scholarship project paper 2019 good news travels slowly and bad news has wings: can news-based...

www.set.or.th/setresearch

Disclaimer: The views expressed in this working paper are those of the author(s) and do not necessarily represent the Capital Market Research Institute or the Stock Exchange of Thailand. Capital Market Research Institute Scholarship Papers are research in progress by the author(s) and are published to elicit comments and stimulate discussion.

Scholarship Project Paper 2019

Good News Travels Slowly and Bad News Has Wings: Can News-Based Sentiment Predict Stock Market Movements?

Sapphasak Chatchawan Faculty of Commerce and Accountancy

Thammasat University

20 April 2020

Abstract The objective of the study is to investigate effects of news-based market sentiment on the predictive power and market volatility of SET, SET50, SET100 and MAI returns from November 2017- February 2019. The market sentiment is extracted from a corpus of headline news related to the Stock Exchange of Thailand in Thomson Reuters Eikon newswires. We find that the news-based sentiment index has the predictive power and contains significant information of SET, SET50, SET100 and MAI returns. In addition, the positive market sentiment significantly affects the volatility in SET50 and SET100 returns. As the positive sentiment increases, the volatility of SET50 and SET100 returns decreases. JEL Classification: G14 Keywords: News-based Market Sentiment, Computational Linguistics, Financial Market Volatility E-Mail Address: [email protected]

2019 Capital Market Research Institute, The Stock Exchange of Thailand

Content

Page

Chapter 1 Introduction 3

Chapter 2 Literature Review 5

Chapter 3 Research Methodology 6

Chapter 4 Empir ical Strategy 15

Chapter 5 Empir ical Results 19

Chapter 6 Robustness Checks 28

Chapter 7 Discussion 29

Chapter 8 Conclusion 32

References 33

Table

1. Dependent variables 7

2. Descript ive stat ist ics of dependent variables 7

3. Independent variables 9

4. Descript ive stat ist ics o f independent variables 9

5. Binary probit 19

6. Results of ARMA(2,2) 21

7. Results of ARMA (2,2) 22

8. Benchmark EGARCH(1,1) and volume-augmented

EGARCH(1,1) 24

9. Posit ive and Negative sentiment in EGARCH(1,1) 25

10. Posit ive and Negative sentiment in volume-augmented

EGARCH(1,1) 26

Figure

1. A sample of headlines from newswires 11


2. Word frequency in headlines 11

3. WordCloud before TF-IDF 12

4. WordCloud after TF-IDF 13

5. News-based sentiment index 14

6. Negative sentiment 17

7. Posit ive sentiment 18

Appendix

1. Unit root test 36

Acknowledgement 37


3

Chapter 1 Introduction

Financial firms enjoy benefits of real-time newswire services from international news agencies,

such as Thomson Reuters Eikon and Bloomberg. Newswires electronically distribute financial

related press releases, headlines and stories in qualitative textual information and help market

participants monitor real-time market conditions. Newswires spread up-to-the-minute news in

every second and naturally create a voluminous amount of textual data. This paper asks

following empirical questions: to what extent the sentiment of financial news from newswires

predicts the direction of the stock market? how does stock prices react to positive and negative

market sentiment? and, is news-based sentiment better at forecasting stock market volatility

than stock market movements?

A growing body of literature empirically explores causal relations between financial news and

stock market movements. The sentiment extracted from financial newspapers, such as The

Wall Street Journal, The New York Times and The Guardian, also has the predictive power over

stock market movements and corporate earnings (Tetlock, 2007; Tetlock, Saar-tsechansky and

Macskassy ,2008; Ferguson, Philip, Lam and Guo, 2014). The content of news influences

investors’ decision making. The effect of the unfavourable tone of financial news on stock

market prices is more pronounced during the economic recession and has more predictive

power than the favorable tone (Garcia, 2013). In addition to the study of printed newspapers,

there exists a number of papers explore impacts of newswires on financial markets (Boudoukh,

Feldman, Kogan and Richardson ,2012). The sentiment constructed from newswires has the

predictive over forecasts the stock market changes better than macroeconomic variables (Uhl

,2014). The daily sentiment predicts stock market returns for very short period. However, the

weekly is able to forecast returns for a longer period. Moreover, it seems that the sentiment


4

predicts the direction of volatility better than stock prices (Allen, McAleer and Singh,2013;

Heston and Sinha, 2016; Atkins, Niranjan and Gerding, 2018).In addition to the stock market,

news-based sentiment can also predict and explain movements in the term structure of interest

rates (Gotthelf and Uhl, 2018) and in foreign exchange market (Uhl, 2017).

