A content analysis of Dutch financial news media and the interaction with the Dutch stock market

(1)

A content analysis of Dutch financial news media and the

interaction with the Dutch stock market

Abstract

This paper measures the interactions between sentiment obtained from the Dutch financial media and the stock market returns of the AEX. By quantitatively measuring news content of the front page of Het Financieele Dagblad, I create a pessimism variable to proxy for sentiment. Using a specialized software program this paper uses a straightforward and easy replicable approach to test classification, showing an 85% classification accuracy rate. The findings in this paper, against prediction, find no significant causality between sentiment and stock returns, supporting traditional theories that sentiment does not play a role in the determination of asset prices.

JEL classification: G14; D81

Key words: Behavioral Finance, Sentiment, Dutch stock market, Content Analysis, AEX index, Pessimism, the Netherlands.

Degree: MSc BA Finance Date: 28 Augustus 2012 Name: Alwin Sander Tulner Supervisor: Viola Angelini

(2)

Introduction

With his paper, Tetlock (2007) was one of the first researchers that explored the interaction of the content of news and the stock market. By analyzing the content of information, Tetlock (2007) showed that noise trader sentiment can persist in financial markets and impact the pricing of assets. Furthermore, his paper answered a more relevant question; that of how to measure investor sentiment and to what extend this measurement affects the stock markets.

Building on the foundation provided by Tetlock (2007), this paper investigates the interactions between pessimism and stock market returns in the Netherlands. Using pessimism as a direct proxy for sentiment, this research tries to capture the tone of news content classified by either negative or non-negative1. Deviating from the methodology regarding the pessimism media factor used by Tetlock (2007), this research adds to the literature by using the research model of Tetlock (2007) on a different type of financial news media in a non US country, namely the Netherlands. In addition this paper developed a content classification method based on Naïve Bayes, allowing self-developed specialized software to analyze content regardless of language, thereby allowing it to be easy replicable for other languages, supporting comparable research in other non – English speaking countries. Indeed, accuracy tests show that 85% of the content obtains a correct classification, which is an impressive statistic for such a straightforward method. Moreover, by arguing that the content of financial news media can influence a population I generate a pessimism factor by aggregating the classification of the front page news content of Het Financieele Dagblad2. The aggregated pessimism factor thereby captures the most relevant news as indicated by its front page positioning. Furthermore, the sentiment literature suggests causality between sentiment and stock returns. However the direction of the causality is not clear. To answer this question, I employ a vector autoregressive (VAR) framework, in which I use data on both variables for the year 2010. The results for the Netherlands presented in later sections, are not in line with the findings by Tetlock (2007). The model does not show any significant causality between sentiment and stock returns for the Netherlands. Interestingly, it does show a small, yet negligible significant causality pointing in the direction of returns influencing sentiment when lags are introduced to the model.

This papers key contribution is firstly regarding the type of content. By analyzing the content of a country’s leading financial newspaper front page, this paper suggest to draw a

1_{Later on in this paper, also indicated by goodness and badness to articulate the difference with respect to} pessimism induces via negative articles.

2

(3)

measure of pessimism from different content explored in earlier research. Secondly, this research adds to the existing literature by moving the perspective to another region, where current research on this subject predominantly focuses on the United States.

Lastly regarding the content quantifying technique used, where Tetlock (2007) employs a program specific to the English language3, this paper, studying the Dutch market, employs a technique allowing for similar content quantification while being independent on the language of the content. However, the software allowing for the technique to classify content was not written by me, I merely supplied the algorithms. Furthermore, with regard to the quantifying content technique, the method used in this paper, resembles that of Antweiler and Frank (2004), with the difference that this method can be seen as a more simplified method allowing for easy replicable research in other countries and on different type of content. For example content which can be found on the internet.

The remainder of this paper is structured as followed. Section I provides a literature review on sentiment and its impact on stock prices, combined with some background to the concept of sentiment. After setting out the theoretical model and hypotheses, Section II elaborates on the methodology used in order to test for the interaction under investigation. Also, the process of creating a pessimism factor from news content is explained in this section. Section III details the dataset and provides summary statistics on the dependent variables together with an accuracy test of the self-made classification software. After this, the findings are presented in Section IV together with robustness checks. Section V concludes the paper.

I Literature review and background

In 1912, Selden wrote Psychology of the stock market based upon the belief that movement of price on the exchange to a very considerable degree is dependent, on the mental attitude of investors and traders. Many years later, this belief formed the inspiration for a study by DeBondt and Thaler (1985) who studied the effect of unexpected and dramatic news events on stock market returns and found a significant overreaction. Furthermore, their paper gave rise to a new branch of financial research, a field now known as behavioral finance.

With respect to behavioral finance, Paul Slovic once wrote, “a full understanding of human limitations will ultimately benefit the decision-maker more than will naïve faith in the

3

(4)

infallibility of his intellect.”(Slovic 1972,780). With the financial crisis of 2008 fresh in mind, it is not surprising that interest in behavioral finance has picked up. The notion by Slovic (1972) is becoming more a plausible view than that of investors, advisors and bankers, knowing perfectly well what to do. The goal of the following paragraphs is to introduce the concept of behavioral finance to the reader, thereby presenting the reader with the central ideas of behavioral finance, organized around the themes of limited arbitrage and investor sentiment. Once presented with this theoretical framework, models of investor sentiment will be discussed on which predictions regarding the research questions are based. At various points, the paper will touch on findings and results with respect to the field of behavioral finance. However, it will not be a survey, but more a presentation of themes to create a perspective for the reader to place the research topic, the role of media, and in specific, the role of financial news media content, within the concept of investor sentiment.

A. Behavioral Finance4

Advocates of behavioral finance dispute the standard view that, dictated by modern neoclassical finance theory, financial decision making is rational. Instead, they suggest that poorly informed and unsophisticated investors lead to the view that financial markets are inefficient. The debate between neoclassical and behavioral finance is wide ranging and can be explained by the way both views model behavior in terms of individual decision making. Traditionally, financial economist model behavior of financial decision making based on the notion of rationality. In short, it is assumed that investors make rational decisions indicating that pricing in the market is rational. Moreover, markets are efficient and prices fully reflect their fundamental values.

