1 The Effect of News on Aggregate Investor Sentiment in Turkey Kristina Vasileva, Senior Lecturer, University of Westminster, London, UK Tel: +44 20 7911 5000 Ext: 66771 E-mail: [email protected]Belma Ozturkkal, Assistant Professor of Finance, Department of International Trade and Finance Kadir Has University, Cibali, Istanbul Turkey Tel: +90-212-5335765 Ext: 1603 E-mail: [email protected]
21
Embed
The Effect of News on Aggregate Investor Sentiment in Turkey ANNUAL...1 The Effect of News on Aggregate Investor Sentiment in Turkey Kristina Vasileva, Senior Lecturer, University
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
The Effect of News on Aggregate Investor Sentiment in Turkey
This study investigates the role of news, both positive and negative on the returns of a big
emerging market such as Turkey. Stock price movements have long been studied and with the
establishment of behavioral finance in the early 1980s, alternative explanations have been put
forward to explain excess price volatility that cannot be simply explained by dividend return or
other financial and economic factors. Investors often overreact (Hirshleifer, 2001) to external
influences that often have nothing to do with company fundamentals.
Investor sentiment has been a topic in behavioral finance since the 1990s (Morck,
Shleifer, Vishny, Shapiro and Poterba, 1990) with an increasing importance in the past two dec-
ades. Its origins are in noise trading where investors react to factors not related to stock funda-
mentals and this drives the prices in a certain direction (Black, 1986). It is usually broadly de-
fined as the general feeling or mood in the market about future cash flows and risks without nec-
essary support for this from stock fundamentals (Baker and Wurgler, 2007). These influences
have the power to distort market prices as a consequence.
Sentiment in the market can be viewed and analyzed from different perspectives. Initially
it was studied in the context of asset prices, demonstrating that sentiment affects the Efficiency
Market Hypothesis (EMH) and can explain at least in part, why we do not observe efficient mar-
ket pricing. In this vain, early studies by Black (1986) and DeLong et al. (1990) introduced the
notion of noise trading.
Later studies looked at the role of market reaction, both over and under-reaction.
Barberis, Shleifer and Vishny (1997) designed a model of investor sentiment where they exam-
ined the underreaction and overreaction to market news and the implications for market liquidity.
Baker and Wurgler (2006) look at a cross section of market returns and find that sentiment has a
greater effect on stocks which are more difficult to evaluate. Baker and Wurgler (2007) focus on
investor sentiment affecting the stock prices and find that low capitalization and high volatility
stocks are more sensitive to investor sentiment changes. We hypothesize that there is a difference
between developed and developing markets in terms of reaction to sentiment and our results will
show the nature of this difference in behavior. Kearney and Liu (2014) summarize different
methods used in sentiment analysis and conclude that there is still some way to go to improve on
the methodological approach to sentiment and its effect on returns. Financial statements provide
4
investment information that shows if aggregate corporate investment is higher or lower during
peaks and troughs of market activity therefore demonstrating that managers either ‘cater’ to the
market exuberance or are themselves caught up in it. (Arif and Lee, 2014). Cross sectional stock
returns are analysed in Stambauch et al. (2012) to find the effect of investor sentiment. They find
that due to restrictions in short-selling, a lot of prices reflect investors who are too optimistic.
This overpricing is more likely during periods of high investor sentiment.
The role of sentiment isn’t always clear-cut. Brown and Cliff (2004) investigate the role
of investor sentiment in the near-term returns and counter to many literature findings, show that
even though sentiment levels and changes are strongly correlated with contemporaneous market
returns, they have little predictive power for near-term future returns. The reason for this could
be the measure for investor sentiment or the effect of investor sentiment itself. While most com-
monly used investor measures are survey based, direct measures, they also test some indirect
measures such as: recent market performance ratio; certain types of variables relating to the type
of trading; derivatives trading activity and similar proxies. None of these include the role of news
a valid case can be made in defining investor sentiment. Survey based sentiment measures are
another good way to measure sentiment. There are some commonly known surveys such as: the
American Association of Individual Investors (AAII) which collects polling data on samples
from its members; Investors Intelligence (II) survey, which collects and analyses market views
from independent investor newsletters; UBS/Gallup survey of investor sentiment asking individ-
ual investors to rate the stock market performance over the next year. These and other surveys
have been used in a number of studies on sentiment such as Qiu and Welch (2004); Graham,
Harvey and Huang (2009); Fisher and Statman (2003); Schmeling (2009).
