key: cord-0051359-xcyqlwwd
authors: Cookson, J Anthony; Engelberg, Joseph E; Mullins, William
title: Does Partisanship Shape Investor Beliefs? Evidence from the COVID-19 Pandemic
date: 2020-09-29
journal: Rev Asset Pricing Stud
DOI: 10.1093/rapstu/raaa018
sha: 4a6951400479c3359889c6696c18b27a88304706
doc_id: 51359
cord_uid: xcyqlwwd

We use party-identifying language—like “liberal media” and “MAGA”—to identify Republican users on the investor social platform StockTwits. Using a difference-in-difference design, we find that partisan Republicans remain relatively unfazed in their beliefs about equities during the COVID-19 pandemic, while other users become considerably more pessimistic. In cross-sectional tests, we find Republicans become relatively more optimistic about stocks that suffered the most during the COVID-19 crisis, but more pessimistic about Chinese stocks. Finally, stocks with the greatest partisan disagreement on StockTwits have significantly more trading in the broader market, explaining 28% of the increase in stock turnover during the pandemic.

The COVID-19 pandemic presented investors with a complex valuation problem. With the first sign of community spread in the United States, investors were faced with a series of questions: How quickly would the virus spread? How deadly would it be? How would the government respond? How long would the pandemic last? How quickly would the economy recover? Could the pandemic present an opportunity for some firms? With fundamental uncertainty about each of these factors, even among experts, investors had to update their expectations about firms' future prospects.

In this paper, we show these investor expectations in the wake of COVID-19 can be predicted from their political identity as measured before COVID-19. Specifically, we find that partisan Republicans become more optimistic than other investors when the crisis begins and remain more optimistic through the end of April 2020. We also show that political identity shapes views among the cross-section of stocks during the pandemic:

Republicans become more pessimistic about U.S.-listed Chinese firms (e.g., Baidu and Alibaba), while remaining more optimistic about firms that experienced the greatest losses.

We also find that stocks with the greatest partisan disagreement see the greatest increase in stock turnover during the COVID-19 period.

Partisan differences influence beliefs across a host of issues (Milner and Judkins, 2004; Gaines et al., 2007) and have increased dramatically in the last 30 years (Bishop, 2008; Abramowitz and Saunders, 2008; Gentzkow, Shapiro, and Taddy, 2019; Kaplan, Spenkuch, and Sullivan, 2019) . It is nonetheless surprising to see partisan identity matter when forming stock expectations. Investors have a strong financial incentive to form correct beliefs about a stock's future cash flows regardless of political affiliation. If partisan identity does not help investors update expectations about a firm's prospects, then their political identity should be ignored when forming these expectations. Nevertheless, we find a widening and sustained difference in investor beliefs between partisan Republicans and other investors beginning with the onset of the COVID-19 pandemic.

We employ novel data from StockTwits, a popular investor social network, to measure partisan identity at the individual level, and observe investment beliefs at a daily frequency.

StockTwits users explicitly stamp individual messages with bullish or bearish sentiment, which provides a direct measure of their investment beliefs. We observe precisely when these declarations of sentiment are made, allowing us to track the evolution of investor beliefs through the COVID-19 pandemic. Critically, to link investor beliefs to partisanship, we observe individuals' partisan affiliation as revealed by their use of political language in StockTwits posts prior to the pandemic. Our classification, which follows an approach similar to Gentzkow and Shapiro (2010) , cleanly identifies partisan Republican individuals who use the platform. 1 We identify how partisanship shapes investor beliefs during the COVID-19 period using a difference-in-difference design. We compare the difference in optimism between Republicans and non-Republicans (the first difference) before and during the COVID-19 period in the United States (the second difference). In support of our empirical approach, we observe parallel pre-trends for investment beliefs in the pre-COVID-19 period (e.g., Figure 4 ), followed by a sharp divergence in investor beliefs after the first suspected case of community spread of the virus in the United States.

Our core finding is that partisan Republicans remain, on average, more optimistic about equities than other users during the COVID-19 period. The optimism of partisan Republicans closely tracks that of other users from October 2019 through February 2020. However, after the first case of U.S. community spread of COVID-19 in late February, partisan Republicans became significantly more optimistic than other StockTwits users, a difference of 2-3 percentage points across all stocks and 4-5 percentage points for firms in the S&P 500.

Because we include user-security fixed effects, these estimates reflect changes in optimism within user about the same security through the COVID-19 period. This partisan optimism 1 We also validate our individual classification of partisan users using belief updating around the 2016 election. Consistent with the evidence in Meeuwis et al. (2019) , we find that partisan Republicans become significantly more optimistic than other StockTwits users around the 2016 election of Donald Trump. Moreover, this partisan gap in optimism rises at critical junctures during the Trump presidency: the onset of the U.S.-China trade war, the 20% market drawdown of late 2018, and the primary focus of this study, the COVID-19 crisis in the United States. gap held throughout the COVID crisis, as market valuations reached their bottom in late March and began to recover in April. 2 We also perform heterogeneity tests to illuminate the underlying source of partisan investor optimism during the pandemic. First, we evaluate whether partisan Republicans became more optimistic than other users about stocks that suffered the largest losses during the COVID-19 period. Greater optimism about stocks that lost value during the crisis reflects a belief in a quick stock market recovery. Consistent with this view, our tests reveal that partisan Republicans are disproportionately more optimistic about stocks that fell the most during the COVID-19 period.

