Is (net) payout yield a good predictor for stock returns during the financial crisis of 2008? Empirical study performed on US stock market
Author:
D.T.F. Ploegmakers
Student number:
10653422
Thesis supervisor: R. C. R. van Lamoen
Finish date:
07/2018
2 NON-PLAGIARISM STATEMENT
By submitting this thesis the author declares to have written this thesis completely by himself/herself, and not to have used sources or resources other than the ones mentioned. All sources used, quotes and citations that were literally taken from publications, or that were in close accordance with the meaning of those publications, are indicated as such.
COPYRIGHT STATEMENT
The author has copyright of this thesis, but also acknowledges the intellectual copyright of contributions made by the thesis supervisor, which may include important research ideas and data. Author and thesis supervisor will have made clear agreements about issues such as confidentiality.
Electronic versions of the thesis are in principle available for inclusion in any EUR thesis database and repository, such as the Master Thesis Repository of the Erasmus University Rotterdam
3
ABSTRACT
In this thesis the predictive power of dividend yield, payout yield (dividends plus repurchases) and net payout yield (dividends plus repurchases minus issuances) on excess market returns will be examined, highlighting the effect of the financial crisis of 2008 on the predictability. This thesis will contribute to existing literature by looking at a unconventional type of period, namely a crisis period and uses besides annual data, also monthly data. The results of the annual data analysis provide evidence that during the financial crisis, predictability for all measures disappears. After the financial crisis, the results show dividend yield and net payout yield have predictive power in-sample. However, the out-of-sample results indicate both measures are outperformed by the historical mean. The results of the monthly data analysis provide similar evidence that during the crisis predictability disappears for both measures. But, after the financial crisis, predictability restores for dividend yield, which is supported by the out-of-sample results.
4
TABLE OF CONTENTS
TABLE OF CONTENTS 4 LIST OF TABLES 5 LIST OF FIGURES 6 CHAPTER 1 Introduction 7CHAPTER 2 Literature review 10
2.1 Dividend and payout yield 10
2.2 SEC rule 10B-18 and its effect on share repurchases 11
2.3 Financial Crisis 13
2.4 Time-Series prediction 14
CHAPTER 3 Data and methodology 15
3.1 Data 15
3.2 Methodology 17
CHAPTER 4 Results 19
4.1 Analysis of the annual data 19
4.2 Crisis based on annual data 23
4.3 Analysis of the monthly data 25
4.4 Crisis based on monthly data 28
4.5 Robustness check 30
CHAPTER 5 Conclusion and discussion 33
REFERENCES 35
APPENDIX A.1: Annual Variables 39
APPENDIX A.2: Monthly Variables 40
APPENDIX B: Index returns 41
APPENDIX C.1: Summary statistics 42
5
LIST OF TABLES
Table 1: Return predictability using annual data. 22
Table 2: Return predictability in the period of the financial crisis, using annual data. 23
Table 3: Return predictability using monthly data. 27
Table 4: Return predictability in the period of the financial crisis, using monthly data. 29
Table 5: Return predictability around the oil crises (1973-1982). 31
Table 6: Detailed calculations of the annual dataset. 39
Table 7: Detailed calculations of the monthly dataset. 40
Table 8: Summary of descriptive statistics based on the annual data set. 42
Table 9: Summary of descriptive statistics based on the monthly data set. 42
Table 10: Correlation matrix based on the annual data set. 43
6
LIST OF FIGURES
Figure 1: Evolution of different forms of payout. 8
7
CHAPTER 1 Introduction
Since the existence of stock markets every investor wants to be able to predict stock returns so that he can make additional profits. Throughout the years there have been many asset pricing models using different variables to try to predict stock returns. In 2015, Harvey, Liu and Zhu had counted 313 different models that had proven to be capable of predicting returns. But how predictive are these models in times of financial crisis and how accurate is their prediction? In this thesis the predictive power of different payout yield measures around the time of the financial crisis of 2008 will be explored.
Moreover, this thesis also contributes to existing literature by examining an additional dataset, namely a dataset based on monthly observations. One of the most popular predictive variables is dividend, examined by among others, Campbell, Lo and MacKinlay (1997) and Cochrane (2001). Boudoukh, Michaely, Richardson and Roberts (2007) used different variations of payout yields on top of dividend yield to explain and predict returns in time-series and explain returns in the cross-section.
Boudoukh et al. (2007) argue that dividend payout has lost its power as key variable in asset pricing models over the last couple of decades. However, in 2008 the financial crisis struck, which had a major impact on the financial world. Stock prices fell hard during the financial crisis (Am Al-Rjoub and Azzam (2012)). Although companies tried to keep dividends relatively constant throughout the financial crisis, share repurchases drop more quickly during financial distress (Bliss, Cheng and Denis (2015)). This is illustrated in the graph below. Figure 1 show the aggregated levels of common dividends, common share repurchases and share issuances throughout the years by nonfinancial firms. The graph is based on annual data from Compustat. Striking is the decline in share repurchases during the beginning years of the financial crisis in 2008.
Using annual data from Compustat and return data from Center for Research in Security Prices (CRSP), different yield variables will be constructed. The yield variables will be dividend yield, payout yield, which also takes share repurchases into account and net payout yield, which takes on top of dividends and repurchases, share issuances into account as well. Payout yield will be calculated in two ways. The first way is based on the cash flows of repurchases and the other way accounts for changes in treasury stock, resulting in two different measures of net payout yield.
In addition to the analysis of the annual data, the effect of the financial crisis on predictability will also be examined based on monthly data acquired from CRSP. By looking at monthly data, the number of observations increase and leads to more precise results (Wang, Zhang and Fu (2014)). However, due to data limitations, only dividend yield and net payout yield can be examined.
8
Figure 1: Evolution of different forms of payout.
The figure shows data on common dividends, purchase of common and preferred stock and sale of common and preferred stock, collected from the CRSP/Compustat Merged database. Only nonfinancial firms are included in the sample (i.e. firms with a SIC between 6000-6999 are excluded from the dataset). The bars represent the aggregated cash flow from or towards shareholders. The white bars represent the issuance of stocks, which is a negative cash flow to shareholders. The light gray bars represent the common dividends, which is a positive cash flow to shareholders. The dark gray bars represent the repurchase of shares by firms, which is also a positive cash flow to shareholders.
The decline in share repurchases, shown in the figure, had led to the research question of this thesis:
Is (net) payout yield still a good predictor for stock returns during the financial crisis of 2008?
Empirical study performed on US stock market.
Further questions that will be answered in this thesis are:
How is the predictability of (net) payout yield before and after the financial crisis? Is there a difference?
Will (net) payout yield still outperform dividend yield in the years after the research of Boudouhk et al. (2007)?
When the levels of repurchases restore after the decline in the beginning of 2008, will (net) payout yield still outperform dividend yield in predictability?
