The Impact of Reduced In-Person Contact on Information Quality: The Case of Analyst Forecast Accuracy During COVID Lockdowns

Abstract

This paper examines whether and how in-person communication influences the quality of information available to market participants. We utilize the lockdown periods during the COVID-19 pandemic as a proxy for an exogenous reduction in in-person communication, and analyst forecast accuracy as a proxy for information quality. We document that when firms’ headquarter states undergo lockdowns, analyst forecasts for these firms issued in the lockdown periods are significantly less accurate than for other firms or other periods, with an average difference of about 8% of the mean value of absolute forecast errors. The difference is also greater when analysts have less firm-specific experience or cover more industries, or when firms are smaller or younger. Supplemental analyses show greater differences when companies convert from in-person to virtual events despite holding more of the latter. In addition, analysts input more effort and display more divergent opinions during the lockdown periods. The collective results indicate that when there is less in-person communication, forecast quality is lower despite more effort and similar or higher frequency of virtual communication.

Keywords

in-person communication information quality lockdowns analyst forecast accuracy D82 G10 G24 M41

Introduction

Prior research indicates that face-to-face interactions communicate information that is difficult to articulate verbally or in writing, especially tacit information relying on body language or eye contact (Rhoads, 2010). Senju and Johnson (2009) find that eye contact modulates social cognitive processes and moderates information that could be stifled by virtual communication. However, in-person communication tends to be costly for both companies and participants, and limits participation. Computer-mediated or virtual communication offers a less costly and potentially effective alternative to a broad base of market participants, and the saved time and resources can be utilized in generating higher-quality work (Baek, 2023; Ferrazzi & Zapp, 2020). Some psychologists suggest that virtual collaboration may lead to more focused and comprehensive decisions as it could remove noise associated with body language or speech (Armstrong, 2006; Rhoads, 2010). In addition, technological development in recent years, such as live video conferences, provides features that enable more information-enhancing interactions among the participants (Brucks & Levav, 2022). This raises the question of whether in-person contact influences the quality of information available to capital markets when virtual communication potentially serves as a substitute.

The question is also motivated by the rising popularity of remote channels, such as teleconferences and webinars, that could replace or reduce in-person communication (Brochet et al., 2023; Brodsky & Tolliver, 2022; IR Magazine, 2023).¹ The switch to remote channels saves time and cost, and provides more convenient access to a larger audience and executives who could not attend in-person (Baek, 2023; Bloom et al., 2015). Examining firms that switch for this reason may bias toward finding improvement in information quality after the switch. On the other hand, management may willfully choose virtual communication to avoid shareholder scrutiny and become less transparent, for example, through allocating less time to shareholder questions (Brochet et al., 2023). Such switches can bias toward finding deterioration in information quality. Accordingly, it is important to account for the endogeneity in the switch from in-person to virtual channels.²

We avoid this endogenous choice by exploiting an exogenous reduction in in-person communication during the period when lockdown and stay-at-home orders were issued in early 2020 across most U.S. states due to the COVID-19 pandemic. These lockdowns occurred over a staggered period, largely influenced by the rate of COVID transmission and political pressure within each state (Jacobsen & Jacobsen, 2020). During the lockdown periods, varying from 24 to 251 days across states, most firm events or other opportunities for communication were forced to be in a virtual format, resulting in an exogenous reduction in affected firms’ in-person communication to the markets. To gauge information quality, we use analyst forecast accuracy (De Franco et al., 2011; Lang et al., 2003) because analysts are sophisticated players in the market and provide a powerful setting for identifying information distributed by companies.³ In addition, analysts’ output is measurable, a feature that facilitates the empirical analysis of information quality.

Our main sample includes 166,302 analyst forecasts issued in 2019 and 2020. We analyze analyst forecast accuracy with a generalized difference-in-differences research design that controls for firm, year, and week fixed effects, in addition to specific characteristics of analyst, forecast, and firm. Firm characteristics include changes in business fundamentals, management disclosure levels, and internal information quality. The fixed effects are feasible due to the staggered nature of the lockdowns and the large variation in lockdown period across states. They enable us to compare the effect of lockdown on forecast accuracy across time within firms, during the same time between states with different lockdown status in 2020, while using forecasts for the same firms in 2019 as a counterfactual. Further analyses also add analyst fixed effects and document similar effects for the same analysts forecasting firms in states under lockdown versus other firms not under lockdown. These fixed effects mitigate the concern that macroeconomic conditions or uncontrolled analyst or firm characteristics – such as variation in lockdown status among analysts or changes in business fundamental – drive the differences in forecast errors. We document significantly larger absolute errors of forecasts issued when firms’ headquarter states undergo lockdowns, by about 8% of the mean value. These results are consistent with deteriorating information quality when in-person communication is likely to be reduced.