Second, we examine Republicans' relative beliefs about U.S.-listed Chinese stocks (e.g., Baidu or Alibaba) during the COVID-19 crisis. President Trump repeatedly referred to COVID-19 as the "Chinese virus" in public statements, and singled out China's lack of forthright communication about the seriousness of the virus in the early stages of the pandemic (Higgins, 2020) . Consistent with politically driven negativity about China affecting the investment beliefs of partisan Republicans, we find that they are significantly more pessimistic about Chinese stocks during the outbreak in the United States. The timing of Republican beliefs about Chinese stocks is instructive: they did not become more pessimistic about these stocks during the COVID-19 outbreak in China; instead, Republicans became pessimistic about them in mid-March, when new cases in China had fallen, but the crisis in the United States was deepening. This points to a political, rather than economic, model for their beliefs. 3

The political divide in investor beliefs during COVID-19 coincided with an enormous increase in trading volume: at the height of the pandemic, abnormal daily stock turnover 2 Our difference-in-difference design recovers the partisan gap in investment beliefs that emerges during the pandemic. We do not, however, take a position on whether Republicans or non-Republicans have correct beliefs during the pandemic. Our goal is to identify the role of political identity in shaping investor beliefs during the COVID-19 pandemic.

3 Pessimism about Chinese firms could reflect the belief that Western politicians would place economic restrictions on these companies. However, in accordance with this view, non-Republicans should also become more pessimistic about Chinese stocks. Viewed in this light, our finding of differential pessimism by Republicans is difficult to reconcile with purely economic beliefs. had increased by approximately 36% relative to pre-pandemic levels. In our final set of tests, we measure partisan disagreement at the daily stock level and relate abnormal stock turnover to partisan disagreement among the cross-section of stocks. Our estimates imply a tight connection: a one standard deviation increase in partisan disagreement during the pandemic leads to 10% more abnormal stock turnover, which is 28% of its increase during the COVID-19 period, and is greater than the trading implied by a standard deviation increase in overall disagreement.

Our analysis primarily focuses on how partisan investment beliefs diverge during the COVID-19 period, but our measure of partisan Republican sentiment accurately reflects political events outside of this period. We find that Republicans became relatively more optimistic about equities After the 2016 Presidential Election and during the U.S.-China Trade War that began in early 2018 and culminated with a near 20% market drawdown in December 2018. These key moments of politically based divergence in investor beliefs on StockTwits also appear in Internet search volume. Google Search Intensity for "Trump" together with "stock market" spikes for each of these events: the 2016 election, the U.S.-China trade war and the COVID-19 pandemic. In fact, Google Search Intensity for "Trump and stock market" is greater during the COVID-19 period than around the 2016 election, suggesting that the pandemic is an especially important period to investigate partisan disagreement in investor beliefs. Moreover, the fact that these spikes in Google search volume coincide with a widening of the partisan optimism gap on StockTwits provides support for the external validity of our partisan classification, as does our finding of a clear link between partisan disagreement and stock market turnover. 4

Our central contribution is to show an individual link between partisan identity and investor beliefs. Our findings most closely relate to Meeuwis, Parker, Schoar, and Simester (2019) , which shows that individuals in Republican areas invest their retirement assets 4 Buckman et al. (2020) provides a daily news sentiment index that extends through the COVID-19 period. In Figure A .7 in the Internet Appendix, we find that the expectations of the majority of StockTwits users closely track this index during 2020 and for several years before (a correlation of around 0.8). more aggressively after the 2016 election, consistent with partisan investment optimism.

Our analysis of investment beliefs during the pandemic period is substantively different in at least two ways. First, the COVID-19 period was not only a shock to the level of economic activity, 5 but also a shock to uncertainty (e.g., Baker et al. 2020a ). Thus, our finding of partisan optimism in the face of the pandemic is different than showing that Republican investors view a Republican policy regime as favorable to stock market valuations, and suggests that people default to core identities, such as political affiliation, when facing significant uncertainty. 6 Second, our evidence is from direct observation of investment beliefs, expressed on StockTwits, whereas Meeuwis et al. (2019) infer differences in investment beliefs from changes to portfolio holdings.

Our research also contributes to the emerging literature on how partisan identity shapes financial beliefs (e.g., Kempf and Tsoutsoura, 2019) . Our party identification and sentiment measures from StockTwits have four advantages. First, our use of high-frequency data on belief revisions enables us to describe the partisan dynamics at a daily or weekly frequency, cleanly identifying the effect of an event on beliefs. Second, we provide an individual link between partisanship and direct declarations of investment beliefs rather than relying on geography to infer partisan identity. Geography is correlated with many variables besides political party, whereas individual declarations, such as #Trump2020, are unequivocal. 7 Third, we show important cross-sectional differences in partisan belief formation: our finding of Republicans' pessimism about Chinese stocks shows that these partisan investor beliefs do not reflect general economic optimism, but instead a more nuanced alignment of investor beliefs with partisan philosophy. Finally, we connect these 5 U.S. gross domestic product decreased at an annual rate of 32.9% in the second quarter of 2020, according to the BEA advance estimate. 6 The uncertainty identity theory of Hogg (2007) posits that people cling to their social identity as a way to reduce uncertainty. Relatedly, Bénabou and Tirole (2011) develop a model in which investments in one's identity (e.g., political identity) are important for shaping beliefs. 7 Using geography to proxy for partisan identity risks conflating partisanship with other omitted factors that relate to investor belief formation, for example, social connections (Bailey et al., 2018b) . Meeuwis et al. (2019) addresses this measurement concern by ruling out hedging needs and initial differences and use survey data to show that Republicans are more optimistic about the national economy, but not about their own economic situations. beliefs to outcomes in the overall market, specifically the extent of daily stock turnover.

Our research also relates to the literature on belief formation (Bailey et al., 2018a) and sources of disagreement (Cookson and Niessner, 2020) . Our main findings suggest that partisan identity affects how people update their market beliefs upon the arrival of new public information. In showing that a noninformational factor drives differential belief updating, we provide new evidence that investors apply different models to interpret market information (e.g., Kandel and Pearson, 1995) . The political friction we identify is distinct from other views about inefficient belief updating, such as extrapolation (Bordalo, Gennaioli, and Schleifer, 2018) or motivated beliefs (Brunnermeier and Parker, 2005; Benabou, 2015) , and is complementary to work on selective exposure to confirmatory information (Cookson, Engelberg, and Mullins, 2020) . Relative to existing work on how investors form beliefs, we show that partisan identity can lead to significant differences, which is surprising given recent evidence that typical investor characteristics do not explain much variation (Giglio et al., 2020a) . 8 A survey-based literature in political science robustly shows that people express greater optimism about the economy when their partisan position matches that of the Government (e.g., Bartels (2002) ; Evans and Andersen (2006) ). A related series of papers argues that this link between partisanship and economic perceptions affects outcomes such as spending or economic activity Huber, 2009, 2010; Gillitzer and Prasad, 2018; Benhabib and Spiegel, 2019) , but recent work has challenged this claim (McGrath, 2017; Mian, Sufi, and Khoshkhou, 2017) . Our paper contributes to this debate on the real effects of partisan perceptions by providing evidence that partisan disagreement materially increases stock market trading volume.