On the other end of the financial decision making spectrum, the field of behavioral finance challenges the neoclassical school of thought by arguing against the concept of ‘rationality’. The starting point for the contrast lies in one of the major foundations on which behavioral finance theory rests, namely that of sentiment. Behavioral finance supporters argue that sentiment impacts the prices of all assets, thereby driving the difference between what behavioral and neoclassical finance tell us about the relationship between risk and return (DeBondt et al. 2008). With regard to sentiment, a good starting point is the theory by psychologist and sociologist Herbert Simon (1957), who found that individual decision

(5)

making, is affected by behavioral preferences, which are affected by a number of limitations. Simon (1957) elaborates by proposing that when people formulate their judgment, they use all available information, making judgment itself limited by the scope of this available information. Furthermore, Simon (1957) introduced the idea that certain cognitive limitations existing in the mind of people, affect judgment. 5 The theory was summarized under the term ‘Bounded Rationality’, and now forms one of the major underpinnings of sentiment theory. Moreover, based on this theory behavioral finance argues that with regard to financial decision making an individual investor often forms market expectations with limited information. These limitations bring risk and uncertainty to the decision making process, and challenge the notion in traditional finance theory by assuming that people make rational decisions to maximize their wealth in the face of risk and uncertainty (Nofsinger, 2002). In reality, it seems that whenever money is involved, logic not always overcomes emotion and psychological biases.

Moreover, when faced with the difficult task of judging the likelihood of possible events occurring, psychology dictates that people are limited to a number of strategies, called heuristics, to simplify their decision making process. The field of behavioral finance links these heuristics, to investor sentiment. One of those biases particularly interesting to this research is the availability bias. The availability bias, or availability heuristic as it was labeled by Tversky and Kahneman (1982), shows that investors seem to overweigh information that is easily accessible. In addition, a study by Barber and Odean (2008) find that people are more likely to remember events that receive a lot of attention by the media, arguing that the media can induce certain feeling states.

Since this research is interested in the interaction between financial news media and stock returns, I will elaborate more thoroughly on the notion of sentiment in the following section. In addition, the role of financial news media in the context of investor sentiment is discussed.

B. Sentiment

One of the major foundations of behavioral finance is investor sentiment, which is best explained as the theory of how people actually form their beliefs and valuations, and more generally, how investors form their demand for securities. The literature surrounding the sentiment theory became more popular in periods in which the stock markets were

5

(6)

experiencing highly volatile episodes. Example of such periods are the Dot Com bubble early 2000 and late 1990s, or more recently the sub-prime mortgage market failure and the credit crisis in 2008 that was the result of this. Moreover, these events have formed a challenge to the rationality assumption in the classic literature. To understand these wild movements in stock prices, which seem to move prices away from fundamentals, recent theoretical models of security markets relax the ‘rationality’ assumption and turn to sentiment to explain the price movements.

While the media frequently cite investor sentiments as a major force in the stock market (Richard, 2010), current literature provides no universal agreement on what constitutes sentiment, as there can be multiple factors influencing the behavior of an investor. A good example of this is the ever growing branch of literature investigating investor mood on asset prices. The mood factor used in this line of research is used as a proxy for sentiment suggesting returns are related to changes in human emotions. An interesting research is done by Hirschleifer and Shumway (2003) who find that sunshine is strongly correlated with stock returns. They argue that sunlight affects the mood of people, and that people tend to evaluate future prospects more optimistically when they are in a good mood. On the other end, just as an investor can misattribute any positive feelings from sunlight, they can also misattribute negative feelings from other factors in their environment. Take for example the paper by Edmans et al. (2007) who composed a mood indicator from sport results and found a significant effect on a country’s stock market after a loss of the national team of the country. Edmans et al. (2007) explain their findings by stating that the win or a loss by a country’s national team can evoke large mood swings within the population of a country, thereby lending itself to be a good mood indicator. However, Edmans et al. (2007) do not find evidence of a similar effect after a win of the country’s national team.

B.1. Sentiment: The role of financial news media

(7)

picking up. Where quantifying the qualitative nature of this type of information has formed the main barrier (see Cutlet et al., 1989), this barrier seems to deteriorate slowly, thereby bringing research closer to solving this interesting puzzle.

Take for example two recent studies, summarized later on in this paragraph, that have found interesting relationships regarding news and stock market price movements. To overcome the barrier of quantifying the content of financial news media these researchers began to measure linguistic data contained within the articles using computational linguistics. The first, a study by Antweiler and Frank (2004) focuses on the content of internet news, in specific the information found in chat rooms regarding stocks. Within these chat rooms, buy, sell or hold recommendations for specific stocks were discussed and in this same logic, Antweiler and Frank (2004) classified these messages. Based on textual analysis, they created a quantitative bullishness factor. In turn, the bullishness factor was tested for predictability on stock market price changes. They found that stock market messages help predict market volatility, although they do not find a statistically significant effect of “bullish” messages on returns. Another study quantitatively measuring the relation between the media and the stock market is that by Tetlock (2007), who creates a measure of pessimism from a popular news column in the Wall Street Journal using a completely automated program known as the General Inquirer (GI) to convert words into numerical values which represent a media pessimism factor. Using this media pessimism factor as a proxy for negative sentiment, Tetlock (2007) finds that high media pessimism predicts downward pressure on the market prices followed by a reversal to fundamentals, and that unusually high or low pessimism predicts high market trading volume. In a later paper by Tetlock et al. (2008), previous research was extended and it was argued that the content and in specific, the tone of the financial news articles, captured information beyond fundamentals. Given that individual investors form market expectations to make financial decision with limited information, and that these investors possibly suffer from a belief bias, I argue that the media can induce feeling states, and in specific, the feeling of pessimism, which in turn influences the decision making of individual investors.

The paper by Tetlock (2007) forms the guideline of this research, with the exception6 that a part of the used methodology will resemble that of Antweiler and Frank (2004). More

6

(8)

important, the theoretical framework, from which the statistical hypothesizes are drawn, will have its roots in the paper by Tetlock (2007).

Going back to the effects of financial news media, I argue that a series of negative financial news reports can induce general pessimism among the population of investors in a country. As a leading financial newspaper in a country will reach a large group of the investor population, it is able to facilitate mood swings among investors by inducing pessimism. Furthermore, I argue that general pessimism will alter the risk perception of this group of investors giving rise to sentiment, in turn depressing a nation’s stock market returns.

