Firm-specific information and systemic risk

Journal Pre-proof Firm-specific information and systemic risk A.E. Clements, Y. Liao PII:

S0264-9993(19)31430-0

DOI:

https://doi.org/10.1016/j.econmod.2019.11.031

Reference:

ECMODE 5089

To appear in:

Economic Modelling

Received Date: 8 September 2019 Revised Date:

28 November 2019

Accepted Date: 28 November 2019

Please cite this article as: Clements, A.E., Liao, Y., Firm-specific information and systemic risk, Economic Modelling (2020), doi: https://doi.org/10.1016/j.econmod.2019.11.031. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Published by Elsevier B.V.

Firm-Specific Information and Systemic Risk

Abstract Although there is substantial literature linking news to the asset return volatility of a single asset, little attention has been paid to how news influences the relationships between firms. This paper addresses this issue by examining how firm-specific scheduled and unscheduled news arrivals influence the systemic risk of individual firms based on a sample of 47 US financial institutions. Whereas negative surprises from scheduled news announcements and a higher rate of unscheduled news both increase the systemic risk of a firm, positive news surprises decrease this systemic risk. In addition, negative scheduled news and a higher rate of unscheduled news across the sector increases the total connectedness or systemic risk across the sector as a whole. These effects are magnified when the market is already in distress. The results indicate that regulators should consider more than volatility and pay attention to the news flow when monitoring systemic risk.

Keywords Information flow, volatility connectedness, network, information uncertainty JEL Classification Numbers C22,G00.

1

Introduction

There is a long history of research into the relationship between news arrivals and asset return volatility, predominantly in the context of a single firm or market, motivated by the theory of (Clark, 1973, Tauchen and Pitts, 1983). In contrast, little attention has been paid to how the news flow relating to a specific firm influences its relationships with other firms, such as industry peers. By modeling the interactions between the volatilities in a novel network framework (Diebold and Yilmaz, 2014), this paper considers how news arrivals relating to individual firms impact their systemic risk within a larger system. The main results are that positive (negative) surprises from scheduled news decrease (increase, resp.) the systemic risk of a firm and that firms experiencing unscheduled news also experience greater systemic risk. These findings are important for policy makers and regulators who are concerned about the stability of financial markets and, hence, the manner in which shocks are transmitted between assets. The results suggest that regulators should monitor the rate and nature of the news flow when monitoring the stability of financial markets. While theoretically appealing, there has been a great deal of debate in the literature regarding the appropriate measures of information flow. Early studies had a tendency to focus on scheduled firm announcements such as earnings announcements. Seminal works such as Ball and Brown (1968) and Beaver (1968) consider how prices, volume and volatility react to earnings announcements in the US stock market. Patell and Wolfson (1981) investigate the evolution of the implied standard deviation (ISD) near annual earnings announcement dates on a sample of 28 firms quoted on the Chicago Board Options Exchange. Isakov and Perignon (2000) consider the dynamics of implied volatility around earnings announcements in the Swiss market, and Donders and Vorst (1996) investigate the same issue in the Dutch market. In recent years, there have been rapid advances in the capture, storage and analysis of large volumes of unscheduled news items. This trend has led to the development of a new strand of literature linking electronic news arrivals and market behavior. Kalev, Liu, Pham, and Jarnecic (2004) and Kalev and Duong (2011) investigate the impact of firm-specific news arrivals as a proxy for the public information flow in the context of Australian equities. Groß-Klußmann and Hautsch (2011) use news data for individual equities traded on the London Stock Exchange sampled at high frequency to examine how news arrivals influence the trading activity in individual stocks. Riordan, Storkenmaier, Wagener, and Sarah Zhang (2013) focus on the impact of newswire messages on the intraday price discovery, liquidity, and trading intensity of Canadian stocks. Smales (2014) uses the news headlines relating to the constituents of the S&P 500 to examine

the relationship between the aggregate news sentiment and changes in the implied volatility index (VIX index). In addition, a number of recent studies (Shen, Zhang, Xiong, Li, and Zhang, 2016, Shen, Li, and Zhang, 2018, Zhang, Feng, Jin, Shen, Xiong, and Zhang, 2014, Zhang, Shen, Zhang, and Xiong, 2013) have relied on textual analysis to measure the impact information flow from internet activity on stock market behavior. On the other hand, given the rapidly increasing degree of financial integration and connectedness in modern markets, it is important to move beyond a single firm and consider the linkages between stock return volatilities across different firms. There is a vast literature that explores the transmission of volatility across stock markets; see Gagnon and Karolyi (2006) for an extensive review. Surprisingly though, the question of how news or information flow affects volatility linkages has received little attention. Connolly and Wang (1999) and Jiang, Konstantinidi, and Skiadopoulos (2012) examine the relationship between macroeconomic news announcements and stock market volatility spillovers in developed markets, and Hanousek, Kocenda, and Kutan (2009) and Hanousek and Kocenda (2011) analyze the same issues in emerging markets. Connolly and Wang (1999) uses volatility measured from historical data (a conditional volatility model) to examine the impact of macroeconomic news on the volatility linkages among US, UK and Japanese markets. Jiang et al. (2012) investigate the effect of macroeconomic news on implied volatility spillovers across US and European stock markets. In terms of the impact of firm-specific news announcements, Ball and Kothari (1991) examine how the average betas across firms behave relative to earnings announcements, and Patton and Verado (2012) investigate how the dynamics of the realized beta using high-frequency data respond to earnings announcements. Moving beyond simple measures of volatility comovements, this paper estimates measures of the directional volatility connectedness within a network framework and examines how the resulting measures of systemic risk respond to firm-specific news flow. From a modeling standpoint, while multivariate GARCH-style models that capture the correlation between stock returns have become extremely popular in the context of volatility transmission, they are of limited use in examining the broader linkages between the volatility of high-dimensional stock portfolios. Recent advances in network analysis offer a natural framework within which the interconnections between the volatilities of a potentially large number of stocks can be analyzed. While network analysis has been used in a range of areas in science, it has only gained popularity in the finance literature since the GFC, with Billio, Getmanski, Lo, and Pellizzon (2012) and Diebold and Yilmaz (2014) developing network models of systemic risk and with Diebold and Yilmaz (2014) proposing a number of summary measures of the network structure. These measures relate to the importance of an individual firm or stock relative to the

broader market or network (known as centrality), how other firms or stocks within the network affect an individual firm or stock (known as fragility), and the structure or strength of the connections of the overall network. Here, these measures are used to examine how scheduled and unscheduled firm-specific information flows influence the interactions between individual firms and the overall market (and vice versa). How these impacts vary with the financial market and macroeconomic conditions is also considered. In particular, this paper focuses on the realized volatility (RV) connectedness of 47 US financial institutions. Given the role of financial institutions in the global financial crisis (GFC) of 2007-2009, and the fact that understanding financial institution connectedness is key for understanding financial crises and their evolution, there is a burgeoning literature that studies the systemic risk within the financial sector. Diebold and Yilmaz (2016) provide a detailed study of connectedness both within and between US and European financial institutions from 2004 to 2014, which includes all phases of the GFC. Moreover, Demirer, Diebold, Liu, and Yilmaz (2018) investigate the static and dynamic patterns of the global bank volatility connectedness. However, these studies did not consider the impact of the information flow on the volatility connectedness. Therefore, this paper fills the void to examine the following: 1) whether firm-specific news announcements influence the total systemic risk in the financial industry; 2) whether specific news announcements increase the systemic importance of the announcing institution; 3) whether an individual institution’s volatility increases when there are information flows from other institutions; 4) whether the impact of firm-specific news announcements changes over time; and 5) whether there are “bellwether firms” whose news releases are more informative and have stronger effects on the system. The answer to these questions is of particular importance, as they will shed light on whether firm-specific news releases lead to the resolution or creation of market uncertainty and whether the systemic risk changes. The empirical findings here suggest the following patterns: 1) scheduled news containing positive surprises leads to a decrease in both the total industry systemic risk and the systemic risk of the announcing firms; and 2) negative scheduled news surprises and unscheduled news increase the total industry systemic risk and the systemic importance of the announcing firm. In other words, the results show that positive scheduled news surprises resolve information uncertainty, whereas negative scheduled news surprises and unscheduled news create information uncertainty. These results are consistent with the recent findings of Jiang et al. (2012), who investigate the effects of scheduled and unscheduled macroeconomic news announcements on implied volatility spillovers and conclude that scheduled (unscheduled) macroeconomic news announcements resolve (create, resp.) information uncertainty. These news effects are magnified when the economy is in distress and