Though both political and investment beliefs have been studied in the context of COVID-19, our work is the first to connect the two. For example, several articles have shown that 8 Our research also relates to, but is distinct from, the literature on media slant (Gentzkow and Shapiro, 2010) . Goldman, Gupta, and Israelsen (2020) show Republican firms are covered more favorably in the Wall Street Journal than in the New York Times. Political bias in media coverage may contribute to our observed differences in investor beliefs. political beliefs affect real activities, such as social distancing compliance (Allcott et al., 2020; Barrios and Hochberg, 2020; Painter and Qiu, 2020 ) and subsequent infection rates (Burstyn et al., 2020) . In addition, an emerging literature studies how COVID-19 has affected household consumption (Baker et al., 2020b) , risk preferences (Bu et al., 2020) , expectations (Hanspal, Weber, and Wohlfart, 2020) , and belief updating (Giglio et al., 2020b) . Our findings draw a connection between political and investment beliefs during COVID-19 by showing a strong divergence in stock market beliefs between Republicans and non-Republicans, and a relationship between this partisan-based disagreement and the cross-section of trading volume.

In this section, we describe the StockTwits data, describe our approach to identifying partisans among StockTwits users, and provide some initial evidence linking partisanship and investor beliefs.

We employ message-level data from the investor social network, StockTwits. Founded in 2008, StockTwits claims to be the "largest social network for investors and traders, with over two million registered community members and millions of monthly visitors." The platform is similar to Twitter. Panel A of Figure 1 shows the user interface. Users post messages of up to 1,000 characters and use "cashtags" with the stock symbol (e.g., $AAPL or $BTC for Apple or Bitcoin) to link the user's message to a particular company. Cashtags allow users to aggregate opinions about particular stocks or other assets in a broader discussion, just like hashtags on Twitter.

Although StockTwits users are not a fully representative sample of investors, the opinions expressed on StockTwits have been shown to have external reliability. For example, prior work has linked dispersion of opinion on StockTwits to overall trading volume in stocks: both Cookson and Niessner (2020) and Giannini, Irvine, and Shu (2018) show that proxies for dispersion of sentiment relate to market-level trading volume. Because of its unique data features, such as social connections and high-frequency belief updating, profile information on location, and investor approaches, StockTwits has begun to attract academic attention (Giannini, Irvine, and Shu, 2017; Cookson, Engelberg, and Mullins, 2020) .

We have the full history of messages posted to StockTwits through April 2020. We restrict attention to messages that mention only one ticker to focus on sentiment that can be directly linked to a specific stock. Panel A of Table 1 A valuable feature of StockTwits for academic research is that the platform encourages users to self-classify whether their sentiment is bullish or bearish for each message. Users can by click on a prominently displayed button on the StockTwits interface before posting (e.g., see panel A of Figure 1 ). Following Cookson, Engelberg, and Mullins (2020) , we aggregate posts to the user-stock-day level to prevent a user with multiple posts about a security on a single a day from having inordinate weight. In addition, old messages cannot be deleted on StockTwits. This feature not only preserves the incentives of users to post truthful best forecasts for their follower base but also ensures that the data we extract from StockTwits reflect an unselected view of how users viewed the market at each date in our sample. Using a battery of analyses of text and market events, Cookson and Niessner (2020) 

In addition to the self-classified sentiment about investments, StockTwits users sometimes discuss other topics, including politics. These additional posts are useful for measuring individual users' partisan affiliation. We follow Gentzkow and Shapiro (2010) and identify a list of keywords that flag posts as political.

We begin by identifying partisan Republicans. The process is as follows: we consider all posts before 2020 that expressed direct support for President Trump's reelection via the terms #Trump2020 and #MAGA. For the users who created these posts, we examine all of their other posts, looking for terms that this group uses at least 20x more frequently than other users. From these posts, we select purely political terms that meet this condition, such as "Stupid dems" and "Leftists." Then, we identify users who employ this expanded set of terms frequently. Adding these users to our set of partisan Republicans, we identify additional political terms that this expanded group uses more frequently than other users.

We continue to iterate in this way until no new political phrases emerge. We take a similar approach to identify partisan Democrats, beginning with the terms "Idiot Trump" and "Stupid Trump." Panel A of Table 2 reports the final list of keywords, which contain distinctively partisan Republican or Democrat language, such as "Liberal Media" and "Russia Hoax" among Republicans, and "Faux News" and "#ImpeachTrump" among Democrats. The vast majority of the posts containing keywords are unambiguously partisan. Panel B of Table 2 reports several examples of partisan messages. To be clear, this classification of users is "out of sample" in two respects: (1) partisan users are identified via pre-2020 messages, distinctly prior to our COVID-19 sample period and (2) the sentiment-stamped messages in our analysis rarely contain political language. Because the classification assigns partisanship at the user level, we do not rely upon the combination of messages that contain sentiment and political statements. Table 1 presents summary information on our classification procedure. Applied to the 179.2 million StockTwits messages before 2020, a total of 70,031 messages contain partisan keywords (18,371 are stamped with bullish or bearish sentiment). Of the 780,908 users who were active on StockTwits prior to 2020, we identify 34,284 users who make or like at least one partisan message.

