Bad News, Good News: Coverage and Response Asymmetries

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Bad News, Good News: Coverage and Response Asymmetries Luca Gambetti, Nicolo` Maffei-Faccioli, Sarah Zoi 2023-001 Please cite this paper as: Gambetti, Luca, Nicol`o Maffei-Faccioli, and Sarah Zoi (2023). “Bad News, Good News: Coverage and Response Asymmetries,” Finance and Economics Discussion Series 2023-001. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2023.001. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

* Bad News, Good News: Coverage and Response Asymmetries Luca Gambetti(cid:132) Nicolo` Maffei-Faccioli(cid:133) Universitat Auto`noma de Barcelona, BSE, Norges Bank Universita` di Torino and Collegio Carlo Alberto Sarah Zoi§ Federal Reserve Board October 2022 Abstract We study the dynamic link between economic news coverage and the macroeconomy. We construct two measures of media coverage of bad and good unemployment figures based on three major US newspapers. Using nonlinear time series techniques, we document three facts: (i) there is no significant negativity bias in economic news coverage. The asymmetric responsiveness of newspapers’ coverage to positive and negative unemployment shocks is entirely explained by the effects of these shocks on unemployment itself; (ii) consumption reacts to bad news, but not to good news; (iii) bad news is more informative to the agents and affects their expectations more than good news. JEL classification: C32, E32. Keywords: News Coverage, Agents’ Information, Business Cycles, Asymmetry, Threshold-SVAR. *This paper should not be reported as representing the views of Norges Bank or the Board of Governors of the Federal Reserve System. The views expressed are those of the authors and do not necessarily reflect thoseofNorgesBankoroftheFederalReserveSystem. ThispaperbenefitedfromdiscussionswithKnutAre Aastveit, Guido Ascari, Drago Bergholt, Jeffrey Campbell, Fabio Canova, Francesco Furlanetto, Domenico Giannone,LuigiIovino,VegardH.Larsen,ElmarMertens,MichelePiffer,GiorgioPrimiceri,GiuseppeRagusa, Alex Tagliabracci, Christian Wolf and participants at conferences and seminars. (cid:132)Departament d’Economia i d’Historia Economica, Edifici B, Office B3-1130, Universitat Auto`noma de Barcelona, 08193 Barcelona. Phone: (+34)935814569. E-mail: luca.gambetti@uab.es. Luca Gambetti acknowledges the financial support from the Spanish Ministry of Science and Innovation, through the Severo OchoaProgrammeforCentresofExcellenceinR&D(CEX2019-000915-S),thefinancialsupportoftheSpanishMinistryofScience,InnovationandUniversitiesthroughgrantPGC2018-094364-B-I00,andtheBarcelona School of Economics Research Network, and the financial support from the Italian Ministry of Research and University, PRIN 2017 grant J44I20000180001. (cid:133)NorgesBank,Bankplassen2,P.O.Box1179Sentrum,0107Oslo. Phone: (+47)40641754. E-mail: nicolomaffei.faccioli@norges-bank.no §FederalReserveBoardofGovernors,20th&ConstitutionAve. NW,mailstop: K-3620,20551Washington DC. Phone: (+1)2024408608. E-mail: sarah.zoi@frb.gov 1

1 Introduction Expectationsaboutcurrentandfutureeconomicconditionsareattherootofagents’decisionmaking process and are an important source of business cycle fluctuations. Under the fullinformationrationalexpectationsparadigm,agentsformexpectationsandtakedecisionswith perfect knowledge of the economy. In practice, agents’ information may differ substantially from the comprehensive representation of the economy assumed by the theory.1 Agents acquire signals through a variety of channels which mediate the informational content available to them. News media represent one of these channels and a major source of economic information (see Blinder and Krueger (2004)). This establishes a potentially important link between economic news coverage, agents’ information and expectations, and macroeconomic dynamics. Thispaperempiricallyinvestigatesthislinkwithaspecificfocusonthequalitativecontent of economic news. In particular, we test for two potential asymmetries in the way in which newspapers and agents, respectively, respond to positive and negative economic information. First, weinvestigateifthereisanegativitybiasineconomicnewscoverage. Suchbiasimplies that media cover negative economic events more than positive ones (see Soroka (2006)). A higher number of negative news during bad times and a subdued amount of positive news in good times influence agents’ information accordingly and may result in overpessimistic and underoptimistic views during recessions and expansions. The second asymmetry is related to agents’ response to bad and good news. Agents may attach, everything else equal, higher informational value to negative than to positive economic news and revise their expectations and consumption decisions differently depending on the type of news they receive.2 In this case, agents would react asymmetrically to good and bad news with clear implications for economic outcomes. We test both asymmeteries within a unified econometric framework. Webeginbyconstructingtwonovelmeasuresofmediacoverageofbadandgoodeconomic events using three major US newspapers.3 Our measures of bad and good news represent the total number of articles each month reporting increases or high values of unemployment and decreases or low values of unemployment, respectively. We focus our attention on un- 1See, among others, Mankiw and Reis (2002), Sims (2003) and Woodford (2002). 2Such asymmetry can arise, for instance, when agents are more concerned about their income losses than abouttheirgains,asforexampleintheoreticalmodelswithlossaversion(seeKahneman(1979))orwithrisk averse and rational inattentive agents (see Tutino (2013)). 3Textualanalysishasbecomeapowerfultoolnotonlyformacroeconomicanalysisasitisusedhere,butalso for forecasting. For instance, Larsen and Thorsrud (2019) shows that many news topics are good predictors of economic variables in Norway. Mueller and Rauh (2018) builds a large dataset of news to construct an indicator for predicting armed conflicts. 1

employment since this variable is a major cyclical indicator and is central to the economic news selection process (see Fogarty (2005)). We compare our bad and good news measures with their counterparts obtained from the Michigan Survey of Consumers: the portion of respondents reporting they recently heard bad and good news about unemployment and employment, respectively. We find that both our measures strongly correlate with their counterparts from the survey. The importance of this finding is twofold. On the one hand, it validates our indicators by proving their consistency with respect to other measures of information about unemployment. On the other hand, it suggests that newspaper information is relevant to the agents and could thus represent an important element in shaping their expectations and decisions. We use our news measures to build two standard indicators of information: negativity, obtained as the difference between the number of bad news and good news, and total information, obtained as the sum of the two. The former represents the prevailing (negative) tone of news about unemployment, while the latter its overall coverage. Both our indicators are highly correlated with the unemployment rate. We then use a Threshold Structural Vector Autoregression (TSVAR) model to assess the dynamic link between media coverage of the economy and agents’ consumption, information andexpectations. Thisdynamic, multivariatemodelallowsustoaddresspotentialendogeneity issues and account for much richer (non-linear) dynamics when compared to simple linear regressions. First, we use such framework to study the dynamic response of negativity and total information to positive and negative unexpected changes in the unemployment rate, and thus assess if there is any asymmetry in the news reporting process of the economy. Second, we use the model to study the dynamic responses of agents’ consumption, information and expectations to bad and good news shocks, as measured by positive and negative unexpected changes in negativity. The methodological novelty of our approach is to allow for asymmetries in the impulse response functions based on the sign of the shock considered. Ourfirstcontributionistoshowthatthereisnonegativitybiasineconomicnewscoverage once we allow for asymmetries in the dynamics of the economy. A bad economic shock which unexpectedly increases the unemployment rate generates a larger and more persistent effect on negativity than a good shock. This is in line with previous evidence from the political science literature which points towards the existence of a negativity bias in media coverage of economic events, as measured by changes in the unemployment rate (see e.g. Soroka (2006) and Soroka (2012)). However, the shock has also a substantial nonlinear effect on the unemployment rate: a bad shock generates larger and more persistent effects than a good shock. If the effects on negativity are normalized by the effects on the unemployment rate, 2

asymmetries in media coverage vanish. Indeed, the response of negativity becomes extremely similar for negative and positive shocks. This result represents evidence against the existence of a negativity bias in media coverage of the economy. The same finding is obtained for total information. We conclude that the negativity bias previously found in the literature is attributable to asymmetries in the dynamics of unemployment. Our second contribution is to document a significant asymmetry in the agents’ response to bad and good news shocks. A bad news shock which unexpectedly increases negativity decreasesconsumptionsubstantially,especiallyofdurablegoods,whileagoodnewsshockhas essentially no significant effect. To better understand this result, we estimate the responses of agents’ information and expectations to the two shocks and find that these are consistent with the responses of aggregate consumption. The fraction of informed individuals from the Michigan Survey increases (reduces) in response to a positive (negative) shock to negativity. Moreover, agents agree more about economic outcomes and change their expectations more markedly facing an increase in the negativity than a reduction. These results point to a substantially higher information content of bad news compared with good news and suggest a rationale for the stronger response of aggregate consumption to bad news. Our findings are robust to several changes in the model specification and, in particular, to the exclusion of the Great Recession, which represents an unprecedented period of bad news reporting about the economy. Based on these results, we draw an interesting conclusion. While there is no significant negativity bias in economic news reporting, there exists a significant bias in the way agents weight the qualitative content of economic news when they form expectations and take consumption decisions. To the best of our knowledge, we are the first to provide this evidence. The asymmetric behaviour of expectations and consumption that we document in this paper cannot be explained by a higher number of negative news relative to positive news (for which we control here) and thus is at odds with both the standard Permanent Income Hypothesis and models of sticky expectations (Carroll (2003)). On the contrary, the type of asymmetry we document suggests that, given an equal number of good and bad news items, agents give greater weight to negative rather than positive economic information. Media coverage of economic events has been studied, to some extent, in the economics literature (see Mullainathan and Shleifer (2005) and Nimark and Pitschner (2019)), but the bulk of contributions comes from the political science literature. The key finding in this field istheexistenceofanegativitybiasineconomicnewsreporting: negativeeventsreceivehigher media attention than positive events, see Goidel and Langley (1995), Fogarty (2005), Soroka 3