9 This thesis has the following layout. Chapter 2 will give an overview of the existing literature regarding the payout yield measures, fundamentals about the financial crisis, the implementation of the SEC rule 10B-18 and predictability in general. Chapter 3 will describe the dataset and the methodology that has been used in this thesis. Chapter 4 will discuss the results of the analysis. Chapter 5 will give a summary of the findings of the conducted research in this thesis, along with a discussion section, in which the strengths and limitations of this study will be discussed, along with some recommendations for future research.
10
CHAPTER 2 Literature review
This chapter will give an overview of the current literature regarding the payout measures. Section 2.1will discuss research on dividend yield and different forms of payout yield as a predictive variable. The next section discusses the evolution of dividends and how the amount of repurchases has changed since the introduction of Securities and Exchange Commission (SEC) rule 10b-18. Section 2.3 explains how the financial crisis occurred and what its effect was on the financial world, in particular on payout policies. Finally, section 2.4 will review predictability characteristics in general.
2.1 Dividend and payout yield
Paying dividends is a way for companies to signal investors about their financial well-being (John and Williams (1985), Bernheim (1991), and Allen, Bernardo, and Welch (2000)). It also provides investors with a secure income. There is a lot of existing literature on dividends. Dividend has been a very popular variable used to help to explain and predict stock returns, even though Miller and Modigliani (1961) argued that, considering their irrelevance theorem, dividends play no role in determining returns. However, this theorem does not tell us anything about using dividends to explain and predict stock returns. Most theories and models assume that markets are efficient, but if stock returns are in any way predictable, this efficiency does not hold. Moreover, risk premia change over time due to, among others, the state of the economy (e.g. a recession). So, the variations in risk premia over time give support for asset pricing models. Support for asset pricing models considering dividends as their main explanatory variable in time-series came from Fama and French (1988), who used dividend yield to help explain returns. They found that when considering short-horizon returns (e.g. monthly or quarterly returns), dividend yield only explained less than 5% of the variations in returns. But, when looking at two- to four-year returns, dividend yields explained more than 25% in the variance, which was
attributed by Fama and French (1988) to the fact that high positive autocorrelation causes the variance of expected returns to grow more than in proportion to the horizon of returns. So, when horizons became larger, dividend yield explained a larger fractions of the variances in returns, which is very logical. For example, it is much easier to forecast seasonal weather circumstances than it is predicting a day-to-day change. Therefore, t-statistics and R-squared mechanically increase when the horizon becomes larger (Cochrane (2001)).
Since the last two decades a debate has been going on about how well dividends still proxy for total payout? Does this weakening proxy lead to mismeasurement and what are the implications? Proponents of this claim are Fama and French (2001), who argue that dividend payment has been substituted by repurchases. This argument has been supported by Grullon and Michaely (2002), Dittmar and Dittmar (2002), Brav et al. (2005) and Boudoukh et al. (2007). According to Stambaugh (1999), Valkanov (2003), Lettau and Ludvigson (2002), Cochrane (2001), and Goyal and Welch (2003)), dividend
11 yield has lost its power as key variable in asset pricing models. Boudoukh et al. (2007) argue that this loss in predictability is related to the definition of payout. Another form of payout is through share repurchases. Buying back shares is favourable for shareholders, because shares outstanding decrease and therefore EPS increase. Moreover, the repurchase of share can also prevent an increase in shares outstanding by, for example, offsetting share remuneration policies (Senate Report No. 550 (1967). According to Boudoukh et al. (2007), share repurchases should also be considered in the payout variable. One of the arguments to ignore repurchases in the payout yield is the fact that it has been neglected in the past, because share repurchase activity by firms was very low and close to zero due to the anti-fraud provisions of the Securities Exchange Act of 1934. However, in 1982 rule 10b-18 was adopted by the SEC. This rule stated that under certain conditions, firms are allowed to buy back a certain amount of their shares. This has a significant impact on repurchase activity by firms, which can also be seen in figure 1. The advent and the effects of SEC rule 10b-18 will be further discussed in section 2.2.
Furthermore, Boudoukh et al. (2007) also included issuances into the payout yield variable, which they referred to as net payout yield. Issuances (e.g. seasoned equity offerings) can be seen as a negative cash flow to shareholders, because it is a cash flow from investors to the firm. Through issuances, shares outstanding increase, earnings per share (EPS) and dividends per share (DPS) decrease, which is also not beneficial for shareholders. Although, issuances is a negative form of payout for investors, it should be included into the payout yield variable, because of the possibility that cash is raised so that firms are able to maintain their dividends. Therefore, it can be seen as a correction for true dividends (Allen and Michasely (2003)).
2.2 SEC rule 10B-18 and its effect on share repurchases
For decades, US firms preferred paying out cash dividends over share repurchases, despite the tax advantage capital gains have over ordinary income. In 1986, the Tax Reform Act (TRA) relatively reduced the tax advantage of capital gains relatively to ordinary income. However, the marginal rate on ordinary income remained higher than the marginal rate on capital gains. In 2001, the top marginal rate on ordinary income was 39.6%, relative to only 20% on capital gains, implying there is still a relative tax advantage on capital gains. Moreover, share repurchases allow investors to postpone the realization of capital gains and therefore tax payments as well, which is an advantage over dividends (Grullon and Michaely (2002)).
Before the TRA of 1986 was introduced, relative tax advantages on capital gains were much larger. So, why was repurchase activity so small before the mid-80s? The explanation of Grinblatt and Titman (1998) is that firms were just simply wrong for paying so much in dividends. On the other hand, John and Williams (1985), Bernheim (1991), and Allen, Bernardo, and Welch (2000) argue that firms use
12 dividends, instead of share repurchases, to signal the financial health of the firm to investors, meaning dividends and repurchases are not substitutes. Another explanation for the repurchase activity being so low could be the fear of firms to violate anti-fraud provisions, implemented by the Securities Exchange Act (SEA) of 1934 to prevent firms from buying back shares extensively. The SEA was introduced, because according to Senate Report No. 550 (1967), repurchasing shares might have a disturbing effect on the financial order of financial markets. This might explain why firms became very reluctant to participate in share repurchase activity, they were afraid to be accused of illegal market manipulation by the SEC, which could lead to very large direct and indirect costs regarding a regulatory inquiry (Feroz, Park and Pastena (1991), Karpoff and Lott (1993), Nourayi (1994), and Beatty, Bunsis, and Hand (1998)). However, until 1982 there were no explicit rules in the US that gave firms guidelines on buying back shares. The SEC was aware of this problem and started to propose different guidelines on the matter. The first proposition of the SEC was Rule 10b-10, released in 1967, which stated that firms were
required to disclose information about repurchase activity and to follow certain rules. However, this rule got rejected. After the rejection of Rule 10b-10, other proposals followed. For instance, Rule 13e-2 in 1970 and variations of these rules in 1973 and 1980. Nevertheless, all got rejected. Finally, in 1982 the SEC adopted Rule 10b-18. This rule provides firms with specific guidelines for repurchasing shares without violating the anti-fraud provisions of the SEA of 1934. Rule 10b-18 states that repurchasing firms are allowed to use only one broker on a single trading day. Moreover, repurchasing activity must be avoided during the beginning and closing hours of the market. Also, daily volume is limited to a certain amount each day. Grullon and Michaely (2002) found that the aggregated amount of cash used to repurchase shares has tripled in just one year after the introduction of Rule 18. Before Rule 10b-18 was introduced, Grullon and Michaely (2002) observed that annual expenditure on share repurchases was 5.5 billion dollars. After the adoption of the rule, they found that the annual expenditure on share repurchases had increased to 62 billion dollars, which is an increase of more than a thousand percent. Moreover, the ratio of firms that repurchase shares as a form of payout, has increased dramatically as well. For example, the ratio of firms that participate in repurchase activity in 1972 was 26.6. This has increased to 84.2 in 2000.