In cross-sectional tests, we document larger forecast errors for analysts with less firm-specific experience or broader industry coverage and for younger or smaller firms. These analysts presumably face more challenges to read nonverbal cues due to having fewer interactions with the management prior to the lockdown periods or being stretched over heterogenous industries. Younger or smaller firms tend to have sparse information available and the reduction in in-person communication has a relatively higher impact on their overall information environment. The variation in results further supports the adverse influence of reduced in-person communication on information quality.

We conduct two supplement tests and four robust checks to validate the main results. The first supplemental test focuses on the sample of forecasts issued in the 2020 lockdown period and those for the same set of firms during the corresponding period in 2019 (hereafter, the pseudo-lockdown period). We manually collect in-person and virtual events that are held or participated by firms and explicitly consider the modality of events preceding analyst forecasts in both periods. We document an increase in forecast errors when firms switch from in-person events to virtual events, irrespective of whether more virtual events are held. In contrast, in a falsification test that compares forecasts made after virtual events in both periods for the same firms, we do not find a significant increase in forecast error. Although low power associated with the small number of virtual events in the pseudo-lockdown periods could potentially explain the insignificance, we note that the coefficient is about 55% of the documented effect when firms switch from in-person to virtual events. Overall, these results strengthen our argument that reduced in-person communication from the firms is the mechanism underlying increased forecast errors.

We next compare forecast errors by analysts located in states under lockdown with those by other analysts covering the same firm-years. We find that the former are significantly larger than the latter. These results based on analyst lockdown status triangulate our main results based on firm lockdown status.

We further conduct four tests to check the robustness of the main results, including the application of robust regressions to mitigate the influence of outliers; adjusting the lockdown indicator for forecasts issued after in-person events held right before the lockdown; separately considering firms that issued guidance during the 2019 pseudo-lockdown period but discontinued it during the 2020 lockdown period; and analyzing only industries of essential business whose business fundamentals are less affected by COVID-19.⁴ All four tests continue to support our inferences that reduced in-person communication leads to significantly larger forecast errors.

In light of reduced forecast accuracy, we further examine analyst effort and forecast dispersion to more fully gauge the impact of reduced in-person communication on analyst activity and forecast quality. We find that analysts significantly increase forecast frequency during the COVID lockdown periods, which suggests increased effort input by analysts (Jacob et al., 1999). Additionally, in the lockdown periods, the dispersion of analyst earnings estimates is significantly higher, consistent with more disagreement among analysts when in-person communication is more limited (Diether et al., 2002; Thomas, 2002). Taken together, these results suggest that as compared to virtual communication, in-person communication reduces analyst effort and enhances analyst forecast quality.

This paper contributes to the literature by furthering the understanding of channels through which information is distributed to capital markets, in particular whether the in-person format aids the distribution incrementally to the virtual format. Our analyses answer the call by Bushee et al. (2011, p. 1188) for future research to “provide new insights into not only the market response to voluntary disclosure, but also the choice of voluntary disclosure venues.” Our finding that reduced in-person contact is related to significantly lower analyst forecast accuracy suggests a benefit of in-person contact that is not fully attained through virtual contact. Given the rise in virtual communication, our findings offer an important aspect to consider when making choices of communication modality (Brochet et al., 2023; Brodsky & Tolliver, 2022; IR Magazine, 2023).

In addition, Soltes (2014) and Brown et al. (2015) document the importance of private communication for analyst research. We extend this literature by documenting the importance of in-person communication to their forecast quality. We find that despite more frequent virtual communication and increased analyst effort, when in-person communication is limited, forecast accuracy deteriorates and forecast dispersion increases.