I find support in the seminal work of Johnson and Tversky (1983), who find that after participants read a newspaper article containing negative news, these participants gave a higher likelihood to various other negative events happening. The authors argued therefore that negative emotional reactions to articles could influence people’s general risk estimates. In turn strengthening my reasoning behind the concept of financial news media able to induce general pessimism, when it reports on varies of financial topics. As for the reasoning behind the ability of general pessimism depressing a nations stock return, I argue that the emotional reaction to negative newspaper content influences the investor’s general risk estimates by inducing fear. As many individual investors use their knowledge of the stock market to obtain either income now or in the future, in the form of wages or a pension, fear might induce a change in the allocation of income. I argue that increased uncertainty that accompanies these financial decisions, influenced by pessimism leads to actions in order to minimize uncertainty. This concept is also known as “willingness to take a safer bet”, introduced first by Raghunathan and Pham (1999), who argue that fearful people will take cognitive behavioral actions to minimize uncertainty and the unpleasant state of fear. In the next section, I have set out the model of investor sentiment in which I will place my study on the impact of financial news media content, its induced pessimism and the stock market.

C. Research question and hypotheses

(9)

is important regarding sentiment theory, both investor types can be separated based on how they form their beliefs. The first type of investors is called the rational arbitrageur, who forms rational expectations about assets returns, being immune to sentiment. The second type is the irrational investors, in the literature often identified as noise traders, who form their expectations either too positive or too negative, relative to the rational arbitrageurs. This is in accordance with the idea that the difference in beliefs between both types of investors can be seen as sentiment.

The model with both these types of investors leads to a situation in which equilibrium arises where the beliefs of noise traders about future returns influence current returns. More specific, when the noise traders experience a negative belief shock and give in to overall pessimism about the future, they respond by selling stocks in favor of less risky investments like bonds. In this situation, the arbitrageurs buy these stocks, which will temporarily depress returns. With regard to possible hypotheses that can be extracted from above described model, it has to be noted that the timing of pessimism induced by the media is important to estimate the result of its effect. In short, whether sentiment influences stock returns, or whether stock returns provoke sentiment.

I investigate if a series of negative news media reports, inducing pessimism, used as a proxy for sentiment, results in decreasing returns on a country’s national stock index. Taken that the direction of the prediction by the model is unclear, it is unknown whether the pessimism factor generated from the financial news media forecasts future investors sentiment in the form of high pessimism through news or past investor sentiment. This leads to two different hypotheses to account for this causality question. The first hypothesis predicts that in periods of high pessimism it is expected to find low short-term returns and a reversal to fundamentals in the long term. The second hypothesis predicts high media pessimism following low past returns, and expecting high future returns. Since this research focuses on the short run effects of interactions between sentiment and returns, using one year of data, the reversal prediction will not be investigated.

(10)

By building on the theoretical model used by Tetlock (2007), this research will test the model’s prediction on the interaction between the Dutch financial news media and its effect on the Dutch stock market AEX. The research question is as follows:

- Does general pessimism induced by Dutch financial news media (content), reflecting investor sentiment, have a negative effect on the Dutch stock market index AEX?

The research question allows me to formulate two testable hypotheses regarding the interaction between financial news content and stock market returns, with special regard to the fact that the timing of both factors influences the outcome of hypothesized predictions as stated under the theoretical model. The formulated hypotheses take the following form: ଵ: High pessimism, reflecting negative sentiment, predict negative short-term AEX index returns

ଶ: Negative short-term AEX index returns predict high pessimism, reflecting negative sentiment

Using the same method as Telock (2007), I employ a vector autoregressive model to test my hypothesis. The VAR model is explained in the methodology.

II Methodology

In this section, I will elaborate on the specialized software to convert textual data into numerical data, and discuss the mathematical algorithms used. This section will also contain a formalization of the research question into a statistical hypothesis and the statistical methods used to test for the interaction between the constructed pessimism factor and index returns.

A. Financial news content classification

(11)

this process for each word found in articles, the program generates a probability of an article being ‘negative’. Mathematically it looks as followed:

Consider a stream of words _௜ that is found in either message type T or its anti-type, were T stands for positive messages and for negative messages. Furthermore, let m denote the number of occurrences of word ௜ in message type T and let denote the number of occurrences of word ௜in message type so that;

= (|௜), and = (|௜) (1)

For words found in the training set the ‘goodness’ and ‘badness’ of words can now be calculated according to the following formulas.

௜ = | ௜) |௜) + (|௜) (2) Where ௜ = ௜) |௜) + (|௜) (3)

Once the level of goodness or badness for each word has been calculated, the formula used to classify an article in either good or bad is as followed;

When, ∑ _௜ < ∑ ௜ , Article is classified as negative (4) When, ∑ ௜ > ∑ ௜ , Article is classified as positive (5)

Using each individual article classification, I create a daily pessimism factor by taking the absolute number of ‘negative’ classified articles on a daily basis. This factor varies between zero and four, where four articles is the average of articles relevant to this research published on the front page of the newspaper ‘Het Financieele Dagblad’. Lastly, due to the arbitrary aggregation process, results have been tested using different aggregation methods. Using differently generated pessimism factors I obtained qualitatively similar results. The results of using a differently constructed pessimism variable are briefly discussed in the robustness section.

B. Naïve Bayes content classification

(12)

media into one pessimism media factor on a daily basis. The method is simple in that it works by looking at the words that make up the article individually, and counting the number of positive and negative words in each of the articles. This method is sometimes referred to as a ‘bag of words’ approach, since word location is completely disregarded. Furthermore, the method employed in this research is very similar to that of Naïve Bayes in that it treats every word completely separate from any other word appearing in the articles. This treatment strokes with the key assumption underlying the Naïve Bayes classification which assumes that words are independent of each other. Even though the aforementioned assumption seems highly unrealistic, Naïve Bayes performs well in practice.7

C. Vector autoregressive (VAR) model

To investigate the interaction between the content of financial news media and the AEX index returns I argue that the tone of financial news content can induce certain feeling states. Specific to this research, I study the impact of negative financial news content, aggregated into a single daily pessimism factor, as a proxy for negative investor sentiment. Furthermore, I take into account that causality might exist between news induced sentiment and the stock market returns by the use of a vector autoregressive (VAR) approach. This approach allows me to effectively estimate the directional causality between stock market returns and news induced sentiment simultaneously. The model, as earlier indicated, is based on the paper by Tetlock (2007), and has the following form:

௧ = + ௝

௟

௝ୀଵ

∗ ௧ି௝+ ∗ ௧+ ௧ (6)

Where, _௧ is a two variable vector; _௧ and !"_௧. Pfactor represents the daily constructed pessimism factor at time t, and AEXd represents the daily return on the AEX index at time t. X at time t is the parameter for control variables. In addition, all VAR estimates include # lags8 representing the number of days before market activity. The endogenous variables in the model are ௧ and !"௧. To correct for any

7_{As an empirical matter it has been found that a surprisingly small amount is gained at substantial cost by} attempting to exploit grammatical structure in the algorithms (Manning and Schutze , 1999)

(13)

heteroscedasticity and autocorrelations in the residuals of the standard errors, I employ Newey and West (1987) robust standard errors.