(2006) and Soroka (2012). This is typically shown in the context of simple linear regressions where the tone of unemployment news is regressed on positive and negative changes in the unemployment rate, together with an additional set of controls.4 We document the absence of any bias in news reporting once the effects of economic shocks on economic variables are explicitly taken into account. News coverage reacts significantly and very similarly, both qualitatively and in terms of magnitudes, to positive and negative economic shocks. Our paper also closely relates to a vast literature studying how news affects macroeconomic outcomes. News shocks to productivity have been documented to be an important driver of the business cycle.5 With respect to this literature, we make three main contributions. First, we do not limit our attention to news about technology, but rather consider general news about future unemployment developments. Second, we use a measure of news constructed from newspaper articles rather than relying on a theory-based identification.6 Third, and most importantly, we allow bad and good news to have asymmetric effects on the economy. Several studies have focused on the link between the media and consumers’ expectations, seeforinstanceLarsenetal.(2020). Empiricalevidencesuggeststhatagentsupdatetheirexpectations more frequently during periods of high news coverage, typically during recessions, see Doms and Morin (2004) and Carroll (2003). Also, bad news is found to have larger effects than positive news on consumers’ opinion and confidence, see Soroka (2006) and Soroka (2014) for a review. Our results largely confirm this finding. With respect to this literature, we make two main contributions. First, we use a dynamic, multivariate model which, with respect to simple regressions, addresses potential endogeneity issues and is able to account for nonlinear dynamics. Second, we study the role of bad and good news for the expectation formation process by focusing on agents’ information and its implications for consumption. The remainder of the paper is organized as follows. Section 2 describes our measures of bad and good news and their relation to the unemployment rate and measures of news from the Michigan Survey of Consumers. Section 3 discusses the extent of news coverage of economic events. Section 4 presents how agents’ information, agents’ expectations and agents’ consumption respond to bad and good news. Section 5 discusses the robustness of 4An exception is Casey and Owen (2013). In this paper, the opposite conclusion is reached: news significantly respond to positive forecast of GDP growth but not to negative forecast, a sort of positivity bias. 5ApartiallistofcontributionsincludesBeaudryandPortier(2004),BeaudryandPortier(2006),Cochrane (1994), Den Haan and Kaltenbrunner (2009), Forni et al. (2017), Jaimovich and Rebelo (2009), Barsky and Sims (2011), Schmitt-Groh´e and Uribe (2012)), Barsky and Sims (2012). 6In this respect, our work is closely related to Larsen and Thorsrud (2019) and Chahrour et al. (2021), which use textual information from newspapers to identify the news shock. 4

our main findings. Section 6 concludes. 2 The U-news indexes This section describes the construction of our news variables and discusses their time series properties. 2.1 Constructing the indexes We construct two novel measures of newspaper coverage of bad and good unemployment figures, which we refer to as the U-news+ index and the U-news− index, respectively. For thispurpose, weuseDowJonesFactiva, acomprehensivedatabaseofnewsarticles, andfocus our analysis to three major newspapers in the United States for the period from June 1980 to December 2019: The New York Times, The Wall Street Journal and The Washington Post. The choice of the three outlets is motivated by the fact that they have consistently appeared among the largest US newspapers by circulation during the period of interest and all aim for national audiences. We focus our attention on articles about the unemployment rate since this variable represents a major cyclical indicator and its fluctuations are closely monitored by the news media (see Fogarty (2005)). We construct the time series of bad news, U-news+, by counting the number of articles each month in which the word “unemployment” appears near to another word denoting an increase or high level. Similarly, for the good news variable, U-news−, we count the number of articles in which the word “unemployment” appears close to words denoting a decrease or low level.7 We then clean these two measures by subtracting from each of them the number of articles which are selected under both good and bad criteria. Thus, we explicitly exclude thosearticles(approximately6%ofthetotalsample)thatcannotunambiguouslybeclassified in one of the two categories. We acknowledge the fact that this class of news can also be of some interest, since this news may convey information about periods of relatively stable unemployment or reflect mixed signals about the labor market. However, for the purpose of the present study, which is concerned with potentially asymmetric effects of good and bad 7The index is similar in spirit to the R-Word index constructed by The Economist and to the media coverage series used in the seminal paper Soroka (2006). The difference with the R-Word index is that our searchisbasedonthewordunemploymentanddifferentiatesbetweenpositiveandnegativenews. Adetailed explanation of the search queries is included in the Online Appendix. Notice that our news variable is not a sentiment-basedindicator. Webelieveitwouldbeinterestingtotrytoconstructsuchanindicatorandstudy potential differences with ours. We plan to do this in the future. 5

news, it is of primary importance to have a clear measure of news polarization.8 The final dataset includes a total of 35933 bad news items and 22317 good news items over the period considered. Using the two raw indexes, we construct two additional variables. The first, which we call negativity, is the difference between the two indexes of bad and good news: U-tone = U-news+ −U-news−. If U-tone is positive, newspaper coverage of unemployment figures is prevailingly negative, and vice-versa. The variable is expected to be positively correlated with the unemployment rate and its average depends on the averages of good and bad news. The second variable is a measure of total information and it is defined as the sum of good and bad news: U-total = U-news++U-news−. 2.2 Descriptives In the left-hand column of Figure 1 we report our two news indexes (blue lines) together with the unemployment rate (red lines). The averages of bad and good news are, respectively, 76 and 47 articles per month, and the standard deviations are 46 and 20. News reporting of bad unemployment figures is, on average, higher and more volatile than the reporting of good unemployment figures. The most striking difference between the two indexes, however, is in terms of the correlation with the unemployment rate: 0.78 for bad news and -0.28 for good news. This is also clear from a simple visual inspection of the pattern of comovement of the two indexes with the unemployment rate. The measure of bad news, Unews+, tracks the unemployment rate extremely closely, with two major spikes of similar magnitude in correspondence of the early 1980s recession and the Great Recession. On the contrary, and quite surprisingly, the measure of good news, U-news−, seems largely unrelated to the unemployment rate, except in three episodes: the end of the 1980s, the end of the 1990s and after 2015. The news reporting of negative economic events appears substantially more cyclical than the coverage of positive economic events. We report the U-tone index in the bottom right-hand panel of Figure 1. As expected, negativity has a high correlation with the unemployment rate (0.79). The average negativity is 29 and statistically different from zero. An interesting feature of the U-tone index is that it leads the unemployment rate: negativity tends to anticipate both increases and decreases in the unemployment rate. This suggests that the articles we consider have informational content not only about the current, but also prospective developments in the unemployment 8The results presented below are robust to the inclusion of this ambiguous news. The reason is that this set of news is relatively small over the sample considered. 6