Furthermore, Grullon and Michaely (2002) found that since the adoption of Rule 10b-18, firms rather increase share repurchases than dividends to increase their payout ratios. The preference of increasing share repurchases is attributed to the tax advantage that capital gains still has over ordinary income. These results are consistent with the findings of Lie and Lie (1999) and Sarig (2000), who had also found that firms rather increase share repurchases than dividends, due to the relative tax advantage.
13 2.3 Financial Crisis
In 2008, the financial crisis struck, which had a massive impact on the financial world. The crisis, also known as the subprime mortgage crisis was caused by default on subprime mortgages. Before the financial crisis, the housing market was doing very well. Prices were high, and it was easy for individuals to obtain a mortgage, causing a housing bubble, where prices kept rising, while income stayed the same. Consequently, bad loans remained unnoticed (Barth, 2009). People simply paid off mortgage payments by taking on another mortgage, while they offered their houses as collateral. This situation could not continue forever. Eventually, the bubble had to burst, and so it did. House prices dropped dramatically and people defaulted on their mortgage. Banks foreclosed the houses and suffered huge losses by doing so. Some banks even collapsed, for example Lehman Brothers. Besides banks, investments companies and investors suffered as well, due to mortgage-backed securities (Greenspan, 2007). The crisis also had consequences for payout policies. Floyd, Li, & Skinner (2015) examine the payouts of industrials
compared to banks. A key feature of dividends is the implied commitment, which means that managers are reluctant to cut dividends (Lintner (1956), Brav, Graham, Harvey, and Michaely, (2005), DeAngelo, DeAngelo , and Skinner (2009)). This commitment helps to explain dividends in two ways. First, dividends are used to help address agency costs of free cash flows (Jensen, 1986). Second, dividend payments can be used by firms to signal their profitability and financial strength (Miller and Rock (1985), Baker, Mendel and Wurgler (2016)). So, although most firms are affected by a crisis, they try to maintain their levels of dividends. The first explanation is more important to industrials and therefore the
aggregate real industrial dividends declined by 5% during the beginning of the financial crisis. For banks the second explanation holds more, especially during a crisis. Many banks received funding under the Capital Purchase Program (CPP), which is a program that is designed to bail out the financial sector in case of need. Although, banks under the CPP could not increase their dividends, they were not urged to decrease them, which they were also reluctant to do. Floyd, Li, & Skinner (2015) found that most large banks that received funding under the CPP did not decrease their dividends during the first part of the crisis. The largest banks, like Goldman Sachs, kept dividends constant during the entire crisis, while other banks had to cut dividends from the beginning of 2009. Some small banks even had to eliminate dividends entirely. Repurchases on the other hand are not necessarily being used as signals, because the magnitude of repurchases is hard to asses. Furthermore, repurchases are not specifically allocated to certain periods. First, an announcement is made, followed by the implementation. The latter can take up to three years. Therefore, it is hard for investors to take any signal from this (Ikenberry and
Vermaelen (1996), Vermaelen (1981)). Since there is less commitment in repurchases, firms rather cut repurchases over dividends, which is also confirmed by the data. During the financial crisis, industrials cut dividends by 5,4%, while they cut repurchases by 71% (Floyd, Li, & Skinner (2015)). Dividends and repurchases follow a similar pattern for commercial banks. During the crisis the banks tried to cut
14 dividends slowly, while cutting repurchases much faster and eliminating them almost completely at the end of the crisis.
Moreover, Bliss et al. (2015) examine the effect of the 2008 financial crisis on payout in form of dividends and share repurchases on a corporate level. They found that both dividends and share repurchases declined during the crisis. Even though, firms are reluctant to cut dividends, Bliss et al. (2015) found that, between 2006 and 2009, the number of firms that reduce (or even eliminated) dividends increased with 19%. For repurchases this increase was almost double. So, just like in the work of Brav et al. (2005) and Jagannathan, Stephens, and Weisbach (2000) , this indicates that firms rather cut share repurchases than dividends. Furthermore, Bliss et al. (2015) found that the aggregated dollar amount of payout between 2006 and 2009 had declined 58%. This percentage is attributed to the fact that during the crisis it became more costly for firms to borrow. Cornett, McNutt, Strahan and Tehranian (2011) report a significant decline in lending activity by constrained banks, due to the crisis. Therefore, firms reduced payouts as a form of internal funding for investments and to build up cash reserves. This view is also supported by Campello, Graham and Harvey (2009), who surveyed CFOs and concluded that a large proportion experienced increasing difficulty and costs in borrowing funds during the crisis. Finally, Bliss et al. (2015) report that firms who are highly levered, have growth options and possess little cash are more likely to have reduced their payouts during the crisis.
2.4 Time-Series prediction
Since the existence of stock markets every investor wants to be able to predict stock returns so that he can make additional profits. However, according to theory, markets are informationally efficient, which means that prices reflect most information about the fundamental value (Fama (1971)). So, if new information becomes available investors will immediately trade on it and prices would quickly reflects the new information. Predictability does not necessarily imply that markets are inefficient. Even in efficient markets, conditional moments can change over time. For instance, during a financial crisis expected returns increase, due to a higher premium as compensation for the additional risk during the crisis (Am Al-Rjoub and Azzam (2012)). Also, returns on assets follow real economic movements, and are therefore subject to business cycle fluctuations. This shows that risk premia are varying over time, which give implications for asset pricing models and predictability. Meanwhile, there are many researchers that state they found predictability in their models. In 2015, Harvey et al. counted, up till that moment, 313 different models that had proven to be capable of predicting returns. Two of the most common research methods in asset pricing are based on time-series analysis and cross-sectional analysis. Because the focus of this thesis lies on predictability, only the former will be discussed. In the methodology section, a more elaborate view on the setup will be discussed.
