Sentiment spillover effects for US and European companies

Journal of Banking and Finance 106 (2019) 542–567

Contents lists available at ScienceDirect

Journal of Banking and Finance journal homepage: www.elsevier.com/locate/jbf

Sentiment spillover effects for US and European companies Francesco Audrino, Anastasija Tetereva∗ Chair of Mathematics and Statistics, University of St Gallen, Bodanstrasse 6, St Gallen 9000, Switzerland

a r t i c l e

i n f o

Article history: Received 20 August 2018 Accepted 30 July 2019 Available online 2 August 2019

a b s t r a c t The fast-growing literature on news analytics provides evidence that financial markets are partially driven by sentiments. In contrast with previous studies that have almost exclusively focused on the direct effects of the news related to single companies or sectors, we investigate the time-varying dynamics of news’ cross-industry influences for a set of US and European stocks over a period of 10 years. The graphical Granger causality of the news sentiments-excess return networks is estimated by applying the adaptive lasso. We find significant spillover effects and show the importance of sentiments related to certain sectors for the whole cross-section of stocks. © 2019 Elsevier B.V. All rights reserved.

1. Introduction The influence of the news and social media in politics and economics has grown consistently over the last decades. Researchers have made numerous efforts to understand the limits of arbitrage using the synthesis of social science and financial economics. The availability of real-time online sources and recent developments in machine learning algorithms have made the news relevant for the area of quantitative finance. The news influences the opinions and expectations of investors, which find expression in sentiments. A growing number of agencies are developing news indices, which can potentially help improve trading strategies, as recent research shows that the behavior of market participants may be highly influenced by the news. Some authors have explored the sensitivity of stock returns to stock-related news; see, for example, Fang and Peress (2009), Peress (2014), Akyildirim et al. (2015), Narayan and Bannigidadmath (2015), Ding et al. (2015) and Luss and d’Aspremont (2015). Akyildirim et al. (2015) describe the role of firm-specific public announcements on liquidity, price and the volatility of individual stocks. Chan (2003) finds a significant influence of bad news on the monthly returns of small and illiquid stocks. Hwang (2011) shows that the views on some countries affect firms’ investment decisions. Moreover, the author finds empirical evidence that countryspecific sentiment causes the deviation of securities’ prices from their fundamental values. Benhabib et al. (2016) suggest to employing sentiment shocks for asset pricing over business cycles.

∗

Corresponding author. E-mail addresses: [email protected] (F. Audrino), anastasija.tetereva@ unisg.ch (A. Tetereva). https://doi.org/10.1016/j.jbankfin.2019.07.022 0378-4266/© 2019 Elsevier B.V. All rights reserved.

Allen et al. (2015) have analyzed how the performance of the GARCH, GJR and EGARCH can be improved by including sentiment data in the model. Additional, general conclusions have been obtained by Cahan et al. (2009) who have empirically shown that the information coming from the news can be seen as an additional factor in the Fama French factor models. Beber et al. (2015) suggest extracting real-time macroeconomic factors from the news and demonstrate their ability to forecast future changes in economic conditions. The most recent research by Borovkova et al. (2016) makes an effort to construct a systematic risk indicator based on the news related to the biggest financial companies. They show that the proposed risk measure outperforms the conditional capital shortfall measure of systemic risk (SRISK) by Brownlees and Engle (2015) and the CBOE volatility index (VIX) by Brenner and Galai (1989) in signalling periods of financial stress. Most of the above-mentioned studies concentrate, however, on the direct effects of the asset-related news on the price and do not consider spillover effects. There is a vast literature discussing the presence, size, and relevance of spillover effects in returns and volatilities for the pricing of securities, hedging and other trading strategies, and for regulatory policies. Most of these studies have focused mainly on spillovers among markets; see, among others, King et al. (1994), Booth et al. (1997), Forbes and Rigobon (2002), and Diebold and Yilmaz (2009). In contrast, there are only a few analyses that have investigated spillover effects among industry sectors; see, for example, Hammoudeh et al. (2009) who considered spillovers among the service, banking, and the insurance sectors, and Chan et al. (2016) who analyzed spillover between the real estate sector and other industry sectors. Along this second strand of the literature, we extend previous studies analyzing the presence and

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

relevance of spillover effects among industry sectors using news sentiment data. In this paper, we study the news spillover effects among sectors based on the news data provided by Thomson Reuters. The daily sentiment indices for 10 US sectors, 10 non-US sectors and 5 countries are considered together with the prices of more than 100 stocks. The analyzed time interval ranges from January 1, 2005 to December 31, 2014 and covers the global financial crisis, the US debt-ceiling crisis, and the European sovereign debt crisis. We conduct a rolling window analysis of the cross-industry influence of news sentiments. The influence of news on the stock excess returns is defined by means of the graphical Granger causality, which is estimated by constructing a sparse network. In order to reduce the number of false positive edges of the network, the adaptive lasso methodology is applied for the estimation of the networks with a related testing procedure introduced in Audrino and Camponovo (2018). Two characteristics describing the relevance and strength of the news coming from an individual sector or a specific country are suggested, and their dynamic behavior is studied. The news sentiments coming from US and nonUS industrial sectors show similar behavior and are highly correlated; for this reason, the spillover effects are analyzed separately for the US and European companies, and the results are compared. In this study, we provide strong empirical evidence that the class of stock-relevant sentiments is wider than just the stockspecific announcements and the news coming from the related sector. We show that the returns of the whole industry are driven by the news coming from several sectors. Moreover, we investigate how the importance of the news changes over time, getting stronger just before periods of financial turbulence, which can be seen as an early warning signal for investors. The paper is organized as follows: The first section features an introduction to the Thomson Reuters MarketPsych news data, and Section 2 discusses how the signal can be extracted from noisy news. The methodology related to the estimation of the graphical Granger causality by means of sparse regression models is provided in Section 3. Full results are presented in Section 4. Section 5 contains an empirical illustration of how the predictive power of time series models can be improved by augmenting the conditional mean equation by specific sentiment indices. Finally, we summarize the main contribution of the paper. 2. Sentiment data 2.1. Understanding the Thomson Reuters MarketPsych index The sentiment data used in the current paper are provided by Thomson Reuters and include the MarketPsych indices (TRMI) ranging from 2005 till 2014. In this section, we provide more details on the construction of the MarketPsych index and its characteristics. TRMI is an advanced linguistic index which scores the online media sources specific to companies, currencies, commodities and countries. It employs machine learning techniques to account for variation in data sources and to consider changes in a word’s meaning over time. In contrast to the most popular approaches, the TRMI is sensitive to grammatical structures and accounts for correlations among the words. Moreover, TRMI has better coverage than other data used in the academic community. Real-time updated TRMI indices monitor 24 different emotions going beyond positive or negative sentiments based on 50,0 0 0 professional news sources. The algorithm starts with the search of keywords associated with the asset of interest and looks for positive and negative words associated with it. In the next step, the algorithm maps the keywords to various asset-related variables. A numerical value which

543

accounts for the tense, proximity, and many other multipliers is then assigned to each variable. Variables referring to the same subject are aggregated to a sentiment score. Additionally, such characteristics as relevance and novelty are taken into account when the variables are summed up in the index. The sector-related sentiment is a composite index grouping the sentiments of the most influential companies in the sector. The weights of all constituents are constantly recalculated, giving bigger coefficients to the more influential companies. In the current paper, we make use of the TRMI sentiment index for 10 industries and 5 countries. Moreover, S&P 500 and Euro Stoxx 50 market-wide sentiment indices are considered as market-wide sentiment indicators. All indices are listed in Appendix A (Table 4). This analysis considers a daily frequency, as we are interested in the analysis of the global financial crisis, which goes back to 2008 when tick-by-tick data were not available for all the asset classes we are interested in.

2.2. Extracting signal from noise As can be seen from Fig. 1, the news data are very noisy and cannot be directly used for modelling. Two main approaches for extracting the information from the noise which appeared in the recent sentiment literature could be potentially used. The first approach is the moving average convergence/divergence oscillator (MACD). It was first developed by Appel (2003) and has been applied to sentiment data by, for example, Peterson (2016), Kirange et al. (2016) and Lugmayr and Gossen (2013). Becker (2016b) shows that 10–30 MACD of the TRMI sentiment about Starbucks has an influence on its price. Becker (2016a) points out the connection between Volkswagen share prices and the 30–200 MACD of the TRMI Media sentiment. In the same spirit, in the current study we follow a more parsimonious standard moving average approach. The size of the rolling window is set to 200 trading days. In other words, the smoothed sentiment at time t is calculated as the average sentiment from t − 200 till t. The moving average is applied to the daily values of the news sentiment. Covering the previous 40 weeks of trading, the 200-day simple moving average is considered to be a key indicator by traders and market analysts for determining the overall long-term market trend. In a second approach, we employed the Kalman filter first introduced by Kalman (1960) and previously discussed by Borovkova and Mahakena (2015) and Borovkova et al. (2016) for applications to sentiment data. For more details on the estimation of unobserved sentiment by means of the Kalman filter we refer to Appendix B. One main drawback related to the application of this smoothing technique in the context of the sentiment data is due to the required assumption of a Gaussian distribution. The sentiment indices which always take values between minus one and one do not fulfill this assumption. To address this shortcoming, one could employ a simulation-based approach. Nevertheless, such a computationally demanding approach is out of the scope of the current study. The results of our estimations showed that for most indices the moving average of the sentiment and the Kalman filtered version of it are very closely related. As an illustrative example, Fig. 2 shows both versions of the financial sector-related sentiment. It is evident that the moving average version can be seen as some sort of lagged version of the Kalman smoothed sentiment. Moreover, the similarity becomes more obvious with a decrease of the size of the rolling window: In fact, both versions are almost identical for the 30-day moving average financial sector-related sentiment. For the sake of brevity, given that results of the whole analysis obtained with the Kalman smoothed version of the sentiment are

544

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Fig. 1. News sentiment for the US Energy and Non-Cyclical Consumer Goods and Services and the smoothed news sentiment.

Fig. 2. Kalman smoothed (solid line) and moving average (dashed line) news sentiment for the US financial sector.

Fig. 3. Moving average news sentiment for the US (solid line) and the non-US (dashed line) financial sector.

qualitatively similar to those discussed for the moving average approach, we summarize them in Appendix G (Tables 11 and 12). The original data and moving average versions for Energy and Non-Cyclical Consumer Goods and Services sectors are shown in the Fig. 1. It is worth noting that for most sectors, news sentiments of the US and non-US sectors behave in a very similar way. The moving average versions of US and non-US financial news sentiments are given in Fig. 3. It is evident that the two series overlap significantly and it would be difficult to disentangle the influence of the US and non-US news on the different industries under investigation in a linear regression framework. The problem is not so relevant when considering industry sector sentiment series within a common market: In this case we did not find any empirical evidence of a common trend, and correlations among residuals from an individual ARIMA-GARCH estimation on the sector sentiment series are quite low and close to zero. We decided, therefore, to analyze the US and the European markets separately.

3. Modelling spillover effects In the current study, we are interested in analyzing whether and under which circumstances the individual time series of the companies’ prices can be potentially influenced by the news on sectors and countries. Cross-industry news spillover effects have not been directly addressed in the literature thus far. Most of the recent studies investigate the direct influence of the sentiment about an asset on the asset itself and do not consider that, for example, news coming from the energy sector can influence financial companies. Borovkova and Mahakena (2015) and Borovkova (2015) have shown that extreme positive and extreme negative sentiment days influence the future price momentum of natural gas and that there is a complex relationship between the arrival of the news and the price jumps. Similar results have been developed for the energy markets in Borovkova and Lammiman (2010). Erawan (2015) uses the sentiment data for different

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

sectors as an input for classification and regression trees and finds empirical evidence that the trading strategy can be improved by including the sector-specific news data in the model. However, no spillover effects are studied. Borovkova and Mahakena (2015) show that abnormal stock returns calculated from Fama-French and Carhart factor models might be explained by the specific sentiment data. The recent study by Borovkova et al. (2016) makes an effort to construct the risk measure based on the financial news sentiment data only and tests for the ability of this indicator to forecast the stress in the market. 3.1. Penalized estimation of the sentiment networks As mentioned above, the aim of this work is to conduct a systematic empirical investigation of the spillover effect of the sentiment coming from different industries and countries on individual stock prices. For this purpose, we employ sparse graphical models in order to obtain insight into the joint causal relationship between the individual time series. First, we introduce the notation specific to the network literature. In the first step, a graph G = V, E  is considered, where V is the set of p nodes corresponding to each variable X 1 , X 2 , . . . , X p and E ⊂ V × V is the set of the edges corresponding to the pairwise association of all variables. Causal relations are usually represented by directed graphs: for example, if the variable Xi is assumed to be influenced by a set of variables Xj , j = 1, . . . , p, j = i, the associations j → i are studied. Thus, the dependence of Xi and a set of the nodes is stated as:





X i = fi X 1 , X 2 , . . . , X p + εi , i = 1, . . . , p, j = i.