rate. The top right-hand panel of Figure 1 reports the U-total index together with the unemployment rate. The information content is countercyclical, with a correlation of 0.64 with the unemployment rate. This result could be seen as prima facie evidence supporting a larger degree of news coverage of bad economic events than good events. We explore the issue more formally in the next section. Apotentialconcernrelatedtotheconstructionofthenewsindexescouldbethatthethree newspapersconsideredmaycoverunemploymentdevelopmentsdifferently,dependingontheir political view. Figure 9 in the Online Appendix reports our news indexes disaggregated by newspaper. Overall, the coverage of both bad and good unemployment developments is remarkably consistent across different newspapers. Indeed, all of the indexes track each other very well over the sample period. The finding rules out the existence of a relevant political bias in the unemployment news reporting for the newspapers considered. Apossiblereasonfortheabsenceofarelevantnegativecorrelationbetweengoodnewsand theunemploymentratemightbethefactthatpositivenewsreferstoincreasesinemployment rather than decreases in unemployment. In Online Appendix A.2 we discuss the construction ofanalternativemeasureofgoodnews,theE-news+ index,basedontheword“employment”. Inasimilarwaytotheothertwomeasures,weselectarticlesinwhichtheword“employment” appears within a specified distance of another word denoting an increase or high level. We report this alternative measure together with the unemployment rate and with the U-news− index in Figure 10 in the Online Appendix. The correlation among U-news− and E-news+ is 0.14, while the correlation of E-news+ and unemployment is even positive (0.28). This suggests that the unemployment-based measure is more reliable and that its small negative correlation is not the result of a poor search strategy. 2.3 U-news indexes and consumer survey information At first glance, the relatively small procyclicality of the U-news− index might be puzzling, since a priori it would be reasonable to expect a pattern close to the reverse of the U-news+ index. In what follows, we compare our news indexes with other measures of news taken from the Michigan Survey of Consumers in order to assess the consistency of our measures with the information of the agents from the survey. The survey provides a wide variety of variables that reflect agents’ information and expectations about the current and future state of the economy. The variable NEWS in the survey corresponds to the percentage of individuals who recently heard of any favorable or unfavorable changes in business conditions. Question A6 of the questionnaire asks the fol- 7

lowing: “During the last few months, have you heard of any favorable or unfavorable changes in business conditions?”. There are two possible answers: “Yes” and “No, haven’t heard”. If the individual answers “Yes”, then the second question is A6a: “What did you hear?”, which is an open-ended question. The Michigan Survey provides few variables constructed on the basis of the type of answer to these two questions. Among those, we focus on the following variables: “No News”, which is the percentage of respondents choosing the corresponding option in question A6; “Favorable” and “Unfavorable”, which correspond to the percentage answeringpositivelyandnegativelytoquestionA6a; and“Favorable: employment”and“Unfavorable: unemployment”, corresponding to answers to question A6a which are specifically related to positive and negative evaluations of, respectively, employment and unemployment figures. While our indicators of bad and good news represent objective measures of the amount of negative and positive published news items related to unemployment figures, the corresponding two variables from the Michigan Survey represent the subjective information that theagentsperceivefromthemediaoralternativesourcesofinformation. Inprinciple, agents’ subjective information may not coincide with our measures of objective information. For example, agents may mostly get informed through other channels (TV, social networks, etc.) or they may be rational inattentive even in information-rich environments (see Sims (2003), Nimark and Sundaresan (2019)). The first column of Figure 2 illustrates our U-news+ and U-news− indexes together with the corresponding measures in the Michigan Survey of Consumers, namely the “Unfavorable: unemployment” and “Favorable: employment” items of NEWS. We uncover an interesting finding: both indexes track the corresponding variables of the Michigan Survey extremely closely over the sample considered. The correlation between “Unfavorable: unemployment” and U-news+ is 0.68, and the correlation between “Favorable: employment” and U-news− is 0.46. Overall, our indexes and the survey measures are consistent with each other. This suggests that newspaper information is a relevant channel for agents’ information. It could thus be important for shaping agents’ expectations and decisions. The second column of Figure 2 reports our measures of negativity and total information together with their counterparts constructed using the variables of the Michigan survey. As far as total information is concerned, the correlation between the two variables is 0.61, while the correlation is 0.65 for negativity. This again confirms the consistency between our newspaper measures and the survey measures. 8

3 Asymmetric coverage of economic events This section studies how news reporting relates to economic events. More specifically, we investigate how negativity and total information of unemployment news respond to positive and negative changes in the unemployment rate. To study asymmetries, we use a Threshold SVAR model (TSVAR). With respect to the simple regressions used in the political science literature, this type of model allows us both to address potential reverse causality issues and to capture interesting non-linear dynamics. The model per se is standard, but the way we use it is innovative. The main novelty is represented by the fact that the state variable in the model depends on the sign of the shock itself. Therefore,shocksofdifferentsignsimplydifferentdynamicssincethethresholdvariable is different. This feature, absent in standard TSVAR, is our methodological contribution and is discussed in detail below. 3.1 The model Let y be a time series vector including the variables of interest following t y = (1−F(z ))[a+A(L)]y +F(z )[b+B(L)]y +ε (1) t t t−1 t t−1 t where ε ∼ WN(0,Σ) is a vector of white noise residuals, A(L) = A +A L+...+A Lp−1 t 1 2 p and B(L) = B +B L+...+B Lp−1 are matrix polynomials in the lag operator L, z is a 1 2 p t scalarvariable, F(·)isafunctiontakingvaluezeroorone, andaandbarevectorsofconstant terms. We start from a minimal specification which includes in y , in this order, the unemt ployment rate change and either negativity (Section 3.2) or total information (Section 3.3) of unemployment news. In the robustness section, we use richer specifications and the results are largely unchanged. The state variable is the lag of the change in the unemployment rate, z = ∆U , whereU denotestheunemploymentrate. Thisensuresthatz isexogenouswith t t−1 t t respect to ε . We then set F(z ) = 0 if ∆U ≤ 0 and F(z ) = 1 if ∆U > 0. The choice t t t−1 t t−1 of the threshold variable is motivated by the fact that we are interested in understanding potential asymmetries in news dynamics to increases and reductions in the unemployment rate. Thus, A(L) are the VAR parameters governing the dynamics of the system of variables when the first lag of the unemployment rate change is negative, while B(L) are the VAR parameters in place when the change is positive. Under these assumptions, the model can be simply estimated using OLS. To test whether increases and reductions in the unemployment rate receive asymmetric 9

news coverage, we investigate the impulse response functions of either negativity or total information to an unemployment shock. To identify such shock, let S be the Cholesky factor of Σ, i.e. S is lower triangular and SS(cid:48) = Σ, and let u = S−1ε be a vector of t t orthonormal shocks. The first shock, u , is the innovation in the unemployment rate change 1t which is orthogonal to u , and it captures any factor that changes the unemployment rate 2t unexpectedly. Such a shock does not have any structural interpretation. It is a combination of the different structural disturbances that drive the one-month-ahead forecast error in the unemployment rate change. The impulse responses to this shock represent how the system of variables evolves if the unemployment rate change in the next month is higher or lower than expected. The fact that the unemployment shock has no structural interpretation does not represents a limitation from our perspective, since our aim is just to understand whether news coverage reacts differently to positive and negative innovations in the unemployment rate change, regardless of the nature of the underlying shock.9 Notice that, with this model specification, the sign of the innovation in ∆U becomes the t relevant state for the impulse response functions. To better understand the point, let β(L) = (I −B(L)L)−1S = β +β L+β L2+... 0 1 2 be the moving average representation of the model when ∆U > 0 and t−1 α(L) = (I −A(L)L)−1S = α +α L+α L2+... 0 1 2 when ∆U < 0. Call β˜(L) and α˜(L) the coefficients associated with u , i.e. the first row t−1 1t of β(L) and α(L) respectively. Due to our identification strategy, the impact effects are the same across regimes and do not depend on the sign of the shock, i.e. α˜ = β˜ = S , where 0 0 1 S is the first column of S.10 For the generic horizon h > 0, the responses to the shock will 1 be α˜ if the change in the unemployment rate in h−1 is negative, and β˜ if positive. If the h h responses of the change in unemployment rate are sufficiently persistent, then one can simply condition, as we do here, on the sign of the impact effect and the responses are β˜(L) for a positive shock and α˜(L) for a negative shock. To construct the confidence bands of the impulse responses, we use the bias-corrected 9OurapproachissimilartoDelNegroetal.(2020)thatidentifyanorthogonalinnovationinunemployment to study the effects of real business cycle shocks on economic variables in the context, however, of a linear VAR model. 10The assumption α˜ = β˜ = S is made for sake of interpretability of the results. The results are very 0 0 1 similartothoseobtainedintherestrictedmodelwhenwerelaxthisassumptionandweallowfortwodifferent impact effects. 10