The paper attempts to quantify the daily market sentiment from Reuters Eikon’s newswires by

employing methods in computational linguistics. The sentiment index is computationally

extracted from a collection of more than 1,000 news headlines. The whole corpus consists of

more than 20,000 words. This paper empirically analyzes the predictive power of news-derived

sentiment, which is characterized into both negative and positive sentiment, on the direction of

prices and volatility in the Stock Exchange of Thailand (Hereafter, SET) from 2017 to 2019.

Research Objective:

1. Does news-based market sentiment have predict ive power over

stock market returns?

2. How dose new-based market sentiment affect f inancial market

volat i l i ty in Thailand?

3. How do negative and posit ive sentiment variables affect market

volat i l i ty?

Expected outcomes

1. News-based market sentiment is l ikely to contain meaningful

information for predict ing direct ions of the Stock Exchange of Thailand.

2. Posit ive and negative market sentiment variables may affect stock

market volati l i ty in different ways.


5

Chapter 2 Literature Review

The paper contributes to the existing literature in three aspects. First, the study employs the

user-defined dictionary method as presented in Louhran and McDonald (2011) to characterize

the market sentiment into positive and negative sentiment. The technique is automated and

replicable. Unlike Heston and Sinha (2016), Uhl (2014), Uhl (2017) and Gotthelf and Uhl (2018)

that use Thomson Reuters sentiment index, the market sentiment for SET is constructed from a

collection of all news headlines relevant to SET from Thomson Reuters Eikon’s newswires. This

can be achieved by searching “BKstmad.BK”. Thus, the approach is more flexible than using

Reuters sentiment index, which was released until 2012. Second, the study investigates the

predictive power of news-based sentiment index on both stock market prices and volatility. This

is to confirm findings in Allen, McAleer and Singh (2013), Heston and Sinha (2016), Atkins,

Niranjan and Gerding (2018), which study the same aspect in the U.S. market. Third, to my

knowledge, the paper is the first attempt to explore the effects of news-based sentiment on

movements and volatility of prices and returns and the effects of sentiment on the trading

volume in SET.


6

Chapter 3 Research Methodology

Sources and Data description

All financial data are collected from Thomson Reuters Eikon. The sample begins from

9/11/2017 to 8/02/2019, which is the longest range of the sample period in which news

headlines were possibly scraped from Thomson Reuters Eikon’s newswires. All financial news

headlines were compiled by Routers Instrument Code (RIC): “BKstmad.BK”. Textual analysis

will be employed to extract market sentiment from those headlines, which are specific to the

Stock Exchange of Thailand. The major reason why the paper does not use Thomson Reuters

Eikon sentiment score is that the score had been published until 2012. It seems that the service

is no longer available1. There are 308 observations of both financial and news-based sentiment

variables. Data are daily time series and pass the unit root test at 1% level of significance. That

is, empirical results drawn from the data rule out the possibility of spurious relationship among

variables. The unit root tests are presented in table.A1 in Appendix 1.

Dependent variable

There are four independent variables, which are stock market returns, in the study. All

dependent variables are the percentage change of the Stock Exchange Thailand index (SET

index) and the Market for Alternative Investment (MAI index). Stock market returns can be

computed from equation 1.

𝑅𝑒𝑡𝑡 = ln (Xt

𝑋𝑡−1) × 100

1 The service is no longer available.

(1)


7

𝑋𝑡 represents SET, SET50, SET100 and MAI index. In table 1, SET_RET, SET50_RET,

SET100_RET and MAI_RET are stock market returns. Table 2 shows descriptive statistics for

dependent variables.

Table 1: shows dependent variables in the study.

Variable Description

SET_RET SET index return

SET50_RET SET50 index return

SET100_RET SET100 index return

MAI_RET MAI index return

Source: Author’s calculation

Table 2: Descriptive statistics of dependent variables

SET_RET SET100_RET SET50_RET MAI_RET

Mean -0.01 0.00 0.00 -0.13

Median 0.05 0.06 0.06 -0.01

Maximum 2.27 2.71 2.81 1.61

Minimum -2.42 -2.65 -2.60 -2.65

Std. Dev. 0.71 0.80 0.80 0.72


8

Skewness -0.32 -0.22 -0.17 -0.79

Kurtosis 4.12 4.26 4.31 4.13

Jarque-Bera 21.51 23.07 23.53 48.64

Probability 0.00 0.00 0.00 0.00

Sum -3.74 -1.21 1.14 -38.64

Sum Sq. Dev. 156.21 195.01 195.92 160.56

Observations 308 308 308 308


Independent variables

In the study, independent variables are classified into 2 groups, namely financial data and

textual data. Financial data are trading volume in SET, SET50 and SET100. Trading volume

data enter into EGARCH (1,1) in the mean equation in the form of percentage changes. The

textual data in the study are the news-based market sentiment, the positive sentiment and the

negative sentiment. All of which is extracted from the corpus of headline news scraped from

Thomson-Reuters newswire. Table 3 and table 4 depict the variable description and descriptive

statistics of independent variables.