when the announcing firms are “bellwether firms” whose information offers greater potential for investors to learn about the rest of the economy. This work contributes to the literature from three perspectives. First, this work is related to the empirical literature on information spillovers and contagion. Examples include stock return comovements across markets in relation to changes in macroeconomic conditions (see Shiller (1989), Karolyi and Stulz (1996) and Connolly and Wang (2003), for example), and stock return comovements across firms in response to the release of a firm-specific information flow (Patton and Verado, 2012). The current work adds to this literature by linking the comovement in volatility rather than returns to the release of firm-specific information flows, and it provides insights into how granular firm-specific information changes the price uncertainty of related stocks. Second, this work is related to a number of empirical studies that examine the volatility spillover coinciding with news announcements; see Gagnon and Karolyi (2006) for an extensive review. Whereas the previous studies mainly focus on the effects of macroeconomic news announcements on the volatility spillover across different markets, e.g., Jiang et al. (2012), this work complements the literature by adding an analysis of the effect of firm-specific news announcements in general and unscheduled news arrivals in particular on the systemic risk of individual firms. Finally, this study contributes to the recent literature on cross-firm learning and pricing, e.g., Hameed, Morck, Shen, and Yeung (2015) and Patton (2011). The analysis here differs from those articles in that the focus is on the reaction of stock volatility, rather than the returns or beta; furthermore, we use recent advances in network modeling to identify the industry “bellwether firms”. In common with these previous studies, though, is the important role that cross-firm pricing plays: the volatility connectedness revealed here may be explained by learning by investors across different individual companies, and particularly the bellwether firms within the same industry. The paper proceeds as follows. Section 2 outlines the data used for generating the volatility estimates along with the various sources of news flow. Section 3 outlines the econometric methodology for estimating the volatility connectedness between firms. Then, Section 4 presents the empirical methodology and results to examine how news impacts the connectedness in a variety of different ways. Section 5 presents a number of robustness tests in terms of the analysis at different data frequencies and using different news announcements. Lastly, Section 6 provides concluding comments.

2

Data

This study considers 47 large US financial institutions for the period from 4 January 2003 to 30 December 2016, which represents 3519 trading days. These firms were chosen from the list of S&P 500 composite index constituents at the beginning of 2017. We focus on financial institutions because an understanding of the information transmission and systemic risk within the financial industry is conducive to comprehending the role of information in the recent GFC. We originally selected the top 50 financial institutions from the S&P 500 to represent the network of systemically important financial institutions, but three were removed due to a lack of unscheduled news data. A list of the firms and their tickers can be found in Appendix A. A description of both the intraday price data required for the estimation of volatility and the firm-specific scheduled and unscheduled news flow now follow.

2.1

Estimating Volatility

Split and dividend-adjusted prices are sampled at a 5-minute frequency to construct estimates of daily volatility. Given these prices, we define the j − th discrete-time return within day t as rtk = p(t−1)+ k − p(t−1)+ k−1 , k = 1, 2, ....M, M

(1)

M

where M refers to the number of intraday equally spaced returns over trading day t. As noted in Andersen and Bollerslev (1998), Andersen, Bollerslev, Diebold, and Labys (2003), and Barndorff-Nielsen and Shephard (2002), the quadratic variation of the process can be estimated by the realized volatility (RV ), which is defined as the sum of the intraday squared returns, i.e., RVt ≡

M X

rt2k .

(2)

k=1

A series of daily RV estimates for each stock is constructed, and the network structure underlying these estimates is examined in the subsequent empirical analysis.

2.2

Firm-Specific Information Flow

To capture the unscheduled public news flow, preprocessed news data from the Thomson Reuters News Analytics database are used. This dataset is the same source of news data used by GroßKlußmann and Hautsch (2011) and Riordan et al. (2013). The text of the news items broadcast over the Reuters network is analyzed using a linguistic pattern recognition algorithm. This analysis produces a number of characteristics relating to each news item including the sentiment

or tone, which are measured as -1, 0 or 1 for negative, neutral and positive tones, respectively. In this paper, the volume of firm-specific news for each day and its associated sentiment are employed. The news for day t is aggregated over the period 4:00 PM on day t − 1 to 4:00 PM on day t for each firm. N Fi,t and N Si,t represent the total number and the average sentiment of the news items for firm i on day t, respectively. Figure 1 displays the time series plots of the total volume of news flow (aggregating the N Fj,t across the firms) and the average sentiment (average of N Sj,t across the firms). The top panel shows that the volume of news flow increased somewhat from roughly 2005 and reached a peak in 2008-2009 during the Global Financial Crisis. Although the average sentiment in the lower panel shows that the average sentiment of the news flow was approximately neutral, during 2008-2009, it is clear that the news flow was consistently negative on average. Information relating to scheduled news, namely, annual earnings announcements in this case, was obtained from the Thomson Reuters I/B/E/S database. Standardized earnings surprises relating to the annual earnings announcements of the individual firms were also obtained, and they will be denoted as EN Si,t . The summary statistics of the news data can be found in Table 2.

3 3.1

Systemic risk Estimating systemic risk from volatility connectedness

In this section, the econometric methodology for estimating the volatility connectedness and systemic risk is introduced, with more detail provided in Appendix B. This framework provides estimates of total systemic risk (total connectedness) of the whole system, an estimate of the systemic importance of an individual firm (total directional connectedness to other firms) and the total directional connectedness from others to an individual firm on each day. These estimates permit an analysis of how the volatility shocks originate and are transmitted within a system and how the whole system is connected by volatility spillover effects. Diebold and Yilmaz (2014) show how the traditional vector autoregressive (VAR) model and associated variance decompositions provide a natural and insightful framework to measure the network connectedness of a panel of financial time series. By treating RV as the object of direct interest, the time-varying connectedness in the volatility of the 47 financial firms can be estimated by a time-varying parameter VAR (TVP-VAR) model using the method of Gary and Dimitris (2013):

yt = Zt βt + εt ,

(3)

βt+1 = βt + ut ,

(4)

and

where yt is a 47 × 1 vector containing observations of the time series of RVs and Zt is a 47 × k matrix defined so that each TVP-VAR equation has an intercept and p lags of the 47 variables. Thus, k = 47(1 + 47p). Following Diebold and Yilmaz (2014), p is set to p = 3. εt is i.i.d. N (0, Σt ), ut is i.i.d. N (0, Qt ), and εt and ut are independent ∀t. Because the parameter vector βt follows an AR(1) process that allows for time-varying coefficients, time-varying estimates of the network structure can be derived1 . This modeling framework allows an estimate of the connectedness in volatility to be generated by assessing the shares of the forecast error variation in one firm’s RV due to shocks arising from other firms’ RVs. This connectedness is related to the familiar econometric notion of variance decomposition in which the forecast error variance of variable i is decomposed into parts that are attributable to other variables in the system. The ijth H-step variance decomposition is 0 denoted by dH ij , which measures the fraction of variable i s H-step forecast error variance due

to shocks in variable j and takes the following form: −1 PH−1 0 2 σjj,t h=0 (ei,t At,h Σt ej,t ) H dij,t = PH−1 0 , 0 (e A Σ A e ) t i,t t,h h=0 i,t t,h