(1)

In (1), ε i is an error term. The function fi is usually assumed to be linear. In this way, the estimation of the network can be reduced to the estimation of the individual regressions

Xi =



β i j X j + εi ,

(2)

j=i

where the associations j → i are expressed in terms of the coefficients β ij . The individual regressions can be estimated by OLS (ordinary least squares) or more sophisticated methods. For example, Peng et al. (2009) suggest estimating nonzero partial correlations by joint sparse regression models. Where the causal relationship among the individual time series is important, it is useful to consider the concept of Granger causality, first introduced by Granger (1980) and now widely discussed j in the literature. Per definition, the time series process {Xt }, t = 1, . . . , T Granger causes the time series process {Xti }, t = 1, . . . , T if j

the regression of Xti on the past values of Xti and Xt gives a better fit than the regression of the past values of Xti . One way to define Granger causality for a network where Xti is the response time series which is caused by a high-dimensional [−i] p time series Xt = Xt1 , Xt2 , . . . , Xt , t = 1, . . . , T , j = i is mentioned in Lozano et al. (2009) and Arnold et al. (2007) in the context of bioinformatics. It has been proposed to test for Granger causality within the network by applying some kind of variable selection j j procedure. In particular Xt is Granger causing Xti if the lags of Xt are selected by some sparse estimation procedure for any time lag l ∈ L, j = 1, . . . , p, t = 1, . . . , T . Formally speaking, the following regression model is estimated:

Xti =

p  

j βli, j Xt−l + εti i = 1, . . . , p.

(3)

j=1 l∈L

In practical applications, the networks often appear to be highdimensional and the number of nodes can exceed the sample size. This fact leads to inaccuracy and overfitting of the estimates. The

545

remaining part of this section discusses how these drawbacks can be overcome by means of sparse estimation methods. A natural way to improve the performance of the estimator in a high-dimensional regression model is to introduce a regularization penalty. Many new regularized methods have been developed in the literature over the last decades, including the least absolute shrinkage and selection operator (lasso) estimator by Tibshirani (1996), SCAD by Fan and Li (2001), elastic net by Zou and Hastie (2005), fused lasso by Tibshirani et al. (2005) and extensions of the lasso models, such as adaptive lasso by Zou (2006) and group lasso by Yuan and Lin (2006). The application of sparse estimation methods for the estimation of high-dimensional networks in computer biology is widely discussed in the literature; see, for example, Gustafsson et al. (2005), Shimamura et al. (2007), Li and Li (2008), Friedman et al. (2008) and Jacob et al. (2009). The present work employs the lasso approach originally formulated by Tibshirani (1996), i.e.

 T



t=max (L )+1

β(λ )=argmin

Xti −

p



j=1

l∈L

j βli, j Xt−l

T − max (L )

β



2

+λ β 1 , (4)

p  i, j where β 1 = j=1 l∈L |βl | and λ > 0 is a penalty parameter. As a result of l1 penalization, the lasso solution is sparse, i.e. some coefficients are set exactly to zero. However, lasso has been shown to lack the consistency for selecting the relevant variables and to produce small false positive non-zero coefficients. The two-stage adaptive lasso procedure introduced by Zou (2006) corrects the behavior of the lasso and reduces the number of false positives by re-weighting the penalty function, i.e.:

β

adaptive

 T

(λ ) = argmin β

+λ

p   j=1 l∈L



t=max (L )+1



Xti −

p



j=1

l∈L

j βli, j Xt−l

2

T − max (L )

|βli, j | , i, j |βinitial | ,l

(5)

i, j where β , j = 1, . . . , p, l ∈ L is the initial estimator from the initial , l i, j first step of the procedure. Thus, if β is large, the adaptive initial , l i, j

lasso employs a small penalty for the βl , j = 1, . . . , p and improves the estimation of the effective variables. The simple OLS, the ridge regression or any other consistent estimator can be used to obtain the initial estimates. The theoretical properties and the details of the estimation algorithms can be found in Bühlmann and Van De Geer (2011). In the current work, we include several lags of each regressor in the regression model in order to check for Granger causality. In particular, we are interested in whether the news for a given sector and not the specific lags are Granger causing the price of the asset. We implement the adaptive lasso procedure with the OLS coefficients as the initial estimators. In addition, we employ the testing procedure based on the finite sample properties of the adaptive lasso developed by Audrino and Camponovo (2018) to reduce the number of false positive selected variables. In the current framework, the causality of the news data and its lags on the prices of the assets in different sectors needs to be estimated. For this reason, we define two sets of variables. Let {Xti }, t = 1, . . . , T , i = 1, . . . , p1 be the set of the prices or excess returns j and let {Xt }, t = 1, . . . , T , j = p1 + 1, . . . , p be the set of the news i } be the corresponding lags. Thus, the network sentiment and {Xt−l consists of p nodes. As we are interested in Granger causality of the news rather than returns of other assets, the edges connecting j the price variable Xti to the lags of other prices Xt−l , j = 1, . . . , p1 ,

546

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

j = i are set to zero. This restrictive assumption is introduced because the sentiment data of the industries contain the price information of the biggest companies. If both price and sentiment lags are included in the regression, the estimated coefficients might be misleading. The combination of the Granger causality concept with the adaptive lasso variable selection procedure results in the following procedure: 1. Define the input as the p1 -dimensional vectors of returns {Xti }t=1,...,T , i = 1, . . . , p1 and p2 -dimensional vector of senti-

{Xtj }t=1,...,T ,

ment j = p1 + 1, . . . , p. 2. Set the adjacency matrix corresponding to the network G = V, E  equal to the zero matrix. 3. For i = 1, . . . , p1 : • apply the adaptive lasso variable selection with its related testing procedure for false positives to the model

Xti =

 l∈L

  p

i βli,i Xt−l +

j βli, j Xt−l + εti ,

(6)

j= p1 +1 l∈L

where L is the set of the predefined lags. j • If for Xti at least one lag of Xt is selected as significant, place j i the edge X → X into E, l ∈ L. 3.2. Graphical Granger causality Employing the notation introduced above, we say that the news from one industry is Granger causing the excess returns in the other industry if the lags of the sentiment on one sector are selected by the adaptive lasso procedure (6) for the excess returns of the assets from the other industry, i.e. X j −→ X i , i = 1, . . . , p1 , j = p1 + 1, . . . , p. In the current studies, we distinguish between direct effects and spillover effects, i.e. if the sentiment of a particular sector is Granger causing the excess returns of the assets from the same sector, we call them direct effects, whereas in case the assets belong to different sectors we call them spillover effects. This study focuses on the information spillover effect due to sentiment. For this reason, the network at time t, t = 1, . . . , T is constructed by regressing the individual excess returns Xti on their i , l ∈ L and the set of the news sentiment lags on the own lags Xt−l j

sectors and the countries Xt−l , j = p1 + 1, . . . , p, l ∈ L. As explained above, lags of excess returns of other stocks are not taken into consideration in the individual regressions. The estimation is performed using the R package glmnet by Friedman et al. (2009). In this study, we suggest including 1 day, 1 week and 1 month lags in the individual regressions, i.e. l ∈ {1, 5, 22}. Sentiment data of the same day (0 lag) are not included in the analysis because they might contain the information about the prices and thus overestimate the causality of the news. The choice of 1 month as mentioned above is motivated by the recent studies in the sentiment literature which reveal that the shocks in the sentiment can drive the price of the underlying during the period of the next month; see Borovkova et al. (2016), for instance. 3.3. Relevance and strength of the sentiment In order to analyze the spillover effects over time, one needs to introduce some measures of connectedness among the excess returns and the news sentiments of the sectors and countries. Dynamic analysis of these measures can provide insight into the importance of particular sectors and countries during the global financial crisis starting in 2008 and the subsequent European sovereign debt crisis. In this study, we are interested in two characteristics of the sentiments: the relevance and the strength. We propose to measure the relevance of each sentiment by the number of outgoing edges. This measure reflects the share of regressions which have selected the considered sentiment index as an

active variable. Therefore, the relevance is not informative for expressing the strength of the causality. For this reason, we define the second characteristic (strength) as the mean absolute value of the outgoing edges. Formally speaking, the overall relevance of the sentiment variable Xj can be defined from the estimated coefficients over p1 regressions of the form (6):





R X j → {X 1 , . . . , X p1 } =

 p1

i=1

 1{|β 1

i, j

i, j | + |β5i, j | + |β22 |}

p1

,

(7)

j = p1 + 1, . . . , p,



1{x} =

1 if x > 0, 0 otherwise

is an indicator function. The overall strength of the lth lag of the sentiment variable Xj can be defined as the average absolute value of the coefficients of the outgoing edges:

 p1 i, j   |β | Sl X j → {X 1 , . . . , X p1 } =  p i=1 l i, j , l ∈ {1, 5, 22}, 1  |} 1{|β





i=1

(8)

l

if R X j → {X 1 , . . . , X p1 } is different than zero and is zero if the variable is irrelevant. Thus, this characteristic represents the averi, j in the regressions of the age absolute value of the coefficients β l excess returns of company i for the sentiment j. If only the sum of the absolute values were considered, it would be impossible to distinguish between the nodes with many links with small coefficients and the nodes with a small number of links and bigger coefficients. Speaking in terms of graphical representation, the relevance characterizes the average number of the outgoing links and the strength characterizes the average width of the link. If the sentiment (node) has small relevance and high strength, it is selected as significant in a small number of regressions but estimated with relatively high coefficients. In definition (8), the strength of the regulator is defined separately for each lag. Similarly, the same measures can be defined for particular groups of companies. In this case, the average is taken over the companies contained in the specific group. For example, if the average is taken over the companies corresponding to the financial sector with the Industrials sentiment as the variable under consideration, the relevance and the strength will show the spillover effect of the Industrials-related news to the financial sector. 3.4. Spillover effects and economic factors It is important to mention that the scores used for the analysis are aggregated and processed scores. In particular, major news shocks contribute simultaneously to the aggregated sentiment scores of all companies and sectors. For this reason, the omission of a market-wide sentiment might cast doubts on the validity of the spillover effects. If this is the case, after controlling for the market-wide sentiment, most spillover effects found in the previous section should be significantly reduced or fully eliminated. We verify whether the market-wide sentiment is the true channel causing the observed spillover effects by including two possible sources of market-wide information in the model. First, we control for the country-specific factor by including US, China, Germany, Italy, and Greece sentiment data in the analysis. These countries gained wide attention in the news during the considered time period due to the crisis situation. Failing to consider this information could lead to a biased estimation of the parameters. Second, the market-wide sentiment of the S&P 500 index or of the Euro Stoxx 50 index is used in the US and European stock market analysis to control for systemic general effects, respectively.

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

547

Fig. 4. Moving average news sentiment for S&P 500 index (left) and Euro Stoxx 50 index (right).

Smoothed sentiments of the above-mentioned indices are shown in Fig. 4. Moreover, in an additional robustness check we replicated the analysis for the US market adding the VIX volatility index and its corresponding lags as additional covariates. The VIX – CBOE volatility index – measures the expected volatility implied by the S&P 500 index option and therefore reflects uncertainty in the market. This index is often called a ’fear index’ due to its ability to mimic investors’ sentiments. Since the results remain almost unchanged when the VIX index is added to our analysis, we summarize them in Appendix D (Figs. 12 and 13). The fact that the spillover effects remain qualitatively unchanged after controlling for market-wide economic factors is an indication that the aggregation scheme that is used to construct the sentiment indices for the different sectors does not have a significant impact on the main conclusion of the analysis. 4. Results In this section, the news spillover effects of 10 industrial sectors are analyzed using the methodology of the graphical Granger causality. The data set contains the stock excess returns and TRMI sentiment indices on 10 sectors and two market indices listed in Appendix A (Table 4) and 5 countries, namely, US, China, Germany, Italy, and Greece. 78 US companies and 78 European companies are investigated, which corresponds to approximately 8 companies per sector. The full list of the companies can be found in Appendix C (Tables 5 and 6). In our analysis of cross-industry spillover effects, we focus on the largest businesses by market capitalization and the most liquid stocks. In fact, it has been previously empirically shown that small- and micro-sized companies and illiquid stocks heavily depend on firm-specific announcements; see, for example, Chan (2003). The daily data ranging from the January 1, 2005 to December 31, 2014 are used. In order to extract the signal from the noisy sentiment data, the moving average approach described in Section 2.1 is applied. All the data are standardized after applying the rolling window average. As mentioned above, the sentiments for the US and non-US industries show similar trends over time. Therefore, the analysis is performed separately for the US and European market. The rates of returns in excess are calculated

by applying the CAPM model by Sharpe (1964) using as a benchmark the S&P 500 index for the US market and the Euro Stoxx 50 index for the European market. The choice of this parsimonious model is motivated by the arguments discussed by Harvey (2017), who states that the statistical significance of other factors in the ’factor zoo’ is questionable if the multiple testing problem is not controlled for and p-values and thresholds are not adjusted accordingly. The preliminary analysis showed that the results for the prices and excess returns using different benchmarks coincide, which is consistent with the discussion in Erawan (2015). A daily rolling window approach is employed in order to analyze how the connectedness between the news and the excess returns changes over time. The size of the rolling window is set to be equal to 200 trading days, which is the shortest possible size addressed in the literature; see Audrino and Knaus (2016) for more details. The small size of the rolling window is motivated by the possible time breakpoints. The examples of the networks obtained by the adaptive lasso procedure for the US stocks on January 7, 2008 and January 7, 2013 are presented in Figs. 5 and 6, respectively. The industries and countries are grey-colored in the graph, the abbreviations are given in Appendix A (Table 4), and the companies are whitecolored. The whole network is constructed by regressing excess returns of each company on the past values of excess returns and lags of the sentiments of the sectors and countries. Therefore, the connections coming from the sentiment (grey nodes) to the returns (white nodes) represent the Granger causality of the sentiment indices to the excess returns. If the node is self-connected, the lags of the excess returns of the company itself were selected by the adaptive lasso procedure. The date always corresponds to the last day of the rolling window. It is evident that the number of connections within the network changes significantly over time. The number of edges reaches its maximum during the 2008 global financial crisis. This finding coincides with the results by Chunxia et al. (2016) who find that the amount of information flow between industries is growing significantly as crisis intensifies. Similarly Baur (2012) shows increased contagion between financial sector stocks and the real economy during the crisis. For European companies, the number of connections increased significantly during the global financial crisis and European sovereign debt crisis in 2010.