estimator described in Kilian (1998), where we bootstrap the threshold variable, ∆U , t−1 together with the other regressors. 3.2 U-tone In the first specification, we set y = [∆U U-tone ](cid:48) and p = 2, as suggested by the BIC t t t criterion.11 The first two rows of Figure 3 report the results. The left-hand panels show the responses to negative (blue lines) and positive (red lines) shocks to the unemployment rate change. The solid lines are point estimates, while the dashed-dotted lines are 68% confidence bands. The right-hand panels report the sum of the impulse response functions (black lines) to positive and negative shocks. The solid line is the sum in the point estimates, while the dashed-dotted lines are the 68% confidence bands. This sum can be interpreted as a measure of asymmetry. Under perfect symmetry of the responses, the sum is zero. The larger (in absolute value) the sum is, the larger the degree of asymmetry is. Negativity reacts more, and with a higher degree of persistence, to an increase in the unemployment rate than to a reduction. Indeed, the asymmetry index is positive and significant over the horizon considered. The magnitude of this asymmetry is sizable. An increase in the unemployment rate of 0.15 percentage points on impact generates, on average over the horizonconsidered, about5morebadnewsitemsthangoodnewsitemspermonth. However, a reduction of the same magnitude generates less than one good news more than bad news items per month. This suggests that the negativity of media coverage reacts asymmetrically to economic developments, giving a substantially greater weight to negative events than to positive events. This result is in line with the findings in Soroka (2006) and Soroka et al. (2018). If our analysis was to stop here, we would confirm the existence of a negativity bias in newspaper coverageofeconomicevents. However,asnoticeablefromthefirstrowofFigure3,thereisalso a sizable and significant asymmetry in the effects on the unemployment rate change: positive shocks have larger and more persistent effects than negative shocks. So, when comparing the effects on media negativity of increases and decreases in the unemployment rate, the different dynamics of unemployment should be taken into account. Indeed, the larger response of negativity to an increase in the unemployment rate could simply be due to a larger and more prolonged effect of the positive shock on unemployment. We therefore compute a dynamic Media Multiplier of economic fluctuations. The multiplier is constructed as the cumulative sum of the impulse response functions of negativity 11Using the levels of the unemployment rate or using more lags yields very similar results. 11

divided by the cumulative sum of the changes in the unemployment rate at every horizon. For instance, at a horizon of 48 months ahead (the last horizon of the impulse response functions), the multiplier can be interpreted as the total number of bad news items in excess of good news items produced over four years following a 1 percentage point change in the unemployment rate. The responses are shown in the two bottom panels of Figure 3. The multipliersforincreasesandreductionsintheunemploymentrateareextremelysimilar, with no significant asymmetries. At the four year horizon, a 1 percentage point increase in unemployment generates 305 bad news items in excess of good news items, while a decrease of the same magnitude generates 292 good news items in excess of bad news items. The result suggests that, when nonlinearities in the dynamics of the unemployment rate are taken into account, the media bias towards bad events disappears. The result is new and contrasts with the evidence pointing to the existence of a negativity bias in economic news coverage (see Soroka et al. (2018) for a review). The reason our result differs substantially from previous findings in the literature is the fact that none of the earlier studies accounted for the asymmetry in the dynamics of unemployment. 3.3 U-total We repeat the analysis of the previous subsection, using model (1) with a different variable specification. Now, y = [∆U U-total ](cid:48). Apart from this, the model specification is identical t t t to the previous one. The first two rows of Figure 4 report the results. The left-hand panels report the responses to negative and positive shocks to the unemployment rate change. The right-hand panels report the sum of the impulse response functions to positive and negative shocks. The asymmetry between positive and negative shocks is clear. Shocks that push up unemployment increase total information substantially more, and with a higher degree of persistence, thanshocksthatimproveunemploymentfigures. Theasymmetryindexisalways significantoverthehorizonconsideredandthedifferencesaresizable. A0.15percentagepoint increase in the unemployment rate on impact generates up to 25 news items more than a 0.15 percentage point reduction. However, the shock, as for negativity, generates a marked nonlinearity in the response of the unemployment rate change, which is much more persistent for badshocksthanforgoodshocks. Asbefore,wecomputetheMedia Multiplier,i.e. were-scale the cumulative impulse response functions of total information by the cumulative change in unemployment. The responses are reported in the third row of Figure 4. When taking into account the dynamics of unemployment, the asymmetries in the news reporting process are 12

substantially dampened, the responses of total information to positive and negative shocks being essentially the same and the asymmetry index being never significantly different from zero. The conclusion of this first part of the analysis is that the apparent asymmetry in the news reporting process of economic events found in previous work does not depend on media bias per se. It depends on the large non-linearity in the unemployment rate response to economic shocks. Unemployment responds more, and with a higher degree of persistence, to bad shocks, i.e. shocks that imply an increase in unemployment. This triggers an important asymmetry in both negativity and total information of unemployment news. Tounderstandwhetherourresultsareconsistentwiththeevidencefrompreviousstudies, we run two simple linear regressions where the dependent variables are, respectively, our measures of negativity and total information of news, and the regressors are the current valueofpositiveunemploymentchanges,thecurrentvalueofnegativeunemploymentchanges and four lags of the dependent variable. This specification closely resembles the regression in Soroka(2006). TheresultsofthetworegressionsaredisplayedinTable1. Inbothregressions the coefficients associated with increases in unemployment are larger than those associated with a reduction, and only the former are significant. So, by neglecting the non-linearity in the response of the unemployment rate change, one would conclude, as previously done in the literature, in favor of a negativity bias in news reporting of economic events. Above we showed that the conclusion is different if asymmetries in the response of unemployment are also considered. 4 Asymmetric responses to news We now focus on the second type of asymmetry we want to test for: the response of agents’ consumption, information and expectations to bad and good news shocks, as measured by positive and negative unexpected changes in negativity. 4.1 The model The first problem we have to confront when assessing the role of bad and good news is that negativity is highly correlated with the unemployment rate: unemployment increases (reduces) and negativity increases (reduces). This implies that potential asymmetries could mistakenly be attributed to a different response of economic agents to bad and good news, while these actually arise simply because agents’ responses differ in the face of bad and good 13

economic shocks. To cope with this issue, we focus on changes in the component of negativity that are orthogonal to contemporaneous and past changes in the unemployment rate. We use the results obtained from Section 3.2 to obtain such component. There, the shock u has the 2t interpretation of a news shock: it triggers a change in negativity with a zero impact effect on the unemployment rate.12 Thus, the component of negativity generated by this shock is unrelated to current or past changes in the unemployment rate. This component is obtained from the TSVAR of Section 3.2 by filtering the shock u with the corresponding impulse 2t response functions of the two regimes:13 x = (1−F(z ))α (L)u +F(z )β (L)u , (2) t t 22 2t t 22 2t Notice that the news component of negativity, x , takes into account the asymmetric effects t of the unemployment shock documented in Section 3.2. That is, it controls for the fact that unemployment responds more, and with a larger degree of persistence, to positive unemployment shocks. This is particularly important since, as we have seen in the previous section, the negativity bias in economic news coverage disappears after controlling for this asymmetry. Using this component will therefore allows us to avoid confounding asymmetries due to news with other types of asymmetries associated with positive and negative changes in the unemployment rate. WethenestimateanewTVARmodel(1)withanalternativevariablespecificationsetting y = [∆x w ](cid:48), where w is a vector of time series of interest. Again, we select two lags of t t t t the dependent variable using the BIC criteria. The state variable is now the difference of the news component of negativity, z = ∆x . We define F(z ) = 1 if the change in t t−1 t negativity is positive, ∆x > 0, and F(z ) = 0 if the change in negativity is negative, t−1 t ∆x ≤ 0. The choice of the threshold variable is motivated by the fact that we are t−1 interested in understanding potential asymmetries to increases and reductions in negativity. With this specification, the coefficients A(L) in (1) are the VAR parameters governing the dynamics when the first lag of the difference in negativity is negative, while B(L) are the VAR parameters in place when the difference is positive. To test whether increases and reductions in the news components of negativity, ∆x , have t 12The identification of the news shock follows the seminal paper by Beaudry and Portier (2006), the main difference being that we use unemployment rather than TFP, so that the resulting shock can be interpreted as an unemployment-related news shock. 13Recallthatα (L)andβ (L)aretheelements(2,2)of,respectively,theimpulseresponsefunctionsα(L) 22 22 and β(L) obtained using the specification of Section 3.2. 14

