Equity Market Prediction. Evidence from Netherlands

Abstract

This thesis contributes to the international debate on stock return predictability by investigating the equity market of Netherlands. Therefore, several predictors of the four categories - market valuation, trend, sentiment, and macroeconomic (macro) -are examined. The present study leads to the result that the book-to-market ratio (BMR), a market valuation variable, is a strong positive predictor of future excess returns of the equity market of Netherlands over one-month to four-year forecasting horizons. It shows both in-sample and out-of-sample predictive power. The BMR even predicts future excess returns in the presence of the other variables investigated. Furthermore, it is observed that all variable categories provide some evidence that stock markets are predictable as they capture different information relevant for predicting the stock market.

Table of Contents

List of Figures

List of Tables

Table of Attachments

List of Abbreviations

List of Symbols

1 Introduction

2 Literature review and variable motivation

2.1 Literature review
2.2 Variable motivation

3 Data and summary statistics
3.1 Data
3.2 Summary Statistics

4 Methodology
4.1 Predictive regression framework (In-sample)
4.2 Out-of-sample methodology
4.2.1 Econometric specification
4.2.2 Forecast evaluation

5 Empirical Analysis
5.1 In-sample return prediction
5.1.1 Univariate regression results
5.1.1.1 General analysis
5.1.1.2 Regression results with market valuation variables
5.1.1.3 Regression results with trend variables
5.1.1.4 Regression results with the sentiment variable
5.1.1.5 Regression results with macro variables

5.1.2 Multivariate regression results
5.1.2.1 Bivariate regression results with market valuation variables
5.1.2.2 Bivariate regression results with trend variables
5.1.2.3 Bivariate regression results with sentiment variables
5.1.2.4 Bivariate regression results with macro variables

5.2 Out-of-sample return prediction

6 Conclusion

Appendix

References

List of Figures

Figure 1: Out-of-sample performance over time. 36

Figure 2: Fluctuation of the earnings-price ratio compared with the total return index. 47

Figure 3: Out-of-sample performance over time of the book-to-market ratio together with the term spread. 48

Figure 4: Volatility in the forecasting period. 48

Figure 5: Book-to-market ratio over time. 49

List of Tables

Table 1: Summary statistics of the Data 1. 14

Table 2: Summary statistics of the Data 2. 17

Table 3: Univariate regression results. 25

Table 4: Bivariate regression results. 31

Table 5: Out-of-sample R-squared statistics. 38

Table 6: Univariate regression results for the PY not in logarithm form. . 47

Table of Attachments

Appendix A: Data construction 43

Appendix B: Robustness test of the univariate regression of PY 47

Appendix C: Figures with complementary information 47

List of Abbreviations

Abbildung in dieser Leseprobe nicht enthalten

1 Introduction

Equity Market Prediction is an quite interesting topic for investment bankers and the academia. It plays an important role in topics like asset allocation, asset pricing, risk management and capital budgeting. Being able to predict the capital markets would result in a huge gain for investors. Even companies may benefit from equity market prediction, because they could time the market by deciding for example the optimal time of an initial public offering (IPO) or pricing this IPO correctly without leaving money on the table. Therefore, this bachelor thesis examines different predictor variables, that are grouped into market valuation, trend, sentiment, and macroeconomic (macro) variables.

Predictor variables are variables that are said to be able to predict the equity market. To test the predictability of these predictors this thesis runs several in-sample and out-of-sample prediction trials with a defined regression framework. In-sample, both univariate as well as multivariate regressions are carried out. Out-of-sample, the predictive power of each predictor is tested stand-alone and compared to a simple benchmark model. In the end a trading strategy resulting from these return predictions may be evaluated.

The focus of this bachelor thesis is the equity market of the Netherlands. The Amsterdam Stock Exchange is one of the oldest or even the oldest stock exchange of the world. Several interesting companies like Adyen (fintech company) and ASML (semiconductor company) are listed at the Netherlands market. However, this thesis is not about predicting individual stock returns, but about predicting the Netherlands stock market in general, and therefore, a broad stock index (the Netherlands-Datastream Market) is investigated, that contains (nearly) every stock of the Netherlands.

The dependent variable of the regressions made in this study is defined as the stock market returns of the Netherlands market in excess of the risk-free rate. The risk-free rate is in the finance world a common used term for a risk-free government bond. The obtained returns are frequently also called excess returns, equity market risk premium or comparable throughout the literature and this bachelor thesis. The predictors examined in this thesis are both, already well-known predictors, as well as predictors that have not yet been tested that frequently, as they have been proposed by literature published more recently.