548

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Fig. 5. The graphical Granger network for the set of US assets on Jan. 7, 2008.

Fig. 6. The graphical Granger network for the set of US assets on Jan. 7, 2013.

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

549

Table 1 Mean relevance of the news sentiments of the US sectors and selected countries for the US companies by sector ranging from Aug. 1, 2006 to Dec. 31, 2014.

asset MPTRXUSFIN MPTRXUSTEC MPTRXUSIND MPTRXUSMAT MPTRXUSUTL MPTRXUSCOM MPTRXUSHLC MPTRXUSNCY MPTRXUSENE MPTRXUSYCY US IT GR DE CN MPTRXUS500

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.92 0.38 0.09 0.20 0.07 0.07 0.07 0.06 0.06 0.41 0.09 0.16 0.11 0.15 0.10 0.04 0.17

0.91 0.33 0.06 0.09 0.07 0.05 0.07 0.04 0.06 0.32 0.06 0.09 0.05 0.13 0.08 0.03 0.11

0.92 0.39 0.07 0.13 0.06 0.05 0.06 0.06 0.03 0.35 0.06 0.08 0.07 0.13 0.10 0.04 0.12

0.93 0.39 0.08 0.14 0.07 0.08 0.09 0.07 0.07 0.40 0.10 0.11 0.12 0.16 0.10 0.04 0.13

0.85 0.44 0.13 0.17 0.11 0.11 0.10 0.12 0.08 0.36 0.11 0.13 0.13 0.16 0.17 0.09 0.18

0.91 0.44 0.07 0.14 0.11 0.09 0.11 0.11 0.10 0.45 0.10 0.17 0.12 0.20 0.11 0.06 0.19

0.93 0.32 0.07 0.13 0.08 0.08 0.06 0.08 0.06 0.36 0.09 0.11 0.08 0.14 0.12 0.05 0.15

0.82 0.29 0.04 0.07 0.06 0.04 0.06 0.03 0.05 0.27 0.05 0.07 0.07 0.10 0.06 0.04 0.08

0.91 0.40 0.09 0.12 0.07 0.08 0.09 0.11 0.07 0.38 0.08 0.12 0.13 0.15 0.13 0.02 0.14

0.82 0.32 0.04 0.10 0.07 0.06 0.08 0.05 0.04 0.31 0.05 0.09 0.05 0.12 0.10 0.03 0.08

4.1. US results To begin with, we discuss the average news spillover effects within the US market. In Table 1 we present the cross-industry mean relevance of the news over the full time period under consideration. For example, the number 0.09 in the first column and the eleventh row means that YCY news is Granger causing the excess returns of 9 companies from 100 in the financial sector on average. It is worth noting that the stocks are mostly self-connected, i.e. the own lags of the excess returns of the stocks are chosen to be relevant for the future excess returns by the adaptive lasso procedure (see the numbers in the first row). From Table 1, one can observe that the relevance of the country-related news spreads almost uniformly among the sectors. There is evidence of a bigger influence of financial and energy news for all sectors. It is worth mentioning that the results of Table 1 should be interpreted with caution, as the average is taken over a long period of time. The time interval of almost 10 years might contain several structural breakpoints in the relevance of the news. The dynamics of the relevance over time becomes clearer from Fig. 7 where the rolling window relevance is presented for the selected countries and industries. It can be concluded that, in general, the relevance is not constant and has a fluctuating behavior. For some sectors, several breakouts can be observed. The relevance of the own lags of the stock drops during the period of the global financial crisis. This finding coincides with the results shown by Audrino and Knaus (2016) in the context of the HAR volatility model. However, the relevance of the S&P 500 market-wide sentiment increases considerably during the crisis, showing that financial markets are more connected during periods characterized by market turbulence and that the investors’ moods about the broad market conditions become one of the most important drivers of the network dynamics. Similarly, the relevance of the country-specific news shows a fluctuating behavior. The exception is the growing relevance of the US-related news around 2008 and the higher relevance of the Germany-related news around 2008–2010 and 2013. These results support the common belief of the global leading role played by the US information flow during the financial crisis and by the Germany-related news (as the leading European country) during the financial and attributed European sovereign debt crises. Similarly, Bekaert et al. (2011) find evidence of contagion from US markets and from the global financial sector during the recent global financial crisis. China’s increase in relevance in 2012 can be due to its abnormal economic growth and increasing role in the world’s economy.

In contrast to the analysis of relevance, a closer look at the strength of the sector-related news gives better insight into the cross-industry news spillover effects. The mean cross-industry strength of the 1, 5 and 22 lags of the news sentiments is presented in Appendix E (Table 9). The results clearly show that the news of Financials and Energy have the strongest causal influence on the excess returns in all sectors. In the same vein, Arouri et al. (2011) discuss the existence of significant volatility spillover between oil prices and sector stock returns in both the US and Europe. This view is supported by Ewing and Malik (2016) who find volatility spillovers between oil prices and US stock market. Another study by Algieri and Leccadito (2017) reports that a distress in the oil market has the largest negative consequences for the rest of the economy and explains this finding by the fact that the sample period includes the global financial crisis which caused a significant decrease in oil demand. In the same vein, Wen et al. (2012) show the existence of contagion phenomenon between crude oil and stock prices. Such a high strength of the financial news could be explained by the fact that the global financial crisis and the European sovereign debt crisis are part of our sample period. As pointed out by Tamakoshi and Hamori (2016) for the UK market, the banking sector was a dominant transmitter of shocks during the financial crisis. However, other financial companies also transmitted shocks for some periods, highlighting the importance of regulation in the financial sector. Similarly, Chunxia et al. (2016) find that the financial sector always has a large output transfer entropy to the nondaily consumption, the energy, the raw material, and the industrial sectors. Interestingly, the 5- and 22-day lag coefficients related to the stock itself are considerably smaller than the 1-day lag coefficients. This is not true for the lags of the sentiment data. Therefore, it can be concluded that sentiment data can provide more timeous information for the prediction of the excess returns. This finding supports the result that the past values of the news contain additional information about the future stock returns; moreover, the importance of the news rises during financial turbulence and macroeconomic recessions. To get a deeper understanding, we present the rolling window results averaged over all stocks in Fig. 8 for one-day lag variables: results for other lags are similar. The strength of the financial news starts rising in 2007, showing that the growing influence of the news can be considered an early warning of future instability. The increase of strength coincides with the first phase of the crisis (according to Baur (2012)) which is described as initial financial turmoil and allows to predict the next phase of the crisis described

550

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Fig. 7. The overall relevance of the news sentiments of the selected sectors and countries on the US stocks ranging from Aug. 1, 2006 to Dec. 31, 2014.

as sharp financial market deterioration. Higher average strength of YCY news is observed during the first half of the sample and decreases thereafter. This might be explained by the fact that the aggregated consumption of US households started to grow rapidly at the end of 2005, went through recession in 2008, and continued to increase in 2009. Consumption has shown stable upward trends since 2009; therefore, the news on the YCY sector is less important in the second half of the period. Moreover, it is observed that the strength of Industrials grows right after the recession period; this sector contributed notably to the recovery process and gained additional attention of the media. Besides the financial sector, the energy sector was closely related to other sectors during the global financial crisis. This can be explained by the fact that energy consumption is linked to build-

ings and the residential sector. Monthly declines of 10 percent in electricity and gas consumption compared to the same periods in 2008 were observed in 2009. In early 2010, as many industrial sectors began to recover, there was a small growth in consumption compared to the same period in 2009. Therefore, the strength of this sector decreased in 2010. However, since the recovery was limited, the influnce of this sector remained remarkable. S&P 500 market-wide sentiment was spreading the spillovers to other sectors during the global financial crisis. As mentioned above, the investors were mostly concerned about the market-wide conditions during the market turbulence caused by the financial sector. The rapid increase in the strength of S&P 500 market-wide sentiment in 2012 can be explained by the downgrade of the US rating by Standard& Poor’s for the first time in the history. The

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

551

Fig. 8. The overall strength of the 1 day lags of the news sentiments on the selected US sectors and countries for the US companies ranging from Aug. 1, 2006 to Dec. 31, 2014.

long-term rating decrease was caused by the concerns about budget deficits. It is important to note that this period of time coincides with the increasing relevance of China-related news when the imbalances between the big creditor (China) and big debtors (US) became more severe. Fig. 8 depicts the strength of selected countries, and it appears that the US news shows relatively high overall strength and relativity. On the other hand, short peaks with low persistence in the strength of the news related to Germany are observed. The fluctuations in Germany’s strength can be explained by the global financial crisis of 2008, the European debt crisis of 2010 and the decision of Germany in 2012 to provide more support to its European partners for more centralized control over the Euro-

zone. The results for other sectors and countries are presented in Appendix F (Figs. 14 and 15). A detailed analysis of the strength by sector can be obtained from the authors upon request. As an additional robustness check, the mean values of the relevance and strength of all sectors after controlling for the VIX influence are calculated and presented in Appendix D (Tables 7 and 8). These results confirm the previous finding that the biggest spillover effects originate in the financial and energy sectors. The main difference in the VIX augmented model is that the adaptive lasso procedure selects the lags of the assets’ returns less frequently, while VIX lags are relevant for 85% of the assets on average. However, due to the small values of the estimated coefficients, the overall strength-related results remain unchanged.

552

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567 Table 2 Mean relevance of the lags of the news sentiments of the European sectors and selected countries for the European companies by sector ranging from Aug. 1, 2006 to Dec. 31, 2014.

asset MPTRXFIN MPTRXTEC MPTRXIND MPTRXMAT MPTRXUTL MPTRXCOM MPTRXHLC MPTRXNCY MPTRXENE MPTRXYCY US IT GR DE CN MPTRXEU50

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.95 0.24 0.12 0.16 0.09 0.12 0.08 0.10 0.11 0.25 0.13 0.24 0.15 0.15 0.21 0.07 0.17

0.95 0.38 0.14 0.24 0.14 0.16 0.08 0.16 0.11 0.26 0.13 0.32 0.24 0.19 0.26 0.09 0.19

0.73 0.31 0.15 0.20 0.14 0.16 0.09 0.14 0.13 0.25 0.13 0.29 0.19 0.15 0.23 0.12 0.21

0.84 0.29 0.09 0.17 0.12 0.14 0.07 0.10 0.09 0.22 0.11 0.24 0.19 0.17 0.18 0.08 0.17

0.93 0.41 0.24 0.29 0.23 0.25 0.17 0.25 0.20 0.40 0.26 0.43 0.30 0.29 0.24 0.16 0.30

0.90 0.36 0.18 0.19 0.14 0.18 0.13 0.15 0.12 0.30 0.17 0.31 0.26 0.22 0.24 0.11 0.26

0.95 0.31 0.15 0.22 0.12 0.20 0.07 0.15 0.10 0.29 0.14 0.27 0.22 0.17 0.15 0.09 0.22

0.95 0.28 0.08 0.13 0.07 0.07 0.04 0.08 0.07 0.20 0.08 0.20 0.13 0.08 0.13 0.04 0.14

0.95 0.32 0.11 0.22 0.11 0.16 0.08 0.11 0.11 0.27 0.12 0.32 0.18 0.20 0.21 0.10 0.21

0.89 0.32 0.13 0.19 0.12 0.16 0.09 0.13 0.14 0.26 0.10 0.29 0.16 0.15 0.16 0.10 0.17