asymmetric effects, we identify a news shock in this second specification. Since ∆x already t represents the news component of negativity, the news shock here is imply identified as the orthogonal innovation in ∆x . The implementation again entails a recursive decomposition. t Let S be the Cholesky factor of Σ, i.e. S lower triangular and SS(cid:48) = Σ, and let u = t S−1ε . The first shock, u , is the innovation in negativity which is orthogonal to u . Again, t 1t 2t conditional on a shock, the sign of the shock becomes the relevant state. When the shock u 1t is positive, ∆x is positive, and the relevant impulse response functions are the first column t ofβ(L) = (I−B(L)L)−1S, callitβ (L). Whentheshockisnegative, ∆x isnegativeandthe 1 t impulse response functions will be the first column of α(L) = (I−A(L)L)−1S, call it α (L). 1 Notice that our procedure is equivalent to an internal instrument approach using ∆x as an t instrument for the news shock, see Plagborg-Møller and Wolf (2021), since the instrument satisfies both the relevance and the exogeneity condition. Again, the Cholesky decomposition is just a statistical device to obtain the orthogonal innovation to the news variable. The shock admittedly lacks of any structural interpretation: it could be an economic news shock, a shock capturing any distortions in journalists view, a fake news shock, etc. However, independently on its nature, the shock represents an unexpected change in the negativity of unemployment news which is orthogonal to current and past unemployment. This is precisely the component we aim at disentangling in order to study the causality link from news to economic variables. One potential concern could be that other shocks which affect unemployment with a delay could be reflected into this component. In the robustness section, we estimate a richer specification which includes additional macroeconomic indicators or forward-looking variables in the model of Section 3.2 and show that the results are largely unchanged. To construct confidence bands for the impulse responses, we use the bias-corrected estimator described in Kilian (1998) and we bootstrap the threshold variable, ∆x , together t−1 with the other regressors. We use three different TVAR models to study the effects on personal consumption expenditures, consumers’ information and expectations. The choice of not using a single model with all of the variables is driven by parsimony considerations and to avoid the curse of dimensionality. 4.2 Consumption Inthissubsection, weuseourmodeltotestforpotentialasymmetriesintheresponsesofconsumption to good and bad news. To test for this asymmetry, we include in w the logarithms t 15

of real total personal consumption expenditures (PCE), real durable goods consumption expenditures (PCE Durable) and real non-durable goods consumption expenditures (PCE Nondurable). Figure 5 reports the effects of positive and negative shocks in negativity of news coverage on consumption. Figure 6 shows the asymmetry indexes. As before, solid lines are point estimates, while the dashed-dotted lines are 68% confidence bands constructed using the Kilian (1998) bias-corrected bootstrap. A clear-cut result emerges. A bad news shock significantly and persistently reduces the three types of consumption, especially of durable goods, while a good news shock has essentially no effects. The three asymmetry indexes are significantly negative at almost all of the horizons considered. We now re-scale the responses for the cumulative effect on negativity to take into account the potential non-linearity in the response of news itself. Again, the differences could simply be due to a larger increase in negativity following a bad shock. Figures 7 and 8 plot the normalized responses of the three types of consumption to negativity shocks and the corresponding asymmetry indexes. Asymmetries are still apparent, with the asymmetry indexes significantly negative over the horizons considered. Consumption reacts asymmetrically to positive and negative shifts in negativity. This result echoes previous evidence which documents a stronger response of consumption growth to predictable income declines than increases in the context of a single regression setup (see Shea (1995) and Bowman et al. (1999)). Inthenexttwosubsections, weshowthatthisresultisconsistentwiththeresponses of indicators of agents’ information and expectations. 4.3 Consumers’ information In our second specification, we include in w three variables of the Michigan Survey of Cont sumers related to consumers’ information (Questions A6 and A6a). The first variable is simply the difference between the percentage of “unfavorable” and “favorable” responses to question A6a. The second variable is the percentage of “No news” to question A6. This second variable measures the percentage of individuals who have not heard any news about current economic conditions and can therefore be interpreted as a proxy of the inverse of information. The third measure is the entropy associated to the answers in question A6 and A6a. Entropy can be interpreted as a proxy for consumers’ agreement about news and is constructed as follows. Let P be the sum of responses “No, haven’t heard” in question A6, t “Favorable” and “Unfavorable” in question A6a. Let p be the proportion of “Favorables” 1t 16

over P at time t and p the proportion of “Unfavorable” over P . Entropy is constructed as t 2t t e = −(p log(p )+p log(p )+(1−p −p )log(1−p −p )) t 1t 1t 2t 2t 1t 2t 1t 2t Thelargertheentropyis,thelargerthedisagreementamongagentsaboutthenewsheard,and vice-versa. Maximumdisagreementisreachedwhenrespondentsallocatethesameproportion (1/3) to each answer from the survey. Figure 5 reports the impulse response functions of the three variables, and Figure 6 reports the asymmetry indexes. Conditional on being informed, agents’ information reacts quite symmetrically to positive and negative changes in negativity. Indeed, the response of “Unfavorable” minus “Favorable” to a positive shock is essentially the mirror image of the response to a negative shock. This is reflected in the asymmetry index for this variable, which is mostly insignificant over the horizon considered. Following a bad news shock more respondents report that they heard unfavorable news relative to favorable news, and viceversa for good news shocks. Thekeydifferenceistheresponseof“Nonews”. Apositiveshiftinnegativitysignificantly increases the number of informed consumers. Indeed, the percentage of consumers reporting “No, haven’theard”decreases. Anegativeshift, ontheotherhand, significantlyincreasesthe numberofindividualswhohavenoinformation. Altogether, thisevidencesuggeststhatwhile bad news is informative, good news is not. A similar indication is obtained by inspecting the response of entropy. Bad news shocks increase agents’ agreement, while good news shocks increase disagreement.14 In conclusion, a rise in negativity of unemployment news increases consumers’ information and agreement, while a reduction has the opposite effect. Figures 7 and 8 plot the normalized responses and the corresponding asymmetry indexes. The main results are unchanged, confirming the above evidence suggesting that bad news is more informative and agents agree more in response to bad rather than good news. 4.4 Consumers’ expectations and confidence In the third specification, we add the logarithms of the current economic conditions index (ICC) and of the index of consumer expectations (ICE) from the Michigan Survey of Consumers in vector w . The two indexes are constructed using survey variables relative to t 14The same conclusion is reached if we construct entropy based on favorable and unfavorable news heard only, thus excluding the percent of respondents which state that they heard no news. In this case, maximum disagreement corresponds to an equal proportion (1/2) of respondents reporting that they heard unfavorable and favorable news. Results are available upon request. 17

expected current and future, personal and general economic conditions, and are components of the index of consumer sentiment. Figure5reportstheimpulseresponsefunctionsofthetwovariablestobadandgoodnews shocks. Figure 6 reports the asymmetry indexes. An increase in negativity has larger and more persistent effects on the two indexes of consumer sentiment than a reduction. Indeed, the asymmetry indexes reduce significantly and persistently over the horizon considered. Agents’ expectations react more to bad news than to good news. Figures 7 and 8 plot the normalized responses and the corresponding asymmetry indexes. Once we re-scale the responses for the potentially non-linear effect on negativity, the main resultsareunchanged. Thisfindingconfirmstheaboveevidence,suggestingthatexpectations indeed react more to a rise in negativity than to a decline. This result is in contrast to those obtained in Casey and Owen (2013) who find that the exogenous components of good and bad news have no effect on consumer confidence, but confirm the findings of Soroka (2006). The results presented in this section and Section 4.3 are consistent with the findings discussed in Section 4.2. A rise in negativity is much more informative than a decline, it makes agents revise their expectations more deeply and, consequently, their consumption path. Thisexpectationrevisioncannotbeexplainedbyahighernumberofnegativenews(for which we control here), as previously documented in the literature and as implied by models of sticky expectations (Carroll (2003)). On the contrary, the type of asymmetry we document suggests that, given an equal number of good and bad news items, agents give greater weight to negative information than positive information. The existence of a negativity bias in consumers’ response to news has been extensively discussed and studied in political science, biology and psychology (see Soroka (2014) and Baumeister et al. (2001)). In economics, the idea that agents may value losses more than equivalent gains is formalized in the concept of loss aversion. This could explain why agents are more attentive to signals (news) reporting a higher risk of utility losses than gains and react more to the former. Our findings are also consistent with the implications of the model in Tutino (2013) which features risk-aversion in an otherwise standard rational inattention setup. Risk-aversion implies that individuals in the model are more concerned with future decreases in their wealth than increases so that they allocate more attention and react faster and stronger to bad news than to good news. 18