Furthermore, this bachelor thesis contributes to the ongoing debate of stock market predictability by examining the Netherlands market. Therefore, it expands the horizon of the debate, as academic literature mainly focused on the US stock market in the past. As a consequence, it might be possible that results found in this bachelor thesis may differ from results for predictors tested in the US or other stock markets.

The thesis first summarizes the evidence of stock market predictability by giving an overview of already published literature to explain the necessity of out-of-sample tests above in-sample tests. Furthermore, the choice of the predictor variables investigated in the thesis is explained (Chapter 2). In Chapter 3 the construction of the variables used for the regressions is presented and the data is described with the usage of summary statistics to gain a better understanding of the data. The methodology for in- and out-of-sample return predictions is described in Chapter 4 to outline how the results presented in Chapter 5 can be obtained and reproduced.

Chapter 5 first starts with the univariate in-sample regressions to determine which predictor or predictors shows the strongest forecasting power. As there are good results obtained for the book-to-market ratio (BMR), the multivariate regression than checks whether the BMR still predicts the equity market risk premium when controlling for the other predictors. Moreover, the multivariate regressions strengthen the evidence whether the different categories of predictors explain variation of excess returns that is not already explained by the BMR.

Finally, the out-of-sample return prediction discusses if the predictors tested in-sample also show out-of-sample predictive power, and therefore shows whether the underlying models can be applied also in the future. This is done by using a regression methodology that ensures that the models have not yet seen the data of the out-of-sample period and by comparing the predictor models with a simple benchmark model (the historical average).

The conclusion (Chapter 6) assesses the thesis from a critical point of view and outlines future research questions. Furthermore, it proposes an approach for a trading-strategy in the equity market of Netherlands.

2 Literature review and variable motivation

The following section gives an overview of already published literature and explains the choice of predictor variables used in this thesis. Furthermore, the necessity of testing variables out-of-sample is outlined.

2.1 Literature review

It is obvious, that the literature is now trying since several decades to find a predictor variable that is capable to predict future returns at a high level, so that practitioners can rely on it. In the past the equity market prediction literature started with valuation ratios such as the dividend yield (e.g., Fama and French (1988)¹ ; Campbell and Shiller (1988)² ), earnings-price ratio (e.g., Basu (1977)³ ; Welch and Goyal (2008)⁴ ; Campbell and Thompson (2008)⁵ ) and book-to-market ratio (e.g., Kothari and Shanken (1997)⁶ ; Pontiff and Schall (1998)⁷, Campbell and Thompson (2008)⁸ ) and macro variables such as inflation and interest rates (e.g., Fama and Schwert (1977)⁹ ; Campbell and Thompson (2008)¹⁰ ; Rapach, Strauss and Zhou (2010)¹¹ ). However, these variables, that were already tested several times in the US, showed only medium in-sample predictability and low, if any, out-of-sample predictability as Welch and Goyal (2008) conclude.¹² So most models are unstable or even spurious, according to them.¹³ This means the models were supposed to be somewhat useless.

So, at the end of the first decade of the 21th century the evidence of predictability of stock market returns was heavily questioned. The idea that stock returns are not predictable was based on the Efficient Market Hypothesis, that suggests that the prices deliver the only source of information and resulting from that academia should not be able to find predictors that show predictability. Consequently, investment managers should not be able to outperform the market. Otherwise, markets would be inefficient.¹⁴

In the more recent past new variables were developed that strengthen again the evidence for stock market predictability. One of these variables is the implied cost of capital (ICC) that was examined, for example, by Li et al. (2013) and outperformed every other variable in their paper used for predicting the US market returns.¹⁵ Despite the ICC, which is a valuation variable, also variables in the trend and sentiment category were examined. Neely et al. (2014) introduced the first technical variables by constructing three trend-following strategies, that deliver “buy” or “sell” when certain criteria are fulfilled.¹⁶

In the sentiment category several sentiment indexes were generated, for example Baker and Wurglers (2012) investor sentiment index, or quite recently the manager sentiment by Jiang et al. (2019).¹⁷ These newer variables often aggregate information into one variable and perform better as for example the ICC of Li et al. (2013)¹⁸. For example the sentiment index of Baker and Wurgler (2012) combines four different proxies for sentiment.¹⁹ In their initial paper Baker and Wurgler (2006)²⁰ even used six different proxies for investor sentiment. After all, they found a negative relationship between stock market returns and their sentiment index.²¹ The manager sentiment index of Jiang et al. (2019) also negatively predicted excess returns, with large in-sample and out-of-sample predictive power in their study.²²

Hence, investors as well as managers seem to be overoptimistic about future stock returns and maybe sentiment variables in general are negatively correlated with the equity market risk premium.