4.2. EU results Comparing the news spillover effects in the European market to the US market, we observe the similarity in terms of the relevance of the news. However, the relevance of the financial and energy sectors is smaller in comparison to the US market. The strength of the news sentiments is less strong on average; see the results summarized in Table 2. The fluctuating behavior of the average relevance of some sectors for all stocks is shown in Fig. 9. The increase of the relevance of the Euro Stoxx 50 sentiment can be related to the global financial crisis and the European debt crisis. The second peak in the relevance can be related to the concerns that the Greek government would be unable to deal with the fiscal crisis without outside help. The relevance decreased significantly after Greece got a loan co-financed by the European Union and the International Monetary Fund. This view is supported by Bhanot et al. (2014) who find that the changes in Greek sovereign yield spreads caused abnormal returns of stocks for a sample of Eurozone countries during the Greek debt crisis. The analysis of the strength of the sectors in terms of the average absolute lasso estimated coefficients coincides with the results for the US market. The news related to Financials and Energy seems to be important for all other sectors; see Appendix E (Table 10). Similarly to the US market, the strength of the financial news rises before the crisis and reaches a maximum during the global financial crisis, see Fig. 10. The results for other sectors and countries are presented in Appendix F (Figs. 16 and 17). Full detailed results of the analysis can be obtained from the authors upon request. It can be concluded that the news sentiments related to several important sectors provide additional information about the future stock prices. For each market, several important sectors can be defined. The news spillover effects coming from these influential sectors seem to be at least as important as the direct effects. Moreover, the relevance of these spillover effects increases during periods characterized by general economic instability and/or financial market turbulence. 5. An empirical illustration The results of the analysis presented in the previous section show that the sentiment data of several sectors have a significant information component, which can be used to improve the prediction of the assets’ returns. This section illustrates the impact of the various sentiment series on the conditional mean and the conditional variance equations of ARMA-GARCH models. It is important

to note that the purpose of this section is to provide an insight into possible practical applications of the results presented earlier rather than draw general conclusions about sentiment-augmented time series models. Further on, a parsimonious ARMA(1,1)-GARCH(1,1) model with Gaussian innovations is applied to the series of log returns of several firms. The choice of the model is due to the fact that lower order GARCH models are used in most applications. Moreover, this model has been selected according to the Akaike information criterion in the majority of cases. This benchmark model is compared to extensions in which the lag of the sentiment index is included in the model as an exogenous variable. The considered model has the following specification:

Yt = α0 + α1Yt−1 + β1 εt−1 + γ1 St−1 + et , 2 2 Var (et |Ft−1 ) = a0 + a1 σt−1 + b1 et−1 + g1 St−1 ,

(9)

where et = σt εt , with ε t being a sequence of iid random variables with mean 0 and variance 1, Yt is the series of the log returns, St−1 is the sentiment index of the pre-defined sector, t = 1, . . . , T , and α 0 , α 1 , β 1 , γ 1 , a0 , a1 , b1 , g1 are the parameters to be estimated. For further details on time series models, we refer to Tsay (2005), Andersen et al. (2009), Francq and Zakoian (2011) and original work on the GARCH model by Bollerslev (1986). The specification (9) results in three competing models. If γ1 = 0 and g1 = 0, the ARMA(1,1)-GARCH(1,1) model without exogenous sentiment variables is considered. If g1 = 0, the sentiment variable is included in the conditional mean equation. If both γ 1 and g1 need to be estimated, the sentiment index is included in the conditional mean and the conditional variance equations. The analysis performed in the previous sections has shown that the sentiment spillover effects dominate the direct effect during periods of financial turbulence. It has been empirically shown that the sentiment data of the financial and energy sectors drive the market during certain periods. In order to validate this conclusion, we separately estimate the models with the firm-specific sector’s sentiment and the sentiment data coming from the financial and energy sectors. After estimating the models based on 200-day rolling windows and obtaining the one-day-ahead forecast series, the models are compared in terms of their predictive power. Their predictive power is expressed in terms of the mean squared prediction error (MSE). The two-sided test of Diebold and Mariano (1995) (DM test) is applied to the squared errors to check whether the model provides a statistically significant improvement in comparison to the ARMA(1,1)-GARCH(1,1), which is always used as the benchmark.

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

553

Fig. 9. The overall relevance of the news sentiments of the selected sectors and countries on the European stocks ranging from Aug. 1, 2006 to Dec. 31, 2014.

Four companies, belonging to two industrial sectors (Basic Materials and Healthcare), are used for the illustrative purposes of this section: International Flavors & Fragrances Inc. (IFF), Williams Companies Inc. (WMB), Chevron Corporation (CVX), and Cigna Corporation (CI). The selection of the time intervals is motivated by Fig. 11. It is evident that the energy news has a strong effect on the companies related to Basic Materials in 2007 and 2010 and is less important for the returns of this sector in 2006 and 2012. The financial news is of highest importance in 2008–2012 and is less influential at the end of the period (2013). In contrast, Basic Materials-related sentiment has only minor influence on the excess returns of the companies belonging to this sector. The same empirical evidence is observed for the companies related to the healthcare sector.

The mean squared prediction errors and the p-values of the two-sided DM test for the log returns of all considered companies are presented in Table 3. The comparison is always made with respect to the ARMA(1,1)-GARCH(1,1) model, i.e. model (9) with g1 = γ1 = 0. It is evident that including the exogenous variable of financial news in the conditional mean equation improves the predictive power of the ARMA-GARCH model for IFF and WMB in 20 08–20 09, which corresponds to the period of strong influence of the financial sector. An improvement is not observed during the year before and the year after the abovementioned period. In the majority of the analyzed cases, exogenous news indices do not provide any further improvement of the predictive power of the model. Including sector-related news (Basic Materials) in the model does not significantly im-

554

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567 Table 3 MSE of the sentiment-augmented time series models and the corresponding p-values (DM test compared to the ARMA(1,1)GARCH(1,1)). MSE IFF (Basic Materials)

γ1 = 0 , g 1 = 0 g1 = 0, S = MPTRXUSFIN S = MPTRXUSFIN g1 = 0, S = MPTRXUSMAT S = MPTRXUSMAT WMB (Basic Materials)

γ1 = 0 , g 1 = 0

g1 = 0, S = MPTRXUSFIN S = MPTRXUSFIN g1 = 0, S = MPTRXUSMAT S = MPTRXUSMAT CVX (Basic Materials)

γ1 = 0 , g 1 = 0

g1 = 0, S = MPTRXUSENE S = MPTRXUSENE g1 = 0, S = MPTRXUSMAT S = MPTRXUSMAT CI (Healthcare)

γ1 = 0 , g 1 = 0 g1 = 0, S = MPTRXUSENE S = MPTRXUSENE g1 = 0, S = MPTRXUSHLC S = MPTRXUSHLC

DM p-value

1.06.2007-1.02.2008 1.383 · 10−2 2.789 · 10−2 1.305 · 10−2 1.440 · 10−2 1.305 · 10−2

3.149 · 10−1 3.198 · 10−1 7.238 · 10−1 9.818 · 10−2

1.05.2007-1.01.2008 2.310 3.058 2.361 2.314 4.033

3.204 · 10−1 2.737 · 10−1 7.096 · 10−1 3.096 · 10−1

1.04.2006-1.12.2006 1.550 · 10+1 1.386 · 10+1 1.404 · 10+1 1.358 · 10+1 1.368 · 10+1

3.760 · 10−1 4.600 · 10−1 3.015 · 10−1 3.397 · 10−1

1.01.2006-30.06.2006 −3

4.147 · 10 4.167 · 10−3 4.301 · 10−3 4.346 · 10−3 4.244 · 10−3

MSE

DM p-value

1.06.2008-1.02.2009 1.381 · 10−3 1.287 · 10−3 1.291 · 10−3 1.349 · 10−3 1.354 · 10−3

8.110 · 10−3 1.106 · 10−2 2.064 · 10−1 2.507 · 10−1

1.05.2008-1.01.2009 1.183 · 10−3 1.151 · 10−3 1.160 · 10−3 1.173 · 10−3 1.168 · 10−3

7.508 · 10−3 3.660 · 10−2 4.516 · 10−1 3.529 · 10−1

prove the predictive power when the financial news is driving the market. The results for the conditional variance equation are not unambiguous and further research should address to this question. For example, including financial sentiment data in the conditional variance equation of IFF in late 2007 and early 2008 improves the forecasting accuracy of the model in comparison to the model when only the conditional mean equation is augmented by the financial sentiment index. However, this improvement is not significant in comparison to the ARMA(1,1)-GARCH(1,1) model. In contrast, including basic materials-related sentiment in the conditional variance equation for WMB in the second half of 2007 almost doubles the mean squared prediction error of the model. This increase is insignificant in comparison to the ARMA(1,1)-GARCH(1,1) model according to a two-sided DM test. Similar results are observed for CVX and CI; see Table 3. It is evident that including energy-related news in the conditional mean equation improves the predictive power of the ARMA-GARCH model when the energy news has significant influence on the stock returns. The conditional variance-related results differ from the results for IFF and WBM. The MSE remains almost unchanged when the sentiment data are included in the conditional variance model. The results for the considered firms support the conclusion of the previous section and suggest that sentiment spillover effects dominate the direct effects in periods when the market is driven by the news coming from a different sector. However, one should be careful when using these results in practical applications. The sentiment data can improve the forecasting power of the model or increase the mean squared prediction error. Each case should be considered individually. This empirical evidence is in line with the general discussion by Farmer et al. (2019), who claim that stock returns exhibit short periods with significant predictibility followed by long periods with little or no evidence of predictibility. The general methodology presented in this paper suggests a way of detecting these “pockets of predictibility” in sentiment-augmented time series models. Therefore, the time series models augmented by including the news of influential sectors might show better predictive power for

DM p-value

1.06.2009-1.02.2010 7.238 · 10−2 7.103 · 10−2 7.265 · 10−2 7.384 · 10−2 7.279 · 10−2

5.012 · 10−1 8.810 · 10−1 6.498 · 10−1 9.050 · 10−1

1.05.2009-1.01.2010 1.020 · 10−1 1.027 · 10−1 1.023 · 10−1 1.019 · 10−1 1.032 · 10−1

2.755 · 10−1 3.030 · 10−1 8.642 · 10−1 3.656 · 10−1

1.04.2007-1.12.2007

1.04.2008-1.12.2008

3.227 1.678 1.964 3.221 3.219

9.998 · 10−3 1.002 · 10−2 1.002 · 10−2 9.994 · 10−3 9.892 · 10−3

1.591 · 10−4 1.503 · 10−2 8.506 · 10−1 7.991 · 10−1

1.07.2006-31.12.2006 −2

9.371 · 10−1 5.245 · 10−1 3.059 · 10−2 2.700 · 10−1

MSE

8.248 · 10 7.655 · 10−2 7.595 · 10−2 9.444 · 10−2 9.684 · 10−2

6.031 · 10−6 4.268 · 10−7 1.844 · 10−2 1.545 · 10−2

1.167 · 10−1 1.399 · 10−1 9.372 · 10−1 1.693 · 10−1

1.01.2007-30.06.2007 1.227 · 10−3 1.200 · 10−3 1.265 · 10−2 1.225 · 10−3 1.213 · 10−3

3.290 · 10−1 3.220 · 10−1 9.119 · 10−1 4.503 · 10−1

returns. The periods when the sentiment indices of particular industries are more informative can be found by applying the procedure described in Section 3. The illustrative examples discussed here support the results of Section 4, which suggests the importance of the information contained in sentiment data for asset pricing theory. Conclusion The goal of this paper is to investigate the cross-industry patterns of the news and stock returns, and in particular to analyze how the news about one industry influences the stock returns in the other industries. For this purpose, the graphical Granger model has been applied to sentiment data on 10 US and 10 non-US industries and on the excess returns of 78 US and 78 European companies. The sentiment data on several countries and market indices have been included in the analysis to control for country-wide and market-wide general sentiment effects. The adaptive lasso procedure has been applied to estimate the return-news networks. Network-based measures reflecting the relevance and strength of each news source have been proposed and analyzed over a period of 10 years by employing a rolling window approach. We found empirical evidence that the relevance of the news coming from different sectors shows a fluctuating behavior and spreads evenly among the industries. Moreover, our results show that the strength of the influence of the news on some sectors grows just before periods of economic and financial instability and reaches a maximum during crises. Interesting patterns are observed in the causality of the financial and energy sentiments. These sentiments can be seen as the most influential, and the spillover effects from the sectors dominate the direct effects. Estimation results show that the overall connectedness among the stock returns and the news is stronger for the US market than for the European market. The importance of sentiment spillover effects has been empirically illustrated. Unavoidably, some spillovers in terms of returns and volatilities across companies are present in our industry sentiment indices. However, disentangling the part of the industry sentiment

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

555

Fig. 10. The overall strength of the 1 day lags of the news sentiments on the selected European sectors and countries for the European companies ranging from Aug. 1, 2006 to Dec. 31, 2014.

spillover effects that is due to price dynamics, e.g. returns, volatility, and liquidity from the part that is due only to investors’ moods is difficult. Some recent papers argue that business activities tend to follow, rather then lead, social mood. Following this reasoning, the TRMI indices we use in our analysis extract sentiment directly from news and social media posts, which are expected to at least partially anticipate investors’ actions. Therefore, what we can fairly say is that most of the spillover effects we find are driven by investors’ beliefs, as when investors become more optimistic about the general economic outlook or perceive an ongoing overvaluation of a stock in the market, generating a more pessimistic outlook, at least for the 1-lag results. In fact, according to this strand of the literature, only one period later one will observe the correspond-

ing changes in the price dynamics and the possible spillover effects due to investors’ trades. It was not the goal of the current paper to enter into this debate, but the question is important and could be a topic for a future analysis. In future research we also plan to relax the assumption of predefined lags in the graphical Granger model and to apply the same adaptive lasso methodology to test the significance of arbitrary lags. This could be especially interesting to show the persistence of the news coming from different sectors. Moreover, the study of the nonlinear or quantile cross-industry dependencies among the news could yield further insights into the analysis of the impact of direct sentiment effects as well as spillover sentiment effects on companies’ excess returns.