(5)

where ej,t is a selection vector with the jth element being unity and zeros elsewhere at time t, At,h is the coefficient matrix of the h-lagged shock vector in the infinite moving-average representation of the VAR model, ΣH t is the covariance matrix of the shock vector in the VAR, and σjj,t is the jth diagonal element of Σt . Because the shocks are correlated here, the sums of the forecast error variance contributions are not necessarily unity. Therefore, dH ij,t is normalized ˜ by to dH ij,t

dH ij,t ˜ = dH PN H . ij,t j=1 dij,t

(6)

The structure in the network can be summarized in a number of connectedness measures that ˜ can be constructed from dH ij,t . More detailed descriptions of the estimation and variance decomposition procedures are provided in Appendix B. To aid the interpretation of the connectedness measures, it is useful to summarize the variance decomposition in a network connectedness table such as Table 1. In Table 1, the main upper-left 1

We do not report the TVP-VAR model estimation results, as the model parameters are time-varying. How-

ever, these results are available upon request.

˜ 47 × 47 block contains the variance decomposition and is denoted by DH = [dH ij,t ], where H is the forecast horizon. The whole connectedness table simply augments DH with the rightmost column containing row sums, the bottom row containing column sums, and the bottom-right element containing the grand average in all cases for i 6= j. The off-diagonal entries of DH are the important elements of the forecast error variance decompositions, as they measure the pairwise directional connectedness. Hence, the pairwise directional connectedness from j to i is defined as H Ci←j = dH ij .

(7)

H 6= C H , so there are 472 −47 separate pairwise directional connectedNote that in general, Ci←j j←i

ness measures. Here, we consider not the individual elements of DH but rather its off-diagonal row or column sums, which are known as the total directional connectedness measures. That is, we define the total directional connectedness from other indices to i as Ci←• =

N X

˜ dH ij,t ,

(8)

j=1,j6=i

and the total directional connectedness to those from j as C•←j =

N X

˜ dH ij,t .

(9)

i=1,i6=j

In fact, Ci←• reflects how the volatility shocks that occur in other firms influence the volatility of the ith firm by identifying the degree of fragility (systemic exposure) of a firm in the system. C•←j reflects how the volatility shocks in the jth firm will influence the volatility of all other firms by identifying the degree of centrality (systemic risk) of a firm in the system. Finally, the grand total of the off-diagonal entries in DH measures the total connectedness: CH =

1 N

N X

˜ dH ij,t .

(10)

i,j=1,i6=j

This measure distills the total structure in the system into a single number reflecting the total systemic risk in the system. Ci←• and C•←j are denoted as fi,t and ci,t for the ith firm at time t, and C H is denoted as T Ct at time t in the following analysis. These volatility network risk measures are more sophisticated than the classical network measures. First, the connection measures are weighted. The matrix is not filled simply with 1/0 entries but with a specific value reflecting the strength of the connection. Second, the connection measures are not symmetric. That is, the strength of link ij is not necessarily the same as that of link ji. These measures are simply a transformation of the model parameters, so time-varying parameters in the model effectively allow for time-varying connectedness measures

that can be used to analyze the impact of exogenous information such as the news flow on connectedness. Following Diebold and Yilmaz (2014), a VAR(3) model with 12-step forecast error variance decomposition is used to construct these measures.

3.2

Static Analysis of Systemic Risk

This section provides a discussion of static volatility connectedness that moves from total connectedness (T Ct ) to total directional connectedness (ci,t and fi,t ) and to pairwise connectedness H ). (Ci←j

The first few rows of Table 2, which are labeled as Total connectedness, Centrality and Fragility, show summary statistics of the static volatility connectedness measures that capture the systemic risk for the 47 US financial institutions in the full sample. It is clear that none of the data series are normally distributed. Moreover, the p-values of the ADF test show that all of the volatility connectedness measures are not stationary. Therefore, the first differences of the measures are adopted in the subsequent regression analyses. Figure 2 further shows the average of the total directional connectedness across the full sample period. The top panel shows that there is clear variation in the estimates of centrality across the 47 financial institutions, with JPMorgan Chase & Co (JPM), Kimco Realty Corp (KIM) and Wells Fargo & Co New (WFC) as the most systemically important in transmitting volatility shocks to other firms, followed by Bank of America Corp (BAC), Citigroup Inc (C), Comerica Inc (CMA), Keycorp New (KEY), Lincoln National Corp (LNC), PNC Financial Services GRP Inc (PNC), Regency Centers Corp (REG), Suntrust Banks Inc (STI) and Vornado Realty Trust (VNO). In contrast, in the lower panel, little variation is observed in the fragility measures across the 47 US financial institutions, with Block H & R Inc (HRB) and Humana Inc (HUM) being two relatively resilient firms. Last, Figure 3 uses a heatmap to present the pairwise connectedness across the 47 US financial institutions. The values in each column of the heatmap indicate the pairwise connectedness from the institutions labeled on the x-axis to the ones on the y-axis. The deepest color in the colormap indicates the strongest pairwise connectedness from the x-axis institutions to the y-axis institutions. It is clear that the three firms JPMorgan Chase & Co (JPM), Kimco Realty Corp (KIM) and Wells Fargo & Co New (WFC) have relatively deep colors through their whole columns in the map, indicating that they have the strongest effects on others among the 47 financial institutions. Those firms are followed by Bank of America Corp (BAC), Citigroup Inc (C), Comerica Inc (CMA), Keycorp New (KEY), Lincoln National Corp (LNC), P N C

Financial Services GRP Inc (PNC), Regency Centers Corp (REG), Suntrust Banks Inc (STI) and Vornado Realty Trust (VNO) in terms of these effects. In contrast, the columns of B B & T Corp (BBT) and Blackrock Inc (BLK), followed by Block H & R Inc (HRB) and Humana Inc (HUM), have the lightest colors in the map, indicating the weakest effects of these firms on others. Unsurprisingly, these findings are consistent with the observations from Figure 2, as the two figures provide the information relating to the firm’s centrality in different ways.

3.3

Dynamic Analysis of Systemic Risk

As the volatility connectedness and, hence, systemic risk may vary over time, this section provides a dynamic analysis of the total connectedness (T Ct ) and total directional connectedness (ci,t and fi,t ). Figure 4 depicts the time-varying total connectedness or systemic risk of the 47 financial institutions. It is very clear that since the onset of the GFC, the total degree of connectedness has increased and has generally remained at these higher levels even after the overall volatility subsided. Figures 5 and 6 plot the time-varying total directional connectedness, that is, the centrality (ci,t ) and fragility (fi,t ) of the individual firms. In terms of the time series behavior of the centrality of individual firms, the results are mixed. While many firms’ centralities increased at roughly the time of the GFC, the centralities of many firms have subsequently fallen. In contrast, the fragility of all firms has increased significantly since the GFC. Clearly, an implication of the greater connectedness between firms is that overall they are more fragile to shocks across the market.

4

The Impact of Information Flow on Systemic Risk

In this section, the links between firm-specific news flow, systemic risk at the total industry level (TC), firm-level systemic risk (centrality) and fragility are considered.