Another group of sentiment studies uses news and media sorties as means to quantify the
investor sentiment and analyze its effect on stock markets. We approach sentiment from this per-
spective. These studies tend to evolve along with technology changes. Tetlock (2007) measures
the interaction between media and the stock market using data from the Wall Street Journal and
finds that pessimistic media coverage predicts low market trading volume. This is a finding
which is consistent with the experience in the subsequent market crisis of 2008. News sentiment
effect on market return is further examined in Kräussl and Mirgorodskaya (2017). They investi-
gate the impact of news sentiment on return and volatility in the long run. They calculate a
monthly indicator of positive and negative news alerts and find negative news increase volatility
5
and decrease return. They confirm bubbles and crisis in the last 20 years. Das and Chen (2007)
develop methodology for extracting small investor sentiment from stock message boards and
find that their aggregated sentiment broadly tracks the index returns. Similarly, Renault (2017)
constructs a sentiment measure from a lexicon of words found on a microblogging platform
where investors share opinions and ideas. This study shows that this aggregated online sentiment
helps forecast intraday stock index returns. Chan (2003) uses company level data to analyse the
effect of news on monthly returns by comparing companies of similar returns that have and do
not have news published about them. He establishes a link between stocks with news and mo-
mentum and a strong drift after bad news as well as a link between news and price reversal after
extreme price movements when they are not accompanied by public news.
Heston and Sinha (2017) investigate the role of news in predicting the stock returns. They
find that firms with neutral news outperform firms without any news and that after controlling
for this effect, positive news affects the stock price for one week. To the contrary, negative news
affects the price for a whole quarter. They interpret this as consistent with short-sale restrictions
which that delay the incorporation of bad news but we would argue it is also consistent with the
disposition effect – the tendency for investors to sell winners quickly but hold on to loser-stocks
for longer (Shefrin and Statman, 1985). Heston and Sinha (2017) confirm general literature con-
sensus that daily news predicts stock returns in the short term (few days) while weekly news has
a longer, quarterly effect. Our data is based on a collection of daily news which is aggregated in
a monthly period and we aim to look at the relationship that aggregate news have on the general
monthly investor sentiment in the Turkish stock market rather than on a company level.
New media types – the internet and social media have been considered in studies for their
influence on the stock market sentiment. Older studies in terms of linking the use of the internet
and stock markets such as Antweiler and Frank (2004) show a link between internet stock mes-
sage boards and the stock market. While the magnitude of the effect was debatable, they clearly
show that it impacts market trading in terms of volume of trading (adding to volatility) and next
day returns. Kim and Kim (2014) also look at investor sentiment from internet message postings.
Unlike Antweiler and Frank (2004) however, they do not find any predictive power of these
message boards over volatility and trading volume. In the social media front, studies have gener-
ally focused on Twitter as a news venue which can influence stock trading. Da, Engelberg and
Gao (2015) investigate the role of household internet search volume as a proxy for market level
6
sentiment and construct an index that predicts short term return reversal and volatility increases.
Even the effects of some sports news such as league football (soccer) are also known to affect
investor sentiment (Dimic, Neudl, Orlov and Aijo, 2017). Housing media sentiment has also
been found to impact the real-estate markets in local areas across the U.S. (Soo, 2015).
Financial news are found to impact stock price returns (Li et al., 2014). They analyze the
effect of news in a generic stock price prediction framework using dictionary generated senti-
ment dimensions and find that sentiment analysis helps prediction accuracy however they do not
find it helpful to predict using sentiment polarity (good/bad news). Public information which is
relevant to the financial markets such as monetary policy announcements, International Mone-
tary Fund related news or other public and political news are known to impact sentiment with
special attention to the emerging markets (Brzeszczyński, Gajdka and Kutan, 2015).