9

Table 3: Independent variables

Variable Description

SET_VOLM Percentage change of trading volume in SET

SET50_VOLM Percentage change of trading volume in SET50

SET100_VOLM Percentage change of trading volume in SET100

MKT_SEN News-based market sentiment

POS Positive sentiment

NEG Negative sentiment


Table 4: Descriptive statistics of independent variables

SET_VOLM SET50_VOLM SET100_VOLM MKT_SEN POS NEG

Mean 0.14 0.17 -0.01 -0.02 0.03 0.05

Median -0.89 -1.10 -2.12 0.00 0.00 0.04

Maximum 70.22 123.78 85.29 0.36 0.36 0.29

Minimum -65.66 -89.31 -72.07 -0.29 0.00 0.00

Std. Dev. 20.35 32.13 27.77 0.08 0.05 0.06


10

Skewness 0.22 0.20 0.33 0.14 2.22 1.05

Kurtosis 3.83 3.41 3.10 4.16 11.03 3.66

Jarque-Bera 11.42 4.35 5.88 18.18 1080.08 62.06

Probability 0.00 0.11 0.05 0.00 0.00 0.00

Sum 43.71 52.98 -4.47 -5.79 9.84 15.63

Sum Sq. Dev. 127097.50 316928.30 236801.70 2.09 0.70 0.99

Observations 308 308 308 308 308 308


Textual analysis

Analytical pre-processing

Financial news headlines are unstructured data in nature. That is, numbers, punctuations,

alphanumeric and non-alphanumeric characters are mixed together in haphazard way. Text

mining is the process of transforming unstructured to structured and quantifiable data. That is,

we are going to quantify text. In the natural language processing, textual pre-processing is the

procedure of cleaning text and transformation of textual characters into the bag-of-words model.

In this paper, the corpus is a daily collection of financial headline news from Reuters Eikon

newswires. One day is equivalent to one document in the corpus. For example, in figure 1,

there are two documents. One is 8-Feb-19 and another is 7-Feb-19. Text data pertaining to

each day will be used in textual analysis.


11

Figure 1 Financial news headlines from Reuters Eikon ‘s newswire.


The ultimate goal of pre-processing is to eliminate the redundant dimensionality in the bag-of-words in order to filter noises out of the document. In this paper, the procedure of textual pre-processing includes word concatenation, removing white space, numbers, punctuation marks, non-alpha numeric (symbols), unicode characters and ‘English’ stopwords, stemming words with Porter stemming and tokenization. The financial corpus contains more than 20,000 terms before analytical pre processing.

Figure 2: shows the number of words in the headline news from Thomson Reuters Eikon newswire.


0

10

20

30

40

50

1

10

19

28

37

46

55

64

73

82

91

10

0

10

9

11

8

12

7

13

6

14

5

15

4

16

3

17

2

18

1

19

0

19

9

20

8

21

7

22

6

23

5

24

4

25

3

26

2

27

1

28

0

28

9

29

8

30

7

31

6

The Number of Words in each Financial Headline News

Words 5-day moving average


12

Some terms in financial headlines occurred at a high frequency rate but they did not contribute

meaningful information. According to Shannon entropy, terms with high frequency carry little

information with them. Therefore, I employ the term-frequency inverse document frequency (TF-

IDF) weighting scheme as the means to drop those terms out of the document. TF-IDF is the

weighting scheme, which is alternative to term-frequency (TF), based on how important words

explain the document. Words that rarely occur in the document will score the high weight

because they contain meaningful information. I will drop terms that their TF-IDF values are zero

out of the document and keep terms that their values are greater than zero.

Figure 3: Word Cloud of the corpus of headline news before TF-IDF scheme.



13

Figure 4: Word Cloud of the corpus of headline news after TF-IDF scheme


Dictionary method and market sentiment

Dictionary method ensures the replicability and the consistency of the result. This paper is not

the first that uses this method. The dictionary method is also used in Tetlock (2007), Tetlock et

al. (2008) and Loughran and McDonald (2011).

In order to construct the market variable, I employ a list of terms in finance used in Loughran

and McDonald (2011) as a user-defined dictionary. Then, simple counting method is used to

compute the market as presented in (2).

𝑀𝐾𝑇_𝑆𝐸𝑁𝑇𝑡 =𝑛𝑑,𝑡

𝑃𝑜𝑠𝑡 − 𝑛𝑑,𝑡𝑁𝑒𝑔

𝑛𝑑,𝑡 (2)


14

"𝑡" is a trading day. 𝑀𝐾𝑇_𝑆𝐸𝑁𝑇𝑡 is the news-based market sentiment extracted from

Reuters Eikon ‘s newswires. 𝑛𝑑,𝑡𝑃𝑜𝑠𝑡 is the number of positive-tone words in a document 𝑑 on

date "𝑡". 𝑛𝑑,𝑡𝑁𝑒𝑔 is the number of negative-tone words in a document 𝑑 on date "𝑡". 𝑛𝑑,𝑡

is the total number of terms in each document.