4.1

The effect of news on total systemic risk

The analysis begins with investigating whether news effects the total market connectedness, leading to hypothesis H1:

H1: Firm-specific news announcements affect sector-wide systemic risk (total volatility con-

nectedness). H1 is tested using the following regression: T Ct = γ1 RVindex,t + γ2 ItEN S

+

N X

EN Sj,t + γ3 ItEN S

−

j=1

+ γ5 ItN S

+

N X

N Sj,t + γ6 ItN S

j=1

−

N X j=1

N X

N Sj,t + εt ,

EN Sj,t + γ4

N X

N Fj,t

j=1

(11)

j=1

where RVindex,t is the market-level volatility that is proxied by the S&P500 index RV . Fur+ PN thermore, ItEN S j=1 EN Sj,t indicates when earnings surprises are positive on average across − PN the 47 firms, and ItEN S j=1 EN Sj,t denotes when the earnings surprise is negative on averPN PN P + N S+ N S − = 1 if age. ItN S = 1 if N j=1 N Sj,t < 0. Hence, It j=1 N Sj,t j=1 N Sj,t > 0 and It represents the aggregate positive unscheduled news sentiment of the 47 firms at time t, and − PN ItN S j=1 N Sj,t represents the aggregate negative unscheduled news sentiment of the 47 firms P at time t. N j=1 N Fj,t denotes the total volume of news flow (number of news items). The estimation results for equation 11 for the total market connectedness are reported in the second column of Table 3. The corresponding t-statistics based on standard errors that are robust to heteroscedasticity and arbitrary within-cluster correlation are also reported, with coefficients significant at 5% marked by an asterisk. First, the significant positive coefficient of the market index RV confirms that the market exhibits greater systemic risk when the volatility increases. More importantly, the coefficients on both scheduled news with negative surprise and unscheduled news with negative sentiment are significantly negative, which implies that as negative news arrives, the volatility connectedness and, hence, the systemic risk within the industry, increases. These findings are in line with the creation of the uncertainty hypothesis of Ederington and Lee (1996) according to which the occurrence of a major negative (regardless of whether it is scheduled or unscheduled) news release causes high unexpected volatility, which increases the information uncertainty of the market participants as well as the connectedness of different firms. Last, the estimates of the coefficients on the volume of news and positive news are very small and insignificant.

4.2

The effect of news on centrality and fragility

This section moves from the aggregate to the individual firm level to examine the effects of connectedness at the firm level, the degree of systemic importance (centrality) and the fragility of individual firms. First, the effect of news on the degrees of centrality of individual firms is tested under the following hypothesis:

H2a : Firm-specific news announcements increase the systemic importance (measured by the degree of centrality) of announcing firms in transmitting volatility shocks. This hypothesis is examined using the following regression: EN Si+

ci,t = α1 RVindex,t + α2 It N Si+

+ α5 It EN Si+

where It

N Si−

N Si,t + α6 It

EN Si−

EN Si,t and It N Si+

surprises and It

EN Si−

EN Si,t + α3 It

EN Si,t + α4 N Fi,t

N Si,t + εi,t ,

(12)

EN Si,t denote the respective positive and negative earnings

N Si−

N Si,t and It

N Si,t denote positive and negative unscheduled news ar-

rivals for firm i at time t, respectively. N Fi,t denotes the number of these unscheduled news items. Second, the effect of the news on the degree of fragility of individual firms is tested under the following hypothesis:

H2b : Firm-specific news announcements increase the fragility of non-announcing firms in volatility spillovers. This hypothesis is examined using the following regression: fi,t = β1 RVindex,t +

N X

+ EN S6= i β2 It

EN Sj,t +

− EN S6= i β3 It

+ β4

N Fj,t +

N S+ β5 It 6=i

j=1,j6=i + EN S6= i

PN

N X

N Sj,t +

N S− β6 It 6=i

j=1,j6=i − EN S6= i

EN Sj,t

j=1,j6=i

j=1,j6=i N X

N X

N X

N Sj,t + εi,t ,

(13)

j=1,j6=i

PN

denote when firms other than firm i N S+ P experience positive and negative earning surprises on average, respectively, and It 6=i N j=1,j6=i N Sj,t − P N S6= and It i N j=1,j6=i N Sj,t denote when firms other than firm i have positive and negative unwhere It

j=1,j6=i EN Sj,t and It

j=1,j6=i EN Sj,t

scheduled news, respectively. Estimates of the coefficients in equations 12 and 13 are obtained using both pooled OLS and panel regression with firm-year fixed effects, with the results reported in the last four columns of Table 3. The first two columns contain the results for the firm’s degree of centrality, and the second two columns contain the results for the firm’s degree of fragility with the corresponding tstats based on standard errors that are robust to heteroscedasticity and arbitrary within-cluster correlation. Several noteworthy findings emerge from the results of Table 4. First, significant positive estimates of the coefficients on RVt in both models indicate that the two firm-level connectedness measures increase when the market is more turbulent. Second, with respect to the impact of news on ci,t , estimates of the coefficient on N Fi,t are significantly positive, indicating that a firm’s centrality increases when the firm in question is associated with more EN Si−

unscheduled news flow. Moreover, the significant negative coefficients on both It N Si−

and It

EN Si,t

N Si,t show that both scheduled and unscheduled negative news increase the centrality

of the announcing firms, suggesting that when firms release negative news, the announcing firms become more systemically important due to the creation of information uncertainty and investors EN Si+

learning from the announcing firm. Last, the significant negative coefficient of It

EN Si,t

indicates that scheduled news events with positive surprises decrease the systemic risk of the announcing firm. This finding implies that positive news resolves information uncertainty and reduces investors’ attention on the announcing firm, which leads to the comovements between it and the other firms in the industry. P In terms of the firm fragility fi,t , significant positive estimates on N j=1,j6=i N Fj,t and significant − − + PN P P N S6= EN S EN S6= N N i 6=i i EN S and I EN S , I negative estimates on It j,t j,t t t j=1,j6=i N Sj,t j=1,j6=i j=1,j6=i are found. These results imply that a firm’s fragility increases when other firms either experience a greater volume of news or undergo negative scheduled or unscheduled news. Hence, a firm’s fragility is more susceptible to shocks from the other firms. Again, the firm’s fragility is reduced when other firms have positive scheduled news.

4.3

Does the impact of news change over time?

Given the significant effect of news on volatility connectedness and systemic risk, this section examines whether this effect changes over time. During periods of crisis, stock market volatility increases and connectedness and systemic effects are expected to become more pronounced. Hence, investigating the role of news announcements during periods of market distress is of particular interest. Accordingly, the third hypothesis considered is as follows:

H3: The impact of scheduled or unscheduled news on market connectedness and systemic risk changes over time in response to the market conditions.

Here, the market conditions are measured by the volatility of the S&P500 Index returns RVt . To capture changes in the market conditions, an indicator variable is defined as follows: I(RVt > RV ) indicates whether the market volatility is above its average level, which would reflect periods of market stress. The set of regressions used to examine this question are augmented versions of equations 12 and 13, which include an additional interaction term I(RVt > RV ) along with all the original right-hand-side variables, except for RVt , as this variable is replaced by the indicator variable I(RVt > RV ). Table 4 contains the regression results. The total market connectedness regression results show that the connectedness and, hence, the systemic risk increase further when the market is experi-

encing negative scheduled news surprises. Consistent with the earlier results, both the centrality and fragility increase during periods of stress with higher market volatility. While the release of positive earnings news continues to decrease the fragility and centrality, this effect significantly decreases when the market volatility is already relatively high. The unscheduled news flow still has a positive impact on the centrality and fragility, and these effects are magnified when the market volatility is high. In addition, the arrival of negative scheduled and unscheduled news still increases the centrality and fragility, with this effect being magnified during periods of high volatility. Overall, during market distress, the arrival of negative firm-specific news will create larger information uncertainty that further intensifies the systemic risk at both the industry and individual firm level. In contrast, the resolution effect of positive scheduled news on the information uncertainty appears to weaken relative to the effect in normal market conditions.