15
CHAPTER 3 Data and methodology
3.1 Data
For the analysis in this thesis, two datasets will be used. One based on annual data and one based on monthly data. The data for both datasets will be collected from the Wharton Research Data Services (WRDS) database. The methodology used in the paper of Boudoukh et al. (2007) will serve as a guideline during the construction of the datasets. The sample period will be from 1926 till 2017. This way the sample has a large enough look-back window to perform the out-of-sample regression. In the sample only nonfinancial firms are taken into account (i.e. firms with a SIC of 6000-6999 are excluded), because financial firms operate very differently compared to other firms. Financial firms have higher leverage than nonfinancial firms. Therefore results from financial firms cannot be interpreted in the same way as nonfinancial firms (Fama and French (1992). In the first part of the time-series regression annual data will be used. The dataset will be constructed by using data from the CRSP and CRSP/Compustat merged databases. First, the excess return will be calculated, which is defined as the market return minus the risk free rate. The market return will be collected from CRSP by taking the value weighted return (including distributions) based on the NYSE, AMEX and NASDAQ. as a proxy for the risk-free rate, the annual return on a 90-day treasury bill will be used, instead of the return on a 1-year bond. Data on the 1-year bond has only been reported since 1940 instead of 1926, the start of our sample. This small modification should have no impact on the results (Goyal and Welch (2003)). The excess return is acquired by taking the natural logarithmic difference between the market return and the risk-free rate. Next, the yield variables will be constructed. To do so, information about dividends, share repurchases and share issuances is required. The dividend yield is the total dividends per year divided by the closing share price, or simply the natural logarithmic difference between the value weighted return including distributions and the value weighted return excluding distributions. The payout yield can be calculated in two ways. One way focuses on share repurchases based on cash flows, while the way other focuses on the change in treasury stock. For the first measure, share repurchases are defined as the total expenditure on the purchase of common and preferred stocks (Compustat item #115) plus any reduction in the value of the net number of preferred stocks outstanding (Compustat item #56). To obtain the payout yield, the natural logarithm of dividends will be added to the natural logarithm of share repurchases dividend by the corresponding market capitalization. Repurchase data is only available from 1971 onwards. So, prior to 1971 repurchase are assumed to be zero. This is a reasonable assumption, since share repurchase activity was very small and almost zero before the early 70s
(Boudoukh et al. (2007), Grullon and Michaly (2002)). The second measure of payout yield focuses more on the distinction between actual share repurchases and exercises on share payout options. Some firms use stock option as part of their remuneration policy. At some point in time these options can be exercised. To avoid fluctuations in share prices, firms anticipate these exercises by buying back shares.
16 So, share repurchases based on treasury stock is defined as the change in treasury stock (Compustat item #225). The payout yield is then constructed by adding the natural logarithm of dividends to the natural logarithm of share repurchases, based on the change in treasury stock, divided by the corresponding market capitalization. Treasury stock data is also not available at the beginning of the sample, and is therefore also assumed to be zero prior to 1982. Finally, there is net payout yield, which yield takes equity issuances into account as well. Equity issuances are defined as the sale of common and preferred stock (Compustat item #108) minus any increase in the value of the net number of preferred stocks outstanding (Compustat item #56). To construct the net payout yield, the natural logarithm of the total dollar amount of equity issuances per year divided by the corresponding market capitalization is subtracted from the two payout yields, respectively. Data concerning issuances, calculated as stated above, are available since 1971. It would be wrong to assume that issuances were zero prior to 1971, just like assumed for repurchases. Therefore, a different measure of issuances is used to cover repurchase and issuance activity, based on the work of Fama and French (2005). This measure considers the net value of repurchases or issuances. If the outcome is smaller than zero it means that a firm has done more repurchases than issuances and vice versa. For a specification on the net equity issued (or repurchased), a full overview of the variables and how they are constructed, you are referred to Appendix A.1: Annual Variables.
To sum up, a total of five different yield variables will be used for the annual time-series analysis.
To construct the dataset based on monthly data, the work of Fama and French (2005) and Boudoukh et al. (2007) will be used as a guideline. Due to data limitations, only measures for dividend yield and net payout yield will be examined. The data will be collected from CRSP through the WRDS database. Excess return is specified in the same way as before, but now based on monthly observations. Dividend yield is specified differently, because changes made by CRSP, which account for unreliable differences between the value weighted index return including- and excluding distributions. This problem was acknowledged by WRDS. Dividend yield is now calculated by taking the natural logarithm of the twelve-month
aggregated dividends and divide it by the end-of-period aggregated monthly market capitalization (i.e. dividend yield of December 2000 is the natural logarithm of the aggregated dividends of January 2000 through December 2000 divided by the aggregated market capitalization of December 2000). Net payout yield is defined as the natural logarithm of the twelve-month aggregated net equity issuances divided by the end-of-period aggregated monthly market capitalization. For a specification on net equity issuances and a full overview of the variables construction, you are referred to Appendix A.2: Monthly Variables.
A summary of the descriptive statistics on both datasets is presented Appendix B. Moreover, the correlation matrix on both datasets can be found in Appendix C.
17 3.2 Methodology
There are different ways to evaluate time-series models. The most basic form of time-series analysis concerns estimation. With estimation, you check how well the model its parameters explain variations in the dependent variable using data of multiple periods. However, to find predictability you have to look at future values. This can be done by doing an in-sample-regression. With an in-sample-regression the variable of interest is used to predict, in this case, future period excess returns, depending on the horizon. In this thesis, the excess return of one period ahead will be predicted using different payout yield measures. But, it is also possible to predict, for instance, the three-year ahead return, which implies that the horizon becomes larger. The focus of this thesis will be the predictability of the different payout yield measures around the financial crisis of 2008. When looking at longer horizons the number of observations decrease by the addition in horizon at the end of the sample, which leads to a decrease in the number of observations for post-crisis period. Therefore, only predictions for one period ahead will be made. Moreover, if predictability is found at a short horizon, predictability automatically increases when horizons become larger and therefore have less empirical value. (Cochrane (2001)). So, if predictability is found when predicting one period ahead, the predictive power would increase when the same analysis is done over longer horizons. The in-sample regression is as follows:
r𝑡+1 = 𝛼 + β1 · 𝑍𝑡+ 𝜀𝑡+1
Where r denotes the excess return at t+1, Z is one of the different payout yield measures, respectively, at time t and ε is the error term at t+1.