Figure 5: presents the News-Based Sentiment index from headline news.


-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

1

10

19

28

37

46

55

64

73

82

91

10

0

10

9

11

8

12

7

13

6

14

5

15

4

16

3

17

2

18

1

19

0

19

9

20

8

21

7

22

6

23

5

24

4

25

3

26

2

27

1

28

0

28

9

29

8

30

7

31

6

Text-based market sentiment index

Market sentiment 5-day moving average


15

Chapter 4 Empirical strategy

Auto regressive model

In order to explore whether the sentiment contains the meaningful information that can be used

to explain stock market movements, I retreat to a basic multiple regression to investigate the

relationship. The econometric specification is closely related to McLaren and Shanbhogue

(2011), which analyses the predictive power of Google search data in U.S. labour and housing

markets. The baseline model is just a standard ARMA(p,q) model.

𝑋𝑡 = 𝛽0 + 𝛽1𝑋𝑡−1 + 𝛽2𝑋𝑡−2+. . . +𝛽𝑝−1𝑋𝑡−𝑝+1 + 𝛽𝑝𝑋𝑡−𝑝 +

𝜀𝑡 + 𝛾1𝜀𝑡−1 + 𝛾2𝜀𝑡−2+. . . + + 𝛾𝑞𝜀𝑡−𝑞

From equation 3, 𝑋𝑡 is a dependent variable. Nine benchmark equations, which are SETI,

SETI50, SETI100, SET_RET, SET_RET50, SET_RET100, SET_VOL, SET50_VOL and

SET100_VOL, are estimated. "𝑝" is the order of lagged variable and is optimally chosen by

the Akaike Information Criteria (Hereafter, AIC). 𝑋𝑡−1, 𝑋𝑡−2, … , 𝑋𝑡−𝑝 are autoregressive

variables. 𝜀𝑡 is an independent, identical and normal disturbance term .

𝑋𝑡 = 𝛽0 + 𝛽1𝑋𝑡−1 + 𝛽2𝑋𝑡−2+. . . +𝛽𝑝−1𝑋𝑡−𝑝+1 + 𝛽𝑝𝑋𝑡−𝑝 +

𝜀𝑡 + 𝛾1𝜀𝑡−1 + 𝛾2𝜀𝑡−2+. . . + + 𝛾𝑞𝜀𝑡−𝑞 + 𝛼𝑀𝐾𝑇_𝑆𝐸𝑁𝑇𝑡

In order to examine the predictive power of market sentiment, I firstly fit equation 3 and

compare AIC with AIC from benchmark models. If the news-based sentiment contains

meaningful information to stock market variables, the estimated coefficient of 𝛼 will be

statistically significant with a positive sign and AIC in equation 4 will be lower than AIC in

equation 3.

(3)

(4)


16

Exponential General Autoregressive Conditional Heteroscedastic (EGARCH)

Volatility cluster is ubiquitous in high-frequency financial data. In order to explore the effects

of the news-based sentiment index, we employs EGARCH. There are three major reasons why

the paper will use EGARCH. First, GARCH assumes that only the magnitude of unanticipated

excess returns, not the positivity or negativity excess returns, influences the conditional

variance. This assumption rules out the evidence that the volatility rises in response to bad

news and falls in response to good news. Second, the non-negativity constraints imposed on

GARCH parameters to ensure that the conditional variance is positive. Third, it is difficult to

interpret the persistence in the conditional variance. Thus, Nelson (1991) proposes Exponential

GARCH model (EGARCH) as an alternative to GARCH.

EGARCH model has an advantage over GARCH because it ensures that the conditional

variance is positive and allows for the asymmetric response of the volatility to good and bad

news. EGARCH also incorporate the asymmetric leverage. This paper therefore employs

Nelson‘s EGARCH as a work horse to analyze effects of the semantic similarity of forward

guidance on the volatility in financial markets data

EGARCH (1,1) specification

𝑦𝑡 = 𝛽0 + 𝛽1𝑉𝑂𝐿𝑡 + 𝜔𝑡

𝜔𝑡~(0, ℎ𝑡)

log(ℎ𝑡) = 𝛾0 + 𝛾1 (𝜔𝑡−1

√ℎ𝑡−1

) + 𝛾2 (|𝜔𝑡−1

√ℎ𝑡−1

| − √2

𝜋) + 𝛾3 log(ℎ𝑡−1)

+ 𝛼𝑆𝑡

I will employ a parsimonious EGARCH(1,1) because the model is sufficient to capture the non-

normality of the data in this paper. There will be 9 EGARCH equations as they are consistent

with the study in section 3.23. 𝑦𝑡 is the independent variable. 𝜔𝑡 is the disturbance term. ℎ𝑡

is the variance of 𝜔𝑡 . Equation 5 and 6 are the mean and the conditional variance equations. I

control the effects of the trading volume in each market in the mean equation. In this section,

(5)

(6)


17

the market sentiment is classified into positive and negative sentiment. Equation 7 and 8 show

the derivation.