Today the question arises whether the predictability of stock market returns exists also on a global level. There is only a small amount of literature testing the predictability of the equity market risk premium outside the US, and therefore, it is difficult to reach a conclusion at this stage.

Resulting from that, this bachelor thesis wants to make a contribution to this issue by investigating the Netherlands capital market. In the following, the evidence on the prediction of the Netherlands equity market risk premium is summarized.

Assefa, Esqueda and Mollick (2017)²³ have tested the change of interest rate to predict stock market returns on a global level. They divided the countries in developed and developing countries, counting Netherlands to the developed countries. Altogether, they found a negative correlation with stock market returns in both developed and developing countries, but a more significant effect in developed countries, like the Netherlands. Therefore, this thesis will examine short and long term interest rates as well.

Another academic paper from Hjalmarsson (2010) examined the dividend-price-ratio, earnings-price ratio, short interest rate and the term spread for several developed and emerging markets.²⁴ Hjalmarsson (2010) found a rather weak evidence for the possibility that the earnings-price ratio predicts stock returns, and reported an insignificant slope coefficient for the Netherlands for this predictor, as can be seen in table 4 of their paper.²⁵ For the dividend-price ratio Hjalmarsson (2010) reported stronger results than for the earnings-price ratio on global level²⁶, but again not a significant result for the Netherlands market. For both, the dividend-price ratio and the earnings-price ratio, Hjalmarsson (2010) found insignificant positive relationships.

In contrary to the valuation ratios, Hjalmarsson (2010) obtained significant results for the short interest rate in 8 out of 23 developed markets, among them the one of the Netherlands.²⁷ The observed relationship between the short interest rate was often negative, also for the Netherlands. Also for the term spread (defined as log difference between 10 year government bonds and short interest rates in their paper), they found significant results for a bunch of developed countries, again among them the Netherlands, but this time the slope coefficients reported by them mostly show a positive relationship with the respective market returns.²⁸ In addition, they also examined the out-of-sample power for the above mentioned variables and found that significant results in-sample are often associated with out-of-sample predictive ability, in particular for the term spread.²⁹ Reasoning from that, this paper concludes that a predictive component for stock market returns might be related to interest rate variables, like the short interest rate or the term spread.³⁰

To summarize, one can see that already some work has been done for examining the capital markets on a global level, but more variables have to be investigated and more research is necessary to foster the evidence of equity market prediction on the global level.

2.2 Variable motivation

The following section outlines the motivation for the variables used to predict excess returns throughout the present thesis.

This bachelor thesis investigates variables, that have been investigated traditionally, but also investigates new variables, that came along in the more recent past, with the aim to explain variation of stock market returns that has remained unpredictable with traditional variables. Therefore, the investigation of market valuation, trend, sentiment, and macroeconomic (macro) variables seems to be reasonable.³¹

Abbildung in dieser Leseprobe nicht enthalten

Where is the current stock price, the cost of equity and dividend at time . This model discounts expected cashflows (dividends) by using the cost of equity of the respective listed company as discount rate.

The cost of equity is often calculated with the capital asset pricing model (CAPM) formula:³²

Abbildung in dieser Leseprobe nicht enthalten

Where is the cost of equity of a company, the risk-free rate, the risk factor, that measures the market risk in relation to equity risk, and the return of a market index. is the excess return, also called equity market risk premium throughout the thesis. Based on this theoretical framework, it makes a lot of sense to include predictor variables in the analysis that should influence the dividend discount model or are somehow connected to it. So, it is reasonable to look at the dividend yield (DY) as potential predictor as dividends are directly reflected in the DDM. The same applies for the payout yield (PY) which incorporates another component of cash flows that an investor receives by investing into a stock. Valuation ratios like the earnings-price ratio (EPR) or book-to-market ratio (BMR) are more indirectly connected, as for example, the book-to-market ratio opposes book value of equity to the market value of equity, and the market value of equity is connected with the price of a stock.