The dependent variable is the excess market return on the NYSE, AMEX and NASDAQ. The independent variable Z will be one of the different yield measures. There are five measures for the annual analysis and two measures for the monthly analysis. To do the in-sample regression the first lagged values of the explanatory variables will be taken to see how well the different yield measures explain future returns, which is the same as taking the lead value of return. To correct for potential serial correlation, Newey West standard errors will be used. To evaluate the in-sample-regression, the R2s and the coefficient have to be used. However, in-sample predictability does not say much about the performance of the model when using a different dataset. Therefore, out-of-sample regressions are very important. The prediction model is as follows:
𝑟̂
𝑡+1= 𝛼̂
𝑡+ 𝛽̂
𝑡 · Z𝑡The dependant variable is the predicted excess market return at time t+1 and the independent variable is one of the different payout yields at time t.
18 The out-of-sample regression evaluates the forecasting performance of the model, because in-sample models are more sensitive to outliers and data mining. For the out-of-sample analysis the parameters of the model that is used to make the prediction have to be estimated. This regression is repeated multiple times, based on a set number of observations, which is referred to as the look-back window. In this thesis the look-back window for the annual analysis will be set 60 observations (Boudoukh et al. (2007)). For the monthly analysis the look-back window will be expanded to 240 observations Campbell and Thompson (2008). Based on the look-back window you predict future returns.
To evaluate the results, the framework of Goyal and Welch (2003) will be used. They state that the root-mean-squared-error differential (dRMSE) has to be calculated, which is an evaluation measure for the out-of-sample results. The dRMSE is calculated by using the following formula:
𝑑𝑅𝑀𝑆𝐸 = √∑ (𝑟𝑡+1− 𝑟̅1:𝑡) 2 𝑇 𝑡=60 𝑇 − √ ∑𝑇𝑡=61(𝑟𝑡+1− 𝑟̂𝑡+1)2 𝑇
The dRMSE is the difference between the root-mean-squared-error (RMSE) of the prediction model and the RMSE of the benchmark. The benchmark in this thesis will be the historical mean of the risk
premium (Goyal and Welch (2003, 2008)). The aim is to find a dRMSE greater than zero, which means the prediction model outperforms the benchmark.
Another measure to evaluate the prediction model is the out-of-sample R2 by Campbell and Thompson (2008). If the R2 based on the prediction model is smaller than zero, it also means that the historical mean is a better predictor than the variable of interest.
𝑅
𝑂𝑂𝑆2= 1 −
∑
(𝑟
𝑡+1−
𝑇
𝑡=60
r̂
𝑡+1)
219
CHAPTER 4 Results
In this chapter the results of this research will be discussed. First, the results based on the timeframe of Boudoukh et al. (2007) will be compared to the results based on the full dataset (i.e. 1926-2018). Second, a closer look will be taken at the financial crisis. These results will be based on annual data. Third, the results based on monthly data will be discussed. Finally, some robustness checks will be done by using an alternative time-frame.
4.1 Analysis of the annual data
For decades, dividend yield was the most important finding for predicting time-varying expected returns. However, when in 1982 the 10B-18-rule was adopted by the SEC, the composition of payout changed for firms. This change in composition led to a discussion about the relevance of dividend yield as main predictor. In 2007, Boudoukh et al. examined different forms of payout yield, compared to dividend yield. In the following paragraph the results of the different measures of payout, based on the annual dataset will be discussed. Table 1 presents the results of the performed analysis, where the dependent variable is the excess market return and the independent variables are the different payout yield measures.
Panel A shows the results of the timeframe used by Boudoukh et al. (2007). Their sample covers a dataset from the 1926 to 2003.
For dividend yield a coefficient of 0.116 with a t-statistic of 2.21 was found, which is significant at a 5% confidence level. Furthermore, dividend yield has a R-squared of 5.5%. So, dividend yield explains 5.5% of the variance of future excess market returns. It is clearly visible that the other payout yield measures do a better job explaining future excess market returns, based on the results of Table 1. For instance, the cash flow-based measure of payout yield has a coefficient of 0.209 with a t-statistic of 3.35, which is significant at a 1% confidence level. Also, it has a higher R-squared of 9.1%. So, all other payout yield measures are significant at a higher confidence level than dividend yield and they have higher R-squared. Furthermore, the t-statistic of the remaining payout yield measures are all significantly higher than the t-statistic of dividend yield.
Moreover, the R-squared of both net payout yield measures is remarkably high. According to the results are the best in explaining future excess market returns. The results of panel A are interpreted by
Boudoukh et al. (2007) as evidence for the ongoing discussion about dividend yield losing its value as main predictor of returns and different measures of payout yield taking its position. Based on the evidence given in the first panel of Table 1, the treasury stock-based measure of net payout yield is best in predicting (in-sample) excess market returns, due to its high R-squared.
Panel B of Table 1 shows the results of the same regression, using the same measures, only now for a larger and more recent dataset, with a sample period 1926-2018. What is remarkable, is that almost all
20 values decrease in the new dataset. First, the coefficients become smaller and for some measures the significance of the coefficients become smaller as well. For instance, dividend yield is now significant at a 10% confidence level instead of a 5% level. Also, the confidence level of the treasury stock-based payout yield measure went from 1% to 5%. The confidence levels of the remaining payout yield
measures stay the same. However, all measures remain significant, based on the in-sample regression. A possible explanation for the negative change in statistical values for the different payout yield measures might be the financial crisis. It is a period of economic downtrends, where values of payout in firms tend to change (Bliss et al. (2015)). According to the in-sample results of panel B it cannot be concluded yet if the predictive power of all measures decreases.
Goyal and Welch (2003) argue that in-sample results do not say much about the predictive power of a regressor. To check for predictability, the out-of-sample results must be examined. A measure to check for out-of-sample is the root mean squared error differential (dRMSE). This measure compares the root mean squared error of the model, based on the different payout yield measures, to the root mean squared error of a benchmark. In this thesis the benchmark is the historical mean of the excess market return based on the data that is available up to the observation, before prediction. If the dRMSE is smaller than zero, it means the benchmark outperforms the prediction model of interest, which means the model has no significant predictive power, even though the model does have strong predictive power in the in-sample analysis. The dRMSE is stated in the eighth row of each panel of Table 1. For the sample period 1926-2003, dividend yield has a significant in-sample coefficient, but has a negative dRMSE. So, even though the in-sample coefficient of dividend yield is strongly significant, the prediction model is outperformed by the historical mean and therefore has no significant predictive power. The same holds for the treasury stock-based measure of payout yield. According to the in-sample analysis payout yield (TS) also has significant predictive ability, but the dRMSE is negative. Therefore, payout yield (TS) also has no significant predictive power. The remaining measures have a positive dRMSE and are considered good predictive measures for excess market return, where the cash flow-based measure of net payout yield is the best measure, based on its dRMSE.