5 Robustness We perform three main robustness checks on the models of Section 3 and Section 4. First, we estimate the models excluding the Great Recession period, using data up to December 2007. Results are reported in Figures 11-14 in the Online Appendix. The responses of negativity and total information are very similar to those obtained using the full sample. A positive change in unemployment causes a much larger and persistent increase in negativity and total information than a negative change. The two asymmetry indexes are always positive and significant. As far as the re-scaled responses are concerned, asymmetries are again mitigated, althoughfortotalinformationthedifferenceisstatisticallysignificant. ByexcludingtheGreat Recession, media negativity bias seems to be somehow more important when considering total information. The response of “no news”, entropy and the two confidence indexes are qualitatively similar to those obtained in the full sample, conveying the same message: bad news appears to be more informative, reduces disagreement and has a more marked effect on confidence. Total consumption and durable consumption still decrease to a greater extent in the face of negative news, while non-durable consumption is not responsive to negative or positive news. All in all, the results, although with some quantitative differences, depict a similar picture to that arising from the full sample case. Second, we repeat the analysis including in the TSVAR of Sections 3.2 and 3.3 also industrial production growth and PCE inflation, ordered after the unemployment rate and before negativity or total information. This means that the component of negativity we consider for the model of Section 4 is now unrelated to changes in the unemployment rate, as well as industrial production growth and PCE inflation. The rationale for this exercise is that the unemployment rate is a lagging variable, so the estimated news component in our baseline model could still include cyclical shocks which affect unemployment with some delay. Figures 15-18 in the Online Appendix report the results, which are very similar to the baseline specification. Third, we add to the baseline models of Sections 3.2 and 3.3 stock prices growth (as measured by the S&P500 index), ordered last. The component of negativity unrelated to unemployment now takes into account the inclusion of this forward-looking variable in the previousstep. Figures19-22intheOnlineAppendixpresenttheresults, whichareessentially unchanged compared to the baseline specification. 19

6 Concluding remarks We provide novel empirical evidence on the asymmetric relationship between economic news coverage, agents’ information and expectations, and macroeconomic dynamics. Using nonlinear SVAR techniques and two novel measures of newspaper coverage of bad and good economic events, we document three facts: (i) There is no significant negativity bias in newspaper coverage of the economy. News coverage is more responsive to negative than positive economicdevelopmentssbecausebadeconomicshockshavelargerandmorepersistenteffects on economic variables than good shocks; (ii) consumption, especially of durable goods, reacts to bad news but not to good news. This finding can be rationalized by the fact that (iii) bad news is more informative for agents than good news. Indeed, the percentage of informed individuals increases facing a rise in bad news relative to good news, while it decreases for the reverse. Bad news increases agents’ agreement about economic outcomes and modifies their expectations more than good news. A potential explanation for the existence of a negativity bias in the consumer’s reaction to news is loss aversion. In a world where the utility reduction induced by a loss is higher than the utility increase from a gain of the same amount, agents can be more attentive to economic news reporting a risk of losses than a risk of gains. Higher agents’ information can inturnleadtolargerconsumptionfluctuations. Weplantotestthisimplicationinourfuture research. 20

References Barsky, R. B., & Sims, E. R. (2011). News shocks and business cycles. Journal of monetary Economics, 58(3), 273–289. Barsky, R. B., & Sims, E. R. (2012). Information, animal spirits, and the meaning of innovations in consumer confidence. American Economic Review, 102(4), 1343–77. Baumeister, R. F., Bratslavsky, E., Finkenauer, C., & Vohs, K. D. (2001). Bad is stronger than good. Review of general psychology, 5(4), 323–370. Beaudry, P., & Portier, F. (2004). An exploration into pigou’s theory of cycles. Journal of monetary Economics, 51(6), 1183–1216. Beaudry, P., & Portier, F. (2006). Stock prices, news, and economic fluctuations. American Economic Review, 96(4), 1293–1307. Blinder, A., & Krueger, A. (2004). What does the public know about economic policy, and how does it know it? Brookings Papers on Economic Activity, 35(1), 327-397. Bowman, D., Minehart, D., & Rabin, M. (1999). Loss aversion in a consumption–savings model. Journal of Economic Behavior & Organization, 38(2), 155–178. Carroll,C.D. (2003). Macroeconomicexpectationsofhouseholdsandprofessionalforecasters. the Quarterly Journal of economics, 118(1), 269–298. Casey, G. P., & Owen, A. L. (2013). Good news, bad news, and consumer confidence. Social Science Quarterly, 94(1), 292–315. Chahrour, R., Nimark, K. P., & Pitschner, S. (2021, December). Sectoral media focus and aggregate fluctuations. American Economic Review, 111(12), 3872-3922. Cochrane, J. H. (1994). Shocks (Tech. Rep.). National Bureau of Economic Research. Del Negro, M., Lenza, M., Primiceri, G. E., & Tambalotti, A. (2020). What’s up with the phillips curve? Brookings Papers on Economic Activity, 301–357. Den Haan, W. J., & Kaltenbrunner, G. (2009). Anticipated growth and business cycles in matching models. Journal of Monetary Economics, 56(3), 309–327. Doms, M.E.,&Morin, N.J. (2004). Consumersentiment, theeconomy, andthenewsmedia. FRB of San Francisco Working Paper(2004-09). Fogarty, B. J. (2005, 07). Determining Economic News Coverage. International Journal of Public Opinion Research, 17(2), 149-172. Forni,M.,Gambetti,L.,Lippi,M.,&Sala,L.(2017). Noisynewsinbusinesscycles.American Economic Journal: Macroeconomics, 9(4), 122–52. Goidel, R. K., & Langley, R. E. (1995). Media coverage of the economy and aggregate economic evaluations: Uncovering evidence of indirect media effects. Political Research 21

Quarterly, 48(2), 313–328. Jaimovich, N., & Rebelo, S. (2009). Can news about the future drive the business cycle? American Economic Review, 99(4), 1097–1118. Kahneman, D. (1979). Prospect theory: An analysis of decisions under risk. Econometrica, 47, 278. Kilian, L. (1998). Small-sample confidence intervals for impulse response functions. Review of economics and statistics, 80(2), 218–230. Larsen, V. H., & Thorsrud, L. A. (2019). The value of news for economic developments. Journal of Econometrics, 210(1), 203–218. Larsen, V. H., Thorsrud, L. A., & Zhulanova, J. (2020). News-driven inflation expectations and information rigidities. Journal of Monetary Economics. Mankiw, N. G., & Reis, R. (2002). Sticky information versus sticky prices: a proposal to replace the new keynesian phillips curve. The Quarterly Journal of Economics, 117(4), 1295–1328. Mueller, H., & Rauh, C. (2018). Reading between the lines: Prediction of political violence using newspaper text. Mullainathan, S., & Shleifer, A. (2005). The market for news. American Economic Review, 95(4), 1031–1053. Nimark,K.P.,&Pitschner,S. (2019). Newsmediaanddelegatedinformationchoice. Journal of Economic Theory, 181, 160–196. Nimark, K. P., & Sundaresan, S. (2019). Inattention and belief polarization. Journal of Economic Theory, 180, 203–228. Plagborg-Møller, M., & Wolf, C. K. (2021). Local projections and vars estimate the same impulse responses. Econometrica, 89(2), 955–980. Schmitt-Groh´e, S., & Uribe, M. (2012). What’s news in business cycles. Econometrica, 80(6), 2733–2764. Shea, J. (1995). Myopia, liquidity constraints, and aggregate consumption: a simple test. Journal of money, credit and banking, 27(3), 798–805. Sims, C. A. (2003). Implications of rational inattention. Journal of monetary Economics, 50(3), 665–690. Soroka, S. N. (2006). Good news and bad news: Asymmetric responses to economic information. The journal of Politics, 68(2), 372–385. Soroka, S. N. (2012). The gatekeeping function: Distributions of information in media and the real world. The Journal of Politics, 74(2), 514–528. 22

Soroka, S. N. (2014). Negativity in democratic politics: Causes and consequences. Cambridge University Press. Soroka, S. N., Daku, M., Hiaeshutter-Rice, D., Guggenheim, L., & Pasek, J. (2018). Negativity and positivity biases in economic news coverage: Traditional versus social media. Communication Research, 45(7), 1078–1098. Tutino, A. (2013). Rationally inattentive consumption choices. Review of Economic Dynamics, 16(3), 421–439. Woodford, M. (2002). Imperfect common knowledge and the effects of monetary policy. P. Aghion, R. Frydman, J. Stiglitz, and M. Woodford, eds., Knowledge, Information, and Expectations in Modern Macroeconomics: In Honor of Edmund S. Phelps, Princeton: Princeton University Press.. 23