For each user, we count the number of Republican (R) and Democrat (D) messages that they either posted or liked on StockTwits before 2020. We measure each user's partisanship as the difference between the two (R − D). For our main analysis we identify partisan Republicans as those who have an R − D of at least four. In other words, we require that, before 2020, a user has either posted or liked at least four messages with political sentiment and that the number of Republican messages outnumber the number of Democrat messages by at least four. Table A .1 demonstrates that other cutoffs (e.g., R − D ≥ 3 or R − D ≥ 5) yield similar results. Our R − D ≥ 4 constraint helps ensure that the individuals we identify are truly and persistently partisan Republican, and identifies 6,191 users in the pre-2020 period. Restricting attention to the sample pe-riod (Oct 2019-Apr 2020), our sample contains 3,448 Republican-identified users and a comparison group of 115,986 other users who were active during this period. 10 Of course, many of these other users are also Republicans who do not regularly use or like Republican language on StockTwits. To the extent that our comparison group includes nonpartisan users and some partisan Republicans missed by our classification, our estimates potentially understate the true size of the partisan gap. Finally, although partisan Republicans comprise only 3% of the users in our sample, they tend to be more active, making up 12% of the user-symbol-day level observations.

Since the 2016 Presidential Election, President Trump has strengthened the connection between political identity and beliefs about the stock market. He has tweeted about the "stock market" 130 times through May 24, 2020, and often cites the rise of the stock market as a political accomplishment while cheering market milestones. 11 Meeuwis et al. (2019) highlights a connection between partisan identity and investment beliefs, showing that the 2016 Presidential Election led investors from pro-Trump zip codes to invest more aggressively, whereas investors in pro-Clinton zip codes did the opposite.

As validation of our classification of investor partisanship, we evaluate the sentiment of StockTwits messages beginning in January 2015. To construct a time series of sentiment by partisan affiliation, we estimate the following specification:

in which the dependent variable is an indicator for whether user j is bullish about stock s on day t, and η s, j are user-security fixed effects which absorb each user's average senti-10 Most of our regressions consider the set of securities that have a permno and drop singleton observations. With these restrictions, we have 2,754 partisan Republican and 66,634 other users. 11 The tweets can be found at the online Trump Twitter Archive (https://bit.ly/3cBxbmN). it appears unlikely that our classification is picking up the views of an unrepresentative population; it seems more likely that they represent broader partisan views given that they coincide with the revealed interest of the Google-user population (Google had approximately 250 million users in the United States in 2019). 13

We focus on the COVID-19 period for several reasons. First, as we saw in panel A of Figure   2 , the sharpest divergence of partisan investor beliefs occurs during the pandemic. Second, the COVID-19 period also exhibits the largest amount of attention to "Trump and Stock

Market," indicating that the connection between politics and financial markets has been particularly salient during the pandemic. Third, as we show in Section 3, the COVID-19

period exhibited significant market turmoil, with especially high trading volume. Divergence of opinion is one potential explanation for this volume increase, and, as we will

show, partisan differences of opinion are an important explanation for this rise in trading.

To provide evidence on the timing of the pandemic shock, we estimate a version of Equation (1), with daily fixed effects from January 2020 through April 2020. Panel A of Figure 3 presents the estimated daily sentiment fixed effects, after sweeping out user-stock fixed effects. This series shows that partisan Republicans and other users on StockTwits exhibit similar investor belief dynamics from the beginning of January through the beginning of March, but as of early March, partisan Republicans become significantly more optimistic than other users. Panel B shows that the divergence in beliefs corresponds closely to when Google Search volume for "Trump and Stock Market" spikes. 13 Our finding of a strong link between partisan disagreement and stock market turnover (in Table 6 ) also supports the external validity of our partisanship measure. Similarly, in A.7 in the Internet Appendix, we find that the expectations of the majority of StockTwits users closely track the daily news sentiment index from Shapiro, Sudhof, and Wilson (Forthcoming) and Buckman et al. (2020) during the COVID-19 period and for several years before (correlations of around 0.8).

This section presents regression evidence of how the investor beliefs of partisan Republicans diverge from other users on StockTwits through the COVID-19 period.

We estimate the partisan gap in investor optimism during COVID-19 by focusing on the period from October 2019 through April 2020. Using the sample of user-security-day observations of sentiment-stamped declarations about single stocks, we estimate the following monthly difference-in-difference specification:

in which the dependent variable Bull s, j,t is an indicator for whether user j is bullish about stock s on day t (multiplied by 100 to aid interpretation as a percentage). η s, j are usersecurity fixed effects that absorb the average sentiment of each user about each security in the sample period. The month fixed effects (γ B m ) yield time-varying sentiment estimates for the baseline group. The coefficients of interest are β R m , which give the month-by-month differences of partisan Republican investor sentiment relative to that of other users. We cluster standard errors by user to account for serial correlation in sentiment within user. Table 3 , the difference in the beliefs of partisan Republicans in the COVID-19 period is larger when we condition on large capitalization stocks in the S&P 500. 14 Republicans remaining optimistic during the pandemic might suggest that they are optimists in general, which could drive our result. To address this concern, in column 7 of Table A.2 in the Internet Appendix, we classify the 28% of users who always declare bullish sentiment pre-2020 as "Pre-COVID-19 Optimists." We estimate a single difference-indifferences coefficient on 1Covid t × 1PartisanR j , rather than month-by-month coefficients other users from 1.9 percentage points in the baseline specification in column 1 to 2.1 percentage points. 