4.4

Does the impact of news differ across firms?

Given the evidence of news influencing firm-level volatility connectedness, this section further investigates whether cross-sectional differences in changes in the firm-level connectedness measures (that is, centrality and fragility) around news announcements are related to the firms’ characteristics. First, the cross-sectional differences in changes in centrality are examined under the following hypothesis:

H4a : Announcing firms whose news announcements are more informative are associated with greater changes in their systemic importance (measured by firm centrality).

Second, the cross-sectional differences in changes of fragility due to firm news announcements are examined using the following hypothesis:

H4b : Changes in fragility are larger for firms whose news announcements are less informative.

Firms whose fundamentals better predict the fundamentals of other firms in their industry have announcements that are more informative and attract more analyst coverage (Hameed et al. (2015)). Therefore, analyst coverage is used here as a proxy for the informativeness of a firm. Then, the firms are sorted into quintiles based on the analyst coverage, and crosssectional differences in changes in firm centrality around news announcements are analyzed using the cross-sectional regressions of equations 12 and 13, which are estimated for stocks

grouped into the quintiles. Since the number of analysts covering a stock is well known to be positively correlated with market capitalization, following Patton (2011), the effect of market capitalization is controlled for by estimating the following cross-sectional regression each year: ln(1 + nai,t ) = αt + βt ln(capi,t−15 ) + i,t ,

(14)

where nai,t is the number of analysts who have issued a forecast for stock i in the ninety days leading up to an earnings announcement on day t, and capi,t−15 is the market capitalization of stock i measured fifteen trading days before the earnings announcement. Given estimates of the parameters αt and βt and the estimates of i,t , the residual analyst coverage is used in the analysis. As analyst coverage is associated with annual earnings announcements in this paper, the stock quintiles sorted based on analyst coverage are updated annually. The cross-sectional regression estimates are reported in Table 5. The results reveal that i) the previous findings that negative scheduled and unscheduled news increases a firm’s centrality and fragility, and positive scheduled news decreases the two measures are still preserved across all of the stock quintiles. Similarly, the positive effects of the rate of news flow and the market volatility on the two measures are also preserved. ii) The positive scheduled news from the announcing firms with higher analyst coverage resolve more information uncertainty. The results show that a one-unit positive scheduled news surprise decreases the systemic importance (centrality) by −0.0483 (t-stat of −8.0918) for firms with low analyst coverage and −0.4126 (t-stat of −11.6561) for firms in the top quintile of the residual analyst coverage. iii) Negative scheduled and unscheduled news from announcing firms increases the centrality for firms with low analyst coverage by −0.0081 (t-stat of −0.3251) and −0.4678 (t-stat of −6.4078) for firms in the top quintile of the residual analyst coverage. These findings confirm the intuition that information releases on stocks that are more informative offer investors greater potential to learn about the rest of the economy. iv) The centrality also increases more in firms with higher analyst coverage when the market volatility is higher and the number of news events increases. v) The effects of news on firm fragility have less cross-sectional dispersion than do the effects of news on firm centrality. Firms with higher analyst coverage do not react differently from other firms.

5

Robustness checks

The first robustness test is to see whether the empirical findings are sensitive to the modeling framework used to estimate the volatility. The same analysis presented in Section 4 is repeated

with intraday estimates of the volatility. 5-minute returns are used to construct 30-minute RV estimates, which are directly used in the analysis presented in Section 3.1 to construct measures of the connectedness and systemic risk at a weekly frequency. Then, these measures of connectedness are related to the news arrivals observed in the corresponding weekly periods, with the results reported in online Table 1. While this analysis is conducted at a lower frequency and without the use of time-varying parameter models (one fixed parameter model for each week), it is clear that the same impact of the news flow is present. Higher index-level volatility increases both the firm-specific and total market connectedness. Positive earnings surprises continue to reduce the centrality and fragility of firms. The rate of the unscheduled news flow increases the centrality and fragility, while the arrival of both negative scheduled and unscheduled news increases the individual firm-level and market-level systemic risk. As an alternative to earnings announcements, the second robustness check is to present results relating to the impact of scheduled dividend announcements on connectedness and hence systemic risk. Annual dividend surprises were also downloaded from I/B/E/S and used in place of the earnings surprises in equations 11 to 13, with the results reported in online Table 2. Overall, the results are consistent with the case of earnings. Negative estimates of α2 /β2 /γ2 show that positive dividend surprises reduce the network connectedness in the same way as earnings surprises. Moreover, market volatility and unscheduled news continue to have the same impact. In addition to market-level volatility, the third robustness test is to use an alternative, broader measure of the market conditions. Thus, we employ the Aruoba-Diebold-Scotti Business Conditions Index, denoted below as BCt , to capture the state of the broader macroeconomy. This index is based on a combination of the weekly initial jobless claims, monthly payroll employment, industrial production, personal income less transfer payments, manufacturing and trade sales and quarterly real GDP; for more details, see Diebold and Scotti (2009). An indicator variable is defined as I(BCt < 0), and it is used to indicate periods of below average macroeconomic conditions. The results in online Table 3 also indicate that the connectedness changes with broader macroeconomic conditions; in a similar way, the connectedness changes in response to changes in volatility. Both the centrality and fragility increase when the business conditions are poor. The arrival of positive earnings surprises decrease the firm-level market connectedness, with smaller reductions occurring when the business conditions are already poor. Negative scheduled news surprises and unscheduled news (regardless of whether it is good or bad news) continue to have a positive impact on the centrality and fragility. However, when the conditions are poor (that is, when I(BCt < 0)), negative unscheduled news has a greater impact, whereas positive unscheduled news has less of an impact. With regard to the total market systemic risk,

only negative scheduled news surprises will cause a significant increase, and other news has no significant effect.

6

Conclusion

There is a long history of empirical research examining the impact of news arrivals on the volatility of asset returns, with much of this work focusing on the univariate case for a single asset over time. Little attention has been paid to the role played by news in explaining the links between the volatility of individual assets. This paper moved beyond the single asset case and considered how the firm-specific news flow influences the links between assets. Building on traditional time series techniques, a number of summary measures of the network structure can be used to construct measures of systemic risk. These measures relate to the systemic importance of an individual firm relative to the sector as a whole, how the overall sector impacts an individual firm, and the overall structure or strength of the connections in the network. In the context of a sample of large financial institutions, this paper considered the impact of both scheduled and unscheduled news arrivals on network connections and systemic risk. It was found that the total degree of systemic risk has increased since the GFC. Increases in market-wide volatility are found to increase the total systemic risk. Firm-specific news flow is also found to be important in terms of both the market-level systemic risk and the firm-specific centrality and fragility. Specifically, positive surprises in scheduled earnings announcements decrease systemic risk, while the arrival of unscheduled negative news increases the risk, and is amplified during periods of market stress. Similar results are found when considering alternative announcements such as dividends. The importance of the news flow is revealed here, and it suggests that regulators should monitor the rate and nature of the news flow when monitoring the stability of financial markets.

Y1

Y2

···

YN

Y1

dH 11

dH 12

···

dH 1N

Y2 .. .

dH 11 .. .

dH 12 .. .

··· .. .

dH 1N .. .