556

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Fig. 11. The strength of 1 day lags of the news sentiments of the FIN and ENE sectors for the MAT and HLC US companies ranging from Jan. 1 2005 to Dec. 31, 2014.

Appendix A. TRMI sentiment indices Table 4 Thomson Reuters MarketPsych sentiment indices used for the analysis. Asset Code

Description

MPTRXUSENE/MPTRXENE MPTRXUSMAT/MPTRXMAT MPTRXUSIND/MPTRXIND MPTRXUSYCY/MPTRXYCY

US/non-US Energy (ENE) US/non-US Basic Materials (MAT) US/non-US Industrials (IND) US/non-US Cyclical Consumer Goods and Services (YCY) US/non-US Non-Cyclical Consumer Goods and Services (NCY) US/non-US Financials (FIN) US/non-US Healthcare (HLC) US/non-US Technology (TEC) US/non-US Telecommunications Services (COM) US/non-US Utilities (UTL) S&P 500 Euro Stoxx 50 China Germany Greece Italy USA

MPTRXUSNCY/MPTRXNCY MPTRXUSFIN/MPTRXFIN MPTRXUSHLC/MPTRXHLC MPTRXUSTEC/MPTRXTEC MPTRXUSCOM/MPTRXCOM MPTRXUSUTL/MPTRXUTL MPTRXUS500 MPTRXEU50 CN DE GR IT US

Appendix B. State estimation with Kalman filter The real unobserved sentiment μt is extracted from the noisy news yt by applying the Local Level model by Durbin and Koopman (2001). The starting point of the model is the following system of equations:

μt + εt ,εt ∼ N(0, σε2 ), = μt + ηt ,ηt ∼ N(0, ση2 ).

yt =

μt+1

(10)

In (10), the first equation corresponds to the noisy observed news yt , whereas the second equation is the signal equation which corresponds to the unobserved sentiment μt+1 , t = 1, . . . , T . In the sentiment literature, it is assumed that economic agents are at

least as intelligent as animals and aggregate the news as a function of time decay. This fact is motivated by the findings in the behavioral literature; see Nickerson (2011) and Reilly et al. (2012), for example. The noisy observed news yt and the unobserved sentiment μt+1 , t = 1, . . . , T are described by the Local Level model by Durbin and Koopman (2001):

yt = μt + εt ,εt ∼ N(0, σε2 ),

μt+1 = μt + ηt ,ηt ∼ N(0, ση2 ).

The state moments μ ˜ t = E (μt |Ft−1 ) and Pt = Var (μt |Ft−1 ) are computed recursively by solving the following equations:

vt = yt − μ˜ t , Ft = Pt + σε2 , Kt =

Pt , Ft

μ˜ t+1 = μ˜ t + Kt vt , Pt+1 = Pt (1 − Kt ) + ση2 , μ˜ 1 = μ1 ,P1 = e7 .

(11)

Thus, the unobserved state μt is updated each time a new noisy observation yt arrives. The state at time t is calculated by exponentially weighting the previous states. In the next step, the states are estimated by applying the Kalman smoother and solving the following backward recursion equations:

t = μ μ ˜ t + Pt rt−1 vt

+ Lt rt ; Nt−1 = Ft−1 + Lt2 Nt , Ft Lt = 1 − Kt ,

rt−1 =

Vt = Pt − Pt2 Nt−1 ,

(12)

t |D ) is the conditional density where t = 1, . . . , T , rT = 0 and N (μ t = E (μt |Ft−1 ) and Vt = Var (μt |Ft−1 ). For the techof μt with μ nical details of the estimation procedure we refer to Durbin and Koopman (2001).

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Appendix C. List of the companies

Table 5 The US companies used in the current studies. Name

TRBCEconomicSector

Bangkok Airways PCL Asset Acceptance Capital Corp Apple Inc Abcam PLC Abbott Laboratories AES Corp Amgen Inc Amazon.com Inc Anadarko Petroleum Corp Avery Dennison Corp American Water Works Company Inc American Express Co Boeing Co Bank of America Corp Baxter International Inc BlackRock Inc Ball Corp Citigroup Inc Caterpillar Inc Cigna Corp Colgate-Palmolive Co CMS Energy Corp CVS Caremark Corp Cisco Systems Inc CenturyLink Inc Chevron Corp Dow Chemical Co DTE Energy Co Devon Energy Corp Ford Motor Co Frontier Communications Corp Corning Inc General Motors Co Hasbro Inc Harley-Davidson Inc Hewlett-Packard Co International Business Machines Corp International Flavors & Fragrances Inc Johnson & Johnson JPMorgan Chase & Co Coca-Cola Co Eli Lilly and Co Masco Corp Mattel Inc McDonald’s Corp Metlife Inc Merck KGaA Altria Group Inc Merck & Co Inc Morgan Stanley Microsoft Corp Nike Inc National Oilwell Varco Inc Eversource Energy Nucor Corp Oracle Corp PG&E Corp PepsiCo Inc Pfizer Inc Procter & Gamble Co Praxair Inc Sealed Air Corp Schlumberger NV Southern Co AT&T Inc Target Corp Textron Inc United Parcel Service Inc Vulcan Materials Co Verizon Communications Inc Wells Fargo & Co Williams Companies Inc Wal Mart Stores Inc Xcel Energy Inc Exxon Mobil Corp Yahoo! Inc Cambridge Antibody Technology Group PLC Bell Aliant Inc

Industrials Industrials Technology Healthcare Healthcare Utilities Healthcare Consumer Cyclicals Energy Industrials Utilities Financials Industrials Financials Healthcare Financials Basic Materials Financials Industrials Financials Consumer Non-cyclicals Utilities Consumer Non-cyclicals Technology Telecommunication Services Energy Basic Materials Utilities Energy Consumer Cyclicals Telecommunication Services Technology Consumer Cyclicals Consumer Cyclicals Consumer Cyclicals Technology Technology Consumer Non-cyclicals Healthcare Financials Consumer Non-cyclicals Healthcare Consumer Cyclicals Consumer Cyclicals Consumer Cyclicals Financials Healthcare Consumer Non-cyclicals Healthcare Financials Technology Consumer Cyclicals Energy Utilities Basic Materials Technology Utilities Consumer Non-cyclicals Healthcare Consumer Non-cyclicals Basic Materials Basic Materials Energy Utilities Telecommunication Services Consumer Cyclicals Industrials Industrials Basic Materials Telecommunication Services Financials Energy Consumer Cyclicals Utilities Energy Technology Healthcare Telecommunication Services

557

558

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Table 6 The European companies used in the current studies. Name

TRBCEconomicSector

Anglo American PLC Alcatel Lucent SA Antofagasta PLC Anglo Pacific Group PLC Daisy Group PLC Electricite de France SA Eni SpA Glencore PLC Iberdrola SA Industria de Diseno Textil SA Nestle SA Rio Tinto PLC Roche Holding AG SABMiller PLC Swisscom AG STMicroelectronics NV Sirius Minerals PLC Syngenta AG Telecity Group PLC Telenor ASA United Utilities Group PLC Vitec Group PLC Wolfson Microelectronics PLC Abengoa Yield PLC ABB India Ltd Aditya Birla Minerals Ltd ArcelorMittal SA AstraZeneca PLC Banco Bilbao Vizcaya Argentaria SA GlaxoSmithKline PLC Merck KGaA Novartis Banco Santander SA Total Energy Services Inc Anheuser-Busch Companies LLC HSBC Holdings Lloyds Banking Group BNP Paribas Allianz UBS Group Deutsche Bank Logitech International Infineon Technologies SAP SE Siemens Airbus Schneider Electric LINDE Vinci Glencore ThyssenKrupp BASF Anglo Pacific Group EOAN National Grid Enel SPA Engie SA Energie Baden-Wuettenberg DTE Energy Orange Sanofi Bayer Bosch Continental Man SE Peugeot Volkswagen Daimler Sie de Saint-Gobain BMW Royal Dutch Shell British Petroleum Unilever Sabmiller L’oreal Moet Hennessy Louis Vuitton SE Diageo PLC

Basic Materials Technology Basic Materials Basic Materials Telecommunication Services Utilities Energy Energy Utilities Consumer Cyclicals Consumer Non-cyclicals Basic Materials Healthcare Consumer Non-cyclicals Telecommunication Services Technology Basic Materials Basic Materials Technology Telecommunication Services Utilities Technology Technology Utilities Industrials Basic Materials Basic Materials Healthcare Financials Healthcare Healthcare Healthcare Financials Energy Consumer Non-cyclicals Financials Financials Financials Financials Financials Financials Technology Technology Technology Industrials Industrials Industrials Industrials Industrials Industrials Basic Materials Basic Materials Basic Materials Utilities Utilities Utilities Utilities Utilities Utilities Telecommunication Services Healthcare Healthcare Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Energy Energy Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals Consumer Non-cyclicals

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

559

Appendix D. Overall relevance and strength of sectors and countries for US companies controlling for the VIX index Table 7 Mean relevance of the news sentiments of the US sectors and selected countries for the US companies by sector ranging from Aug. 1, 2006 to Dec. 31, 2014 (controlling for the VIX index).

asset MPTRXUSFIN MPTRXUSTEC MPTRXUSIND MPTRXUSMAT MPTRXUSUTL MPTRXUSCOM MPTRXUSHLC MPTRXUSNCY MPTRXUSENE MPTRXUSYCY US IT GR DE CN VIX MPTRXUS500

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.84 0.50 0.14 0.31 0.14 0.12 0.12 0.10 0.12 0.51 0.16 0.23 0.21 0.23 0.18 0.08 0.83 0.26

0.87 0.49 0.11 0.18 0.14 0.09 0.14 0.09 0.11 0.47 0.11 0.17 0.13 0.20 0.17 0.08 0.89 0.19

0.87 0.55 0.14 0.24 0.13 0.10 0.11 0.12 0.11 0.51 0.14 0.17 0.19 0.21 0.18 0.09 0.84 0.24

0.86 0.48 0.15 0.23 0.14 0.11 0.15 0.12 0.12 0.54 0.17 0.19 0.22 0.23 0.18 0.09 0.85 0.23

0.80 0.56 0.18 0.24 0.19 0.17 0.14 0.14 0.14 0.49 0.16 0.19 0.23 0.22 0.24 0.12 0.84 0.26

0.84 0.57 0.14 0.26 0.17 0.15 0.19 0.16 0.19 0.57 0.16 0.24 0.23 0.31 0.21 0.13 0.81 0.27

0.88 0.48 0.12 0.24 0.15 0.17 0.14 0.13 0.13 0.53 0.14 0.18 0.18 0.24 0.22 0.09 0.84 0.26

0.78 0.41 0.08 0.17 0.10 0.07 0.12 0.08 0.09 0.41 0.10 0.12 0.15 0.15 0.12 0.07 0.91 0.15

0.85 0.54 0.15 0.25 0.18 0.15 0.16 0.19 0.17 0.56 0.15 0.25 0.25 0.26 0.21 0.09 0.81 0.25

0.77 0.44 0.10 0.19 0.12 0.11 0.12 0.10 0.08 0.44 0.09 0.16 0.14 0.21 0.15 0.08 0.88 0.18

Table 8 Mean strength of the lags of the news sentiments on the US sectors and selected countries for the US companies by sector ranging from Aug. 1, 2006 to Dec. 31, 2014 (controlling for the VIX index).

assett−1 assett−5 assett−22 MPTRXUSFINt−1 MPTRXUSFINt−5 MPTRXUSFINt−22 MPTRXUSTECt−1 MPTRXUSTECt−5 MPTRXUSTECt−22 MPTRXUSINDt−1 MPTRXUSINDt−5 MPTRXUSINDt−22 MPTRXUSMATt−1 MPTRXUSMATt−5 MPTRXUSMATt−22 MPTRXUSUTLt−1 MPTRXUSUTLt−5 MPTRXUSUTLt−22 MPTRXUSCOMt−1 MPTRXUSCOMt−5 MPTRXUSCOMt−22 MPTRXUSHLCt−1 MPTRXUSHLCt−5 MPTRXUSHLCt−22 MPTRXUSNCYt−1 MPTRXUSNCYt−5 MPTRXUSNCYt−22 MPTRXUSENEt−1 MPTRXUSENEt−5 MPTRXUSENEt−22 MPTRXUSYCYt−1 MPTRXUSYCYt−5 MPTRXUSYCYt−22 USt−1 USt−5 USt−22 ITt−1 ITt−5 ITt−22 GRt−1 GRt−5 GRt−22 DEt−1 DEt−5 DEt−22 CNt−1 CNt−5 CNt−22 VIXt−1 VIXt−5 VIXt−22 MPTRXUS500t−1 MPTRXUS500t−5 MPTRXUS500t−22

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.50 0.12 0.05 1.00 1.13 1.02 0.32 0.32 0.27 0.48 0.32 0.59 0.44 0.42 0.33 0.42 0.27 0.65 0.28 0.35 0.32 0.32 0.18 0.43 0.20 0.39 0.37 0.85 1.00 0.77 0.32 0.34 0.46 0.37 0.65 0.71 0.32 0.34 0.43 0.44 0.55 0.43 0.30 0.31 0.33 0.18 0.20 0.20 0.48 0.24 0.25 0.54 0.42 0.52

