details actual file. Per share in dollars I/B/E/S Summary History- Summary Statistics

ATQ Assets Total Dollars

Compustat-Compustat Montly updates- North America- Fundamentals Quarterly CHEQ Cash and Short term Investments Dollars

Compustat-Compustat Montly updates- North America- Fundamentals Quarterly

DA Discretionary Accruals Dollars

DISTCD distribution code Is in the

5000-5999 range for stock splits CRSP-Quarterly Update-Monthly Stock File

DLCQ Debt in Current Liabilities Dollars Compustat-Compustat Montly updates- North America- Fundamentals Quarterly DPQ Depreciation and Amortization Total Dollars

(32)

Compustat-Compustat Montly updates- North America- Fundamentals Quarterly IBQ Income Before Extraordinary items Dollars

Compustat-Compustat Montly updates- North America- Fundamentals Quarterly LCTQ Current Liabilities Total Dollars

MEDEST Median estimate earnings Per

share in dollars I/B/E/S Summary History- Summary Statistics

NUMEST Number of estimates I/B/E/S

Summary History- Summary Statistics PPENTQ Property Plant and Equipment Total Net Dollars

RECTQ Receivables Total Dollars

(33)

RET Holding period return. Calculated as: (stock price at the end of the month+ cash dividends)/stock price at the beginning of the month. Fraction CRSP-Quarterly Update-Monthly Stock File RET12months Return on 12 months, RET is lagged for

12,11…,0 months. RET12months is created by multiplying

(1+RET)(1+RETqmin1)…..(RET1qmin12)

Fraction

SALEQ Sales/Turnover(net) Fraction

SHRCD Share Code Equals

1 for ordinary shares. CRSP-Quarterly Update-Monthly Stock File SICCD Standard Industry Classification code.

CRSP-Quarterly Update-Monthly Stock File UNMAN Income Before Extraordinary Items minus

Discretionary accruals

Dollars

Appendix B

Table 12

(Dummy variables for meeting or beating the analyst’ consensus forecast for four periods) Label Equals 1 if (B=beat M=miss) BBBB1 BBBB BBBM2 BBBM BBMM3 BBMM BMMM4 BMMM MMMM5 MMMM MBBB6 MBBB MMBB7 MMBB MMMB8 MMMB BMBB9 BMBB MBBM10 MBBM

(34)

MMBM11 MMBM BBMB12 BBMB BMBM13 BMBM MBMB14 MBMB BMMB15 BMMB MBMM16 MBMM Table 13

(Dummy variables indicating the four possibilities of the positivity/negativity of unmanaged earnings and earnings)

Label Equals one if Additional Condition

UNMAN1 Unman negative IBQ positive

UNMAN2 Unman negative IBQ negative

UNMAN3 Unman positive IBQ positive

UNMAN4 Unman positive IBQ negative

Table 14

(Dummy variables to indicate 10 major industries using SIC codes)

Label: Equals one if the Truncated SICCD=

Industry1 1 Industry2 2 Industry3 3 Industry4 4 Industry5 5 Industry6 6 Industry7 7 Industry8 8 Industry9 9 Industry10 0 Table 15

(Dummy variable to indicate stock splits)

Label Equals one if DISTCD is between

Split 5000-5999

Appendix C

Table 16

(Means of all variables used in the regression)

Variablea Mean Tested variables BBBB1 0.4957 MMMM5 0.0199 DAqmin1 0.0022 DAqmin2 0.0021

(35)

DAqmin3 0.0019 DAqmin4 0.0022 unman2 0 .1639 BBBM2 0.0685 BBMM3 0.0311 BMMM4 0.0194 MBBB6 0.0736 MMBB7 0.0312 MMMB8 0.0186 BMBB9 0.0634 MBBM10 0.0210 MMBM11 0.0143 BBMB12 0.0610 BMBM13 0.0207 MBMB14 0.0213 BMMB15 0.0260 RET12months 1.1929 ibqincrease 0.0086 industry1 0.0633 industry2 0.1710 industry3 0.3149 industry4 0.0993 industry5 0.1346 industry6 0.0249 industry7 0.1372

a) All variables that were deleted from the regression due to lack of observations or to control for the dummy variable trap are also deleted from this table.

(36)

(37)