YN

dH 11

dH 12

···

dH 1N

To Others

PN

H i=1 di1 ,

i 6= 1

PN

H i=1 di2 ,

i 6= 2

PN

H j=1 diN ,

i 6= N

Table 1: Network Connectedness Matrix

From Others PN H j=1 d1j , j 6= 1 PN H j=1 d2j , j 6= 2 .. . PN H j=1 dN j , j 6= N P N 1 H j=1 dij , i 6= j N

Mean

Std. Deviation

Skewness

Kurtosis

ADF

0.9486

0.0124

-0.4148

3.4505

0.6268

median

1.0131

0.1749

-0.4843

3.0769

0.4207

25% percentile

0.8172

0.1463

-1.0551

2.3893

0.2738

75% percentile

1.1503

0.2401

-0.2799

4.4077

0.5009

median

0.9517

0.0145

-1.5083

6.1359

0.6316

25% percentile

0.9458

0.0108

-2.0204

3.1398

0.5968

75% percentile

0.9561

0.0224

-0.6688

8.7397

0.6633

median

0.0195

0.4167

8.8903

368.5360

0.0010

25% percentile

0.0123

0.3363

-1.1812

263.0610

0.0010

75% percentile

0.0328

0.7062

16.1887

553.9439

0.0010

median

0.0441

0.3886

0.1824

5.1626

0.0010

25% percentile

0.0051

0.2882

0.0092

3.5045

0.0010

75% percentile

0.0730

0.4528

1.0694

9.4429

0.0010

median

1.1279

2.2622

4.7644

43.1124

0.0010

25% percentile

0.6322

1.5845

3.8576

29.9684

0.0010

75% percentile

1.9881

3.2948

6.1111

83.3989

0.0010

T Ct ci,t

fi,t

EN Si,t

N Fi,t

N Si,t

Table 2: Summary Statistics. This table reports the summary statistics of the firms’ scheduled and unscheduled news data sets and the volatility connectedness measures, including the mean, std. deviation, skewness, kurtosis and augmented Dickey Fuller (ADF) test p-values. The sample spans the period from 04 January 2003 to 05 December 2016 and includes 3519 daily observations. The numbers reported are percentiles (i.e., 25% percentile, median and 75% percentile) of the unconditional measures (that is, the average of the time series measures) across the 47 firms.

Total Market

RVindex,t EN S + EN S − NF N S+ N S−

Centrality

Fragility

OLS

OLS

Panel

OLS

Panel

0.0900

277.7921∗

271.9851∗

216.0484∗

217.6272∗

(1.7814)

(7.9026)

(7.3857)

(35.5803)

(35.7425)

−0.0620

−0.0502∗

−0.0490∗

−0.0004∗

−0.0004∗

(−1.6197)

(−8.4284)

(−8.2282)

(−3.4568)

(−3.6442)

−0.4329∗

−0.2166∗

−0.2126∗

−0.0039∗

−0.0039∗

(−3.8272)

(−15.0533)

(−14.7630)

(−11.2563)

(−11.2366)

0.0014

0.0106∗

0.0179∗

0.0002∗

0.0002∗

(0.4105)

(12.6379)

(16.9053)

(19.0137)

(17.0333)

0.0268

0.0107

0.0071

0.0008∗

0.0003

(0.2492)

(0.6115)

(0.3730)

(2.5742)

(0.8269)

−0.4971∗

−0.0343

−0.0943∗

−0.0045∗

−0.0034∗

(−2.4113)

(−1.6092)

(−4.1036)

(−8.1804)

(−5.3989)

Table 3: The impact of both scheduled and unscheduled news on volatility connectedness. This table reports the estimation results for the regressions in equations (11, 12 and 13). The sample spans the period from 04 January 2003 to 30 December 2016 and includes 3519 daily observations. The corresponding t-stats based on standard errors that are robust to heteroscedasticity and arbitrary within-cluster correlation are reported in the bracket, and the significant coefficients at the significance level of 5% are marked by an asterisk.

Total Market

Centrality

Fragility

RVt

0.1160

1541.2603∗

235.2663∗

1392.9762∗

236.4479∗

(1.8803)

(13.9025)

(6.0533)

(75.0035)

(37.2669)

RVt × I(RVt > RV )

0.0360

1463.0013∗

0.0599∗

1305.8375∗

0.0181∗

(0.2571)

(12.4842)

(3.8409)

(66.2468)

(3.8719)

−0.1592∗

−0.1214∗

−0.1172∗

−0.0014∗

−0.0015∗

(−2.4623)

(−12.2700)

(−11.8354)

(−7.2693)

(−7.4297)

0.1546

0.1119∗

0.1087∗

0.0014∗

0.0016∗

(1.9238)

(9.0498)

(8.7767)

(5.9353)

(6.4386)

−0.8481∗

−0.6321∗

−0.6246∗

−0.0063∗

−0.0083∗

(−3.7055)

(−24.2471)

(−23.9406)

(−9.1816)

(−11.9337)

−0.5519∗

−0.5910∗

−0.5873∗

−0.0036∗

−0.0057∗

(−2.0949)

(−18.9193)

(−18.7948)

(−4.5092)

(−7.1063)

0.0060

0.0193∗

0.0287∗

0.0004∗

0.0001∗

(0.9967)

(16.3465)

(20.0625)

(25.6526)

(2.4417)

0.0053

0.0193∗

0.0198∗

0.0003∗

0.0003∗

(0.6602)

(10.0774)

(11.3878)

(13.5710)

(9.3265)

−0.0119

−0.0064

−0.0032

−0.0017∗

−0.0009∗

(−0.0993)

(−0.3129)

(−0.1464)

(−4.8953)

(−2.4208)

0.1121

0.0407

0.0396

2.9516

2.3988

(0.3464)

(1.0012)

(0.9071)

(−0.0005)

(−0.0034)

−0.5819∗

−0.0381

−0.0791∗

−0.0005

2.3988

(−2.4828)

(−1.4849)

(−2.9217)

(−0.7412)

(−0.0034)

0.4799

−0.0005

−0.0133

−0.0015

0.0021

(0.9623)

(−0.0114)

(−0.2711)

(−1.0963)

(1.4030)

EN S + EN S + × I(RVt > RV ) EN S − EN S −

× I(RVt > RV )

NF N F × I(RVt > RV ) N S+ N S+

× I(RVt < RV )

N S− N S − × I(RVt < RV )

Table 4: The change of the news impact in response to market conditions. The market condition is measured by the volatility of the Dow Jones Index returns, RVt . To capture changes in the market conditions, an indicator variable I(RVt > RV ) is defined in periods when the market volatility is above its average level, thereby capturing periods of market stress. EN S + denotes the arrival of positive earnings surprises, N F denotes the volume of the unscheduled news flow and N S − denotes the arrival of negative unscheduled news. This table reports the estimation results of the augmented versions of the regressions in equations 11 to 13 in which a set of additional interaction terms of the indicator I(RVt > RV ) and all the original right-hand-side variables are included, except for RVt , which is replaced by the indicator variable I(RVt > RV ). The sample spans the period from 28 January 2003 to 30 December 2016 and includes 3519 daily observations. The corresponding t-stats, which are based on standard errors that are robust to heteroscedasticity and arbitrary within-cluster correlation, are reported in the bracket, and the significant coefficients at the significance level of 5% are marked by an asterisk.