0.52 0.14 0.08 0.74 0.71 0.79 0.22 0.20 0.39 0.47 0.29 0.43 0.26 0.41 0.30 0.26 0.28 0.30 0.31 0.19 0.34 0.18 0.34 0.25 0.27 0.46 0.47 0.70 1.07 0.76 0.43 0.26 0.36 0.71 0.49 0.43 0.30 0.31 0.44 0.34 0.57 0.46 0.34 0.34 0.36 0.17 0.27 0.32 0.47 0.31 0.43 0.40 0.50 0.48

0.52 0.11 0.05 1.00 0.78 1.05 0.27 0.34 0.31 0.51 0.37 0.40 0.31 0.27 0.24 0.26 0.24 0.34 0.25 0.23 0.36 0.32 0.21 0.22 0.32 0.43 0.45 0.95 0.76 0.74 0.33 0.30 0.30 0.45 0.76 0.40 0.33 0.24 0.48 0.55 0.41 0.43 0.32 0.40 0.36 0.35 0.24 0.16 0.67 0.32 0.39 0.44 0.42 0.45

0.51 0.12 0.05 1.00 1.06 1.13 0.34 0.52 0.31 0.94 0.46 0.47 0.77 0.49 0.54 0.28 0.31 0.40 0.39 0.41 0.22 0.46 0.32 0.27 0.41 0.40 0.53 1.01 1.15 1.26 0.39 0.47 0.55 0.63 0.64 0.64 0.37 0.32 0.52 0.70 0.39 0.43 0.36 0.37 0.76 0.24 0.28 0.15 0.63 0.33 0.39 0.78 0.46 0.62

0.47 0.11 0.05 1.05 0.87 0.96 0.31 0.27 0.31 0.49 0.48 0.39 0.32 0.52 0.38 0.36 0.52 0.35 0.40 0.18 0.44 0.45 0.57 0.24 0.28 0.29 0.28 0.89 1.00 0.71 0.28 0.47 0.41 0.49 0.73 0.49 0.37 0.32 0.50 0.42 0.59 0.44 0.33 0.41 0.36 0.24 0.25 0.34 0.40 0.38 0.40 0.52 0.54 0.60

0.48 0.10 0.06 1.56 1.11 1.05 0.38 0.25 0.36 0.46 0.59 0.44 0.38 0.33 0.61 0.49 0.85 0.37 0.29 0.28 0.34 0.48 0.34 0.32 0.34 0.30 0.30 0.76 1.32 0.96 0.36 0.37 0.57 0.45 0.97 0.65 0.35 0.40 0.56 0.61 0.59 0.47 0.44 0.58 0.25 0.32 0.73 0.47 0.73 0.51 0.49 0.53 0.42 0.51

0.50 0.12 0.06 0.87 0.64 0.73 0.15 0.13 0.25 0.39 0.24 0.33 0.30 0.24 0.25 0.32 0.30 0.28 0.21 0.16 0.19 0.23 0.19 0.25 0.17 0.20 0.20 0.67 0.82 0.69 0.24 0.20 0.33 0.37 0.36 0.22 0.29 0.20 0.27 0.29 0.39 0.27 0.24 0.30 0.25 0.17 0.13 0.25 0.35 0.33 0.34 0.37 0.37 0.31

0.53 0.10 0.05 0.81 1.09 1.02 0.43 0.43 0.25 0.42 0.25 0.39 0.37 0.51 0.35 0.25 0.34 0.30 0.29 0.26 0.32 0.27 0.27 0.23 0.43 0.37 0.29 0.66 0.77 0.72 0.45 0.24 0.26 0.82 0.41 0.51 0.39 0.34 0.46 0.43 0.34 0.31 0.33 0.33 0.44 0.15 0.33 0.44 0.47 0.33 0.31 0.54 0.27 0.50

0.48 0.09 0.06 0.90 0.98 1.06 0.22 0.37 0.31 0.43 0.34 0.44 0.23 0.33 0.27 0.28 0.31 0.37 0.27 0.28 0.27 0.53 0.27 0.39 0.18 0.35 0.36 0.98 1.02 0.81 0.29 0.27 0.26 0.48 0.59 0.56 0.39 0.33 0.50 0.45 0.38 0.61 0.31 0.31 0.38 0.26 0.20 0.17 0.62 0.33 0.47 0.43 0.43 0.51

0.50 0.10 0.05 0.78 0.55 0.79 0.19 0.16 0.25 0.20 0.31 0.26 0.25 0.89 0.17 0.22 0.43 0.26 0.31 0.23 0.20 0.13 0.15 0.34 0.22 0.22 0.13 0.63 0.51 0.41 0.16 0.20 0.24 0.19 0.47 0.39 0.21 0.22 0.27 0.41 0.44 0.23 0.27 0.56 0.23 0.17 0.11 0.14 0.29 0.36 0.51 0.25 0.29 0.27

560

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Fig. 12. The overall relevance of the news sentiments of sectors and countries on the US stocks ranging from Aug. 1, 2006 to Dec. 31, 2014 (controlling for the VIX index).

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

561

Fig. 13. The overall strength of the news sentiments of sectors and countries on the US stocks ranging from Aug. 1, 2006 to Dec. 31, 2014 (controlling for the VIX index).

562

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Appendix E. Mean strength of the lags of the news sentiments for US and European companies

Table 9 Mean strength of the lags of the news sentiments on the US sectors and selected countries for the US companies by sector ranging from Aug. 1, 2006 to Dec. 31, 2014.

assett−1 assett−5 assett−22 MPTRXUSFINt−1 MPTRXUSFINt−5 MPTRXUSFINt−22 MPTRXUSTECt−1 MPTRXUSTECt−5 MPTRXUSTECt−22 MPTRXUSINDt−1 MPTRXUSINDt−5 MPTRXUSINDt−22 MPTRXUSMATt−1 MPTRXUSMATt−5 MPTRXUSMATt−22 MPTRXUSUTLt−1 MPTRXUSUTLt−5 MPTRXUSUTLt−22 MPTRXUSCOMt−1 MPTRXUSCOMt−5 MPTRXUSCOMt−22 MPTRXUSHLCt−1 MPTRXUSHLCt−5 MPTRXUSHLCt−22 MPTRXUSNCYt−1 MPTRXUSNCYt−5 MPTRXUSNCYt−22 MPTRXUSENEt−1 MPTRXUSENEt−5 MPTRXUSENEt−22 MPTRXUSYCYt−1 MPTRXUSYCYt−5 MPTRXUSYCYt−22 USt−1 USt−5 USt−22 ITt−1 ITt−5 ITt−22 DEt−1 DEt−5 DEt−22 CNt−1 CNt−5 CNt−22 GRt−1 GRt−5 GRt−22 MPTRXUS500t−1 MPTRXUS500t−5 MPTRXUS500t−22

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.73 0.12 0.45 1.12 1.13 1.07 0.31 0.32 0.31 0.45 0.36 0.72 0.54 0.63 0.35 0.24 0.21 0.76 0.29 0.31 0.40 0.37 0.17 0.66 0.24 0.31 0.33 0.76 1.19 0.72 0.50 0.62 0.46 0.44 0.97 1.14 0.35 0.38 0.52 0.24 0.32 0.22 0.23 0.21 0.15 0.65 0.53 0.43 0.70 0.09 0.61

0.75 0.14 0.34 0.71 0.62 0.82 0.22 0.39 0.47 0.53 0.23 0.30 0.33 0.33 0.20 0.23 0.33 0.38 0.21 0.15 0.30 0.29 0.23 0.12 0.35 0.56 0.78 0.71 1.15 0.51 0.42 0.37 0.34 0.49 0.51 0.53 0.46 0.34 0.59 0.29 0.23 0.38 0.09 0.42 0.41 0.52 0.81 0.41 0.26 0.07 0.43

0.76 0.12 0.41 1.09 0.64 1.15 0.32 0.34 0.54 0.45 0.48 0.35 0.31 0.15 0.31 0.31 0.11 0.51 0.11 0.20 0.42 0.26 0.17 0.22 0.34 0.33 0.44 0.84 0.62 0.87 0.47 0.38 0.26 0.27 1.10 0.38 0.38 0.16 0.73 0.42 0.31 0.45 0.52 0.22 0.33 0.73 0.43 0.44 0.43 0.05 0.53

0.75 0.16 0.49 0.87 0.85 1.11 0.39 0.35 0.34 0.59 0.42 0.42 0.66 0.57 0.57 0.30 0.38 0.40 0.38 0.74 0.16 0.68 0.20 0.50 0.37 0.48 0.42 0.86 0.82 1.09 0.62 0.53 0.58 0.69 0.58 0.65 0.41 0.28 0.41 0.39 0.29 1.00 0.29 0.63 0.07 0.57 0.30 0.36 0.76 0.05 0.72

0.70 0.09 0.43 1.22 0.93 1.14 0.42 0.36 0.41 0.68 0.47 0.63 0.47 0.62 0.46 0.59 0.70 0.45 0.63 0.32 0.58 0.56 0.66 0.34 0.44 0.50 0.48 0.94 1.18 0.97 0.29 0.51 0.43 0.59 0.98 0.63 0.38 0.36 0.75 0.44 0.45 0.44 0.33 0.35 0.35 0.46 0.67 0.69 0.73 0.09 0.74

0.72 0.11 0.35 1.54 0.90 1.00 0.44 0.20 0.44 0.54 0.76 1.12 0.43 0.33 0.46 0.69 1.22 0.40 0.34 0.42 0.29 0.51 0.49 0.26 0.39 0.38 0.57 0.65 1.49 0.96 0.37 0.74 0.89 0.72 1.32 1.08 0.28 0.33 0.75 1.12 0.46 0.22 0.36 1.68 0.46 0.67 0.71 0.36 0.59 0.06 0.81

0.74 0.14 0.43 0.85 0.54 0.85 0.12 0.15 0.22 0.37 0.25 0.27 0.29 0.31 0.37 0.30 0.20 0.27 0.20 0.18 0.20 0.23 0.18 0.34 0.19 0.16 0.22 0.57 0.80 0.69 0.33 0.29 0.22 0.30 0.42 0.29 0.25 0.19 0.37 0.26 0.23 0.24 0.17 0.16 0.27 0.32 0.46 0.38 0.40 0.03 0.31

0.76 0.11 0.44 0.71 0.90 1.41 0.50 0.52 0.31 0.64 0.22 0.52 0.28 0.37 0.28 0.30 0.37 0.16 0.19 0.32 0.36 0.31 0.45 0.23 0.48 0.34 0.29 0.56 1.03 0.67 0.39 0.29 0.31 0.89 0.44 0.60 0.31 0.34 0.42 0.53 0.27 0.47 0.17 0.36 0.65 0.49 0.40 0.62 0.39 0.03 0.65

0.74 0.06 0.39 0.97 0.95 1.31 0.26 0.58 0.39 0.49 0.30 0.22 0.42 0.24 0.23 0.38 0.46 0.37 0.44 0.30 0.33 0.56 0.20 0.25 0.23 0.62 0.37 0.83 0.86 0.78 0.23 0.28 0.35 0.54 0.78 0.84 0.45 0.43 0.37 0.30 0.22 0.68 0.17 0.11 0.13 0.38 0.55 0.69 0.42 0.05 0.45

0.74 0.10 0.38 0.81 0.46 0.73 0.17 0.27 0.26 0.15 0.39 0.27 0.22 1.64 0.28 0.17 0.33 0.30 0.21 0.27 0.12 0.14 0.13 0.39 0.30 0.18 0.10 0.65 0.43 0.40 0.19 0.18 0.30 0.20 0.55 0.40 0.17 0.25 0.18 0.33 0.64 0.22 0.13 0.28 0.20 0.32 0.51 0.20 0.36 0.03 0.29

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

563

Table 10 Mean strength of the lags of the news sentiments on the European sectors and selected countries for the European companies by sector ranging from Aug. 1, 2006 to Dec. 31, 2014.

assett−1 assett−5 assett−22 MPTRXFINt−1 MPTRXFINt−5 MPTRXFINt−22 MPTRXTECt−1 MPTRXTECt−5 MPTRXTECt−22 MPTRXINDt−1 MPTRXINDt−5 MPTRXINDt−22 MPTRXMATt−1 MPTRXMATt−5 MPTRXMATt−22 MPTRXUTLt−1 MPTRXUTLt−5 MPTRXUTLt−22 MPTRXCOMt−1 MPTRXCOMt−5 MPTRXCOMt−22 MPTRXHLCt−1 MPTRXHLCt−5 MPTRXHLCt−22 MPTRXNCYt−1 MPTRXNCYt−5 MPTRXNCYt−22 MPTRXENEt−1 MPTRXENEt−5 MPTRXENEt−22 MPTRXYCYt−1 MPTRXYCYt−5 MPTRXYCYt−22 USt−1 USt−5 USt−22 ITt−1 ITt−5 ITt−22 GRt−1 GRt−5 GRt−22 DEt−1 DEt−5 DEt−22 CNt−1 CNt−5 CNt−22 MPTRXEU50t−1 MPTRXEU50t−5 MPTRXEU50t−22