Our study also contributes to the nascent literature on capital market impacts of the COVID-19 pandemic, such as market reactions to earnings news or analyst forecast during the pandemic (D’Augusta & Grossetti, 2023; Fabrizi et al., 2023; Kale & Kale, 2023), analyst forecast frequency, timeliness, and accuracy (Bilinski, 2023; Durney et al., 2024; Hao et al., 2022; Liu et al., 2023), and differing impact by genders (Du, 2023). This paper expands Hao et al. (2022) and Bilinski (2023) by not viewing firms and the pandemic period as homogeneous and considering the variation across firms and within the pandemic period in terms of information communication modality. This paper also complements concurrent working papers by Liu et al. (2023) and Durney et al. (2024). Both document that when analysts rely more on in-person interactions with informed parties – proxied with geographic proximity – their forecast accuracy declines during the lockdown period and several subsequent quarters. Different from those studies, we focus on firms' communication modality during the lockdown period and the overall impact on forecast accuracy for affected firms. Additionally, as Ben-Rephael et al. (2023) indicate, forecast accuracy decline stemming from reduced in-person interactions could potentially be compensated with increased remote working hours. We find that the overall information quality declines for firms experiencing reduced in-person communication, with cross-sectional variation. The decline is greater for smaller or younger firms, and for those that switch from in-person to virtual modality, but not for firms that maintain a virtual modality.

Literature Review and Hypothesis Development

Companies communicate information to market participants through both in-person and remote channels. Examples of the former are investor and analyst days, private meetings, road shows, and examples of the latter are conference calls and annual meetings through webcast (Brown et al., 2015; Bushee et al., 2018; Soltes, 2014). Over time, more events are held remotely instead of face-to-face (Brodsky & Tolliver, 2022; IR Magazine, 2023).

One stream of literature suggests that in-person contact facilitates communication of information that is difficult or infeasible to articulate virtually or in writing (Rhoads, 2010). As a result, reduced in-person communication leads to a loss of information due to the loss of tacit information and nonverbal cues that are obtained through eye contact and body language (Bushee et al., 2011; Gajendran & Harrison, 2007; Senju & Johnson, 2009). Consistent with this literature, analysts surveyed by Deloitte cite face-to-face meetings as one of their preferred ways to find out about management leadership capabilities (Deloitte Global Services Limited, 2012, p. 7). Virtual meetings could also accord management more discretion in selecting shareholder questions and presentation content and suppress information spread (Schwartz-Ziv, 2021). Findings and arguments from the above literature suggest lower analyst research quality when firms eliminate or reduce in-person communication.

Other studies suggest that the remote format does not necessarily lead to information quality deterioration for the following reasons. First, replacing in-person contact with remote communication saves time and costs for the participants. The savings could facilitate more information communication through participation by more executives and investors (Baek, 2023; Ferrazzi & Zapp, 2020), or longer time spent on information acquisition from other channels (Causholli et al., 2022; Rhoads, 2010). Greater attendance could offset the reduction in virtual-meeting activity and result in no loss of information content when adopting a virtual format (Brochet et al., 2023).⁵ As both effort on online information collection and that on in-person soft information acquisition are positively correlated with analyst forecast accuracy (Ben-Rephael et al., 2023), reduced face-to-face meetings coupled with greater effort in collecting information through virtual venues could lead to unchanged or improved information quality.

Second, virtual meetings allow participants to reflect before responding and are less susceptible to biases stemming from factors such as speech volume or meeting room furniture arrangement, as noted by Armstrong (2006). Thus, virtual meetings could lead to more focused and comprehensive decisions (Rhoads, 2010). In addition, Brucks and Levav (2022) suggest that virtual communication does not affect the effectiveness of selecting the most creative ideas as it “requires cognitive focus and analytical reasoning” (p. 108), a feature shared by financial analysts. To the extent that Brucks and Levav’s results could be generalized to analyst activity, virtual communication does not necessarily affect analyst forecast quality. The above discussions suggest that, when in-person contact is limited or eliminated, the affected market participants could replace it with remote channels and produce similar or higher-quality results.

Given the above discussion, ex ante, it is unclear whether and to what extent in-person contact affects analyst forecast quality. If virtual channels do not completely substitute for in-person contact in terms of information communication, we expect to observe lower analyst research quality when firms eliminate or reduce in-person communication. Otherwise, we expect to observe similar or higher analyst research quality. We summarize the above arguments with the hypothesis below.

Hypothesis: Information quality deteriorates with the reduction in in-person communication from firms.

Research Design

The COVID-19 pandemic led to stay-at-home orders across the United States, which concentrated in the period of late March to late May of 2020, and were somewhat staggered in that the ending date of the mandatory order varied by one or more months.⁶ On average, the lockdown period lasted 57 days, with Mississippi at 24 days and New Mexico at 251 days. These orders lead to an exogenous reduction in in-person communication from firms. We use a binary variable LOCKDOWN to capture this period. As a proxy for information quality, we use analyst absolute forecast errors, with higher errors for lower information quality (De Franco et al., 2011; Lang et al., 2003). We start with descriptive statistics to compare forecast errors across time for the same firms and then utilize the multivariate regression below that controls for various analyst, firm, and forecast characteristics, as well as fixed effects for firm, year, and week.