Q1(Low)

Q2

Q3

Q4

Q5(High)

Centrality RVindex,t

102.8905

1.5277

211.6005

284.7548

306.8946

(1.5277)

(6.4488)

(0.1308)

(2.0197)

(1.2948)

EN S +

−0.0483

−0.0536

−0.2792

−0.1215

−0.4126

(−8.0918)

(−3.4808)

(−8.6105)

(−5.3651)

(−11.6561)

−0.0081

−0.0817

−0.2267

−0.2245

−0.4678

(−0.3251)

(−4.1489)

(−18.0334)

(−0.6249)

(−6.4078)

0.0120

0.0164

0.0173

0.0183

0.0191

(2.5097)

(3.3637)

(5.5193)

(2.6352)

(6.0679)

0.0053

0.0056

0.0052

0.0083

0.0070

(1.3209)

(1.2075)

(0.1213)

(0.3048)

(0.0244)

−0.0390

−0.0688

−0.0740

−0.0994

−0.0987

(−5.9523)

(−6.1878)

(−6.8639)

(−3.8181)

(−4.3549)

EN S − NF N S+ N S−

Fragility RVindex,t

146.9898

258.1260

214.5925

270.9919

232.6250

(22.7296)

(14.3155)

(16.3108)

(16.5007)

(18.8979)

EN S +

−0.0004

−0.0007

−0.0005

−0.0005

−0.0004

(−3.0083)

(−2.0949)

(−2.1805)

(−1.7447)

(−1.9014)

EN S − NF N S+ N S−

−0.0001

−0.0002

−0.0002

−0.0003

−0.0002

(−11.8437)

(−7.0878)

(−7.3538)

(−7.8966)

(−9.8282)

0.0006

0.0002

0.0010

0.0004

0.0005

(1.6420)

(0.2321)

(1.2950)

(0.4062)

(0.6568)

0.0009

0.0000

0.0008

0.0032

0.0010

(1.3026)

(0.0210)

(0.5590)

(1.8642)

(0.7906)

−0.0004

0.0002

−0.0011

−0.0011

−0.0012

(−1.0703)

(−1.1061)

(0.2994)

(−1.2078)

(−2.0759)

Table 5: Changes in firm centrality and fragility around information flows by residual analyst coverage. This table presents the estimation results of regressions 12 and 13 which are estimated for stocks grouped into quintiles of residual analyst coverage, defined as the residual from a cross-sectional regression of analyst coverage on market capitalization. t-statistics, shown in parentheses, are computed from standard errors that are robust to heteroskedasticity and arbitrary intraday correlation.

1000

Total number of news across 47 financial firms

500

0 2003 0.4

2005

2007

2009

2011

2013

2015

Average news sentiment across 47 financial firms

0.2

0

-0.2 2003

2005

2007

2009

2011

2013

2015

Figure 1: Top plot: total volume of news flow, which aggregates the N Fj,t across the firms. Lower plot: average sentiment, the average of N Sj,t across the firms.

Figure 2: Full-sample Total Directional Connectedness. The top and lower panels of the bar graph present the centrality and fragility measures of the individual US financial institutions based on the full sample, and the predictive horizon is 12 days. The centrality measure gives the total directional connectedness from firm i to others, and the fragility measure gives the total directional connectedness from other firms to firm i.

Figure 3: Full-sample Pairwise Connectedness. This heatmap presents the full sample average of the time-varying pairwise connectedness of the 47 US financial institutions that are estimated from a TVP-VAR model with a 12-day predictive horizon.

0.98

0.97

0.96

0.95

0.94

0.93

0.92

0.91

0.9

0.89 2003

2005

2007

2009

2011

2013

2015

Figure 4: Time-varying Total Connectedness. This figure plots the time series of the total connectedness of the 47 US financial firms that is estimated from a TVP-VAR model with a 12-day predictive horizon for the underlying variance decomposition.

Figure 5: The total directional connectedness (centrality). This figure plots the time-varying centrality measures for the individual firms, which are estimated from a TVP-VAR model with a 12-day predictive horizon for the underlying variance decomposition.

Figure 6: The total directional connectedness (fragility). This figure plots the time-varying fragility measures for each individual firm, which are estimated from a TVP-VAR model with a 12-day predictive horizon for the underlying variance decomposition.

Appendix A: List of firms considered in this study AET

AETNA INC NEW

AFL

AFLAC INC

AIG

AMERICAN INTERNATIONAL GROUP INC

AIV

APARTMENT INVESTMENT & MGMT CO

AJG

GALLAGHER ARTHUR J & CO

ALL

ALLSTATE CORP

AXP

AMERICAN EXPRESS CO

BAC

BANK OF AMERICA CORP

BBT

B B & T CORP

BEN

FRANKLIN RESOURCES INC

BK BLK C CB CI CMA

BANK NEW YORK INC BLACKROCK INC CITIGROUP INC CHUBB CORP CIGNA CORP COMERICA INC

COF

CAPITAL ONE FINANCIAL CORP

FRT

FEDERAL REALTY INVESTMENT TRUST

HCN

HEALTH CARE REIT INC

HCP

HEALTH CARE PPTY INVS INC

HIG

HARTFORD FINANCIAL SVCS GRP INC

HRB

BLOCK H & R INC

HUM

HUMANA INC

JPM

JPMORGAN CHASE & CO

KEY

KEYCORP NEW

KIM

KIMCO REALTY CORP

LNC

LINCOLN NATIONAL CORP

LUK

LEUCADIA NATIONAL CORP

MAA

MID AMERICA APT COMMUNITIES INC

MAC

MACERICH CO

MMC

MARSH & MCLENNAN COS INC

O

REALTY INCOME CORP

PGR

PROGRESSIVE CORP OH

PNC

P N C FINANCIAL SERVICES GRP INC

PSA

PUBLIC STORAGE INC

REG

REGENCY CENTERS CORP

RJF

RAYMOND JAMES FINANCIAL INC

SPG

SIMON PROPERTY GROUP INC NEW

STI

SUNTRUST BANKS INC

STT

STATE STREET CORP

TMK

TORCHMARK CORP

UDR

UNITED DOMINION REALTY TR INC

UNH

UNITEDHEALTH GROUP INC

UNM

UNUMPROVIDENT CORP

VNO

VORNADO REALTY TRUST

WFC

WELLS FARGO & CO NEW

XL

X L GROUP LTD

Table 6: List of stock tickers and names.

Appendix B: TVP-VAR model estimation and generalized variance decomposition In this section, we provide a detailed description of the approach to estimating the TVP-VAR model using forgetting factors (Gary and Dimitris (2013)). Consider the model described in Section 3.1, yt =

p X

Φi yt−i + N Vt γt + υt ,

(15)

i=1

and Θt+1 = Θt + µt ,

(16)

where yt = (y1t , y2t , ..., ymt ) is a m × 1 vector of jointly determined dependent variables, and N Vt is an m × d matrix containing a set of news-related exogenous variables. Therefore, each TVP-VAR equation has an intercept, p lags of yt and d × 1 exogenous variables. To estimate the model, we make the following standard assumptions: 0

• Assumption 1 E(εt ) = 0, E(εt εt ) = Σt for all t, where Σt is a positive definite matrix, 0

E(εt εt0 ) = 0 for all t = t0 and E(εt |N Vt ) = 0. • Assumption 2 All the roots of |I −

Pp

i=1 Φi z

i|

= 0 fall outside the unit circle.