where the dependent variable Bull s, j,t (multiplied by 100 for ease of interpretation) is an indicator for whether user j is bullish about stock s on day t, and η s, j is a user-stock fixed effect to absorb cross-user heterogeneity in sentiment across each stock. 1Covid t is an indicator for the COVID-19 period (March and April centage points. The triple interaction is also substantially larger than the baseline 2 percentage point optimism gap we observe in the difference-in-differences coefficient 1 Covid t × 1 PartisanR j . 17 16 In support of this, Figure A .2 in the Internet Appendix presents the daily time series for Google Search volume for the term "Chinese virus." The time series has two peaks: One around the time of the Wuhan lockdown (January 25th), likely reflecting curiosity about a then-unnamed virus. The second peak occurred around a series of tweets by President Trump that mentioned the term "Chinese Virus" (March 19th), for example, see Figure A .3. The second peak arguably occurs when China's role became politicized in the United States and coincides with the largest partisan differences in beliefs about Chinese stocks (see Figure  5 ). 17 A potential concern is that our partisan classification selects StockTwits users who have expressed greater interest in China and are abnormally attuned to matters that affect Chinese stocks, which may contribute to the pessimism we observe. To address this concern, we count the number of pre-2020 posts or likes that mention the terms "China," "Chinese," or "Trade War" for each user and add an indicator for users interested in these topics interacted with the COVID-19 dummy. As we report in Table A .2 (columns 2-6) in the Internet Appendix, controlling for users' intensity of interest in China or the Trade War has little impact on In addition, to show when this negativity emerged, we estimate a version of Equation (3) that replaces the 1Covid t indicator with a series of weekly fixed effects. Figure Returning to the potential concern that our classification of partisan Republican could simply proxy for optimism, in Table A .3 in the Internet Appendix we also examine Pre-COVID-19 Optimists' views on U.S.-listed Chinese stocks. We find that they are differentially optimistic about U.S.-listed Chinese stocks, unlike partisan Republicans, who are differentially pessimistic (column 4 vs. column 3). This further supports our conclusion that partisan Republican is not a stand-in for optimism.

In column 2, we report estimates from the specification in Equation (3) Given COVID-19 was a massive, economywide shock, it is interesting to see Republicans' differential optimism concentrated in large-cap stocks. While plenty of small-cap and micro-cap firms with high idiosyncratic volatility are discussed often on StockTwits (e.g., Aurora Cannabis or Virgin Galactic), these were not the stocks where political disagreement manifested during the pandemic. Instead, it appeared in stocks that best represented beliefs about the market in general, like large, bellwether stocks in the S&P 500.

Column 3 confirms this interpretation. At the end of 2019, we run a year-long marketmodel regression with daily returns and recover the fraction of variation explained by the market (R-squared). We then estimate the triple difference specification in Equation (3) with each stock's market model R-squared as the interaction term. The result is clear:

Republican disagreement during the COVID-19 period is concentrated among stocks with the highest share of systematic variance. Stocks with high levels of idiosyncratic variance were not the playing field of partisan disagreement during the pandemic.

Finally, another possible manifestation of the partisan divergence of opinions during COVID-19 is that partisan Republicans expect a faster and more complete economic recovery than other users, so that they would be more optimistic about stocks that lost the most value.

To evaluate this possibility, we estimate the triple difference specification in (3) using the stock return over the preceding month (from 21 trading days before t to the trading day preceding t, denoted month return s,t−1 ) as the interaction variable. If partisan Republicans are more optimistic about stocks that recently lost value, we would expect a negative triple difference coefficient. 18 Table 4 presents the results from estimating this specification in column 3, restricting attention to S&P 500 stocks where partisan differences in opinion are clearest. Consistent with partisan Republican belief in a faster and more complete recovery, we estimate that partisan Republicans are around 10.7 percentage points more optimistic about the worstperforming firms during the This optimism about firms that recently lost value runs counter to the usual relation between recent market returns and investor beliefs on StockTwits, which typically exhibit significant momentum (indicated by the positive and significant estimate on the baseline coefficient month return s,t−1 ). More than leading partisan Republicans to reduce their typical proclivity toward momentum, the results in this table indicate that partisan Republicans' investor beliefs became contrarian with respect to recent market movements. 20

In this section, we connect differences in partisan investor beliefs to daily stock turnover, which increased considerably during the pandemic. Figure 6 plots the daily percentiles (10th, 25th, 50th, 75th, and 90th) of stock turnover from January 2019 through April 2020.

Consistent with the timing of the onset of the COVID-19 crisis, daily stock turnover sharply increased around the beginning of March 2020 and remained high through the end of our sample. In addition to this increase in daily stock turnover, a similarly large increase occurs in the cross-sectional spread in daily turnover across firms.

Next, we turn to relating partisan differences in political beliefs to trading at the stock-day level. For this analysis, we construct a difference of opinion measure between partisan Republicans and other users. Following the approach in Cookson and Niessner (2020) for 19 The estimated coefficient for the triple interaction equals 13.7%, multiplied by the worst-performing firm's loss of 78.3% (Halliburton -see Table A .6), which equals our reported magnitude of 10.7. 20 Table A .6 reports the 10 worst-performing S&P 500 stocks for the period from January 1st through the market bottom on March 23rd. These stocks are in the energy sector (Halliburton and Schlumberger), airlines (United Airlines and Boeing Co.), and major retail (Macy's and Kohl's). the two-group case, we measure partisan disagreement at the stock-day level as 

For our analysis of stock turnover in the cross-section, we focus on a daily panel of stock information from March 2019 through April 2020. Table 5 presents summary information about this sample. We estimate the effect of partisan disagreement on daily stock turnover using the following specification:

AbnormalLogTurnover s,t = β 1 1Covid t + β 2 Overall Disagree s,t and abnormal log turnover for day t − 1. Standard errors are double clustered by stock and day to account for, respectively, serial and within-day correlation in the errors. Table 6 presents the results from estimating Equation (5). In column 1, we present a benchmark specification that quantifies the rise in stock turnover during the pandemic: we estimate that the COVID-19 period has 36% more abnormal stock turnover, after accounting for stock fixed effects, which is consistent with the univariate evidence in Figure 6 . In column (2), we include number of impression fixed effects to absorb user activity at the stock-day level, accounting flexibly for differences in attention (news, press releases, etc.).

Controlling for attention in this way reduces the estimated coefficient on 1Covid t by more than half, to 16%.

In column 3, we also include Overall Disagree s,t and 1Covid t × Overall Disagree s,t .