Consequently, beyond these market valuation variables also interest rate and interest rate related variables are chosen for the research of this bachelor thesis. Interest rates negatively influence the equity market risk premium as the risk-free (interest) rate is directly subtracted in the calculation of it.

Furthermore, variables that influence the interest rates by influencing the overall economy, like the Inflation rate (INFL), are also motivated by the dividend discount model. These market valuation and macroeconomic (macro) variables have already been tested in the US market (see section 2.1).

However, most of these market valuation and macro variables showed poor in-sample and out-of-sample predictive power for the US market as concluded by Welch and Goyal (2008)³³. Therefore, new variables are required like the implied cost of capital (ICC) which is in the first instance equivalent to , which is used as discount rate in the DDM.³⁴

The ICC is then aggregated at market level in order to calculate an implied market return (ICOC). As a predictor the implied equity market risk premium (EICC) is used in publications like in the paper of Li et al. (2013)³⁵. Also for forecasting the equity market risk premium of the Netherlands stock market the EICC is used as predictor in Chapter 5 of this study. In this thesis EICC is calculated by subtracting the risk-free rate from the implied market return (ICOC), that was provided by the Chair of Financial Management and Capital Markets of the Technical University of Munich. A more detailed calculation is given in section 3.1 and in the Appendix A. Under the assumption that EICC is a good proxy for the actual equity market risk premium, it is expected that EICC positively correlates with it. However, the methodology for deriving the ICOC is based on analysts forecasts, and as a matter of the nature of a forecast it can even deviate a lot from reality.

Other variables, like trend and sentiment variables, are not so straightforward motivated, but it is reasonable to consider them as potential predictors as these variables might explain the more psychological part of the stock market. So, for example, the volatility index is seen as a “fear gauge index”³⁶, and a high volatility index is expected to negatively correlate with excess returns³⁷.

As a sentiment variable, this paper therefore investigates AEX VIX as a proxy for an imaginary VIX of the Netherlands-Datastream Market. As trend variables, the two best performing (measured by their in-sample and out-of-sample results) moving average and momentum indicators respectively proposed by the research paper of Neely et al. (2014)³⁸ are examined. Respective variable motivating research papers are added in brackets to the description of the Data in Section 3.1.

3 Data and summary statistics

This Chapter describes the data aggregation process and dicusses some summary statistics of the data used in Chapter 5 - Empirical Analysis.

3.1 Data

The sample period spans from the end of December 1989 to the end of September 2020 and all variables are computed by taking the data of the end of the respective month. Following Li et al. (2013)³⁹ continuously compounded excess returns ( ) are used as dependent variable in the regression. This variable is computed as the difference of continuously compounded return of the NETHERLAND-DS Market - TOT RETURN IND, downloaded from Datastream with the code TOTMKNL and the RI Datatype, and the continuously compounded risk-free rate. The one-month continuously compounded return is calculated as . The underlying Index is a total return index, which means that dividends are included in the index levels.

The risk-free rate (treasury bill) ( ) is constructed as described in the Online Appendix of Schmidt et al. (2019)⁴⁰. In detail, that means that a combination of the Interbank rate (IBR) of the Netherlands and overnight indexed swap rate (OIS) of the Netherlands is used. The IBR data is downloaded from Refinitiv Datastream by using the code HOLIB3M, and the OIS data is obtained with the code OIEUR3M. The construction of the is made as follows: Where only the IBR is available, the IBR is used as proxy, where both are available, is constructed as the minimum of IBR and OIS, and where only the OIS is available, as the IBR is discontinued in the sample period, the OIS is used as a proxy for the risk-free rate. Both IBR and OIS are annual yields and provided in percentage points. The risk-free rate has to be divided by 100 to get „normal“ values again. The monthly continously compounded risk-free rate is then calculated as . So the one-month excess return is calculated as follows:

Abbildung in dieser Leseprobe nicht enthalten

To follow the predictive regression framework of Li et al. (2013)⁴¹, one-period ahead excess returns/future market returns have to be used in the empirical analysis, and this results in the term in the formula.