Another measure to verify the predictive power of a measure is the out-of-sample R-squared. This measure is calculated by dividing the sum of squared residuals (SSR) acquired by the model of interest, by the SSR based on the historical mean. This ratio is then subtracted from one. The out-of-sample R-squared can be found in the ninth row of each panel. Based on the out-of-sample R-R-squared the cash flow-based measure of payout yield and both net payout yields have significant predictive power. When comparing the results of both panels it becomes clear that the value of dRMSE for payout yield (CF), which was positive in the sample 1926-2003, becomes negative in the sample 1926-2017. Also, the out-of-sample R-squared of payout yield (CF) becomes negative in the new sample, compared to the sample of panel A. These negative changes in value indicate that the predictive power of payout yield (CF) disappears in more recent datasets. The results also indicate that for both dRMSE and the
out-of-21 sample R-squared the values become smaller or even (more) negative for every yield measure, which might be an indication of the effect of the financial crisis of 2008 on predictability. According to Bliss et al. (2015), Brav et al. (2005) and Jagannathan et al. (2000) and Figure 1, firms reduced share repurchases mostly during the financial crisis. The combination of inconsistent returns (Am Al-Rjoub and Azzam (2012)) and the reduction in share repurchases could have led to the disappearance of predictability for payout yield.
In the next section a closer look will be taken at the financial crisis to see if this reduction in predictive power was caused by the financial crisis. According to the results of panel B of Table 1, the best predictor of excess market returns is net payout yield based on the cash flow calculation, due to its R-squared and significance.
22
Table 1: Return predictability using annual data.
The results shown here are based on the analysis of the samples using annual data. The sample consist of all non-financial firms (i.e. firms with SIC-code between 6000 and 6999 are excluded from the dataset) from the databases CRSP and COMPUSTAT. The dataset concerns data about excess market returns, dividends paid on common stock, purchases of common stock and sales of common stock. The dependent variable is the value-weighted total return (including distributions) minus the 90-day return on a US Treasury bill, which is denoted by excess market return. The independent variables are natural logarithms of the different payout yield measures, respectively. Dividend yield is computed by the difference between the value-weighted index return including and excluding distributions (available since 1926). Payout yield is the sum of dividend payments and repurchases of shares and is calculated in two ways. First, by adding the purchase of common stock (acquired from the cash-flow statements of COMPUSTAT, available since 1971) dividend by the year-end market capitalization to the dividend yield. Second, by adding the change in treasury stock divided by the year-end market capitalization (available since 1982) to the dividend yield. Repurchases based on treasury stock are assumed to be zero prior to 1982. Finally, net payout yield is defined as payout yield minus the natural logarithm of sales of common stock divided by the year-end market capitalization (available since 1971). Repurchases are not assumed to be zero prior to 1982. Therefore, the monthly change in shares outstanding is used to account for issuances prior to 1982 (available since 1926). For the calculation of this measure, see Appendix A.1.
Panel A presents the results of the analysis based on the timeframe used by Boudoukh et al. (2007). Panel B presents the results of the analysis based on the most recent timeframe available. All standard errors are Newey West standard errors. dRMSE (root mean squared error differential) is a calculation for out-of-sample predictability and uses a look-back window of 60 observations.
Dividend
Yield
Payout Yield
(CF)
Payout Yield
(TS)
Net Payout
Yield (CF)
Net Payout
Yield (TS)
Panel A: 1926-2007
Coefficient
0.116**
0.209***
0.172***
0.759***
0.721***
Constant
0.438**
0.715***
0.610***
1.607***
1.545***
Standard error
0.053
0.062
0.061
0.145
0.143
t-statistic
2.21
3.35
2.82
5.24
5.05
p-value
0.030
0.001
0.006
0.000
0.000
R
20.055
0.091
0.080
0.262
0.264
Observations
77
77
77
77
77
dRMSE
-0.0248
0.0015
-0.0105
0.0101
0.0061
R
2(OOS)
-0.341
0.019
-0.138
0.124
0.076
Panel B: 1926-2017
Coefficient
0.089*
0.197***
0.132**
0.732***
0.677***
Constant
0.360**
0.681***
0.491***
1.548***
1.466***
Standard error
0.046
0.061
0.054
0.143
0.142
t-statistic
1.93
3.23
2.45
5.12
4.77
p-value
0.057
0.002
0.016
0.000
0.000
R
20.038
0.076
0.052
0.224
0.222
Observations
91
91
91
91
91
dRMSE
-0.0272
-0.0023
-0.0123
0.0027
0.0048
R
2(OOS)
-0.193
-0.020
-0.114
0.024
0.043
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
23 4.2 Crisis based on annual data
This section focusses on the financial crisis. The sample has been split into three timeframes. The first timeframe covers the period before the financial crisis (i.e. 1926-2008). The second timeframe focusses on the period of the financial crisis occurs. In this thesis that timeframe is set to 2008-2012. The reason for the enlargement of this timeframe is, because there is no specific date on which the financial crisis ended. According to the National Bureau of Economic Research (NBER) the financial ended on June 2009, while Lin and Yeh (2017) say that the financial crisis only ended in the beginning of 2013 (for Europe). Figure 2 (Appendix B) also shows that the stock market slowly started to recover at the end of 2012. Therefore, to be sure the timeframe would cover the entire financial crisis, it is set on 2008-2012. The final timeframe, which covers the period after the financial crisis, is set on 2013 up to 2018. In Table 2 the results of the prediction analysis using the different yields variables as predictors for excess market returns can be found.
Table 2: Return predictability in the period of the financial crisis, using annual data.
The results shown here are based on the analysis of the samples using annual data. The sample consist of all non-financial firms (i.e. firms with SIC-code between 6000 and 6999 are excluded from the dataset) from the databases CRSP and COMPUSTAT. The dataset concerns data about excess market returns, dividends paid on common stock, purchases of common stock and sales of common stock. The dependent variable is the value-weighted total return (including distributions) minus the 90-day return on a US Treasury bill, which is denoted by excess market return. The independent variables are natural logarithms of the different payout yield measures, respectively. Dividend yield is computed by the difference between the value-weighted index return including and excluding distributions (available since 1926). Payout yield is the sum of dividend payments and repurchases of shares and is calculated in two ways. First, by adding the purchase of common stock (acquired from the cash-flow statements of COMPUSTAT, available since 1971) dividend by the year-end market capitalization to the dividend yield. Second, by adding the change in treasury stock divided by the year-end market capitalization (available since 1982) to the dividend yield. Repurchases based on treasury stock are assumed to be zero prior to 1982. Finally, net payout yield is defined as payout yield minus the natural logarithm of sales of common stock divided by the year-end market capitalization (available since 1971). Repurchases are not assumed to be zero prior to 1982. Therefore, the monthly change in shares outstanding is used to account for issuances prior to 1982 (available since 1926). For the calculation of this measure, see Appendix A.1.
Panel A presents the results of the analysis based on the pre-crisis timeframe. Panel B presents the results of the analysis during the financial crisis and panel C shows the effect after the financial crisis had ended. All standard errors are Newey West standard errors. dRMSE (root mean squared error differential) is a calculation for out-of-sample predictability and uses a look-back window of 60 observations.