There are several studies which investigate investor sentiment in Turkey. Sayim and
Rahman (2015) look at the stock returns and volatility on the Istanbul Stock Exchange (ISE) and
use a monthly Turkish consumer confidence index for the 2004-2010 period. Uygur and Tas
(2014) use an EGARCH model to link investor sentiment with conditional volatility and returns
on the ISE. Canbas and Kandir (2009) employ VAR analysis and Granger causality to investi-
gate the relation between investor sentiment and stock returns on the ISE. All these studies indi-
cate there is a link between sentiment and the Turkish Stock Market however none of them has
looked at sentiment from a news analytics approach.
Most of the literature on sentiment involves the US market as data is more widely availa-
ble for both market performance forecast and investor sentiment. Some international evidence
for sentiment can be found in Schmeling (2009) who investigates expected stock returns in 18
industrialized countries based on measuring investor sentiment through consumer confidence as
a proxy. Their findings show that investor sentiment plays a significant role in expected returns
on average across the 18 countries and it is strongest on short and medium-term horizon which,
intuitively, is to be expected. Chen, Chen and Lee (2013) study sentiment in 11 Asian countries
and several industries and find that optimism leads to industry returns to be overvalued and vice
versa for pessimism. Corredor, Ferrer and Santamaria (2013) analyze four, developed European
markets and show that sentiment has a significant influence on returns, but this varies across
markets. This indicates the need for analyzing sentiment in different geographical areas to under-
7
stand its impact better but also across different classes of markets. Furthermore, a link is estab-
lished with the cultural differences investigated in Chui, Titman and Wei (2010) where the great-
er individualism (as defined by Hofstede (1980)) is linked to greater trading volume and volatili-
ty. Eight emerging markets including Turkey are studied in Daszyńska-Żygadło, Szpulak and
Szyszka, (2014) at a contemporaneous level and they find that markets are sensitive to investor
sentiment or mood especially during periods of negative sentiment. Emerging markets are also
the subject of analysis in Schmeling (2009) who finds that the impact on sentiment is higher in
counties where greater herd-like behavior is exhibited and for countries that have less efficient
regulatory institutions or less market integrity. Turkey, which is the subject of this study, scores
low on the individualism scale (37) and the finding above would lead us to conclude that we
should expect a greater impact of investor sentiment.
We collect data with LexisNexis on Turkey, one of the major emerging markets (a G20
country) for five years 2012-2017. Our approach to market sentiment is to aggregate positive and
negative news. We do this using daily data (for robustness we also consider monthly data).
We contribute by constructing a news based sentiment index for an emerging market.
Turkey is a volatile emerging market and has relatively high level of liquidity. This provides us
the chance to measure the decision making process of the market forces and reaction to the sen-
timent measure in market news. The next section describes the dataset. We continue by discuss-
ing the model and results, and the last section concludes.
Data and Methodology
We search through LexisNexis1 database and restrict our search to one country, Turkey
and two subjects, banking and finance news. This is to exclude all other news which are not rele-
vant to stock market investors. We focus our search on the headline and lead paragraph follow-
ing Kräussl and Mirgorodskaya (2017). This is to ensure that the keywords are featured in a
prominent position in the news articles and visible to a wide investor audience who may not
always read an item of news in its entirety. We perform the search for news in English, for the
past five years, for each sentiment word (positive and negative). The duplicates are grouped to
1 LexisNexis is global provider of information and analytics for professional and business customers across
industries. https://www.lexisnexis.com/
8
avoid exaggeration of the number of news stories which are based on the same story. This can
happen because multiple news outlets write about the same news item. Grouping stories prevents
any overestimation of the amount of news. To select relevant news stories we use the positive
and negative words to individually search for within articles classified in the “Turkey” geograph-
ic region and within the “banking and finance” news industry group. The search effectively re-
stricts and locates these positive and negative words within news relating to Turkey and within
the two relevant industry groups for the stock market. It searches in newspapers, major world
publications, magazines, wire services, blogs, business and industry news, university newspa-
pers, U.S. newspapers and web news which are available in the Lexis/Nexis database. Ultimately
we aggregate the positive and negative words in order to quantify market sentiment.