The excess returns for the longer horizons ( ) are then calculated as the rolling sum of the one-month excess returns for the different time windows of 12, 24, 36 and 48 months with minimum observations. is (12, 24, 36, 48) and denotes the forecasting horizon for the longer horizons. All excess returns ( ) are normalized by dividing ( ) by for the 12, 24, 36 and 48-month horizon. Taking the rolling sum is possible as the one-month return is continuously compounded. The excess returns are then multiplied by 100 each to convert them to units of percent per month, as I want to work with more familiar units. The multiplication by 100 or 1200 for annualized percentages is common in the literature and should not affect the regression results.⁴²

The predictor variables used in this thesis are grouped into the following categories: Market valuation, trend, sentiment and macro variables. For detailed calculation of the variables see Appendix A.

Market valuation:

· Log Dividend yield (DY): Which is the logarithm (log) of the dividend yield (divided by 100) obtained from Refinitiv Datastream as TOTMKNL(DY). TOTMKNL is the index code and DY is the datatype used. [See, e.g., Fama and French (1988)⁴³ ; Campbell and Shiller (1988)⁴⁴ ; Rapach, Strauss and Zhou (2010)⁴⁵ ]

· Log Payout Yield (PY): The PY is calculated as . The NPY is defined as the sum of the dividend yield (from Datstream) and the net share buyback yield. The net share buyback is calculated as the total share repurchases minus share issuances in the respective year.⁴⁶ The dividends used for the construction of the dividend yield datatype of Refinitiv Datastream are also the dividends of the whole corresponding year. I add a little number to the net payout yield (NPY) in order to make the log be defined again, because the log is not defined for negative values. This approach was also done by Boudoukh et al. (2007) and they don´t find the statistical significance affected even if they add larger values like 1.01 to the NPY.⁴⁷ The exact calculation of the NPY can be found in Appendix A. Due to the calculation of the NPY this variable only spans from 1990-12-31 to 2020-09-30. [See, e.g., Boudoukh et al. (2007)⁴⁸ ; Eaton and Paye (2017)⁴⁹ ]

· Log Earnings-price ratio (EPR): Which is obtained by taking the log of the inverse of the price-earnings ratio from Datastream as TOTMKNL(PE). [See, e.g., Basu (1977)⁵⁰ ; Welch and Goyal (2008)⁵¹ ; Campbell and Thompson (2008)⁵² ]

· Book-to-market ratio (BMR): Which is obtained by taking the inverse of the Price to Book Value (TOTMKNL(BP)) from Datastream. [See, e.g, Kothari and Shanken (1997)⁵³ ; Pontiff and Schall (1998)⁵⁴, Campbell and Thompson (2008)⁵⁵ ]

· Implied equity market risk premium (EICC): EICC is calculated by subtracting the risk-free rate from the implied market return (ICOC), that was provided by the Chair of Financial Management and Capital Markets of the Technical University of Munich. The detailed methodology for the determination of the ICOC can be found in Berg et al. (2017)⁵⁶. A detailed description of the computation of the EICC for the different forecasting windows is given in Appendix A. [See, e.g. Hou, van Dijk and Zhang (2012)⁵⁷ ; Li et al. (2013)⁵⁸ ]

Trend variables:

This study investigates moving average (MA) and momemtum (MOM) trading rules. The MA trading rules generate a “Buy”-signal if the short moving average is greater or equal to the long moving average, and a “No Buy”-signal otherwise. The momentum trading rules generate a “Buy”-signal if the price index level is greater or equal to its level n periods ago, and a “No Buy” signal otherwise. In this thesis and the MOM trading rules are denoted as MOM(n).

- MA(1, 12): Moving average trading rule, which compares the short moving average with length 1 and the long moving average with length 12. [See, e.g. Neely et al. (2014)⁵⁹ ]
- MA(2, 12): Moving average trading rule, which compares the short moving average with length 2 and the long moving average with length 12. [See, e.g. Neely et al. (2014)⁶⁰ ]
- MOM(9): Momentum trading rule, which compares the level of a price index Pt with its level 9 month earlier Pt-9 . [See, e.g. Neely et al. (2014)⁶¹ ]
- MOM(12): Momentum trading rule, which compares the level of a price index Pt with its level 12 month earlier Pt-12 . [See, e.g. Neely et al. (2014)⁶² ]

More detailed definitions about the trend variables can be found in Appendix A. Due to the definition of the trend variables, the moving average rules span from 1990-11-30 to the end of the sample period. MOM(9) spans from 1990-09-28 to the end of the sample period, and MOM(12) spans from 1990-12-31 to the end of the sample period.