Dividend
Yield
Payout Yield
(CF)
Payout Yield
(TS)
Net Payout
Yield (CF)
Net Payout
Yield (TS)
Panel A: Pre-crisis (1926-2007)
Coefficient
0.098**
0.201***
0.151***
0.750***
0.675***
Constant
0.383**
0.691***
0.548***
1.589***
1.456***
Standard error
0.047
0.061
0.056
0.144
0.143
t-statistic
2.11
3.28
2.70
5.20
4.73
p-value
0.038
0.002
0.008
0.000
0.000
R
20.045
0.086
0.069
0.258
0.245
Observations
81
81
81
81
81
dRMSE
-0.0330
-0.0006
-0.0143
0.0080
-0.0004
R
2(OOS)
-0.384
-0.008
-0.189
0.099
-0.005
continued24
Panel B: Crisis (2008-2012)
Coefficient
2.339
0.057
0.377
0.566
1.778
Constant
8.945
0.189
1.372
1.131
3.775
Standard error
1.012
0.324
0.566
0.914
1.370
t-statistic
2.31
0.18
0.67
0.62
1.27
p-value
0.104
0.872
0.553
0.580
0.293
R
20.675
0.002
0.022
0.020
0.217
Observations
5
5
5
5
5
dRMSE
0.0007
-0.0052
0.0019
-0.0084
0.0095
R
2(OOS)
0.062
-0.074
0.027
-0.120
0.129
Panel C: Post-crisis (2013-2017)
Coefficient
1.385**
1.092
1.576**
0.953
2.225
Constant
5.389**
3.480
5.814**
2.007
4.872
Standard error
0.266
1.296
0.392
3.709
2.165
t-statistic
5.21
0.84
4.02
0.26
1.03
p-value
0.014
0.461
0.028
0.814
0.380
R2
0.904
0.131
0.578
0.027
0.310
Observations
5
5
5
5
5
dRMSE
-0.0111
0.0059
-0.0106
0.0105
-0.0033
R
2(OOS)
-0.315
0.180
-0.370
0.308
-0.109
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Panel A presents the results of the analysis of the first timeframe (1926-2007). According to the in-sample results, all coefficients of the different measures of payout yield are significant. Apart from dividend yield, which is significant at a 5% confidence level, all coefficients are significant at a 1% level. So, according to the in-sample results, all measures have predictive power before the financial crisis occurs. However, when looking at the out-of-sample measure dRMSE, only net payout yield (CF) has a positive value, which indicates only net payout yield (CF) has predictive power before the financial crisis. When looking at the sample R-squared the same conclusion can be drawn, because the out-of-sample R-squared is the only measure with a positive value. Based on these results, it can be concluded that only net payout yield, based on the cash-flow measure, has predictive power with respect to excess market returns in the period before the financial crisis (1926-2007.
Panel B focusses on the period of the financial crisis (2008-2012). It is apparent that all coefficients become insignificant in this period, as could be expected. During a recession, firms are reluctant to change payout policies, because a change can be interpreted by investors as a signal of firms becoming
25 unhealthy due to the impact of the financial crisis (Miller and Rock (1985), Baker, Mendel and Wurgler (2016)). Indeed, the financial crisis of 2008 had a massive impact on firms and market returns. Dividend payments tended to change the least, because of the negative signal it can have on investors. But, repurchases drop substantially during times of an economic downturn (Bliss et al. (2015)). This drop in repurchase activity is also noticeable in figure 1. During a financial crisis it becomes more difficult for firms to raise funds. This was especially hard during the crisis of 2008, which was caused by defaults on subprime mortgages due to careless lending by banks. So, when in need of funds, a firm must turn to its shareholders by means of share issuances. The financial crisis also influences market return. Investors were more reluctant to invest, due to the high volatility, which led to a decline in prices and therefore returns (Lin and Yeh (2017)). Altogether, the change in payout policies and the drop of returns could be the cause of the disappearance of predictability in excess market returns. Moreover, the out-of-sample results of panel B have little meaning, unless the in-sample coefficients are significant. Since no
coefficient is significant in the in-sample regression, the out-of-sample results can be neglected. So, it can be concluded that during the financial crisis all measures have no predictive power.
Lastly, panel C shows the results of the analysis after the financial crisis (2013-2017). Based on the in-sample results, predictability has been restored for dividend yield and payout yield (TS). All other measures remain insignificant. The results of the after-crisis out-of-sample analysis are interesting. It seems predictability has been restored for dividend yield and payout yield (TS). But, when looking at the dRMSE and out-of-sample R-squared, the values are negative. Dividend yield has a dRMSE of -0.0111, while payout yield (TS) has a dRMSE of -0.0106, which means that the models are outperformed by the benchmark and therefore have no significant predictive power. A potential reason for the recovery of in-sample predictability can be the fact that firms were reluctant to cut dividends during the crisis, so that dividends could restore more quickly after the crisis. Also, returns started to show signs of growth again (see Appendix B). Although it is possible that the different yield measures have no predictive power post-crisis, this can also be due to the small sample size. Therefore, section 4.3 will take a closer look at the dataset, that is based on monthly data.
4.3 Analysis of the monthly data
In the previous section the focus was on annual data, which was based on a small sample when focussing on the effects of the financial crisis on predictability. Therefore, in the following sections the focus will be on monthly data. By using monthly data, the dataset expands substantially. However, due to data unavailability regarding payout yield, only measures of dividend yield and net payout yield can be used in the analysis. This research provides an addition to current literature by using monthly data to give a more precise and meaningful insight on predictability using different payout yields on excess market return. Section 4.1 described that in-sample predictability was found for all regressors, based on
26 the annual data. In this section the in-sample analysis of the monthly dataset will be discussed followed by the out-of-sample analysis. Results can be found in Table 3. Panel A presents the results based on the timeframe of Boudoukh et al. (2007). According to the results of panel A of Table 3 it can be concluded that, both dividend yield and net payout yield have predictive power. Both yield coefficients are significance at a 1% confidence level. The t-statistic of dividend yield is 3.58 along with a R-squared of 2.1%. Based on these results a strong predictive power was found for dividend yield. However, when comparing this with the t-statistic (4.28) and R-squared (2.8%) of net payout yield, it becomes clear that net payout yield has even more predictive power than dividend yield. So, based on the in-sample results it can be concluded that both measures have predictive power concerning excess market returns, where net payout yield has preference. The results of the out-of-sample analysis in panel A of Table 3 provide evidence that support that both dividend yield and net payout yield have predictive power when it comes to predicting excess returns. It is striking that opposed to the negative dRMSE and out-of-sample R-squared for dividend yield when using annual data, the dRSME and out-of-sample R-squared for dividend yield using monthly data is positive. Therefore, predictability for dividend yield is indicated by the monthly out-of-sample results, whereas the annual results indicate the opposite. This difference can perhaps be explained by modifications in the calculation of dividend yield. This calculation of dividend yield might be more consistent throughout the sample period 1926-2003.