Lastly, to better understand the model as described under (6), one can simplify the model by separating the model into two different equations and thereby assuming independence in the error terms. Equation (7) and (8) as described below allows for testing each hypothesis separately using ordinary least square (OLS) techniques.

௧= 1 + ଵ∗ ௧ି௝ ௟ ௝ୀଵ + ∗ ଶ∗ !"௧ି௝ +ଷ∗ ௟ ௝ୀଵ ௧+௧ (7) !"௧= 1 + ଵ∗ ௧ି௝ ௟ ௝ୀଵ + ∗ ଶ∗ !"௧ି௝+ଷ∗ ௟ ௝ୀଵ ௧+௧ (8)

III Data and Descriptive Statistics

To test for the interaction between the content of financial news media, using the country’s leading newspaper in financial news, and the country’s stock market index, I will use an automated software program to convert the qualitative, textual nature of the content of the financial news media into a quantitative pessimism factor to test for index returns. Furthermore, the regional focus of this research lies on the Netherlands, where the focus of earlier regional focus in this strand literature was predominantly on the United States9.

To the best of my knowledge, the non existence of Dutch text analysis software, with the specific ability to separate positive news from negative news, this research will use a self-developed software program to analyze the content of Dutch financial news media. This section reads as follows. First, I will explain the data used for this research after which follows a section dedicated to detail the creation of the daily pessimism variables accompanied by descriptive statistics of the data set comprised by news articles.

9

(14)

A. AEX index and control variables

To test for the interaction between Dutch financial news media and the Dutch stock market, this investigation uses the AEX index to represent the Dutch stock market. The AEX index represents the performance of the 25 largest stocks by market capitalization in the Netherlands. Since this study focuses on the year 2010, AEX index returns are calculated on a daily basis from January 1st, 2010 to December 31st, 2010.

Furthermore, this research focuses on the Netherlands, which is a relatively small economy with respect to the overall world economy. In light of this, movement in the AEX index might be strongly correlated to different economic indicators worldwide. To control for this, I add two control variables. The first control variable is the MSCI World Index for stock markets, as part of the AEX Index is represented in this index as well. The MSCI world index is a stock market index consisting of over 1,600 stocks selected from all developed markets. Thanks to the scale of this index, it provides a good indicator of general economic conditions worldwide. Secondly, to control for economic performance I added real world interest rates, and in specific 10 Year German bond yields. The 10 year German Bond Yields have a triple AAA rating which makes them suitable as a proxy for the risk free rate. For both control variables and the AEX index variable I consulted Thompson Reuters Datastream.

In addition, this research uses a Dutch leading financial newspaper ‘Het Financieele Dagblad’ as source for the textual data needed to analyze and construct a pessimism factor. As a leading newspaper, I argue that a large percentage of the investor population is exposed to this type of media, thereby able to influence and facilitate mood swings within the Dutch investor population. To obtain the articles needed for this investigation, the LexisNexis10 database of news articles is used. The creation of the pessimism factor is detailed in next paragraph.

B. Generating the Pessimism Factor

Covering a time period of one year, the year 2010, I employ a specialized, self-developed software package11 for text classification, which has formulas encoded into an algorithm. The specific used algorithms are explained in the methodology section. To convert text to numbers, the software package allows performing two key procedures. First, I start by manually classifying a training set of 100 articles. Next, the software package processes the

10_{LexisNexis is an international firm, in the possession of a large database on articles collected daily from a} great variety a text media.

(15)

articles into a training data set which allows for the construction of an algorithm that can classify articles, based on the word statistics available in the training data set. In the second step, the classification software allows for the filtering of my entire sample of articles to obtain a ‘good’ or ‘bad’ classification for each article in the dataset. This research uses twelve months of financial news data, of which I selected the first month, January, to serve as the sample data set. For each day in the time period under investigation I gather the articles published on the front page of Het Financieele Dagblad with more than 100 words. This way I obtain all articles suited for classification, and take out possible ‘noise’ from the dataset. Take for example the daily overview of news summarized in small headlines on the front-page. This section provides an oversight; however the articles indicated are not published on the front page and therefore lie outside the focus of this research.

Lastly, this research defines ‘negative’ news as articles containing negative events, indicated by the usage of negative words and ‘positive’ news as any other type of non-negative news. Using two types of classifications, namely ‘positive’ and ‘non-negative’ this method resembles a text classification method used by Antweiler and Frank (2004), who successfully classified internet board room message into ‘Buy’, ‘Hold’ and ‘Sell’ recommendations by using the Naïve Bayes classification method.

Table II provides some summary statistics on the accuracy of the software program, which is built on algorithms based on Naive Bayes theory. As indicated by the last line of Table I, a total of 1146 articles got classified. Overall, more negative news got published than any other news being non – negative indicated by the last row of Table I with 63% of the entire dataset of articles being classified as negative. To assess the efficiency of the content analysis method I compare the results between the manually classified sample set and the classification outcomes of this sample set according to the software program. The results show that the content analysis software classifies a negative article correctly 85% of the time.

A.1. Aggregating content into a single daily pessimism variable

(16)

Table I

Accuracy of text classification software within Sample and Overall Classification Distribution

The table gives an insight into the data generated using the self-developed text analyzer by providing summary statistics on the classified messages of both the sample straining set and the full data set. The first column indicated with percentage, shows the percentage of articles classified by me manually into positive or negative. The Table then shows the accuracy of the content analysis approach by classifying the predefined sample set and comparing these results to the manual outcome.

By Algorithm Classified: by Me % Positive Negative Positive 47 91 9 Negative 53 15 85c Training seta 51 49 All messagesb(1146) 37 63 a

The total number of 98 news articles posted in the month of January on the front-page of ‘Het Financieele Dagblad’

b _{This line gives an oversight on how the total data set is distributed between positive articles and negative} articles. The total data set, from which a daily pessimism variable is constructed, contains 1146 news articles.

c

Showing, with a 85 percent accuracy rate, that this rather straightforward classification method is highly accurate.

Table II

Summary statistics on daily front page news of Het Financieele Dagblad

This table shows the distribution of articles with more than 100 words published on the front page of Het Financieele Dagblad.

Number of Articles on the Front Page

2 3 4 5 6 7 8

Number of Occurencesa 0 29 256 1 1 1 0

Excluding Weekend Editionb 0 19 218 0 1 1 0

a

The total number of occurrences showing up here deviates from the 260 data points in the data set due to holidays, weekend editions, and those days that there was no market activity.

b

This row removes weekend editions, however holidays and days without market activity, still make up for the difference between total occurrences and total data points.