The search is repeated for 27 positive and 27 negative words as selected and used in Kräussl
andMirgorodskaya (2017). A full list of these words are available in Appendix 1. They use
monthly data and have a different market focus while in our case we use daily data as well as
lags up to two weeks to measure the effect of daily news in the market. Our dataset compiles the
positive and negative words on a daily basis between 2nd
January, 2012 and 2nd
January, 2017.
We use these sentiment measures in order to see what effect news sentiment has on the Turkish
market and by extension reach a conclusion about emerging markets sentiment using a news
based measure. There are 11,333 positive articles and 14,601 negative articles. We find 13,667
positive and 17,168 negative observations in five years. These are 1,619 positive and 1,654 nega-
tive days of newspaper articles. One factor affecting aggregate daily data is the weekend. We
add Saturday and Sunday observations to the Monday observations. This is because we expect
that news over the weekend to be incorporated in the Monday trading. After reformatting data
into a five day week, the positive and negative data observations are reduced to 1,257 days.
The news articles as described above are aggregated to obtain daily observations (number of arti-
cles per day). We construct the daily index similarly to Garcia (2013) by taking the ratio of posi-
tive news to the sum of the negative and positive news:
Pt=
(1)
Nt=
(2)
9
Where Pt is the positive sentiment; positive news is the day’s positive news divided by all
news; Nt stands for the percentage of daily negative news divided by all news for day t. The total
period considered is from 2nd
January, 2012 to 2nd
January 2017.
Figure 1 and Figure 2 show the sentiment index, distribution and the summary statistics.
[ insert Figure 1 and 2 here]
As independent variables we use the composite index daily closing values from Borsa Is-
tanbul 100 as our market benchmark. The daily market data is collected from Reuters. From the-
se we calculate the nominal daily returns for our analysis. One-year government bonds data is
collected over the sample period as a proxy for long term interest rates. The stock market return
is the percentage US Dollar return of the BIST 100 of the day in USD (USDdaychange). The
volume of BIST for the day is Vola and this variable is also used in VAR regression.
is the yield of one year government bond and is t is the
squared demeaned residuals of the market index returns.
[ insert Table 1 here]
The summary statistics are reported at Table 1: Panel A. Correlation coefficients are re-
ported at Table 1: Panel B.
A list of dependent and independent variables are presented in the data section and their
descriptions are presented in Appendix 1. To investigate the potential effect of market sentiment
on the Turkish stock market return we employ a vector autoregression regression (VAR) model
following on from Tetlock (2007) and Kräussl and Mirgorodskaya (2017). Similarly in our main
model, we estimate a VAR model with log return of the market indec and the log of the change
of market sentiment as the endogenous variables in the model. As exogenious variableswe
include several market fundamentals as controls. As our data is made up of daily observations
and not monthly as in the case of Brown and Cliff (2005) and Kräussl and Mirgorodskaya
(2017), we include 10 lags which is equivalent to two weeks or 10 working days. We contend
10
that given the daily aggregate news would not have a longer term effect in the market beyond
this time frame. The second VAR model is specified similarly to the first except we include
market volatility instead of the market return.
We specify the following empirical models:
(3)
And
(4)
where is the log return of the Turkish stock market; L10(Xt) is
the lag operator that transforms the variable Xt into a row vector consisting of 10 lags of Xt;
is the log of the change of aggregate positive news as de-
noted in equation (1); is the squared demeaned residuals of the market
index returns; represents several exogenous variables: volatility in the stock market.
is the yield of one year government bond.
We use a unit-root test to check the contemporary (t) values of the endogenious variables.
The test reveals that they are stationary I(0) processes which means that the unrestricted VAR
model is appropriate for this analysis. We use Augmented Dickey Fuller test as shown in Table
1 Panel C and log of change of sentiment rejects unit root for lag 10. The market return does not
11
have a unit root. Bond yields have unit root, therefore we use change of bond yield. Bond volatil-
ity and market volatility do not have unit root.