𝑃𝑂𝑆 =𝑛𝑑,𝑡

𝑃𝑜𝑠

𝑛𝑑,𝑡

𝑁𝐸𝐺 =𝑛𝑑,𝑡

𝑁𝑒𝑔

𝑛𝑑,𝑡

Equation 7 depicts the fraction of the positive term in relation to the total number of terms in the

document as well as equation 8, which shows the fraction of the negative term. Maximum

likelihood2 is the estimation method for EGARCH. I fit the model with different assumptions of

the distribution of residuals.



2 See Tsay (2013)

0

0.05

0.1

0.15

0.2

0.25

0.3

1

10

19

28

37

46

55

64

73

82

91

10

0

10

9

11

8

12

7

13

6

14

5

15

4

16

3

17

2

18

1

19

0

19

9

20

8

21

7

22

6

23

5

24

4

25

3

26

2

27

1

28

0

28

9

29

8

30

7

31

6

Negative sentiment index

Negative sentiment 5-day moving average

(8)

(7)


18



In order to analyze the effect of market sentiment on the volatility SET, in the variance

equation, 𝑆𝑡 ∈ {𝑃𝑂𝑆, 𝑁𝐸𝐺}. Most importantly, the expected signs of parameters are

𝛼𝑃𝑂𝑆 < 0 and/or 𝛼𝑁𝐸𝐺 > 0 . This can be interpreted that the relatively negative

market sentiment will lead to the high volatility. On the contrary, more positive market sentiment

will lower the market volatility.

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

1

10

19

28

37

46

55

64

73

82

91

10

0

10

9

11

8

12

7

13

6

14

5

15

4

16

3

17

2

18

1

19

0

19

9

20

8

21

7

22

6

23

5

24

4

25

3

26

2

27

1

28

0

28

9

29

8

30

7

31

6

Positive market sentiment

Positive sentiment 5-day Moving Average 10-day moving average


19

Chapter 5 Empirical Results

Simple Probit model

Table 5: shows the results of Binary Probit model

𝑀𝐾𝑇_𝑆𝐸𝑁 𝑀𝐾𝑇_𝑆𝐸𝑁𝑡−1 𝑀𝐾𝑇_𝑆𝐸𝑁𝑡−2 𝑀𝐾𝑇_𝑆𝐸𝑁𝑡−3

𝑆𝐸𝑇_𝑅𝐸𝑇 2.06**

(0.89)

1.24

(0.88)

1.33

(0.85)

-0.1

(0.87)

𝑆𝐸𝑇50_𝑅𝐸𝑇 1.95**

(0.89)

1.60*

(0.89)

1.45*

(0.87)

0.41

(0.87)

𝑆𝐸𝑇100_𝑅𝐸𝑇 1.92**

(0.89)

1.06

(0.88)

1.53*

(0.87)

0.30

(0.87)

𝑀𝐴𝐼_𝑅𝐸𝑇 1.81**

(0.88)

-0.08

(0.87)

-0.20

(0.87)

0.94

(0.87)


Results from the simple binary probit presented in table 5 corroborate the fact that the news-

based market sentiment has the predictive power over market returns to some degree. In table

5, lagged variables of the market sentiment significantly determine and predict the probability

of ups and downs of stock market returns.


20

ARMA

In order to explore the predictive power of the news-based market sentiment, the positive and

negative sentiment on stock market returns, the study employs ARMA(p,q) process as a

workhorse used to compare the predictive power among market sentiment variables. From the

benchmark model, ARMA (2,2) is fitted to capture the process of market returns in SET, SET50

and SET100. As for MAI, it turns out that ARMA (1,1) sufficiently captures the data generating

process of the return of MAI index.