(17)

mood of the investor population is assumed to be equal to for instance, the difference between pessimism factors 2 and 3. The same logic applies for the difference for the other pessimism factors within the factor range.

Since the aggregation method possibly influences the outcome of the VAR analysis, this research also constructs an alternative pessimism factor by using a different aggregation method. As an alternative pessimism variable I create a binary factor which equals 0 for when the number of negative articles published on a given day is below 2.49, as indicated by table I being the average amount of negative articles published on a daily basis. The binary variable shows 1 for days on which more than the average number of negative articles is published. In short, the alternative pessimism factor shows 0 for days on which the amount of negative articles published on the front page is below average and shows 1 for days when the number of negative articles published on the front page is above average, indicating pessimism. The test results between the two variables elaborated on in this section are similar, and therefore the output of the alternative pessimism variable is detailed in the Appendix and only shortly discussed in the Findings section.

C. AEX index returns and the pessimism factor

(18)

For more detailed summary statistics on a monthly basis, I refer to Table XIII. Table XIII in the appendix, confirms above inferences showing May having the highest variation for both the AEX returns and the daily pessimism factor, as indicated in the Standard Deviation column.

Figure I

AEX index Versus the Sentiment Proxy Pessimism

This graph depicts the dataset who makes use of data over year 2010. The X-axis represents the 12 months. The top line represents the sentiment proxy pessimism between 1.5 and 4 whereas the bottom line shows AEX index movement over the year 2010, moving between 300 and 360.

Table III

Summary statistics on the main dependent variables

This tables provides summary statistics on the main dependent variables; AEXD as the daily AEX return and Pfactor, as the daily constructed Pessimism factor as a proxy for sentiment.

Mean Median Maximum Minimum Std. Dev. Observations

AEXD 0.01% 3.80% 7.07% -4.34% 1.26% 259

PFACTOR 2.49 3 4 0 1.02 259

IV Findings

This section presents the results on whether financial newspaper sentiment has the ability to predict future stock market returns. In addition, due to the possible reverse causality between sentiment and stock market returns, I employ a Vector Autoregressive (VAR) model to allow for simultaneous testing of both hypotheses. As explained in the data section, the pessimism construction method is rather arbitrary, and therefore above VAR model also tests the hypotheses using the alternative pessimism variable explained in detail in the data section.

(19)

The results however are comparable to that of the main pessimism factor constructed and therefore will only briefly be elaborated on. The results for the alternative pessimism variables can be found in the robustness section and accompanying tables and graphs in the appendix.

Next paragraphs will report results from Granger-Causality tests, impulse responses and forecast error variance decompositions as dictated by standard practice regarding VAR analyses (Stock and Watson, 2001). In the order mentioned above, test results will be reported using graphical representation, followed by an economic interpretation of the findings. Furthermore, also discussed in the data section, the VAR model makes use of two exogenous variables representing the control variables. From an informational point of view both tests, using the main pessimism factor and the alternative pessimism factor are also done on the VAR model without the control variables. The results of these tests can be found in the Appendix.

A. Granger – Causality statistics

As indicated in the methodology, a critical element of VAR models is the determination of lag length, indicated by # in equation (6). To determine the appropriate lag length, I employ a lag length criteria test.12 Next to relying on this lag length criteria test, I rely on theoretical foundations to assess the appropriate number of lags at which the VAR model is minimized. By building on above explained selection criteria, this research sets the number of lags in the VAR model at # is 6. This number is partly strengthened by the lag length criteria test, where the LR test (Sequential modified LR test statistic) indicates that the optimal number of lags in my VAR model is minimized at 6. As indicated, the LR test supports the assessment of optimal lag length number partly, since other criteria13 all indicate the VAR model should be minimized at zero lags. To validate the decision of 6 lags, I build on theory regarding the incorporation of sentiment into stock prices, which seems logically extracted from the sentiment model, not to be instant. Moreover, it is reasonable to assume that pessimism has a delayed effect on the stock returns, or that stock returns have a delayed effect on pessimism. Therefore, combing the findings presented in Table VI, indicating 6 lags usage, and theory detailed above, 6 lags seems appropriate. Nevertheless, to provide a clear overview on how different selected lag operators influence the model, Table IV provides an oversight of

12_{The results of this test can be found in Table VI, which can be found in the appendix.}

(20)

causality between the AEX returns and Pessimism using different numbers of lags in the model. Also, the robustness section reports on the results when 0 lags are selected to perform the VAR.

Table IV

Granger causality between the daily pessimism factor and the daily AEX index returns

Granger – Causality statistics examine whether lagged values of one variable help to predict another variable. Below table summarizes the findings of multiple Granger – Causality tests using different lag operators for the two vector variable VAR. The table shows the p-value associated with their F-value to test whether coefficients on lags in the model are zero. AEXd represents the daily AEX index return and Pfactor represents the daily constructed Pessimism factor as a proxy for sentiment.

Direction of causality # of lags F-Value p-value Decision

AEXd → Pfactor 2 0.81 0.4473 Do not Reject

Pfactor → AEXd 2 1.12 0.3264 Do not Reject

AEXd → Pfactor 4 1.28 0.2797 Do not Reject

AEXd → Pfactor 6 2.47 0.0246a Reject

AEXd → Pfactor 8 2.54 0.0114 Reject

a

At 6 lags the p-value of 0.0246 shows that at the 5% significance level AEX index returns predict the daily pessimism factor. Using 8 lags, the VAR model shows the same result at the 2 % percent significance level.

To help interpret above results, if the constructed daily pessimism factor does not help predict the daily AEX index returns, then the coefficients on the employed number of lags in the model will all be zero. As can be seen from Table IV, the usage of a specific number of lags make a big difference regarding causality findings between the two main dependent variables; AEXd and Pfactor, representing daily AEX index returns and the daily pessimism factor. Starting at 6 lags and continuing to 8 lags, the model indicates at respectively the 5% and the 2% significance level that daily AEX index returns predict the daily pessimism factor. Using less than 6 lags however, does not give findings indicating any causality between the two variables.

(21)

factor used for this research is constructed using the front page of a leading Dutch financial newspaper. Therefore, it gives rise to the notion that technology might cause this news to be obsolete. For example, the internet has the ability to spread the news reported in the newspaper considerably earlier. Next section shows the results of impulse response test which give an indication on the strength of the causality, if there is any.