Sentiment variable:

- Volatility Index (VIX): The AEX Index is an Index that consists of the 25 largest stocks of the Netherlands Market. In the sample period it has a correlation of 0.9942 (rounded) to the total return index. So it is plausible to use the AEX VIX as proxy for a hypothetical volatility index of the Netherlands-Datastream Market index. The AEX VIX is downloaded from Datastream by using the code AEXVOLI. Unfortunately this data only spans from 2000-01-31 to 2020-09-30 as for the time before no data is available in Refinitiv Datastream. [See, e.g., French, Schwert and Stambaugh (1987)⁶³ ; Mollick and Assefa (2013)⁶⁴ ]

Macro variables:

- Inflation (INFL): Which is the year on year percentage change of the consumer price index (CPI) of the Netherlands. The used data is the already calculated Inflation downloaded from Refinitiv Datastream by using the Code NLCONPR%F. I adjust for the reporting lag by downloading the data one month prior to the sample period and using, for example, the November data of 1989 to predict the excess returns of January and so on. In other words that means, is used in equation (4) to account for the reporting lag (see Chapter 4). This is common in the literature and, for example, also done by Rapach, Strauss and Zhou (2010)⁶⁵. [See, e.g., Fama and Schwert (1977)⁶⁶ ; Rapach, Strauss and Zhou (2010)⁶⁷ ; Campbell and Thompson (2008)⁶⁸ ]
- Long term yield (LTY): 10 year continuously compounded government bond yield, divided by 12 to get monthly values and then multiplied by 100 for percentage values. The 10 year government bond yield is provided by the Chair of Financial Management and Capital Markets of the Technical University of Munich. [See, e.g., Rapach, Strauss and Zhou (2010)⁶⁹ ; Campbell and Thompson (2008)⁷⁰ ]
- Short term yield (STY): The short term yield equals the monthly continuously compounded risk-free rate calculated by The STY is also multiplied by 100 for percentage values. [See, e.g, Campbell (1987)⁷¹ ; Ang and Bekaert (2007)⁷² ]
- Term spread (TERM): This is the difference of the LTY and STY described before. [See, e.g, Campbell (1987)⁷³ ; Fama and French (1989)⁷⁴ ; Hjalmarsson (2010)⁷⁵ ]

3.2 Summary Statistics

Abbildung in dieser Leseprobe nicht enthalten Table 1: Summary statistics of the Data 1.⁷⁶

The table displays the mean, standard deviation and autocorrelations of all variables used in the thesis except the trend variables. The implied equity market risk premium (EICC), the excess return (ER), the book-to-market ratio (BMR), the log dividend yield (DY), log earnings-price ratio (EPR), term spread (TERM), Inflation (INFL), long term yield (LTY) and short term yield (STY) are expressed in monthly data and span from the end of December 1989 to the end of September 2020. The log payout yield (PY) spans from December 1990 to September 2020. VIX spans from January 2000 to September 2020. LTY and STY are continously compounded and TERM is the difference of LTY and STY. EICC, ER, TERM, LTY and STY are reported in annualized percentages the other are variables either in logarithm or reported as they are. The autocorrelations of ER represent the sum of autocorrelations up to the given lag, and the autocorrelations for all other variables represent the autocorrelations at the respective lag. The data displayed in the table is rounded to two digits. The exact construction of these variables is explained in section 3.1 and in Appendix A.

Table 1 shows summary statistics of all variables except for the trend variables. The average EICC is 5,48%, which is the average implied excess return and is comparable to the actual one-month average return of 5.38%. The autocorrelations of the implied equity market risk premium are decreasing, but positive, and are still above 0.20 after 60 months, which suggests good persistency for the EICC.

The sum of the autocorrelations of excess return (ER) is negative only at the 60 month horizon, which suggests a mean reversion tendency of stock returns at a rather late horizon.

The mean of the book to market ratio is 0.65, which means that the aggregated book value of equity of the companies in the total return index accounts for 65% of the aggregated market value of equity in average. Therefore, the total return index of the Netherlands stock market would be considered as overvalued in average of the time by investors following the logic of the book-to market ratio.

The autocorrelation of the BMR is at the first lag at 0.97 and hovers above 0.50 even until lag 24. As shown in Table 1, it decreases only slowly with about a 0.20 decrease between the lags and becomes negative only after 60 months. This means that the BMR is quite persistent, especially at the short term horizon, and therefore increased predictive power can be expected.