We use Granger causality test for pairs, and we find statistical significant causality of
bond yields have for sentiment.
Results
We estimated two VAR models as denoted in equations 3 and 4. The model analyses the
effect of positive news on the stock market return in Turkey. The sentiment index is calculated
from agreegate daily positive news in terms of the Turkish market regarding finance and banking
news. Our models estimate the influence of the exogenious variables (lags of market sentiment
and market return and control variables: unemployment rate, consumer confidence, CPI,
industrial production and market volume of transactions). The findings are presented in table 2.
[ insert Table 2 here]
In column (1) of table 2 we present coefficients for the VAR model specified in eqation
1. The results show that positive stock market performance is associated with positive and
statistically significant lags of positive news sentiment in lags 5 and 7. This indicates that
increase in aggregate positive news (as defined in equation 1) in the stock market in Turkey
influences the market returns at time 0. This means that it takes at least a week for positive news
to be incorporated in the market returns. We argue that this can be understood as an efficiency
indicator for emerging markets such as Turkey. From the control variables, we find that they are
not statistically significant, Column 2 of table 2 presents the results for the second VAR model
(denoted in equation 4). We consider the coefficients of the market sentiment when the
dependent variable is the market volatility. We find that sentiment is negative and significant at
lag 1 only. Consistent with expectations, this indicates that positive new media sentiment has a
negative effect on market volatility i.e. it decreases volatility. It can be said that positive news
have a ‘calming’ effect on the markets. From the other control variables we find that the
unemployment rate has a negative and significant effect on the volatility in the markets. This
12
result is counter-intuitive as it indicates that an increase in the unemployment rate decreases the
market volatility. We also find the market volume of transactions to be positive and significant.
In this case it indicates that higher amount of market transactions mean an increase in market
volatility as defined in the previous section.
The findings suggest that the optimistic news have an effect on the Turkish investors’
perception and thereby influence their decisions and choices about a week after the news occur.
While we do not expect to see such a delayed reaction to positive news, this may be a feature of
the emerging markets where investor sophistication is lower and the influences on the markets
are not straightforward. Our results show that the index created from positive news has a
significant negative effect after five and seven business days.
Conclusion
We analyze the role of sentiment which is generated through aggregate daily news on the
stock market performance in an emerging market such as Turkey. We do this by using a VAR
model following on from similar media sentiment studies such as Kräussl and Mirgorodskaya
(2017) and Tetlock (2007). We include Turkish stock market return, market volatility and
positive news media sentiment as dependent (endogenous) variables in the specified VAR
models. As independent (exogenous) variables we include 10 lags of the dependent variables as
well as some macroeconomic variables which serve as controls for the stock market
performance.
Our results indicate that it takes a week or more for the overreaction to positive sentiment
generated through increased positive daily news to be incorporated in the market returns in
Turkey. This may be expalined as an overreaction in the beginning and continuation of the
positive news in the media after the actual fact and then selling in the stock market by the
sophisticated investors afterwards. We contend that the lack of efficiency in emerging markets is
creating the market reaction to positive news.
13
Tables
Table 1. Panel A: Descriptive Statistics
Pospct is sentiment index of positive number of counts per day divided to the sum of positive and negative numbers. Usddaychange is the BIST 100 stock market return. Vola is the
demeaned squared residuals of market return, Unemploym is the unemployment rate, Consconf is consumer confidence rate, CPI is the monthly consumer price inflation rate, Indp is the industrial production rate, USDVolume is the day’s trading volume in USD.
Table 1. Panel B: Correlation Coefficients Pospct is sentiment index of positive number of counts per day divided to the sum of positive and negative numbers. Usddaychange is the BIST 100 stock market return. Vola is
the demeaned squared residuals of market return, Unemploym is the unemployment rate, Consconf is consumer confidence rate, CPI is the monthly consumer price inflation rate, Indp is the industrial production rate, USDVolume is the day’s trading volume in USD.