B. Impulse response statistics

To further test for the strength of possible causality, an impulse response test is carried out on the VAR model. The impulse response test shows what happens to current values and future values of each dependent variable in the model to a one – unit increase in the current value of one of the VAR errors. The impulse response for the VAR model is depicted in Figure II. Figure II shows the effects of an unexpected 1 percentage point increase in AEX returns on the pessimism factor, and in the same way the effect of an unexpected 1 percentage point increase in pessimism on the daily AEX returns. Also shown in figure II are ± standard error bands, showing a confidence interval for each impulse response. The results show a response movement close to zero, which indicates negligible causality if causality existed. These findings show evidence that support the conclusion that both hypotheses should be rejected. Furthermore, impulse response results for the effect of AEX returns on the pessimism factor confirm the findings presented in Table III on Granger causality between AEX returns and Pessimism. A close look, however, shows that the effect of causality is only marginal around 0.1% at the 6 to 8 lag region. However, when looking at the standard deviation error band, the result is significant

(22)

Figure II

Impulse Responses in the AEX return – Pessimism factor VAR model

C. Forecast error decomposition statistics

The forecast error decomposition statistics show the percentage of the variance of the error made in forecasting a variable. The horizon is selected taking into account that the research focuses on short term results using daily data. The horizon selected for this test is set at two, matching the Granger causality test for easier comparison. Table IV and V show the forecast error decomposition of respectively the daily AEX index return and the Pessimism factor. The forecast error composition can be seen as a partial R2 for the forecast error, by forecast horizon.

Table V

Variance decomposition of the daily AEX index return; AEXd

Forecast Forecast Variance Decomposition ( percentage points)

horizon Standard Error AEXd Pfactor

2 0.0068 99.9 0.1

4 0.0068 99.5 0.5

6 0.0069 97.7 2.3

(23)

Table VI

Variance decomposition for the daily Pessimism factor; Pfactor

Variance Decomposition ( percentage points)

Forecast horizon Forecast Standard Error Pfactor AEXd

2 0.9820 99.7 0.3

4 0.9913 99.3 0.7

6 0.9948 98.7 1.3

8 1.0020 98.0 2.0

The results shown in table IV and V confirm earlier findings from impulse response tests suggesting very little interaction between the variables. In line with granger causality results the tables indicate a small increase in interaction when the number of lags in the model is increased, in the same way when the forecast horizon increases. This effect or interaction is negligible, thereby supporting the notion to reject the hypothesis tested for this research.

D. Robustness analysis

In the data section, I indicated that the constructed pessimism factor did not allow for distinguishing between different effects which the pessimism factors in the range 0 to 4 might induce on the investor population. In specific, it is not clear to what extend the relationship between the induced pessimism from the pessimism factors 0 trough 4 is linear or exponential or somewhere in between. To allow for this effect, and test the interaction of each pessimism factor independently with the AEX Index return, I construct dummy variables for each pessimism factor to test for interactions. As table VII indicates, there are not enough days for which the pessimism factor is 0, therefore I create four dummy variables, combining pessimism factor 0 and 1 into one dummy variable.

Table VII

Pessimism factor distribution over the entire data set

Number of negative classified articles on the front page

0 1 2 3 4

Number of Occurences 11 29 83 96 41

(24)

D.1 Granger-Causality statistics

This section has the same order as previous sections, and start with the results of the lag length criteria test detailed in Table XXIII in the appendix, to determine the appropriate number of lags. As with the main model, the lag length criteria test suggests using 6 lags as an appropriate number of lags. Following the test and method used for the main model in this research I opt for 6 lags in this model as well. Table VIII shows how different lag operators influence the causality between the AEX index return and each pessimism factor dummy. The results table VIII shows two interesting results. First, the causality results are comparable to that of the main VAR analysis, where the pessimism factor was not separated into dummy variables. This is shown by the significant causality found for the dummy variable pessimism factor 2 and 3 at the 6 and 8 lag operator number. More interesting is that the test also confirms the findings of the main analysis that show that the AEX index returns influence the level of pessimism. This analysis therefore confirms the direction of the interaction found between the AEX index returns and the pessimism factor. Another interesting result, specific to the dummy variable VAR model, which test for each variables interactions with the AEX index returns in specific, shows that only a significant effect is found for the pessimism factors 2 and 3. Economically, this result might indicate that the pessimism induced by news might be subject to decreasing returns showing a concave relationship between the individual pessimism variables. Next paragraph depicts the impulse response results, for further inferences.

D.2 Impulse response results

(25)

Table VIII

Granger causality between dummy pessimism variables and the daily AEX index returns

Granger – Causality statistics examine whether lagged values of one variable help to predict another variable. Below table summarizes the findings of multiple Granger – Causality tests using different lag operators. The table shows the p-value associated with their F-value to test whether coefficients on lags in the model are zero.

Direction of Causality

Number of

lags

F-Value p-value Decision

AEXd → P2 2 1.96 0.1423 Do not Reject

P2 → AEXd 2 1.13 0.3257 Do not Reject

P4→ AEXd 2 0.28 0.7530 Do not Reject

AEXd → P2 6 2.58 0.0195* Reject

AEXd → P3 6 2.27 0.0373* Reject

AEXd → P2 8 2.29 0.0224* Reject

AEXd → P3 8 1.74 0.0901** Reject

*

(26)

Figure III

Impulse Responses in the AEX return – Pessimism dummy variable VAR model

D.3 Variance decomposition results

To complete the results of the VAR model using dummy variables for each individual pessimism factor this paper reports the Variance decomposition results. In line with above findings, these results confirm that the negligible effects found in the impulse response test are indeed small. Table IX to XII shows the variance decomposition for each the pessimism dummy variable.

Table IX

Variance decomposition for the daily AEX index return; AEXd

AEXd Variance Decomposition ( percentage points)

Forecast

Horizon Forecast Standard Error AEXd P2 P3 P4

2 0.0068 99.5 0.1 0.0 0.4

4 0.0070 96.3 1.9 1.1 0.7

6 0.0070 94.2 2.4 1.5 1.9

8 0.0071 93.0 3.0 1.7 2.3

Table X

Variance decomposition for the daily Pessimism factor; P2

P2 Variance Decomposition ( percentage points)

Forecast

Horizon Forecast Standard Error P2 AEXd P3 P4

2 0.4689 46.0 0.2 44.8 9.0

4 0.4726 45.5 0.8 44.5 9.2

6 0.4778 44.8 1.3 44.4 9.5

(27)

Table XI

Forecast

Horizon Forecast Standard Error P3 _AEXd _P2 P4

2 0.4879 86.6 0.2 12.4 0.8

4 0.4931 85.9 0.5 12.1 1.5

6 0.4973 85.0 0.6 12.9 1.5

8 0.5004 84.0 1.5 12.7 1.8

Table XII

Forecast

Horizon Forecast Standard Error P4 AEXd P3 _P2

2 0.366360 0.6 0.6 0.0 98.8

4 0.374310 2.2 0.6 1.5 95.7

6 0.376958 2.2 1.0 2.3 94.5

8 0.381581 2.7 1.0 2.6 93.7

The tables show results in line with the granger causality results which indicate a small increase in interaction when the number of lags in the model is increased. In similar fashion, table XVIII shows this effect when the forecast horizon increases. Moreover, the interaction is negligible, thereby supporting the notion to reject the hypotheses tested for this research.