Consistent with the literature, we see that overall disagreement correlates strongly with stock turnover: a standard deviation increase in disagreement is associated with 5% greater abnormal turnover outside of the COVID-19 period. However, during the pandemic, the relation between disagreement and turnover more than doubles in magnitude, increasing by almost 9 percentage points.

In column 4, we add Partisan Disagree s,t and 1Covid t × Partisan Disagree s,t to the specification. Our estimates imply that there is no relation between partisan disagreement and trading volume before COVID-19 (est = −0.008, se = 0.009). However, dur-ing the pandemic, a standard deviation increase in partisan disagreement is associated with 10% greater stock turnover, which is 28% of the baseline rise in stock turnover during the COVID-19 period (column 1), and 50% of the effect of attention on turnover. 22 The estimated magnitude of the 1Covid t × Partisan Disagree s,t coefficient is greater than the baseline coefficient for overall disagreement (0.101 vs. 0.058). Moreover, this increase in the sensitivity to partisan disagreement during COVID-19 reduces the magnitude of the 1Covid t × Overall Disagree s,t coefficient, and renders it statistically insignificant.

The remaining columns show that these inferences about the relationship between partisan disagreement and the cross-section of stock turnover are not sensitive to including day fixed effects (column 5) or to adding control variables often employed in the literature (column 6).

Our findings and empirical design draw a tight connection between partisan disagreement and abnormal stock turnover that is unlikely to be driven by other factors. Our main coefficient of interest is β 5 on the 1Covid t × Partisan Disagree s,t term, which compares the abnormal stock turnover of high partisan disagreement stocks to that of low partisan disagreement stocks, before versus during the COVID-19 period. 23 In this context, we observe a strong link between partisan disagreement and abnormal stock turnover only after the emergence of COVID-19. Further, our tests draw a comparison between stock-days after removing differences in financial attention (by including fixed effects for the number of user impressions). Thus, any potential omitted variable must (a) be orthogonal to the number of users who post opinions about a stock on a particular day, and (b) uniquely emerge as a confounder during the COVID-19 period. 24

In our final column we take advantage of two unique characteristics of the increase in 22 From columns 1 and 2, the inclusion of number of impressions fixed effects reduces the magnitude of the estimate on 1Covid t by 20.2 percentage points. The estimated magnitude on the 1Covid t × Partisan Disagree s,t term is 10.1, which is 50% of this drop. 23 Figure A .6 presents a lead and lag plot at the monthly frequency from a specification that replaces the 1Covid t indicator with monthly fixed effects. The plot shows parallel pre-trends, with a positive coefficient that emerges in February, March, and April. 24 Moreover, we address the reverse causality concern that trading today causes disagreement today by estimating a lagged disagreement specification in Table A .5 in the Internet Appendix. The results are similar. partisan disagreement during the pandemic: (a) it reflects additional disagreement among investors, and (b) because it does not generate a trading motive (beyond increasing overall disagreement), it has no direct relationship with turnover. In a standard IV setting, point (a) constitutes the relevance condition and point (b) constitutes the exclusion restriction. Thus, we use partisan disagreement to investigate the causal impact of disagreement on trading volume, removing standard confounders such as the arrival of news. To the extent that the disagreement literature is interested in the disagreement-turnover elasticity to measure the relative contribution of disagreement and liquidity motives in generating trading volume (e.g., Kandel and Pearson, 1995; Kruger, 2020) , our setting provides a unique opportunity to estimate this parameter.

Specifically, we use partisan disagreement as an instrument for overall disagreement, and column 7 of Table 6 reports the second-stage instrumented coefficients. 25 The IV estimate is 0.23 during the COVID-19 period (0.045 + 0.185), significantly greater than the OLS magnitude of 0.142 (0.053 + 0.089) in column 3. Thus, a one-standard-deviation increase in disagreement leads to 23% greater abnormal stock turnover. 26 To estimate an elasticity we run a log-log IV regression in Table A .4 in the Internet Appendix and obtain an abnormal turnover to disagreement elasticity of 0.66, with a standard error of 0.086. 27

Taken together, this table demonstrates that partisan differences in investment beliefs contribute to the sharp rise in trading that emerged during the COVID-19 period. Our findings suggest that partisanship not only shapes investment beliefs but also influences the extent of trading in the broader market. 25 The first-stage regressions and the ordinary least squares analogue are reported in Table A .4 in the Internet Appendix. The relevance condition is strongly satisfied. 26 Our instrumented estimate pre-COVID-19 is 4.5%, consistent with OLS estimates in the literature (e.g., Cookson and Niessner, 2020) . The interaction with the pandemic period indicator suggests a heightened sensitivity of turnover to disagreement. 27 In fact, we use an inverse-hyperbolic-sine transformation of the overall (and partisan) disagreement variables to account for zero-valued observations.

Our paper provides evidence of a partisan divide in investor beliefs that emerges during the COVID-19 pandemic. Using novel data from StockTwits, we find that partisan Republicans remain significantly more optimistic about equities than other investors and that this pattern persists through April 2020. Consistent with the narrative that partisanship shapes investor beliefs, Republicans express a nuanced pattern of investment beliefs: they are more optimistic about stocks that recently lost value but more pessimistic about U.S.-listed Chinese stocks. The partisan disagreement we document explains 28% of the abnormal trading volume during the COVID-19 period.

It should surprise no one that Democrats and Republicans disagree. Partisans predictably disagree about environmental policy, abortion, immigration and gun rights, among other wedge issues. By contrast, partisan disagreement about equities during the COVID-19 pandemic is surprising, particularly given how unhelpful partisan identity is for equity valuation. After all, disagreement about stocks during COVID-19 should reflect disagreement about the virulence of the virus, its rate of spread, likely government response, its effectiveness, and related epidemiological issues.