The autocorrelation of the DY is the same at lag 1 as the one of BMR, but then decreases faster, and it is already below 0.40 at lag 24 and becomes negative already after 48 months. A reason for the positive autocorrelation for DY in the beginning may be that companies tend to increase the dividend yield to signal that they are financial stable. In a recession DY is shortened as earnings break away. Bad earnings often yield in bad earnings the next year again, so this would again imply positive autocorrelation for dividend yields. Nevertheless, in this study only an average persistency for DY is found.

In contrast to the other market valuation variables Table 1 reports very low positive and negative autocorrelations for PY at every lag with an absolute value below 0.10. So, PY is very low autocorrelated, and there is no trend over time compared to the other variables. This results in PY of being the variable with the lowest persistency of the predictors in this study.

Low persistency of predictors is expected to result in low predictive power, and vice versa as concluded by Boudoukh, Richardson and Whitelaw (2008)⁷⁷. Campbell (2001)⁷⁸ introduced this idea and came to the result that an increase in the R2 -statistic with the holding period is possible, when the predictor variable is variable and persistent. The autocorrelation of DY may be high compared to PY as DY might have an omitted variable bias. This omitted variable could have been the net share buyback yield, that reflects the extra cashflow gained by investors from share buybacks or issuances on top of the DY.

In contrary to DY, PY shows a relativly large standard deviation. Consequently, the high standard deviation is induced by the net share buyback yield component.

For the EPR the same pattern in the autocorrelations as for the DY is reported by Table 1, but with a significantly lower autocorrelation at lag 12 and higher autocorrelations at lags 24-60.

After having discussed the valuation ratios, the performance of the sentiment variable and the macro variables is described. The VIX, which is a measure for the volatility of the Index, unfortunately only spans from the year 2000 until the end of the sample period. But irrespective of that the VIX has a mean of 22.01 and a standard deviation of 9.76. The autocorrelation has a value of 0.85 at lag one and then drops rapidly to 0.23 at lag 12 and already becomes negative until lag 24. At lag 36 and 48 it remains almost on the same level, whereas in the end it increases again a little bit, but remains still negative. Thereof, one can record the fact that the VIX variable shows a rapid mean reverting tendency as the direction of the values is changed quite often. So, the VIX does not show much persistency as the autocorrelations are also low in magnitude. However, Boudoukh, Richardson and Whitelaw (2008)⁷⁹ point out that low persistency does not mean that the predictor can´t show predictive power over the holding periods, as the predictors often contain overlapping information across the multiperiod forecasting horizons.

The Term spread (TERM) variable shows a mean of 1.06 annualized percentages and a standard deviation of 1.09. As the mean of the LTY is higher than the mean of the STY, the mean of TERM is greater than zero, which means that a “normal” yield curve in average for the Netherlands market is observed. An inverse yield curve is often seen as a sign for an upcoming recession by practitioners and academia.⁸⁰

[...]

---

¹ Cf. Fama/French (1988), p. 3.

² Cf. Campbell/Shiller (1988), p. 198.

³ Cf. Basu (1977), p. 663.

⁴ Cf. Welch/Goyal (2008), p. 1458.

⁵ Cf. Campbell/Thompson (2008), p. 1511.

⁶ Cf. Kothari/Shanken (1997), p. 169.

⁷ Cf. Pontiff/Schall (1998), pp. 141f.

⁸ Cf. Campbell/Thompson (2008), p. 1511.

⁹ Cf. Fama/Schwert (1977), p. 143.

¹⁰ Cf. Campbell/Thompson (2008), p. 1513.

¹¹ Cf. Rapach/Strauss/Zhou (2010), p. 831.

¹² Cf. Welch/Goyal (2008), p. 1504.

¹³ Cf. Welch/Goyal (2008), p. 1504.

¹⁴ Cf. Elton et al. (2014), p. 410.

¹⁵ Cf. Li/Ng/Swaminathan (2013), p. 420.

¹⁶ Cf. Neely et al. (2014), p. 1775.

¹⁷ Cf. Baker/Wurgler/Yuan (2012), p. 281; Jiang et al. (2019), p. 127.

¹⁸ Cf. Li/Ng/Swaminathan (2013), p. 420.

¹⁹ Cf. Baker/Wurgler/Yuan (2012), p. 274.

²⁰ Cf. Baker/Wurgler (2006), p. 1655.

²¹ Cf. Baker/Wurgler/Yuan (2012), p. 286.

²² Cf. Jiang et al. (2019), p. 127.