21

𝑆𝐸𝑇_𝑅𝐸𝑇𝑡 Benchmark

(1)

𝑆𝐸𝑇_𝑅𝐸𝑇𝑡 Market Sentiment

(2)

𝑆𝐸𝑇_𝑅𝐸𝑇𝑡 Positive Sentiment

(3)

𝑆𝐸𝑇_𝑅𝐸𝑇𝑡 Negative Sentiment

(4)

𝑆𝐸𝑇50_𝑅𝐸𝑇𝑡 Baseline

(5)

𝑆𝐸𝑇50_𝑅𝐸𝑇𝑡 Market Sentiment

(6)

𝑆𝐸𝑇50_𝑅𝐸𝑇𝑡 Positive

Sentiment (7)

𝑆𝐸𝑇50_𝑅𝐸𝑇𝑡 Negative Sentiment

(8) 𝑀𝐾𝑇_𝑆𝐸𝑁𝑡 - 0.97*

(0.50) - - - 1.06*

(0.56) - -

𝑃𝑂𝑆𝑡 -

- 2.26** (0.87)

- - - 2.22** (0.97)

-

𝑁𝐸𝐺𝑡 -

- - -0.65 (0.72)

- - - -0.70 (0.81)

𝐴𝑅(1) -0.44** (0.17)

-0.45** (0.18)

-0.92*** (0.12)

-0.44** (0.17)

-0.48** (0.17)

-0.49** (0.18)

-0.54*** (0.15)

-0.47** (0.18)

𝐴𝑅(2) -0.80*** (0.19)

-0.80*** (0.20)

-0.60*** (0.13)

-0.80*** (0.19)

-0.81*** (0.21)

-0.80*** (0.23)

-0.84*** (0.19)

-0.80*** (0.22)

𝑀𝐴(1) 0.47*** (0.16)

0.48** (0.17)

0.97*** (0.11)

0.47** (0.17)

0.51*** (0.17)

0.52*** (0.18)

0.56*** (0.14)

0.50** (0.18)

𝑀𝐴(2) 0.82*** (0.18)

0.82*** (0.20)

0.69*** (0.11)

0.82*** (0.19)

0.82*** (0.21)

0.81*** (0.23)

0.86*** (0.19)

0.81*** (0.22)

Constant -0.01 (0.04)

0.01 (0.04)

-0.08 (0.05)

0.02 (0.06)

0.01 (0.05)

0.03 (0.58)

-0.07 (0.06)

0.04 (0.06)

Durbin Watson 1.98 1.99 2.02 1.99 2.03 2.04 2.03 2.04

AIC 2.1747 2.1685 2.1631 2.1785 2.4042 2.3987 2.3935 2.4082

Table 6: shows the results of ARMA (2,2) . */**/*** denotes the level of significance at 10%,5% and 1%


22

Table 7: shows the results of ARMA(2,2). */**/*** denotes the level of significance at 10%,5% and 1%


23

In table 6, column 1 and column 5 present benchmark equations for the returns of SET and

SET50. AR and MA terms are statistically significant and AIC values equal 2.17 and 2.40. In

addition, Durbin-Watson statistics indicate that there is no autocorrelation left. When

incorporating the news-based market sentiment into benchmark equations in column 2 and

column 6, coefficients of the news-based sentiment are statistically significant at 10% level

and adding such variables also decreases AIC values. Thus, the news-based market sentiment

improves the predictive power of ARMA (2,2) of SET and SET50 returns.

With regard to the effects of positive market sentiment on SET and SET50 returns in column 3

and 7, the coefficients of the positive market sentiment are significant at 5%. Their AIC values

are lower than those in the benchmark. The positive market sentiment therefore has the

predictive power over SET and SET50 returns. On the contrary, in column 4 and 8, coefficients

of the negative sentiment are not significant. AIC values in column 4 and 8 are greater than

AIC values of the benchmark models. Adding the negative market sentiment does not

significantly contribute to the predictive power of the model. In conclusion, the news-based

market sentiment and the positive market sentiment may contain meaningful information for

returns of SET and SET50.

For SET100 and MAI in table 7, coefficients of the news-based market sentiment in column 2

and 6 are positive and statistically significant. Furthermore, coefficients of the positive sentiment

are statistically significant at 5% level and these coefficients exhibit positive signs. However,

coefficients of the negative sentiment are insignificant in both SET100 and MAI. This may

conclude that the news-based sentiment and the positive sentiment carry some meaningful

information that helps predict stock market returns.

When we incorporate both the news-based market sentiment and the positive sentiment into

the model, values of AIC in ARMA(2,2) and ARMA(1,1) also decrease and these signify that

regression functions may take into account some important information already.


24


25


26


27

EGARCH (1,1)

Table 8 shows results of benchmark EGARCH (1,1) and volume-augmented EGARCH (1,1)

models. Dependent variables in EGARCH (1,1) are SET, SET50, SET100 returns, but, for

volume-augmented EGARCH (1,1), dependent variables are SET, SET50 and SET100 returns.

In EGARCH (1,1), the percentage change of trading volume enters into the model in the mean

equation ;whereas, in volume-augmented EGARCH (1,1) models, the level of trading volume

enters into the models, which share the same spirit of Lamoureux and Lastrapes (1990), in the

variance equation.