The fact that we find a partisan divide in investor beliefs perhaps reflects the unprecedented heights of political polarization we have reached (Abramowitz and Saunders, 2008; Bishop, 2008; Gentzkow, Shapiro, and Taddy, 2019) . Political identity has become increasingly relevant for choices we make (Gerber and Huber, 2009; Chen and Rohla, 2018;  McCartney and Zhang, 2019) and beliefs we hold (Bartels, 2002; Gerber, Huber, and Washington, 2010) . Our analysis begs several questions: Will the partisan divide that emerged during the COVID-19 pandemic continue to shape investor beliefs and market outcomes after the health crisis is over? Or, if partisan investor disagreement subsides, can we expect partisan disagreement to reemerge when investors face the uncertainty of the next crisis?

We leave these questions for future research. This plot shows how differences in optimism between partisan Republicans and other StockTwits users evolve over time from January 2015 to April 2020; the omitted (reference) period is January 2015. Panel A presents estimated monthly fixed effects-separately for partisan Republicans and other users as a baseline-from a model with user-security fixed effects (following Equation (1) This plot shows how differences in optimism between partisan Republicans and other StockTwits users evolve over time from January 2020 until April 2020; the omitted (reference) period is January 1, 2020. Panel A presents the time series of daily fixed effects-separately for partisan Republicans and other users as a baseline-from a model with user-security fixed effects following Equation (1). It also includes the level of the S&P 500 index on the right axis (fixing it at the level of the preceding trading day on nontrading days). Panel B presents the rolling 1-week average of Google Search Intensity for "Trump and stock market" for the same period (January 2020 to April 2020). crisis. Each panel displays the estimated coefficients on the interaction between an indicator for whether a user is a partisan Republican and monthly fixed effects from Equation (2), which includes user-stock fixed effects. The sample follows that used in Table 4 1 M ar 2 0 1 9 1 7 M ar 2 0 1 9 7 A p r2 0 1 9 2 8 A p r2 0 1 9 1 9 M ay 2 0 1 9 9 Ju n 2 0 1 9 3 0 Ju n 2 0 1 9 2 1 Ju l2 0 1 9 1 1 A u g 2 0 1 9 1 S ep 2 0 1 9 2 2 S ep 2 0 1 9 1 3 O ct 2 0 1 9 3 N o v 2 0 1 9 2 4 N o v 2 0 1 9 2 2 D ec 2 0 1 9 (b as el in e) 5 Ja n 2 0 2 0 2 6 Ja n 2 0 2 0 1 6 F eb 2 0 2 0 8 M ar 2 0 2 0 2 9 M ar 2 0 2 0 1 9 A p r2 0 2 0

Week starting Figure 6 : Percentiles of daily stock turnover Tables   Table 1: Summary statistics on StockTwits data   This table presents A presents the list of keywords used to flag partisan Republican tweets on StockTwits, which is the result of an iterative procedure that follows the language-based relative frequency approach of Gentzkow and Shapiro (2010) . Panel B presents six example tweets flagged by this list of keywords: three Republican tweets and three Democrat tweets. In the iterative process that generates this keyword list, we seed the list of Republican keywords with "#MAGA" and "#Trump2020" and add terms to the list if they are commonly used by individuals who write posts containing these initial keywords. If these terms relate to the stock market (e.g., "S&P surging") or are apolitical, we do not add them to the list. We repeat this iterative process to populate the partisan Republican keywords until we obtain a stable set of individuals identified as partisan Republican. We follow the same procedure to construct the list of Democrat keywords, starting instead with "Idiot Trump" and "Stupid Trump." A StockTwits user is identified as a partisan Republican if they post or like at least four more Republican tweets than Democrat tweets.

Republican keywords Democrat keywords "#MAGA" "The Liberals" "Drumpf" "Orange Colored" "Russia Hoax" "Russian Collusion" "Trump Nationalism" "Idiot in Chief" "#TRUMP2020" "Stupid Dems" "Trumptard" "Criminal POTUS" "Hussein Obama" "Leftists" "Trump is a liar" "Trump is an idiot" "Obummer" "Trump Derangement" "Idiot Trump" "Clown Trump" "Fake News Media" "The Socialist" "Faux News" Imbecile Trump "Crooked Hillary" "MAGA 2020" "Clown Child" "Trump is an Imbecile" "Snowflake" "The Commie" "Stupid Trump" Orange Scum" "Liberal Media" "Libtard" "Pig Clown" "Scumpig Clown" "Libs" "Stupid Democrats" "Liar in Chief" "Lying Trump" "Trump Hater" "Sleepy Joe" "Liar Trump" "#IMPEACHTRUMP" "Typical Liberal" "Liberal Democrat" "#F***TRUMP" "Liberal Agenda" "You Liberal" "Your Liberal"

Republican example messages October 10, 2018 "Fox News... This crash will teach those libtards!! $spy October 27, 2019 "Therapy bro, Trump derangement syndrome is no joke. Get some meds" July 8, 2019

"I probably won't be alive to see it but the US is a short step to being a socialistic country. Only one election away. Vote TRUMP 2020 or else"

Democrat example messages January 29, 2018 "$WYNN the only one less popular than Wynn now is the orange colored scumpig clown child masquerading as potus. BEARISH"