D.4 Results excluding control variables

As for the main VAR analysis whose results are presented at the beginning of this section, I also tested the VAR model with dummy variables, with and without exogenous control variables. To test whether the control variables changed the results significantly or had no effect on the results. The output can be found in the appendix where Figure VIII depicts the impulse response output without control variables and Table XXIV to XXVII the variance decomposition without control variables for the main VAR model. The results confirm those found from the main tests that the control variables decrease the already negligible effect even more, however not alter any of the outcomes significantly.

D.5 Lag length and pessimism factors

(28)

and tested it similarly to the main pessimism factor. The result of this VAR analysis strongly resembles the results using the main pessimism variable. The output is shown in the appendix where the lag length criteria test can be found in Table XVII, granger causality results in Table XVIII, impulse response tests depicted in Figure V followed by the variance decomposition results in Table XIX and Table XX. In short, the output is comparable and gives no reason to assume another pessimism variable next to the one specified in the main specification should be used.

As mentioned in previous paragraph, lag length criteria test also indicated the usage of 0 lags. Even though this research opted for the use of 6 lags, for robustness purposes an impulse response test was performed using zero lags. The impulse response test is graphically represented in Figure VII of the appendix. The impulse response test confirms findings from the granger causality results that there is no significant reportable result when 0 lags are used. Moreover, taking a look at the impulse response test shows that there is no direct causal effect.

Furthermore, I tested whether control variables significantly influence the results by testing the alternative VAR model with and without control variables. The results can be found in the appendix showing that the control variables do not significantly change the results, but that they do decrease the already negligible interaction noted in the results and in line with the results from previous section indicating similar findings under the main specification. The output from the alternative pessimism factor in the specification without control variables can be found in Figure VI representing the impulse response test and Table XXI and XXII showing the variance decomposition results. .

The last section of this paper will summarize the research into a conclusion and point to directions for future research.

V Conclusion

(29)

simultaneously test the two directional predictions formulated in the theoretical model. In addition, the VAR model allows for multiple tests to see whether the sentiment measure pessimism, as compiled from the content of front page news articles, shows significant interactions with the stock market returns.

The results of the various tests suggest the rejection of the formulated hypotheses, where the findings show little significant interaction between the pessimism factor constructed and the daily AEX index returns. The shown interaction is economically very small and therefore negligible. However, one interesting result that appeared is that at 6 days before market activity a significant causality is noted between the daily AEX return and the pessimism factor. The test shows that the direction of the causality supports the conclusion that a series of negative stock market returns induce negative news publishing captured by a higher pessimism factor. This notion can be explained by the fact that on the one hand, the newspaper in current technological environment is a slow medium to spread news, or that a series of negative results on the stock market gives rise to the ‘How’ questions, calling for news to explain for these phenomena.

In the context of the literature, this paper’s finding has to conclude that no strong evidence is found that supports the theory that sentiment explains for certain wild movements in the stock market. On the contrary, these findings support the neoclassical view that current stock prices always reflect fundamentals and that sentiment has no role in this.

Furthermore, I like to point out that due to the arbitrary process of creating a pessimism variable as a proxy for sentiment, this research is limited by the assumption that news is able to influence a populations mood by the tone of its news articles. Moreover, due to the time extensive process of content analysis and aggregating these results into a single daily pessimism factor, the research was limited to one year of data.

(30)

Bibliography

Antweiler, Werner, and Murray Z. Frank, 2004, Is all that talk just noise? The information content of Internet stock message boards, The journal of finance 59,1259-93

Barber, Brad M., and Terrance Odean, 2008, All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors, Review of Financial Studies 21 (No.2), 785-818.

DeBondt, Werner F.M., and Richard Thaler, 1985, Does the stock market overreact, Journal of Finance 40, 793-805.

DeBondt, Werner F.M., Shefrin, Hersh M., Muradoglu, Yaz Gulnur and Staikouras, Sotiris K., Behavioural Finance: Quo Vadis?, 2008, Journal of Applied Finance 19, 7-21. Cutler, David M., James M. Poterba, and Lawrence H. Summer,1989,What moves stock

prices?’, Journal of Portfolio Management 15,4-12.

DeLong, J. Bradford, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann, 1990a, Noise trader risk in financial markets, Journal of Political Economy 98, 703– 738.

Edmans, Alex, Diego Garcia, and Oyuind Norli, 2007, Sports sentiment and stock returns, Journal of Finance 58 (No.4), 1967-98.

Hirschleifer, David and Tyler Shummway, 2003, Good day sunshine: Stock returns and the weather,

Journal of Finance 58 (No.3), 1009-32.

Johnson, Eric J., and Amos Tversky, 1983, Affect, generalization, and the perception of risk. Journal of Personality and Social Psychology 45 (No.1), 20-31.

Kahneman, Daniel, and Amos Tversky, 1982, Judgment Under uncertainty: Heuristics and Biases, Cambridge University Press, Ney York.

Manning, Christopher D., Hinrich Schütze, 1999, Foundations of Statistical Natural Language

Processing, (MIT Press, Cambridge, MA).

Raghunathan, Rajagopal, and Michael T. Pham, 1999, All negative moods are not equal: Motivational influences of anxiety and sadness on decision making, Organizational Behavior and Human Decision Processes 79 (No.1), 56–77.

(31)

Simon, Herbert, 1957, A Behavioral Model of Rational Choice in Models of Man, Social and Rational: Mathematical Essays on Rational Human Behavior in a Social Setting, (Wiley, New York).

Slovic, Paul,1972, Psychological study of human judgment :Implications for investment decision making, Journal of Finance 27 (No.4), 779-799.

Stock, James H., and Mark W. Watson, 2001,Vector Autoregressions, Journal of Economic perspectives

Tetlock, Paul C., Maytal Saar-Tsechansky, and Sofus Macskassy, 2008, More than words: Quantifying language to measure firms’ fundamentals, Journal of Finance 63, 1437– 67.