Appendix A.1 - U-news indexes We construct our U-news+ and U-news− indexes using newspaper articles from Dow Jones Factiva. We focus our search to three major US newspapers, in terms of circulation, namely The Wall Street Journal, The New York Times and The Washington Post, and to news related to the US economy over the time period from June 1980 to December 2019. For each newspaper, we look for all the articles, in a given month, in which the word “unemployment” appears within a predetermined distance, in any order, to another word that denotes a negative or positive development. More specifically, we first define two semantic groups, one containing words which share a root denoting an increase or high level (group 1) and another containing words which share a root denoting a decrease or low level (group 2): (cid:136) group 1. The words included in this group have one of the following roots: “high-”, “increas-”, “ris-”, “rose-”, “soar-”, “rais-” or “up-”. (cid:136) group 2. The words included in this group have one of the following roots: “down-” or “low-” or “slow-” or “decreas-”, “drop-”, “fall-”, “fell-”, “slip-”, “declin-”. We classify an article as a bad news item if the word “unemployment” appears within a 5-word distance to a word belonging to semantic group 1, but not within a 1-word distance to a word in semantic group 2. Symmetrically, we define an article as a good news item if the word “unemployment” appears within a 5-word distance to a word belonging to semantic group 2, but not within a 1-word distance to a word in semantic group 1. We choose the 5-word distance criteria to maximize the probability that the corresponding word in group 1 (bad news) or in group 2 (good news) is related to the word “unemployment” and not to other words. We obtain very similar results if we restrict this criteria to 4-word or 3-word distance. Giventhisfirstclassification, wethencleanourtwomeasuresofbadandgoodnews by substracting, for both measures, the number of articles that can be classified as belonging to both groups according to our criteria. In fact, this class of articles cannot be clearly classified as positive or negative, either because these articles deliver mixed signals about unemployment,15 sothattheirresultingtoneisneutral,orbecausetheword“unemployment” is incidentally mentioned close to a word in group 1 and group 2, even if the article does not 15For example, on the 12th of March 2010, The Wall Street Journal writes “[...] initial claims for unemployment insurance dropped to 462,000 in the week ended March 6th, down 6,000 from the week before. Meanwhile,thenumberofpeoplecollectingunemploymentchecksrose37,000to4.6millionintheweekending Feb. 27”. 24

include direct information about unemployment (e.g. articles reporting presidential talks close to the elections). The articles belonging to this last category represent on average 6% oftotalarticlesovertheperiodconsidered. Aftercleaningthemeasures,thenumberofallbad news in a given month is the value of the U-news+ index for that month, while the number of all good news in a given month is the value of the U-news− index for that month. A.2 - Alternative search An alternative measure of good news can be derived based on the word “employment” as opposed to “unemployment”. We define the variable E-news+ as the total number of articles, in each month, in which the word “employment” appears within a distance of 5 words to a word denoting an increase or high level, i.e. to a word belonging to semantic group 1, according to the definition in Appendix A.1. As before, we clean this measure by removing all the articles that are selected under both good and bad search criteria. 25

Tables U-tone U-total Estimate t-stat Estimate t-stat ∆U > 0 29.66∗ 2.37 28.09∗ 2.25 t ∆U < 0 2.30 0.19 14.84 1.18 t Lag 1 0.44∗ 9.30 0.55∗ 11.65 Lag 2 0.30∗ 6.05 0.19∗ 3.57 Lag 3 0.18∗ 3.54 0.07 1.37 Lag 4 -0.01 -0.24 0.08 1.83 Constant 0.53 0.32 12.06∗ 3.49 Note: ∗ means significant at the 5% significance level. Table 1: Regressions of U-tone and U-total on their first four lags and the current positive/negative change in unemployment. 26

Figures 300 250 200 150 100 50 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN 10 9 8 7 6 5 4 % U-news+ and Unemployment 350 300 250 200 150 100 50 1985 1990 1995 2000 2005 2010 2015 U-news+ Unemployment Rate selcitra fo rebmuN 10 9 8 7 6 5 4 % U-total and Unemployment U-total Unemployment Rate 120 100 80 60 40 20 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN 10 9 8 7 6 5 4 % U-news- and Unemployment 250 200 150 100 50 0 -50 1985 1990 1995 2000 2005 2010 2015 U-news- Unemployment Rate selcitra fo rebmuN 10 9 8 7 6 5 4 % U-tone and Unemployment U-tone Unemployment Rate Figure 1: Bad news, good news and unemployment. Upper-left panel: Total number of bad news related to unemployment (U-news+) and the unemployment rate. Bottom-left panel: Total number of good news related to unemployment (U-news−) and the unemployment rate. Upper-right panel: Total information related to unemployment (U-total) and the unemployment rate. U-total is computed as the sum of U-news− and U-news+. Bottom-right panel: Negative tone in news coverage of unemployment (U-tone) and the unemployment rate. U-tone is computed as the difference between U-news− and U-news+. 27

300 250 200 150 100 50 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN 70 60 50 40 30 20 10 stnednopser fo % U-news+ and Michigan Unfavorable Unemployment News 350 300 250 200 150 100 50 1985 1990 1995 2000 2005 2010 2015 U-news+ Michigan Unfavorable Unemployment News selcitra fo rebmuN 70 60 50 40 30 20 stnednopser fo % U-total and Michigan Information U-total Michigan Information 120 100 80 60 40 20 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN 35 30 25 20 15 10 5 stnednopser fo % U-news- and Michigan Favorable Employment News 250 200 150 100 50 0 -50 1985 1990 1995 2000 2005 2010 2015 U-news- Michigan Favorable Employment News selcitra fo rebmuN 60 40 20 0 stnednopser fo % U-tone and Michigan Tone U-tone Michigan Tone Figure 2: Bad news, good news and the Michigan Survey. Upper-left panel: U-news+ and percentage of respondents in the Michigan Survey who heard unfavorable news about unemployment. Bottom-left panel: U-news− and percentage of respondents in the Michigan Survey who heard favorable news about employment. Upper-right panel: U-total and total information from the Michigan Survey. U-total is the sum of U-news− and U-news+. Michigan Information is the number of respondents in the Survey who heard either positive news about employment or negative news about unemployment. Bottom-right panel: U-tone and the negative tone from the Michigan Survey. U-tone is the difference between U-news− and U-news+. Michigan Tone is the difference between the number of respondents in the Survey who heard negative news about unemployment and those who heard positive news about employment. 28

0.15 0.1 0.06 0.05 0 0.04 -0.05 0.02 -0.1 -0.15 0 0 10 20 30 40 0 10 20 30 40 10 10 8 5 6 4 0 2 -5 0 -2 0 10 20 30 40 0 10 20 30 40 100 200 50 0 0 -50 -200 -100 -150 -400 0 10 20 30 40 0 10 20 30 40 Figure 3: Response of news coverage to positive (red) and negative (blue) changes in unemployment - U-tone. Asymmetry Indexes in the right column are computed as the algebraic sumsbetweenIRFstopositiveandnegativechangesinunemployment. TheMedia Multiplier in the bottom left panel is computed as the cumulative sum of the IRFs of U-tone divided by the cumulative sum of the IRFs of the unemployment rate at every horizon. 29

0.15 0.1 0.1 0.08 0.05 0.06 0 0.04 -0.05 -0.1 0.02 -0.15 0 0 10 20 30 40 0 10 20 30 40 30 30 25 25 20 20 15 15 10 10 5 5 0 0 0 10 20 30 40 0 10 20 30 40 500 200 400 150 300 100 200 50 100 0 0 -50 0 10 20 30 40 0 10 20 30 40 Figure 4: Response of news coverage to positive (red) and negative (blue) changes in unemployment - U-total 30

0 0 -0.2 -0.05 -0.4 -0.1 -0.6 0 10 20 30 40 0 10 20 30 40 0.2 1 0 0.5 -0.2 0 -0.4 -0.5 0 10 20 30 40 0 10 20 30 40 4 0.01 2 0 0 -2 -0.01 -4 0 10 20 30 40 0 10 20 30 40 1 0.5 0.5 0 0 -0.5 -0.5 -1 -1 0 10 20 30 40 0 10 20 30 40 Figure 5: Asymmetric effects of news - IRFs to positive (red) and negative (blue) changes in the exogenous component of U-tone. From left to right and from top to bottom: Personal Consumption Expenditure (PCE), PCE of Durable and Non-Durable Goods, respondents in theMichiganSurveywhoheardnonewsaboutunemployment(NoNews), differencebetween numbers of respondents who heard negative and positive news (Unfavorable-Favorable), Entropy of Survey responses, Index of Consumer Expectations and Confidence (ICE and ICC). 31

0 0 -0.2 -0.05 -0.4 -0.6 -0.1 0 10 20 30 40 0 10 20 30 40 0 0.5 -0.1 -0.2 0 -0.3 -0.4 -0.5 0 10 20 30 40 0 10 20 30 40 10-3 1 0 -2 0 -4 -1 -6 -2 -8 0 10 20 30 40 0 10 20 30 40 0 0.2 0 -0.5 -0.2 -0.4 -0.6 -1 -0.8 0 10 20 30 40 0 10 20 30 40 Figure 6: Asymmetric effects of news - Asymmetry Indexes 32