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.74 0.07 0.04 0.49 0.65 0.51 0.29 0.30 0.31 0.35 0.34 0.36 0.25 0.25 0.25 0.33 0.25 0.25 0.33 0.25 0.37 0.22 0.22 0.22 0.26 0.28 0.31 0.31 0.25 0.34 0.22 0.40 0.25 0.46 0.53 0.53 0.37 0.27 0.34 0.33 0.38 0.34 0.31 0.38 0.33 0.19 0.25 0.28 0.38 0.33 0.37

0.69 0.06 0.06 0.87 0.69 0.60 0.35 0.38 0.33 0.47 0.39 0.49 0.28 0.33 0.30 0.44 0.35 0.44 0.25 0.26 0.26 0.34 0.38 0.38 0.36 0.39 0.33 0.51 0.56 0.61 0.31 0.36 0.26 0.58 0.72 0.55 0.46 0.49 0.49 0.49 0.48 0.44 0.37 0.45 0.41 0.26 0.30 0.29 0.49 0.35 0.47

0.67 0.05 0.09 1.36 1.50 1.15 1.19 0.75 0.80 0.75 0.69 0.80 0.72 0.53 0.63 1.04 0.83 1.17 0.92 0.52 0.45 0.75 0.71 0.77 0.61 0.93 0.57 1.17 0.88 1.15 0.57 0.71 0.61 1.60 1.74 1.72 1.41 0.84 1.26 0.90 0.63 0.59 0.77 1.00 0.78 0.74 0.67 0.76 1.09 0.93 0.94

0.72 0.06 0.05 0.75 0.68 0.70 0.41 0.39 0.38 0.44 0.38 0.48 0.28 0.32 0.34 0.50 0.45 0.40 0.24 0.30 0.28 0.35 0.31 0.29 0.29 0.38 0.42 0.57 0.67 0.59 0.30 0.38 0.39 0.54 0.67 0.56 0.50 0.43 0.55 0.49 0.47 0.42 0.43 0.45 0.46 0.42 0.31 0.30 0.43 0.40 0.47

0.59 0.10 0.07 0.71 0.83 0.71 0.54 0.45 0.46 0.57 0.46 0.72 0.37 0.44 0.36 0.44 0.43 0.50 0.49 0.29 0.34 0.41 0.51 0.46 0.44 0.40 0.49 0.73 0.74 0.75 0.41 0.43 0.39 0.84 0.81 0.75 0.50 0.60 0.56 0.58 0.55 0.67 0.68 0.48 0.48 0.42 0.28 0.36 0.60 0.48 0.61

0.66 0.08 0.06 0.71 0.64 0.51 0.39 0.37 0.37 0.42 0.36 0.41 0.30 0.34 0.30 0.38 0.41 0.42 0.35 0.27 0.29 0.35 0.35 0.36 0.33 0.38 0.33 0.56 0.65 0.50 0.34 0.44 0.31 0.65 0.66 0.64 0.46 0.45 0.46 0.38 0.53 0.46 0.39 0.43 0.37 0.26 0.28 0.28 0.47 0.47 0.49

0.68 0.07 0.06 0.49 0.52 0.53 0.30 0.33 0.39 0.39 0.33 0.35 0.23 0.24 0.29 0.33 0.36 0.39 0.20 0.17 0.22 0.28 0.32 0.33 0.28 0.27 0.23 0.50 0.40 0.44 0.29 0.29 0.26 0.46 0.57 0.47 0.39 0.47 0.40 0.37 0.35 0.32 0.31 0.30 0.26 0.22 0.19 0.25 0.33 0.37 0.40

0.78 0.06 0.05 0.57 0.66 0.79 0.44 0.26 0.24 0.33 0.33 0.44 0.28 0.26 0.25 0.32 0.29 0.29 0.29 0.20 0.16 0.37 0.26 0.32 0.37 0.24 0.30 0.45 0.48 0.44 0.28 0.26 0.40 0.53 0.51 0.41 0.33 0.50 0.33 0.29 0.34 0.31 0.31 0.35 0.37 0.40 0.23 0.17 0.38 0.41 0.43

0.72 0.06 0.06 0.66 0.66 0.54 0.39 0.31 0.35 0.41 0.37 0.42 0.29 0.29 0.30 0.40 0.33 0.38 0.24 0.27 0.28 0.31 0.35 0.36 0.31 0.32 0.37 0.48 0.53 0.53 0.25 0.38 0.27 0.61 0.58 0.57 0.40 0.44 0.42 0.41 0.44 0.44 0.38 0.36 0.40 0.21 0.34 0.30 0.45 0.42 0.43

0.68 0.05 0.06 0.54 0.64 0.55 0.32 0.39 0.33 0.31 0.40 0.37 0.29 0.25 0.23 0.32 0.34 0.36 0.25 0.18 0.19 0.31 0.28 0.36 0.31 0.31 0.31 0.48 0.48 0.47 0.21 0.24 0.23 0.56 0.59 0.43 0.39 0.32 0.36 0.35 0.32 0.34 0.31 0.27 0.32 0.26 0.31 0.27 0.34 0.41 0.40

564

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Appendix F. Overall relevance and strength of selected sectors and countries for US and European companies

Fig. 14. The overall relevance of the news sentiments of the selected sectors and countries on the US stocks ranging from Aug. 1, 2006 to Dec. 31, 2014.

Fig. 15. The overall strength of the news sentiments of the selected sectors and countries on the US stocks ranging from Aug. 1, 2006 to the Dec. 31, 2014.

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Fig. 16. The overall relevance of the news sentiments of the selected sectors and countries on the European stocks ranging from Aug. 1, 2006 to Dec. 31, 2014.

Fig. 17. The overall strength of the news sentiments of the selected sectors and countries on the European stocks ranging from Aug. 1, 2006 to Dec. 31, 2014.

565

566

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

Appendix G. Overal relevance and strength of sectors and countries for US companies using Kalman filtered sentiment data Table 11 Mean relevance of the news sentiments of the US sectors and selected countries for the US companies by sector ranging from Jan. 1, 2006 to Dec. 31, 2014. asset

FIN 0.72

TEC 0.77

IND 0.72

MAT 0.69

UTL 0.61

COM 0.71

HLC 0.64

NCY 0.68

ENE 0.66

YCY 0.71

MPTRXUSFIN MPTRXUSTEC MPTRXUSIND MPTRXUSMAT MPTRXUSUTL MPTRXUSCOM MPTRXUSHLC MPTRXUSNCY MPTRXUSENE MPTRXUSYCY US IT GR DE CN

0.16 0.11 0.12 0.10 0.09 0.13 0.15 0.13 0.14 0.12 0.14 0.12 0.17 0.14 0.09

0.17 0.13 0.16 0.07 0.10 0.12 0.14 0.12 0.13 0.10 0.16 0.14 0.15 0.12 0.10

0.17 0.15 0.17 0.15 0.14 0.11 0.21 0.20 0.16 0.12 0.16 0.12 0.16 0.15 0.14

0.19 0.12 0.13 0.13 0.15 0.10 0.14 0.17 0.14 0.11 0.13 0.12 0.16 0.12 0.12

0.17 0.17 0.14 0.13 0.10 0.11 0.12 0.14 0.11 0.12 0.11 0.12 0.15 0.14 0.09

0.17 0.20 0.19 0.11 0.17 0.18 0.16 0.17 0.16 0.14 0.17 0.13 0.17 0.15 0.10

0.16 0.16 0.13 0.12 0.12 0.11 0.12 0.14 0.17 0.11 0.15 0.12 0.11 0.13 0.08

0.18 0.14 0.14 0.12 0.10 0.11 0.14 0.14 0.15 0.12 0.17 0.15 0.16 0.12 0.13

0.22 0.23 0.15 0.18 0.21 0.18 0.13 0.20 0.18 0.16 0.17 0.13 0.14 0.21 0.11

0.15 0.14 0.14 0.11 0.09 0.10 0.13 0.13 0.19 0.11 0.11 0.11 0.14 0.13 0.12

Table 12 Mean strength of the lags of the news sentiments on the US sectors and selected countries for the US companies by sector ranging from Jan. 1, 2006 to Dec. 31, 2014.

assett−1 assett−5 assett−22 MPTRXUSFINt−1 MPTRXUSFINt−5 MPTRXUSFINt−22 MPTRXUSTECt−1 MPTRXUSTECt−5 MPTRXUSTECt−22 MPTRXUSINDt−1 MPTRXUSINDt−5 MPTRXUSINDt−22 MPTRXUSMATt−1 MPTRXUSMATt−5 MPTRXUSMATt−22 MPTRXUSUTLt−1 MPTRXUSUTLt−5 MPTRXUSUTLt−22 MPTRXUSCOMt−1 MPTRXUSCOMt−5 MPTRXUSCOMt−22 MPTRXUSHLCt−1 MPTRXUSHLCt−5 MPTRXUSHLCt−22 MPTRXUSNCYt−1 MPTRXUSNCYt−5 MPTRXUSNCYt−22 MPTRXUSENEt−1 MPTRXUSENEt−5 MPTRXUSENEt−22 MPTRXUSYCYt−1 MPTRXUSYCYt−5 MPTRXUSYCYt−22 USt−1 USt−5 USt−22 ITt−1 ITt−5 ITt−22 DEt−1 DEt−5 DEt−22 CNt−1 CNt−5 CNt−22 GRt−1 GRt−5 GRt−22

FIN

TEC

IND

MAT

UTL

COM

HLC

NCY

ENE

YCY

0.24 0.22 0.19 1.33 0.84 0.90 0.29 0.41 0.30 0.43 0.38 0.57 0.35 0.37 0.32 0.45 0.53 0.48 0.45 0.32 0.33 0.41 0.41 0.40 0.29 0.21 0.30 0.82 1.01 0.78 0.59 0.36 0.39 0.71 0.57 0.41 0.35 0.34 0.34 0.45 0.37 0.47 0.36 0.29 0.34 0.55 0.56 0.51

0.27 0.23 0.18 1.08 0.87 0.94 0.36 0.33 0.33 0.54 0.35 0.47 0.38 0.32 0.35 0.41 0.30 0.31 0.46 0.33 0.43 0.40 0.42 0.38 0.32 0.26 0.34 0.94 0.77 0.70 0.42 0.35 0.30 0.52 0.32 0.46 0.45 0.34 0.38 0.48 0.38 0.39 0.44 0.26 0.31 0.47 0.41 0.54

0.25 0.23 0.20 1.28 1.11 0.94 0.39 0.38 0.34 0.51 0.53 0.46 0.41 0.36 0.42 0.48 0.36 0.41 0.47 0.40 0.46 0.52 0.41 0.38 0.32 0.43 0.35 1.03 1.12 0.88 0.38 0.45 0.44 0.66 0.60 0.59 0.38 0.29 0.38 0.49 0.41 0.47 0.38 0.48 0.37 0.57 0.47 0.56

0.24 0.24 0.18 1.53 1.32 1.28 0.39 0.49 0.39 0.51 0.50 0.64 0.55 0.45 0.45 0.60 0.56 0.47 0.41 0.37 0.63 0.47 0.49 0.49 0.36 0.37 0.35 2.04 0.91 0.88 0.77 0.37 0.54 0.80 0.50 0.68 0.48 0.56 0.58 0.59 0.37 0.43 0.41 0.36 0.55 0.69 0.47 0.51

0.23 0.21 0.18 0.98 0.79 0.95 0.44 0.28 0.39 0.49 0.42 0.46 0.33 0.34 0.38 0.41 0.26 0.42 0.37 0.33 0.34 0.38 0.40 0.34 0.30 0.31 0.28 0.92 0.89 0.87 0.39 0.32 0.54 0.50 0.34 0.45 0.31 0.32 0.35 0.54 0.47 0.48 0.35 0.37 0.26 0.60 0.54 0.37

0.24 0.23 0.18 1.02 1.15 1.01 0.39 0.40 0.41 0.62 0.50 0.61 0.42 0.35 0.45 0.53 0.47 0.46 0.36 0.56 0.41 0.44 0.38 0.40 0.41 0.43 0.39 1.26 0.93 1.04 0.64 0.51 0.52 0.69 0.63 0.66 0.38 0.36 0.51 0.62 0.49 0.56 0.42 0.31 0.35 0.62 0.55 0.58

0.24 0.22 0.19 0.83 1.21 1.07 0.37 0.26 0.41 0.45 0.44 0.43 0.37 0.32 0.41 0.46 0.28 0.39 0.41 0.31 0.32 0.38 0.39 0.48 0.30 0.34 0.26 1.67 1.07 0.87 0.34 0.33 0.52 0.48 0.44 0.67 0.39 0.37 0.63 0.41 0.39 0.43 0.38 0.33 0.34 0.51 0.53 0.48

0.26 0.22 0.18 1.14 0.95 0.78 0.39 0.29 0.28 0.54 0.33 0.51 0.38 0.36 0.28 0.49 0.33 0.36 0.48 0.29 0.29 0.39 0.32 0.34 0.42 0.27 0.29 0.83 0.74 0.92 0.48 0.38 0.40 0.50 0.44 0.47 0.34 0.31 0.34 0.40 0.38 0.35 0.37 0.23 0.35 0.58 0.53 0.48