Table 2. Effect of sentiment on Turkish stock market returns and volatility Pospct is sentiment index of positive number of counts per day divided to the sum of positive and negative numbers. Usddaychange is the BIST 100
stock market return. Vola is the demeaned squared residuals of market return, Unemploym is the unemployment rate, Consconf is consumer confidence rate, CPI is the monthly consumer price inflation rate, Indp is the industrial production rate, USDVolume is the day’s trading volume in USD.
BIST 100 Mar-
ket Return
Volatility
Sentiment(-1) -0.001081 Sentiment(-1) -0.000112
(0.00134)
(0.000049)
[-0.80651]
[-2.26080]***
Sentiment(-2) -0.001745 Sentiment(-2) 0.0000153
(0.00137)
(0.000050)
[-1.27765]
[ 0.30455]
Sentiment(-3) 0.000260 Sentiment(-3) 0.0000599
(0.00136)
(0.000050)
[ 0.19198]
[ 1.19762]
Sentiment(-4) 0.000411 Sentiment(-4) -0.0000483
(0.00137)
(0.000050)
[ 0.30119]
[-0.95903]
Sentiment(-5) -0.003056 Sentiment(-5) 0.0000211
(0.00137)
(0.000051)
[-2.22897]***
[ 0.41762]
Sentiment(-6) 0.001857 Sentiment(-6) -0.0000264
(0.00136)
(0.000050)
[ 1.36081]
[-0.52611]
Sentiment(-7) -0.002795 Sentiment(-7) 0.0000215
(0.00138)
(0.000051)
[-2.02915]***
[ 0.42231]
Sentiment(-8) -0.000750 Sentiment(-8) 0.0000393
(0.00139)
(0.000051)
[-0.54053]
[ 0.76823]
Sentiment(-9) 0.0000919 Sentiment(-9) 0.0000926
(0.00138)
(0.000051)
[ 0.06656]
[ 1.82205]
15
Sentiment(-10) -0.000904 Sentiment(-10) 0.0000180
(0.00137)
(0.00005)
[-0.66031]
[ 0.03583]
UNEMP -0.024297 UNEMP -0.008006
(0.06348)
(0.00239)
[-0.38277]
[-3.34704]***
CONSCONF 0.010463 CONSCONF -0.000677
(0.01436)
(0.00054)
[ 0.72848]
[-1.26260]
CPI -0.031360 CPI -0.000358
(0.09372)
(0.00349)
[-0.33463]
[-0.10282]
INDP -0.048415 INDP 0.001998
(0.02873)
(0.00107)
[-1.68526]
[ 1.86250]
LOG(USDVOLUME) -0.000961 LOG(USDVOLUME) 0.000701
(0.00218)
(8.6E-05)
[-0.44013]
[ 8.18693]***
C 0.008546 C -0.012630
(0.04402)
(0.00170)
[ 0.19412]
[-7.40862]***
R-squared 0.032901 R-squared 0.132770
Adj. R-squared 0.006763 Adj. R-squared 0.109332
F-statistic 1,258,747
F-statistic 5.664592
Log likelihood 2,463,984
Log likelihood 5592.211
16
Figures
Figure 1. Sentiment index
Figure 2. Sentiment index statistics
0
20
40
60
80
100
120
140
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Series: POSPCTSample 1 1257Observations 1257
Mean 0.450827Median 0.444444Maximum 0.954545Minimum 0.045455Std. Dev. 0.172469Skewness 0.130553Kurtosis 2.619804
Jarque-Bera 11.14153Probability 0.003808
17
References
Antweiler, W., & Frank, M. (2004). Is All That Talk Just Noise? The Information Content of
Internet Stock Message Boards. The Journal of Finance, 59(3), 1259-1294
Arif, Salman, Lee, Charles M. C.; Aggregate Investment and Investor Sentiment, (2014) The
Review of Financial Studies, 27 (11), pp. 3241–3279
Baker, Malcolm and Jeffrey Wurgler (2006) ‘Investor Sentiment and the Cross-Section of Stock
Returns’ Journal of Finance, Vol 61(4) pp 1645-1680.
Baker, Malcolm and Jeffrey Wurgler (2007) ‘Investor Sentiment in the Stock Market’ Journal of