²³ Cf. Assefa/Esqueda/Mollick (2017), p. 31.

²⁴ Cf. Hjalmarsson (2010), p. 51.

²⁵ Cf. Hjalmarsson (2010), p. 73; Hjalmarsson (2010), p. 71.

²⁶ Cf. Hjalmarsson (2010), p. 73.

²⁷ Cf. Hjalmarsson (2010), p. 74.

²⁸ Cf. Hjalmarsson (2010), p. 74.

²⁹ Cf. Hjalmarsson (2010), p. 78.

³⁰ Cf. Hjalmarsson (2010), p. 78.

³¹ Cf. Berk/DeMarzo (2017), p. 314.

³² Cf. Berk/DeMarzo (2017), p. 440.

³³ Cf. Welch/Goyal (2008), p. 1504.

³⁴ Cf. Li/Ng/Swaminathan (2013), p. 421.

³⁵ Cf. Li/Ng/Swaminathan (2013), p. 420.

³⁶ Mollick/Assefa (2013), p. 1.

³⁷ Cf. Mollick/Assefa (2013), p. 7.

³⁸ Cf. Neely et al. (2014), p. 1776; Neely et al. (2014), p. 1786; Neely et al. (2014), p. 1775.

³⁹ Cf. Li/Ng/Swaminathan (2013), p. 423.

⁴⁰ Cf. Schmidt et al. (2019), p. 217.

⁴¹ Cf. Li/Ng/Swaminathan (2013), p. 423.

⁴² Cf. Campbell (1987), pp. 375f.

⁴³ Cf. Fama/French (1988), p. 3.

⁴⁴ Cf. Campbell/Shiller (1988), p. 198.

⁴⁵ Cf. Rapach/Strauss/Zhou (2010), p. 830

⁴⁶ Cf. Boudoukh et al. (2007), p. 884.

⁴⁷ Cf. Boudoukh et al. (2007), p. 892; Boudoukh et al. (2007), p. 894.

⁴⁸ Cf. Boudoukh et al. (2007), p. 892.

⁴⁹ Cf. Eaton/Paye (2017), p. 1643.

⁵⁰ Cf. Basu (1977), p. 663.

⁵¹ Cf. Welch/Goyal (2008), p. 1458.

⁵² Cf. Campbell/Thompson (2008), p. 1511.

⁵³ Cf. Kothari/Shanken (1997), p. 169.

⁵⁴ Cf. Pontiff/Schall (1998), pp. 141f.

⁵⁵ Cf. Campbell/Thompson (2008), p. 1511.

⁵⁶ Cf. Berg et al. (2017), pp. 158–165.

⁵⁷ Cf. Hou/van Dijk/Zhang (2012), p. 506.

⁵⁸ Cf. Li/Ng/Swaminathan (2013), p. 420.

⁵⁹ Cf. Neely et al. (2014), p. 1775.

⁶⁰ Cf. Neely et al. (2014), p. 1775.

⁶¹ Cf. Neely et al. (2014), p. 1775.

⁶² Cf. Neely et al. (2014), p. 1775.

⁶³ Cf. French/Schwert/Stambaugh (1987), p. 3.

⁶⁴ Cf. Mollick/Assefa (2013), pp. 2f.

⁶⁵ Cf. Rapach/Strauss/Zhou (2010), p. 831.

⁶⁶ Cf. Fama/Schwert (1977), p. 143.

⁶⁷ Cf. Rapach/Strauss/Zhou (2010), p. 831.

⁶⁸ Cf. Campbell/Thompson (2008), p. 1513.

⁶⁹ Cf. Rapach/Strauss/Zhou (2010), p. 831.

⁷⁰ Cf. Campbell/Thompson (2008), p. 1513.

⁷¹ Cf. Campbell (1987), p. 376.

⁷² Cf. Ang/Bekaert (2007), p. 655.

⁷³ Cf. Campbell (1987), p. 376.

⁷⁴ Cf. Fama/French (1989), p. 26.

⁷⁵ Cf. Hjalmarsson (2010), p. 51.

⁷⁶ Based on Table 1 of Li et al. (2013), p. 423.

⁷⁷ Cf. Boudoukh/Richardson/Whitelaw (2008), p. 1581.

⁷⁸ Cf. Campbell (2001), p. 461.

⁷⁹ Cf. Boudoukh/Richardson/Whitelaw (2008), pp. 1588f.

⁸⁰ Cf. Brettell (2020).