In table 9, which depicts the results of EGARCH (1,1) for negative and positive sentiment

variables, on the left panel, the coefficients of the negative sentiment are insignificant across

SET, SET50, SET100 and MAI returns. However, except for MAI, the log likelihood values of

EGARCH (1,1) augmented with the negative sentiment are higher than those in the benchmark

models. The coefficients of positive sentiment in EGARCH (1,1) are statistically significant and

show negative signs in SET, SET40 and SET100 returns. As the good news is prevalent in the

market, the overall volatility of market returns decreases.


28

Chapter 6 Robustness Checks

Volume-augmented EGARCH (1,1) is employed for the robustness check in table 10. Under the

different set up, some results exist and some are not. On the left panel in table 10, only the

coefficient of the negative sentiment in SET100 returns is statistically significant and shows a

positive sign. This can give rise to an explanation that bad news generates higher volatility in

SET100 returns. However, the result is not that robust after comparing with the estimated

results from EGARCH (1,1).

On the right panel in table 10, the coefficients of the positive sentiment in volume-augmented

EGARCH (1,1) are statistically significant and display negative signs for SET and SET50

returns. The finding is SET100 returns is not that robust.


29

Chapter 7 Discussion

The study constructs the news-based sentiment index from the corpus of financial news specific

to the Stock Exchange of Thailand from Thomson-Reuters Eikon’s media newswire from 2017-

2019. Dictionary technique is a tool that is used to extract the sentiment from the corpus. Not

only the news-based sentiment, the study constructs the negative and the positive sentiment as

well. The news-based sentiment consists of the positive sentiment and the negative sentiment.

Thus, this allows the researcher to examine the effects of the positive and negative sentiment

on market returns in Thailand during 2017-2019.

Does the news-based sentiment from the media newswire have the predictive power over

market returns?

From the Probit model, the news-based sentiment has some predictive power. The index itself

significantly explains the probability of the directions of stock market returns in SET, SET50,

SET100 and even MAI. The 1-day and 2-day lagged sentiment variables have some predictive

power over SET50 returns. In SET100, only 2-day lagged sentiment is statistically significant.

Thus, according to the simple Probit model, the news-based sentiment carries some meaningful

information of the probability of ups and downs of market returns and its lagged values, 1-day

and 2-day, may contain useful information for the probability of the directions of the stock

market returns in SET50 and SET100.

To warrant more empirical evidence of the predictive power inherited in the news-based

sentiment index, the time series models, ARMA (2,2) for SET, SET50 and SET100 and ARMA


30

(1,1) for MAI, are employed as a standard workhorse to examine the predictive power of the

newly introduced index.

News-based market sentiment and its predictive power

Comparing with the benchmark models, adding the news-based sentiment index into ARMA

(2,2) for SET, SET50 and SET100 returns and ARMA (1,1) for MAI returns reduces AIC values

in all equations. Thus, incorporating the news-based sentiment variable improves the model in

terms of the information criterion.

In addition, coefficients of the news-based sentiment are statistically significant. The news-

based sentiment contains significant information for market returns. As for the relationship

between the sentiment index and returns, each market-sentiment coefficient has a positive sign.

The more positive the market sentiment is, the higher the stock market returns across SET,

SET50, SET100 and MAI.

When we separate the news-based sentiment into the positive and the negative sentiment, this

gives more insightful results. Only coefficients of the positive sentiment index are statistically

significant; whereas, coefficients of the negative sentiment index are insignificant. With regard

to the positive sentiment, the coefficients show positive signs. That is consistent with the overall

market sentiment index. Unlike the positive sentiment, signs of negative-sentiment coefficients

are minus. However, they are insignificant.

Above all, the news-based sentiment index and the positive sentiment index contain the

significant information for market prediction.

Effects of the positive and negative sentiment on market volatility.

The positive and the negative sentiment indices are what we obtain from the dictionary method

when calculating the news-based sentiment index. In other words, in calculating the news-


31

based sentiment index, we also obtain the positive and the negative sentiment variables. In this

study, the effects of positive and negative sentiment index on volatility of stock market returns

in SET, SET50 and SET100 are investigated. The parsimonious EGARCH (1,1) models with

the trading volume are estimated and employed to explore the effect of positive sentiment

(good news) and negative sentiment (bad news) on the volatility.

Results show that the coefficients of the positive sentiment in the variance equations are

statistically significant in EGARCH (1,1) for SET, SET50 and SET100 returns. The positive

sentiment tends to have larger effects on the volatility of SET50 returns than SET’s and

SET100’s. Each coefficient exhibits the negative sign, which is in contrast to those in ARMA

model. However, this gives rise to some insightful interpretation. That is, when the positive

sentiment increases, it lowers the volatility in stock markets. The findings seem to be contrast

to the case of the negative sentiment in EGARCH (1,1). Coefficients of the negative sentiment

in the variance equation are insignificant.