0.23 0.22 0.20 1.52 1.15 1.10 0.41 0.54 0.45 0.59 0.51 0.57 0.38 0.38 0.43 0.79 0.55 0.50 0.59 0.51 0.48 0.52 0.49 0.44 0.48 0.42 0.37 2.06 1.33 0.93 0.42 0.49 0.53 0.68 0.51 0.56 0.41 0.50 0.54 0.48 0.58 0.60 0.41 0.35 0.51 0.78 0.64 0.51

0.27 0.23 0.23 1.35 0.88 0.87 0.45 0.46 0.38 0.52 0.54 0.52 0.35 0.30 0.29 0.56 0.47 0.36 0.42 0.39 0.44 0.32 0.43 0.33 0.39 0.28 0.36 0.83 1.13 1.21 0.64 0.44 0.39 0.53 0.54 0.41 0.41 0.64 0.43 0.37 0.54 0.44 0.43 0.32 0.34 0.66 0.49 0.41

F. Audrino and A. Tetereva / Journal of Banking and Finance 106 (2019) 542–567

References Akyildirim, E., Altarovici, A., Ekinci, C., 2015. Effects of Firm-Specific Public Announcements on Market Dynamics: Implications for High-Frequency Traders. Handbook of High Frequency Trading, 305. Algieri, B., Leccadito, A., 2017. Assessing contagion risk from energy and non-energy commodity markets. Energy Econ. 62, 312–322. Allen, D.E., McAleer, M.J., Singh, A.K., 2015. Machine news and volatility: the Dow Jones Industrial Average and the TRNA real-time high-frequency sentiment series. In: The Handbook of High Frequency Trading. Academic Press San Diego, pp. 327–344. Andersen, T.G., Davis, R.A., Kreiss, J.-P., Mikosch, T.V., 2009. Handbook of Financial Time Series. Springer Science & Business Media. Appel, G., 2003. Become your own technical analyst: how to identify significant market turning points using the moving average convergence-divergence indicator or MACD. J. Wealth Manage. 6 (1), 27–36. Arnold, A., Liu, Y., Abe, N., 2007. Temporal causal modeling with graphical Granger methods. In: Proceedings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp. 66–75. Arouri, M.E.H., Jouini, J., Nguyen, D.K., 2011. Volatility spillovers between oil prices and stock sector returns: implications for portfolio management. J. Int. Money Finance 30 (7), 1387–1405. Audrino, F., Camponovo, L., 2018. Oracle properties, bias correction, and bootstrap inference for adaptive lasso for time series m-estimators. J. Time Ser. Anal. 39 (2), 111–128. Audrino, F., Knaus, S.D., 2016. Lassoing the HAR model: a model selection perspective on realized volatility dynamics. Econom. Rev. 35 (8–10), 1485–1521. Baur, D.G., 2012. Financial contagion and the real economy. J. Bank. Finance 36 (10), 2680–2692. Beber, A., Brandt, M.W., Luisi, M., 2015. Distilling the macroeconomic news flow. J. Financ. Econ. 117 (3), 489–507. doi:10.1016/j.jfineco.2015.05.005. Becker, B., 2016a. Sentiment May Have Bottomed For Volkswagen. http: //seekingalpha.com/article/3957529- sentiment- may- bottomed- volkswagen. Accessed: 2017-12-01. Becker, B., 2016b. Starbucks’ Bitter Rewards – Good For Investors?http://lipperalpha.financial.thomsonreuters.com/2016/05/ starbucks- bitter- rewards- good- for- investors/. Accessed: 2017-12-01. Bekaert, G., Ehrmann, M., Fratzscher, M., Mehl, A.J., 2011. Global Crises and Equity Market Contagion. Technical Report. National Bureau of Economic Research. Benhabib, J., Liu, X., Wang, P., 2016. Sentiments, financial markets, and macroeconomic fluctuations. J. Financ. Econ. 120 (2), 420–443. doi:10.1016/j.jfineco.2016. 01.008. arXiv:1011.1669v3. Bhanot, K., Burns, N., Hunter, D., Williams, M., 2014. News spillovers from the Greek debt crisis: impact on the Eurozone financial sector. J. Bank. Finance 38, 51–63. Bollerslev, T., 1986. Generalized autoregressive conditional heteroskedasticity. J. Econom. 31 (3), 307–327. Booth, G.G., Martikainen, T., Tse, Y., 1997. Price and volatility spillovers in Scandinavian stock markets. J. Bank. Finance 21 (6), 811–823. Borovkova, S., 2015. The Role of News in Commodity MarketsAvailable at SSRN: https://ssrn.com/abstract=2587285. Borovkova, S., Garmaev, E., Lammers, P., Rustige, J., 2016. SenSR: A Sentiment-Based Systemic Risk IndicatorAvailable at SSRN: https://ssrn.com/abstract=2759289. Borovkova, S., Lammiman, A., 2010. The impact of news sentiment on energy futures returnsVrije Universiteit Amsterdam, https://sbe.vu.nl/en/Images/paper_ borovkova_tcm258-204330.pdf. Borovkova, S., Mahakena, D., 2015. News, volatility and jumps: the case of natural gas futures. Quant. Finance 15 (7), 1217–1242. Brenner, M., Galai, D., 1989. New financial instruments for hedge changes in volatility. Financ. Anal. J. 45 (4), 61–65. Brownlees, C. T., Engle, R. F., 2015. SRISK: a conditional capital shortfall index for systemic risk measurement. Department of Finance, New York University. Bühlmann, P., Van De Geer, S., 2011. Statistics for High-Dimensional Data: Methods, Theory and Applications. Springer Science & Business Media. Cahan, R., Jussa, J., Luo, Y., 2009. Breaking news: how to use news sentiment to pick stocksMacquarie US Equity Research. Chan, S., Han, G., Zhang, W., 2016. How strong are the linkages between real estate and other sectors in China? Res. Int. Bus. Finance 36, 52–72. Chan, W.C., 2003. Stock price reaction to news and no-news: drift and reversal after headlines. J. Financ. Econ. 70 (2), 223–260. doi:10.1016/S0304-405X(03)00146-6. Chunxia, Y., Xueshuai, Z., Luoluo, J., Sen, H., He, L., 2016. Study on the contagion among american industries. Physica A 444, 601–612. Diebold, F.X., Mariano, R., 1995. Comparing forecast accuracy. J. Bus. Econ. Stat. 13 (3), 253–263. Diebold, F.X., Yilmaz, K., 2009. Measuring financial asset return and volatility spillovers, with application to global equity markets. Econ. J. 119 (534), 158–171. Ding, X., Zhang, Y., Liu, T., Duan, J., 2015. Deep learning for event-driven stock prediction. In: Proceedings of the 24th International Joint Conference on Artificial Intelligence (ICJAIâ15), pp. 2327–2333. Durbin, J., Koopman, S., 2001. Time Series Analysis by Space State Methods. Erawan, S.D., 2015. Essays on Behavioural Approach in Finance. University of St. Gallen Ph.D. thesis. Ewing, B.T., Malik, F., 2016. Volatility spillovers between oil prices and the stock market under structural breaks. Global Finance J. 29, 12–23.

567

Fan, J., Li, R., 2001. Variable selection via nonconcave penalized likelihood and its oracle properties. J. Am. Stat. Assoc. 96 (456), 1348–1360. Fang, L., Peress, J., 2009. Media coverage and the cross-section of stock returns. J. Finance 64 (5), 2023–2052. Farmer, L., Schmidt, L., Timmermann, A., 2019. Pockets of predictability. Available at SSRN 3152386. Forbes, K.J., Rigobon, R., 2002. No contagion, only interdependence: measuring stock market comovements. J. Finance 57 (5), 2223–2261. Francq, C., Zakoian, J.-M., 2011. GARCH Models: Structure, Statistical Inference and Financial Applications. John Wiley & Sons. Friedman, J., Hastie, T., Tibshirani, R., 2008. Sparse inverse covariance estimation with the graphical lasso. Biostatistics 9 (3), 432–441. Friedman, J., Hastie, T., Tibshirani, R., 2009. Glmnet: lasso and elastic-net regularized generalized linear modelsR package version 1. Granger, C.W., 1980. Testing for causality: a personal viewpoint. J. Econ. Dyn. Control 2, 329–352. Gustafsson, M., Hornquist, M., Lombardi, A., 2005. Constructing and analyzing a large-scale gene-to-gene regulatory network lasso-constrained inference and biological validation. IEEE/ACM Trans. Comput. Biol.Bioinf. 2 (3), 254–261. Hammoudeh, S.M., Yuan, Y., McAleer, M., 2009. Shock and volatility spillovers among equity sectors of the gulf arab stock markets. Q. R. Econ. Finance 49 (3), 829–842. Harvey, C.R., 2017. Presidential address: the scientific outlook in financial economics. J. Finance 72 (4), 1399–1440. doi:10.1111/jofi.12530. Hwang, B.H., 2011. Country-specific sentiment and security prices. J. Financ. Econ. 100 (2), 382–401. doi:10.1016/j.jfineco.2010.10.020. Jacob, L., Obozinski, G., Vert, J.-P., 2009. Group lasso with overlap and graph lasso. In: Proceedings of the 26th Annual International Conference on Machine Learning. ACM, pp. 433–440. Kalman, R.E., 1960. A new approach to linear filtering and prediction problems. J. Basic Eng. 82 (1), 35–45. King, M., Sentana, E., Wadhwani, S., 1994. Volatility and links between national stock markets. Econometrica 62 (4), 901–933. Kirange, D.K., Deshmukh, R.R., 2016. Sentiment analysis of news headlines for stock price prediction. Composoft, ‎Int. J. Adv. Comput. Technol. 5 (3), 2080–2084. Li, C., Li, H., 2008. Network-constrained regularization and variable selection for analysis of genomic data. Bioinformatics 24 (9), 1175–1182. Lozano, A.C., Abe, N., Liu, Y., Rosset, S., 2009. Grouped graphical Granger modeling for gene expression regulatory networks discovery. Bioinformatics 25 (12), i110–i118. Lugmayr, A., Gossen, G., 2013. Evaluation of methods and techniques for language based sentiment analysis for DAX 30 stock exchange a first concept of a ”LUGO” sentiment indicator. In: International SERIES on Information Systems and Management in Creative eMedia, pp. 69–76. Luss, R., d’Aspremont, A., 2015. Predicting abnormal returns from news using text classification. Quant. Finance 15 (6), 999–1012. Narayan, P.K., Bannigidadmath, D., 2015. Does financial news predict stock returns? New evidence from islamic and non-Islamic stocks. Pac.-Basin Finance J. 42, 24–45. Nickerson, R., 2011. Mathematical Reasoning: Patterns, Problems, Conjectures, and Proofs. Taylor & Francis. Peng, J., Wang, P., Zhou, N., Zhu, J., 2009. Partial correlation estimation by joint sparse regression models. J. Am. Stat. Assoc. 104 (486), 735–746. Peress, J., 2014. The media and the diffusion of information in financial markets: evidence from newspaper strikes. J. Finance 69 (5), 2007–2043. Peterson, R.L., 2016. Trading on Sentiment: The Power of Minds Over Markets. John Wiley & Sons. Reilly, M., Posadas-Sanchez, D., Kettle, L., Killeen, P., 2012. Rats (Rattus norvegicus) and pigeons (Columbia livia) are sensitive to the distanceto food, but only rats request more food when distance increases. Behav. Processes 91 (3), 236–243. Sharpe, W.F., 1964. Capital asset prices: a theory of market equilibrium under conditions of risk. J. Finance 19 (3), 425–442. Shimamura, T., Imoto, S., Yamaguchi, R., Miyano, S., 2007. Weighted lasso in graphical Gaussian modeling for large gene network estimation based on microarray data. Genome Inform. 19, 142–153. Tamakoshi, G., Hamori, S., 2016. Time-varying co-movements and volatility spillovers among financial sector CDS indexes in the UK. Res. Int. Bus. Finance 36, 288–296. Tibshirani, R., 1996. Regression shrinkage and selection via the lasso. J. R. Stat. Soc. Ser. B 267–288. Tibshirani, R., Saunders, M., Rosset, S., Zhu, J., Knight, K., 2005. Sparsity and smoothness via the fused lasso. J. R. Stat. Soc. Ser. B 67 (1), 91–108. Tsay, R.S., 2005. Analysis of Financial Time Series, Vol. 543. John Wiley & Sons. Wen, X., Wei, Y., Huang, D., 2012. Measuring contagion between energy market and stock market during financial crisis: a copula approach. Energy Econ. 34 (5), 1435–1446. Yuan, M., Lin, Y., 2006. Model selection and estimation in regression with grouped variables. J. R. Stat. Soc. Ser. B 68 (1), 49–67. Zou, H., 2006. The adaptive lasso and its oracle properties. J. Am. Stat. Assoc. 101 (476), 1418–1429. Zou, H., Hastie, T., 2005. Regularization and variable selection via the elastic net. J. R. Stat. Soc. Ser. B 67 (2), 301–320.