(32)

Appendix

Table XIII

(33)

Table XIV

Lag length criteria test

This table shows the results of a lag length criteria test in the VAR model with two variables, Namely the AEX index returns variable and the pessimism variable.

Lag LogL LR FPE AIC SC HQ 0 -6.169.493 NA 0.490656* 4.963739* 5.048013* 4.997653* 1 -6.145.021 4.796776 0.496769 4.976112 5.116569 5.032636 2 -6.113.956 6.039846 0.500327 4.983232 5.179870 5.062364 3 -6.068.008 8.860059 0.497977 4.978492 5.231314 5.080234 4 -6.037.688 5.798187 0.501859 4.986206 5.295209 5.110556 5 -6.008.334 5.566717 0.506171 4.994689 5.359875 5.141649 6 -5.948.006 11.34458* 0.498088 4.978491 5.399860 5.148060 7 -5.927.908 3.747284 0.506116 4.994349 5.471901 5.186528 8 -5.904.330 4.358776 0.512866 5.007434 5.541168 5.222222 * indicates lag order selected by the criterion

LR: sequential modified LR test statistic (each test at 5% level)

FPE: Final prediction error AIC: Akaike information criterion

SC: Schwarz information criterion

HQ: Hannan-Quinn information criterion

Figure IV

(34)

Table XV

Variance decomposition of the daily index return without control variables; Pfactor

Variance Decomposition (percentage points)

Forecast Horizon Forecast Standard Error Pfactor AEXd

2 0.9797 100.0 0.0

4 0.9971 98.2 1.8

6 1.0092 95.9 4.1

8 1.0287 93.5 6.5

Table XVI

Variance decomposition of the daily pessimism factor without control variables; AEXd

Forecast Horizon Forecast Standard Error AEXd _Pfactor

2 0.0127 0.8 99.2

4 0.0128 2.1 97.9

6 0.0130 3.1 96.9

8 0.0131 3.6 96.4

Table XVII

This table shows the results of a lag length criteria test in the VAR model with two variables, namely the AEX index return variable and the alternative pessimism variable. The test shows similar results as for the main pessimism factor.

Lag LogL LR FPE AIC SC HQ

0 557.5805 NA 4.10e-05* -4.426936* -4.398845* -4.415632* 1 559.3824 3.560611 4.17e-05 -4.409.421 -4.325.147 -4.375.507 2 564.9616 10.93625 4.12e-05 -4.422.005 -4.281.549 -4.365.482 3 565.6639 1.365366 4.23e-05 -4.395.728 -4.199.089 -4.316.596 4 568.9496 6.335897 4.25e-05 -4.390.037 -4.137.216 -4.288.295 5 571.7753 5.403690 4.29e-05 -4.380.680 -4.071.676 -4.256.329 6 579.8329 15.28047* 4.16e-05 -4.413.011 -4.047.825 -4.266.051 7 580.6879 1.607820 4.26e-05 -4.387.951 -3.966.582 -4.218.382 8 581.9194 2.296287 4.36e-05 -4.365.892 -3.888.341 -4.173.713

* indicates lag order selected by the criterion

LR: sequential modified LR test statistic (each test at 5% level) FPE: Final prediction error

(35)

Table XVIII

Granger causality between the alternative pessimism factor indicated by Palt, and the AEX index returns.

Granger – Causality statistics examine whether lagged values of one variable help to predict another variable. Below table summarizes the findings of multiple Granger – Causality tests using different lag operators. The table shows the p-value associated with their F-value to test whether coefficients on lags in the model are zero.

Direction of Causality

Number of

lags F-Value p-value Decision

Palt →AEXd 2 2.12 0.2065 Do not Reject

AEXd→ Palt 2 1.16 0.0084 Reject

Palt →AEXd 4 1.56 0.1867 Do not Reject

AEXd →Palt 4 0.97 0.4273 Do not Reject

Palt → AEXd 6 1.42 0.2065 Do not Reject

AEXd → Palt 6 2.96 0.0084 Reject

Palt → AEXd 8 1.04 0.4089 Do not Reject

AEXd →Palt 8 2.37 0.0179 Reject

Figure V

(36)

Table XIX

Variance decomposition of the daily AEX index return; AEXd

Table (XX)

Variance decomposition of the alternative daily pessimism factor; Palt

Forecast Horizon Forecast Standard Error Pfactor(alt) AEXd

2 0.484692 0.1 99.9

4 0.487382 0.3 99.7

6 0.493173 0.9 99.1

8 0.496796 2.1 97.9

Figure (VI)

Impulse Responses in the AEX return – Alternative Pessimism factor VAR model without control variables

Forecast Horizon Forecast Standard Error AEXd Pfactor(alt)

2 0.006792 100.0 0.0

4 0.006838 98.7 1.3

6 0.006905 97.0 3.0

(37)

Table XXI

Variance decomposition of the daily AEX index return without control variables; AEXd

Forecast

Horizon Forecast Standard Error AEXd Pfactor(alt)

2 0.0127 99.6 0.4

4 0.0129 97.7 2.3

6 0.0130 97.3 2.7

8 0.0131 96.6 3.4

Table XXII

Variance decomposition of the alternative daily pessimism factor without control variables; Palt

Forecast

Horizon Forecast Standard Error Pfactor(alt) AEXd

2 0.4849 99.7 0.3

4 0.4899 98.7 1.3

6 0.4991 96.8 3.2

(38)

Table XXIII

This table shows the results of a lag length criteria test in the VAR model with dummy variables, namely the AEX index return variable and the dummy variables representing each individual pessimism variable.

Lag LogL LR FPE AIC SC

0 412.5114 NA 4.53e-07* -3.255071* -3.198888* 1 416.9512 8.702689 4.97e-07 -3.162.958 -2.882.045 2 426.2458 17.92272 5.24e-07 -3.109.528 -2.603.886 3 435.2050 16.99028 5.55e-07 -3.053.426 -2.323.053 4 443.0375 14.60413 5.92e-07 -2.988.347 -2.033.244 5 450.3597 13.41909 6.35e-07 -2.919.201 -1.739.368 6 466.7802 29.57008* 6.34e-07 -2.922.552 -1.517.989 7 476.1672 16.60488 6.69e-07 -2.869.858 -1.240.565 8 485.5845 16.35825 7.07e-07 -2.817.406 -0.963383 * indicates lag order selected by the criterion

LR: sequential modified LR test statistic (each test at 5% level) FPE: Final prediction error