July 20, 2018 $GM drumpf is killing this stock November 2, 2018 "Glad to see the Manipulator in Chief saw Apple earnings on Faux News scrambling for a deal now to try to save markets; very stable genius! This table examines whether partisan Republicans exhibit greater optimism than other users through the COVID-19 crisis. The dependent variable is an indicator (multiplied by 100 to aid interpretation as a percentage) that a user j declares as bullish about stock s on day t. The specification (following Equation (2)) includes monthly fixed effects, and their interactions with an indicator for whether a user is a partisan Republican. The monthly fixed effects show the time series of sentiment for baseline users, whereas the PartisanR j × month interactions show the extent to which partisan Republicans are differentially optimistic in that month. Column 1 is estimated on the top 1,042 securities by message volume on StockTwits (which includes nonstocks, such as Bitcoin); column 2 restricts attention to stocks; column 3 restricts attention to stocks above the 25th percentile of NYSE market capitalization as of December 31, 2019; and column 4 includes only stocks in the S&P 500 as of March 1, 2020. The sample is at the user-security-day level and runs from October 2019 to April 2020. Standard errors clustered by user are reported in brackets. *p <.1; **p <.05; ***p <.01. (3). The dependent variable is an indicator (multiplied by 100 to aid interpretation as a percentage) that user j declares as bullish about stock s on day t. The 1Covid t variable is an indicator equal to one in March and April 2020; the 1PartisanR j indicator is equal to one for partisan Republican users. The Interaction variable varies by column: column 1 examines whether partisan differences in sentiment are different for U.S.-listed Chinese firms; column 2 examines whether they are different for large firms (S&P 500 firms as of February 29, 2020). Column 3 replaces the interaction with a continuous variable, the R 2 from a market model run with daily returns over the whole of 2019, to examine whether differential sentiment was driven by stocks that reflected beliefs about the market in general. Finally, column 4 examines whether partisan differences in sentiment are different for stocks based on their returns over the preceding month (also a continuous interaction). The sample is at the user-stock-day level, runs from March 2019 to April 2020, and covers 930 stocks (the subset of securities that have CRSP permnos out of the top 1,042 StockTwits securities by messages since 2013). Column 4 is run on S&P 500 firms in our sample (148 stocks). Standard errors clustered by user are reported in brackets. *p <.1; **p <.05; ***p <.01. AbnormalLogTurnover s,t = β 1 1Covid t + β 2 OverallDisagree s,t + β 3 (1Covid t × OverallDisagree s,t ) +β 4 PartisanDisagree s,t + β 5 (1Covid t × PartisanDisagree s,t ) + FE + δControls s,t + ε s,t AbnormalLogTurnover s,t is the difference between log turnover on day t and the average log turnover fromt − 140 to t − 20 trading days (6-month period, skipping the most recent month) for stock s. The 1Covid t indicator equals one after February 2020. OverallDisagree s,t is the standard deviation of stamped messages with sentiment (bullish = 1, bearish = −1), while PartisanDisagree s,t is the average divergence in sentiment between partisan Republicans and other users, following Equation (4). Both disagreement measures are normalized to have a mean of zero and a unit standard deviation. Stock (permno) Fixed effects (FE) are included in all regressions; number of impressions and day fixed effects are included in some. Number of impressions s,t is the number of users who tweet with sentiment about each stock each day. Controls include abnormal log turnover on day t − 1; volatility, measured as the standard deviation of abnormal returns over days t − 5 to t − 1; and cumulative abnormal returns measured over days t − 30 to t − 6 and t − 5 to t − 1. Column 7 presents the second-stage estimates from an instrumental variables specification that uses PartisanDisagree s,t as an instrument for OverallDisagree s,t ; first-stage results are available in Table A .4. The sample is at the stock-day level and runs from March 2019 to April 2020. Standard errors separately clustered by stock (permno) and day are reported in brackets; * * and * * * indicate statistical significance at 5% and 1%. 

Is Polarization a Myth?

Polarization and Public Health: Partisan Differences in Social Distancing during the Coronavirus Pandemic

The economic effects of social networks: Evidence from the housing market

Social connectedness: Measurement, determinants, and effects

COVID-Induced Economic Uncertainty

How Does Household Spending Respond to an Epidemic? Consumption during the 2020 COVID-19 Pandemic

Risk Perception through the Lens of Politics in the Time of the COVID-19 Pandemic

Beyond the Running Tally: Partisan Bias in Political Perceptions

The Economics of Motivated Beliefs

Morals and Taboos: Beliefs as Assets

Sentiments and Economic Activity: Evidence from US States

The Big Sort: Why the Clustering of Like-Minded America is Tearing Us Apart

Diagnostic Expectations and Credit Cycles

Optimal Expectations

Risk Taking during a Global Crisis: Evidence from Wuhan

News sentiment in the time of covid-19

Misinformation During a Pandemic

The effect of partisanship and political advertising on close family ties

Why don't we agree? Evidence from a social network of investors

The Political Conditioning of Economic Perceptions

Dissecting Anomalies

Same Facts, Different Interpretations: Partisan Motivation and Opinion on Iraq

What Drives Media Slant? Evidence from U.S. Daily Newspapers

Measuring Group Differences in High-Dimensional Choices: Method and Application to Congressional Speech

Partisanship and Economic Behavior: Do Partisan Differences in Economic Forecasts Predict Real Economic Behavior?

Party Affiliation, Partisonship, and Political Beliefs: A Field Experiment

Nonlocal disadvantage: An examination of social media sentiment. The Review of Asset Pricing Studies

The convergence and divergence of investors' opinions around earnings news: Evidence from a social network

Five Facts about Beliefs and Portfolios

The Effect of Consumer Sentiment on Consumption: Cross-Sectional Evidence from Elections

Political Polarization in Financial News

Income and Wealth Shocks and Expectations during the COVID-19 Pandemic

Trump says China made a 'mistake' and tried to cover up coronavirus outbreak

Uncertainty-Identity Theory

Differential interpretation of public signals and trade in speculative markets

Measuring Geographic Polarization: Theory and Long-Run Evidence

Left Behind: Partisan Identity and Wealth Inequality

Partisan Professionals: Evidence from Credit Rating Analysts

Disagreement and liquidity

Sort Selling: Political Affiliations and Households' Real Estate Decisions

Economic Behavior and the Partisan Perceptual Screen

Belief Disagreement and Portfolio Choice

Partisan Bias, Economic Expectations and Household Spending

Partisanship, Trade Policy, and Globalization: Is There a Left-Right Divide on Trade Policy?

Political Beliefs affect Compliance with COVID-19 Social Distancing Orders

A 57-page memo urged GOP campaigns to blame China for the coronavirus pandemic and insist the term 'Chinese virus' isn't racist

Measuring news sentiment

when good news was reported, the stock market would go up. today, when good news is reported, the stock market goes down. big mistake, and we have so much good (great) news about the economy! Twitter