10-3 2 0 0 -2 -0.02 -4 -6 -0.04 -8 0 10 20 30 40 0 10 20 30 40 0.15 0.02 0.1 0 0.05 0 -0.02 -0.05 0 10 20 30 40 0 10 20 30 40 10-3 0.4 0.2 1 0.5 0 0 -0.2 -0.5 -0.4 -1 -0.6 0 10 20 30 40 0 10 20 30 40 0.1 0.05 0.05 0 0 -0.05 -0.05 -0.1 -0.1 0 10 20 30 40 0 10 20 30 40 Figure7: Asymmetriceffectsofnews-Normalized-Foreachhorizon,IRFsaredividedbythe corresponding response of ∆x to positive and negative changes in the exogenous component t of U-tone. 33

10-3 0 0 -0.02 -5 -0.04 -0.06 -10 0 10 20 30 40 0 10 20 30 40 0.01 0.1 0 0.05 -0.01 0 -0.02 -0.05 0 10 20 30 40 0 10 20 30 40 10-4 5 0 0 -0.2 -5 -0.4 0 10 20 30 40 0 10 20 30 40 0 0.05 -0.02 -0.04 0 -0.06 -0.08 -0.05 0 10 20 30 40 0 10 20 30 40 Figure 8: Asymmetric effects of news - Normalized - Asymmetry Indexes 34

Online Appendix 150 100 50 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN U-news+ 200 150 100 50 1985 1990 1995 2000 2005 2010 2015 U-news+ - NYT U-news+ - WP U-news+ - WSJ selcitra fo rebmuN U-total U-total - NYT U-total - WP U-total - WSJ 50 40 30 20 10 0 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN U-news- 150 100 50 0 1985 1990 1995 2000 2005 2010 2015 U-news- - NYT U-news- - WP U-news- - WSJ selcitra fo rebmuN U-tone U-tone - NYT U-tone - WP U-tone - WSJ Figure 9: Bad news and good news by newspaper 35

60 55 50 45 40 35 30 25 20 15 10 1985 1990 1995 2000 2005 2010 2015 selcitra fo rebmuN 10 9 8 7 6 5 4 % E-news+ and Unemployment 60 55 50 45 40 35 30 25 20 15 10 1985 1990 1995 2000 2005 2010 2015 E-news+ Unemployment Rate selcitra fo rebmuN 120 110 100 90 80 70 60 50 40 30 20 selcitra fo rebmuN E-news+ and U-news- E-news+ U-news- Figure 10: Alternative measure of good news - E-news+ 36

0.1 0.04 0.05 0.02 0 -0.05 0 -0.1 -0.02 0 10 20 30 40 0 10 20 30 40 10 10 8 5 6 0 4 2 -5 0 0 10 20 30 40 0 10 20 30 40 400 150 200 100 0 50 0 -200 -50 0 10 20 30 40 0 10 20 30 40 Figure 11: Response of news coverage to unemployment changes - U-tone - Sample excluding the Great Recession (1980:06 - 2007:12) 37

0.15 0.08 0.1 0.05 0.06 0 0.04 -0.05 0.02 -0.1 -0.15 0 0 10 20 30 40 0 10 20 30 40 15 15 10 10 5 5 0 0 0 10 20 30 40 0 10 20 30 40 500 400 400 300 300 200 200 100 100 0 0 0 10 20 30 40 0 10 20 30 40 Figure12: Responseofnewscoveragetounemploymentchanges-U-total-Sampleexcluding the Great Recession (1980:06 - 2007:12) 38

0.05 0.2 0 0 -0.05 -0.2 -0.1 -0.4 -0.6 -0.15 -0.8 0 10 20 30 40 0 10 20 30 40 2 0.1 0 1 -0.1 -0.2 0 -0.3 0 10 20 30 40 0 10 20 30 40 2 0.01 0 -2 0 -4 -6 -0.01 0 10 20 30 40 0 10 20 30 40 0.5 0 0 -0.5 -0.5 -1 -1 0 10 20 30 40 0 10 20 30 40 Figure13: Asymmetriceffectsofnews-IRFs-SampleexcludingtheGreatRecession(1980:06 - 2007:12) 39

0 0 -0.2 -0.05 -0.4 -0.1 -0.6 -0.15 0 10 20 30 40 0 10 20 30 40 1.5 0.1 0 1 -0.1 0.5 -0.2 0 -0.3 0 10 20 30 40 0 10 20 30 40 10-3 10 2 0 5 -2 0 -4 0 10 20 30 40 0 10 20 30 40 0 -0.2 0 -0.4 -0.6 -0.5 -0.8 -1 0 10 20 30 40 0 10 20 30 40 Figure 14: Asymmetric effects of news - Asymmetry Indexes - Sample excluding the Great Recession (1980:06 - 2007:12) 40

0.06 0.1 0.05 0.04 0 -0.05 0.02 -0.1 0 0 10 20 30 40 0 10 20 30 40 10 10 5 5 0 0 -5 0 10 20 30 40 0 10 20 30 40 400 200 200 100 0 0 -200 -100 0 10 20 30 40 0 10 20 30 40 Figure 15: Response of news coverage to unemployment changes - U-tone - Including Industrial Production growth and PCE Inflation 41

0.1 0.08 0.05 0.06 0 0.04 -0.05 0.02 -0.1 0 0 10 20 30 40 0 10 20 30 40 30 30 20 20 10 10 0 0 0 10 20 30 40 0 10 20 30 40 500 400 200 300 100 200 100 0 0 0 10 20 30 40 0 10 20 30 40 Figure 16: Response of news coverage to unemployment changes - U-total - Including Industrial Production growth and PCE Inflation 42

0.2 0 0 -0.05 -0.2 -0.4 -0.1 0 10 20 30 40 0 10 20 30 40 0.1 0.8 0.6 0 0.4 -0.1 0.2 0 -0.2 -0.2 -0.4 0 10 20 30 40 0 10 20 30 40 4 0.01 2 0.005 0 0 -2 -0.005 -0.01 -4 0 10 20 30 40 0 10 20 30 40 1 0.5 0.5 0 0 -0.5 -0.5 -1 -1 0 10 20 30 40 0 10 20 30 40 Figure 17: Asymmetric effects of news - IRFs - Including Industrial Production growth and PCE Inflation 43

0 0 -0.05 -0.2 -0.4 -0.1 -0.6 0 10 20 30 40 0 10 20 30 40 0.1 0.8 0.6 0 0.4 0.2 -0.1 0 -0.2 -0.2 0 10 20 30 40 0 10 20 30 40 10-3 1 2 0 0 -1 -2 -4 -2 0 10 20 30 40 0 10 20 30 40 0 0 -0.5 -0.5 -1 -1 0 10 20 30 40 0 10 20 30 40 Figure18: Asymmetriceffectsofnews-AsymmetryIndexes-IncludingIndustrialProduction growth and PCE Inflation 44

0.15 0.06 0.1 0.05 0.05 0.04 0 0.03 -0.05 0.02 -0.1 0.01 -0.15 0 0 10 20 30 40 0 10 20 30 40 10 10 8 5 6 4 0 2 0 -5 -2 0 10 20 30 40 0 10 20 30 40 300 100 200 50 100 0 0 -100 -50 -200 -100 -300 0 10 20 30 40 0 10 20 30 40 Figure 19: Response of news coverage to unemployment changes - U-tone - Including Stock Prices growth 45

0.15 0.1 0.08 0.05 0.06 0 0.04 -0.05 0.02 -0.1 -0.15 0 0 10 20 30 40 0 10 20 30 40 30 30 20 20 10 10 0 0 0 10 20 30 40 0 10 20 30 40 500 250 200 400 150 300 100 200 50 100 0 0 -50 0 10 20 30 40 0 10 20 30 40 Figure 20: Response of news coverage to unemployment changes - U-total - Including Stock Prices growth 46

0.05 0 0 -0.2 -0.05 -0.4 -0.6 0 10 20 30 40 0 10 20 30 40 0.2 0.6 0.4 0 0.2 0 -0.2 -0.2 -0.4 0 10 20 30 40 0 10 20 30 40 10-3 2 5 0 0 -5 -2 -10 -4 0 10 20 30 40 0 10 20 30 40 0 0 -1 -1 -2 -2 0 10 20 30 40 0 10 20 30 40 Figure 21: Asymmetric effects of news - IRFs - Including Stock Prices growth 47

0 0 -0.2 -0.05 -0.4 -0.1 -0.6 0 10 20 30 40 0 10 20 30 40 0.1 0.5 0 -0.1 0 -0.2 -0.5 0 10 20 30 40 0 10 20 30 40 10-3 1 2 0 0 -2 -1 -4 -2 -6 0 10 20 30 40 0 10 20 30 40 0 0 -1 -1 -2 -2 0 10 20 30 40 0 10 20 30 40 Figure 22: Asymmetric effects of news - Asymmetry Indexes - Including Stock Prices growth 48