In order to provide further evidence of the findings in EGARCH (1,1), the volume augmented

EGARCH (1,1) is employed as robustness checks. The robustness test confirms the effect of

the positive sentiment on the volatility of SET50 and SET100 returns. The volatility of SET50

returns tends to be the most vulnerable to the positive sentiment from the financial newswire.

However, with regard to the negative sentiments, even though signs of coefficients are positive,

they are insignificant in the variance equation.


32

Chapter 8 Conclusion

The market sentiment is extracted from headline news about the Stock Exchange of Thailand in

Thomson Reuters newswire. The textual analysis, which is the dictionary method, quantifies the

market sentiment from the corpus of headline news in the study. The news-based sentiment

covers the period from 2017-2019 and the positive and the negative market sentiment variables

are also calculated. The news-based sentiment index, together with the positive and the

negative sentiment, contains the critical information in predicting the market returns. In addition,

the positive sentiment index has significant impact on the volatility of market returns. The more

positive the sentiment, the lower the volatility in stock markets.


33

Reference

Allan, D. E., McAleer, M. J. and Singh, A. K. (2015) Machine News and Volatility: The Dow

Jones Industrial Average and the TRNA Real-Time High-Frequency Sentiment Series. In

Gregoriou, G.N. (Ed.), The handbook of high frequency trading (pp.327-344) Amsterdam:

Elsevier.

Atkins A., Niranjan M. and Gerding E. (2018). Financial News Predicts Stock Market Volatility

Better than Close Price. The Journal of Finance and Data Science, 4,120-137.

Boudoukh, F., Feldman, R. Kogan, S. and Richardson, M. (2012). Which News Moves Stock

Prices? a Textual Analysis, NBER working paper series, 18725.

Ferguson, N.J., Philip, D., Lam H.Y.T. and Guo, M. (2015). Media Content and Stock Returns:

the Predictive Power of Press, Multinational Finance Journal, 19(1), 1-31.

Garcia, D. (2013). Sentiment during Recessions. The Journal of Finance. 68(3), 1267-1300


34

Gotthelf, N. and Uhl, M.W., News Sentiment - A New Yield Curve Factor (March 14, 2018).

Journal of Behavioral Finance, Vol. 19 (3), 2018.

Heston, Steven L. and Sinha, Nitish Ranjan, News Versus Sentiment: Predicting Stock Returns

from News Stories (2016-06). FEDS Working Paper. No. 2016-048.

Lamoureux, C., & Lastrapes, W. (1990). Heteroskedasticity in Stock Return Data: Volume

versus GARCH Effects. The Journal of Finance, 45(1), 221-229. doi:10.2307/2328817

Loughran, T., and McDonald B. (2011). When is a Liability not a Liability? Textual Analysis,

Dictionaries, and 10-Ks. The Journal of Finance, 66, 35–65.

McLaren, Nick and Shanbhogue, Rachana, Using Internet Search Data as Economic Indicators

(June 13, 2011). Bank of England Quarterly Bulletin. No. 2011 Q2.

Nelson, D. B. ( 1991) . Conditional Heteroskedasticity in Asset Returns: a New Approach.

Econometrica: Journal of the Econometric Society, 347-370.

Tetlock, P. ( 2007) . Giving Content to Investor Sentiment: The Role of Media in the Stock

Market. The Journal of Finance, 62, 1139–1168.


35

Tetlock, P. , Saar-Tsechansky M. , and Macskassy S. (2008) . More Than Words: Quantifying

Language to Measure Firms’ Fundamentals. The Journal of Finance, 63, 1437–1467.

Tsay, R. (2013). An Introduction to Analysis of Financial Data with R. John Wiley & Sons, Inc.

Uhl, M.W. (2014). Reuters Sentiment and Stock Returns, Journal of Bahavioral Finance,15:4,

287-298.

Uhl, M.W. (2017). Emotions Matter: Sentiment and Momentum in Foreign Exchange., Journal

of Behavioral Finance,18:3,249-257.


36

Appendix I

Variable t-statistic P-value

MKT_SEN -14.27 0.00

POS -14.54 0.00

NEG -14.53 0.00

SET_RET -16.90 0.00

SET50_RET -17.32 0.00

SET100_RET -17.06 0.00

MAI_RET -14.53 0.00

SET_VOLM -15.53 0.00

SET50_VOLM -13.54 0.00

SET100_VOLM -14.43 0.00

SET_VOLM_level -10.47 0.00

SET50_VOLM_level -10.17 0.00

SET100_VOLM_level -8.18 0.00


37

Acknowledgement

I would like to thank the Stock Exchange of Thailand’s Capital Market Research Institute

(CMRI) and I have benefited greatly from the discussion with CMRI’s researchers. The study is

funded by CM research innovation contest.

scholarship project paper 2019 good news travels slowly and bad news has wings: can news-based...

Documents