• Assumption 3 zt and N Vt are not perfectly collinear. Under these assumptions, the above model can be rewritten as the infinite moving average representation yt =

∞ X

At,i εt−i +

∞ X

Γt,i wt−i , t = 1, 2, ..., T,

(17)

i=0

i=0

where the coefficient matrices At,i can be obtained using the following recursive relations: At,i = Φ1 At,i−1 + Φ1 At,i−1 + ... + Φp At,i−p , i = 1, 2, ...,

(18)

with At,0 = Im and At,i = 0 for i < 0, and Γt,i = At,i γt . Next, we outline the forgetting factor method for model estimation. Let Yt = (y1 , y2 ..., yt ) denote the observations through time t. Bayesian inference for Θt involves the Kalman filter, and Kalman filtering proceeds using Θt |Y t−1 ∼ N (Θt|t−1 , Vt|t−1 ),

(19)

where Vt|t−1 = Vt|t−1 + Qt . If we replace the above equation by Vt|t−1 =

1 V , λ t|t−1

(20)

there is no longer a need to estimate or simulate Qt . λ is called a forgetting factor, and it is restricted to the interval 0 < λ < 1. Note that the above equations imply that Qt = λ1 Vt−1|t−1 . Following Gary and Dimitris (2013), we set λ = 0.99. Finally, the model parameters estimated above can also be used in the derivation of the forecast error variance decompositions dH ij,t , defined as the proportion of the n-step-ahead forecast error variance of variable i, which is accounted for by the innovations in variable j in the VAR, and the news-driven forecast error variance decompositions dH news,ij,t , defined as the proportion of the n-step-ahead forecast error variance of variable i, which is accounted for by the innovations H in the jth news variable in the VAR. The terms dH ij,t and dnews,ij,t are respectively expressed as

dH ij,t and dH news,ij,t

−1 PH−1 0 2 σjj,t h=0 (ei,t At,h Σt ej,t ) = PH−1 0 , 0 h=0 (ei,t At,h Σt At, h ei,t )

(21)

PH−1 0 h −1 2 σN h=0 (ei,t Γt ΣN V,t ej,t ) Vjj,t . = PH−1 0 h h0 h=0 (ei,t Γt ΣN V,t Γt ei,t )

(22)

References Andersen, T. and T. Bollerslev (1998). Dm-dollar volatility: intraday activity patterns, macroeconomic announcements and longer run dependencies. Journal of Finance 53, 219–265. Andersen, T., T. Bollerslev, F. Diebold, and P. Labys (2003). Modeling and forecasting realized volatility. Econometrica 71, 579–625. Ball, R. and P. Brown (1968). An empirical evaluation of accounting income numbers. Journal of Accounting Research 6, 159–178. Ball, R. and S. Kothari (1991). Security returns around earnings announcements. The Accounting Review 66 (4), 718–738. Barndorff-Nielsen, O. and N. Shephard (2002). Econometric analysis of realized volatility and its use in estimating stochastic volatility models. Journal of Royal Statisitcal Society, Series B 64, 253–280. Beaver, M. (1968). The information content of annual earnings announcements. Journal of Accounting Research 46, 67–92. Billio, M., M. Getmanski, A. Lo, and A. Pellizzon (2012). Econometric measures of connectedness and systemic risk in the finance and insurance sectors. Journal of Financial Economics 104, 535–559. Clark, P. (1973). A subordinated stochastic process model with finite variance for speculative prices. Econometrica 41, 135–155. Connolly, R. and F. Wang (2003). International equity market comovements: Economic fundamentals or contagion? Pacific-Basin Finance Journal 11 (1), 23–43. Connolly, R. A. and F. A. Wang (1999). Economic news and stock market linkages: Evidence from the u.s., u.k., and japan. In Proceedings of the Second Joint Central Bank Research Conference on Risk Management and Systemic Risk. [23. Demirer, M., F. X. Diebold, L. Liu, and K. Yilmaz (2018). Estimating global bank network connectedness. Journal of Applied Econometrics 33 (1), 1–15. Diebold, F. and C. Scotti (2009). Real-time measurement of business conditions. Journal of Business and Economic Statistics 27 (4), 417–427.

Diebold, F. and K. Yilmaz (2014). On the network topology of variance decompositions: Measuring the connectedness of financial firms. Journal of Econometrics 182, 119–134. Diebold, F. and K. Yilmaz (2016). Trans-atlantic equity volatility connectedness: U.s. and european financial institutions, 20042014. Journal of Financial Econometrics 14, 81–127. Donders, M. and T. Vorst (1996). The impact of firm specific news on implied volatilities. Journal of Banking and Finance 20, 1447–1461. Ederington, L. and J. Lee (1996). The creation and resolution of market uncertainty: the impact of information releases on implied volatility. Journal of Financial and Quantitative Analysis 31, 513–539. Gagnon, L. and G. Karolyi (2006). Price and volatility transmission across borders. Financial Markets, Institutions and Instruments 15, 107–158. Gary, K. and K. Dimitris (2013). Large time-varying parameter vars. Journal of Econometrics 177, 185–198. Groß-Klußmann, A. and N. Hautsch (2011). When machines read the newsd: Using automated text analytics to quantify high frequency news-implied market reactions. Journal of Empirical Finance 18, 321–340. Hameed, A., R. Morck, J. F. Shen, and B. Yeung (2015). Information, analysts, and stock return comovement. The Review of Financial Studies 28 (11), 3153–3187. Hanousek, J. and E. Kocenda (2011). Foreign news and spillovers in emerging european stock markets. Review of International Economics 19 (1), 170–188. Hanousek, J., E. Kocenda, and A. Kutan (2009). The reaction of asset prices to macroeconomic announcements in new eu markets: Evidence from intraday data. Journal of Financial Stability 5 (2), 199–219. Isakov, D. and C. Perignon (2000). Evolution of market uncertainty around earnings announcements. Journal of Banking and Finance 25, 1769–1778. Jiang, G., E. Konstantinidi, and G. Skiadopoulos (2012). Volatility spillovers and the effect of news announcements. Journal of Banking and Finance 36, 2260–2273. Kalev, P. and H. Duong (2011). Firm specific news arrival and the volatility of intraday stock index and futures returns. In G. Mitra and L. Mitra (Eds.), The Handbook of News Analytics in Finance. Chichester, West Sussex, UK: John Wiley and Sons, Ltd.

Kalev, P., W.-M. Liu, P. Pham, and E. Jarnecic (2004). Public information arrival and volatility of intraday stock returns. Journal of Banking and Finance 28, 1441–1467. Karolyi, G. and R. Stulz (1996). Why do markets move together? an investigation of u.s.-japan stock return comovements. Journal of Finance 51, 951–986. Patell, J. and M. Wolfson (1981). The ex ante and ex post price effects of quarterly earnings announcements reflected in option and stock prices. Journal of Accounting Research 19, 434–458. Patton, A. (2011). Volatility forecast comparison using imperfect volatility proxies. Journal of Econometrics 160, 246–256. Patton, A. and M. Verado (2012). Does beta move with news? firm-specific information flows and learning about profitability. Review of Financial Studies 25, 2790–2839. Riordan, R., A. Storkenmaier, M. Wagener, and S. Sarah Zhang (2013). Public information arrival: Price discovery and liquidity in electronic limit order markets. Journal of Banking and Finance 37, 1148–1159. Shen, D., X. Li, and W. Zhang (2018). Baidu news information flow and return volatility: Evidence for the sequential information arrival hypothesis. Economic Modelling 69, 127–133. Shen, D., W. Zhang, X. Xiong, X. Li, and Y. Zhang (2016). Trading and non-trading period internet information flow and intraday return volatility. Physica A: Statistical Mechanics and its Applications 451, 519–524. Shiller, R. (1989). Comovements in stock prices and comovements in dividends. Journal of Finance 44, 719–729. Smales, S. (2014). News sentiment and the investor fear gauge. Finance Research Letters 11, 122–130. Tauchen, G. and M. Pitts (1983). The price variability-volume relationship on speculative markets. Econometrica 51, 485–505. Zhang, W., D. Shen, Y. Zhang, and X. Xiong (2013). Open source information, investor attention, and asset pricing. Economic modeling 33, 613–619. Zhang, Y., L. Feng, X. Jin, D. Shen, X. Xiong, and W. Zhang (2014). Internet information arrival and volatility of sme price index. Physica A: Statistical Mechanics and its Applications 399, 70–74.

Highlights

•

This paper examines how firm-specific news influences the systemic risk of firms

•

Surprises in scheduled news influence systemic risk

•

An increase in the rate of unscheduled news increases systemic risk

•

The effect of news flow is magnified during a time of market distress

•

News flow is important for explaining systemic risk from a regulatory perspective