The results of the analysis based on the full sample can be found in panel B. When using the full sample, the results are striking. Predictability of dividend yield has disappeared completely, opposed to net payout yield, where predictability remained. A reason for this disappearance of predictability in dividend yield could be due to omitted variable bias. Boudoukh et al. (2007) reasoned that omitting repurchases and issuances drives the decline, and in this case disappearance, in dividend yield predictability.
When looking at panel B of Table 3, the dRMSE and R-squared of dividend yield are negligible, because the in-sample coefficient of dividend yield is not significant. The dRMSE and out-of-sample R-squared of net payout yield however, supports the inference of the in-sample analysis of predictability in net payout yield.
By using the most up to date dataset available this research support the conclusions drawn by Boudoukh et al. (2007) that dividend yield loses its value as most important predictor and that net payout yield is a good alternative, because it is a better predictor, having an out-of-sample R-squared of 12.4%.
27
Table 3: Return predictability using monthly data.
The results shown here are based on the analysis of the samples using monthly data. The sample consist of all non-financial firms (i.e. firms with SIC-code between 6000 and 6999 are excluded from the dataset) from the database CRSP. The dataset concerns data about excess market returns, (US dollar) cash dividend paid (based on distribution code 12), net equity issuances and market capitalization. The dependent variable is the value-weighted total return (including distributions) minus the 90-day return on a US Treasury bill, which is denoted by excess market return. The independent variables are natural logarithms of both payout yield measures, respectively. Dividend yield is computed by dividing the twelve-month aggregated dividends (dividends summed over twelve months) by the month-end aggregated market capitalization. Net payout yield is defined as the twelve-month aggregation of net equity issuances (see Appendix A.1 for calculation) divided by the month-end aggregated market capitalization. Data on all variables available from 1926 onwards.
Panel A presents the results of the analysis based on the timeframe used by Boudoukh et al. (2007). Panel B presents the results of the analysis based on the most recent timeframe available. All standard errors are Newey West standard errors. dRMSE (root mean squared error differential) is a calculation for out-of-sample predictability and uses a look-back window of 240 observations.
Dividend
Yield
Net Payout
Yield
Panel A: 1926-2003
Coefficient
0.022***
0.046***
Constant
0.044**
0.069***
Standard error
0.006
0.011
t-statistic
3.58
4.28
p-value
0.000
0.000
R
20.021
0.028
Observations
924
924
dRMSE
0.0041
0.0046
R
2(OOS)
0.116
0.129
Panel B: 1926-2018
Coefficient
0.007
0.038***
Constant
0.001
0.056**
Standard error
0.005
0.010
t-statistic
1.45
3.60
p-value
0.146
0.000
R
20.003
0.018
Observations
1,098
1,098
dRMSE
0.0046
0.0051
R
2(OOS)
0.122
0.135
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
28 4.4 Crisis based on monthly data
In this part the sample will again be split into three different sub-samples to examine the effect of the financial crisis on the predictability of the different yield measures. However, this time monthly data will be used. By using monthly intervals, the sample contains more observations, which increases the accuracy, and making the results more precise. When looking at panel A of Table 4, the pre-crisis period, both dividend yield and net payout yield have positive coefficients and both are significant at a 1% confidence level. Dividend yield has a R-squared of 1.7%, while net payout yield has a R-squared of 2.6% Even though these R-squared values seem low, it can make a considerable difference for investors. Campbell and Thompson (2008) show in their work that a R-squared as little as approximately 0.5% can lead to an increase in portfolio return of more than 30%.
The results of panel A imply predictability, based on the in-sample analysis for both measures. The predictability is confirmed by the out-of-sample measures dRMSE and the out-of-sample R-squared, because they both have positive values for both payout yield measures. Based on the R-squared and the dRMSE net payout yield turns out to be the better predictor.
Panel B of Table 4 shows the results of the analysis during the period of the financial crisis. Not surprisingly, both dividend- and net payout yield have insignificant coefficients , during this period. During the financial crisis volatility in market returns increases, which leads to inconsistency in returns. This inconsistency also occurs in dividends and share issuances during a crisis. Although firms are reluctant to cut on dividends, many firms are forced to do so during economic recessions. The same holds for share issuances. Especially, since the financial crisis was caused by default of subprime mortgages, firms were forced to issue shares to raise funds since a loan could not be acquired through banks (Bliss et al. (2015)). The inconsistency in returns, dividends and share issuances could have led to the disappearance of predictability during the crisis.
Lastly, panel C of Table 4 presents the results for the years after the financial crisis. Surprisingly, only the predictability in dividend yield did restore. After the crisis, the coefficient of dividend yield is 0.112 and significant at a 5% level, while the coefficient of net payout yield remains insignificant. Furthermore, dividend yield has a R-squared of 7.1%. Although net payout yield is significant in the full sample (Panel B, Table 3) and dividend yield is not, this might be evidence that in the years after a big depression (or recession) has occurred, dividend yield is a better predictor for returns than net payout yield.
Other potential reasons could be that the market has not restored completely. Or that net payout yield needs more time to restore. Right now, the results provide evidence that after the financial crisis of 2008 dividend yield retakes its spot as main predictive variable. The out-of-sample R-squared of dividend yield is 48,4%, which is very high. One can argue that the results are therefore misleading and prone to overfitting. However, further research is needed to prove if this is true or not.
29 Table 4: Return predictability in the period of the financial crisis, using monthly data.
The results shown here are based on the of a sample split into three timeframes to cover the effects before, during and after the financial crisis, using monthly data. The sample consist of all non-financial firms (i.e. firms with SIC-code between 6000 and 6999 are excluded from the dataset) from the database CRSP. The dataset concerns data about excess market returns, (US dollar) cash dividend paid (based on distribution code 12), net equity issuances and market capitalization. The dependent variable is the value-weighted total return (including distributions) minus the 90-day return on a US Treasury bill, which is denoted by excess market return. The independent variables are natural logarithms of both payout yield measures, respectively. Dividend yield is computed by dividing the twelve-month aggregated dividends (dividends summed over twelve months) by the month-end aggregated market capitalization. Net payout yield is defined as the twelve-month aggregation of net equity issuances (see Appendix A.1 for calculation) divided by the month-end aggregated market capitalization. Data on all variables available from 1926 onwards.
Panel A presents the results of the analysis based on the pre-crisis timeframe. Panel B presents the results of the analysis during the financial crisis and panel C shows the effect after the financial crisis has passed. All standard errors are Newey West standard errors. dRMSE (root mean squared error differential) is a calculation for out-of-sample predictability and uses a look-back window of 240 observations.