A B S_F E_{ijd} = α_{0} + α_{1} L O C K D O W N_{j d} + \sum_{k = 1}^{n} β_{k} \cdot {C o n t r o l s}_{k} + Fixed effects for firm, year, and week + ε

(1)

As we define in Appendix A, ABS_FE_ijd is the absolute difference between actual earnings and analyst i’s 1-year ahead earnings forecast issued on day d for firm j, scaled by the firm’s stock price at the fiscal year end. LOCKDOWN_jd is an indicator that equals 1 if firm j’s headquarter state is subject to the stay-at-home order when the forecast is issued on day d. Control_k includes analyst, forecast, and firm characteristics. Following Clement (1999), Clement and Tse (2003), and Jacob et al. (1999), we control for the following seven analyst and forecast characteristics: the number of days to the forecasted fiscal year end (FCST_HORIZON), brokerage size (BSIZE), the number of firms and industries covered by an analyst (#FIRMS and #INDUSTRIES), the number of years an analyst has covered the firm (FIRM_EXP), the number of 1-year-ahead earnings forecasts issued by the analyst for the same firm-year (FCST_FREQ), and the number of days since the latest one-year-ahead earnings forecast issued by any analyst for the same firms-year (DAYS_LASTFCST). All are measured in the 1-year period that precedes the current forecast date.

At the firm-year level, we control for changes in business fundamentals and the associated forecast difficulty with the absolute forecast error from a random walk model (ABS(UNEX_EARN)); management disclosure level with the availability of management guidance 1 month prior to analyst forecast date (MGT_EPS_GUIDE); and the quality of overall internal information environment with the speed of earnings announcement from the latest fiscal year end (EA_SPEED). Other controls for the influence of information environment include the number of analyst coverage (#ANALYSTS), total assets (SIZE), and firm age (AGE) (Barry & Brown, 1985; Mikhail et al., 1997; Zhang, 2006). Besides these annual measures of information quality, MGT_EPS_GUIDE also serves as a within-year proxy for information quality, with the issuance indicating higher information quality (Verrecchia, 1990; Waymire, 1985). Regression (1) also includes prior year return on assets (PY_ROA), leverage (LEVERAGE), and the logarithm of stock price (LN_PRICE) to account for variation in forecast errors with them.

We include firm fixed effects to control for other time-invariant firm characteristics not captured by the variables above; and year and week fixed effects for year-over-year economic changes and week-level fluctuations, respectively, in economic fluctuations and general uncertainty. Given the fixed effects, a positive coefficient on LOCKDOWN indicates greater errors of forecasts issued for firms undergoing lockdown relative to those for the same firms in other periods of 2020 or other firms in the same periods, net of weekly variation estimated for the 2019-2020 period.

In additional untabulated tests, we further include analyst fixed effects to account for across-analyst variation beyond the analyst characteristics explicitly controlled for in regression (1). As the lockdown periods are state specific, to account for the potential correlations in forecast errors among firms in the same state, we cluster standard errors by state. To the extent that the control variables in equation (1) fully account for variation in forecast errors with analyst and firm characteristics and economic uncertainty, the larger errors would be attributed to reduced in-person contact during the lockdown periods and support our hypothesis.

Sample Selection

Our sample includes firms with December 31 fiscal year end in Compustat from 2020, the COVID-19 event year, and 2019 as the control year. We require that the firms have headquarter addresses in Loughran and McDonald’s address data, and are covered by at least one analyst according to IBES in both years. As the stay-at-home order applies in the same calendar month for firms in the same state, we include only firms with December 31 as fiscal year ends to mitigate the effect of forecast horizon – which is relative to fiscal year end – on forecast accuracy. We use the headquarter addresses to identify the periods firms are subject to the lockdown orders. At the analyst-firm level, we keep only forecasts issued by analysts that cover a firm in both 2019 and 2020 to mitigate differences in forecast errors across different analysts. These criteria result in 166,302 forecasts made for 2,261 firms by 2,029 analysts in the 2 years.

Table 1 reports descriptive statistics for our sample after winsorizing all continuous variables at 1/99%. Panels A and B include statistics for the full sample and a subsample of analyst forecasts made during state lockdown periods in 2020 and the pseudo-lockdown periods in 2019. Because descriptive characteristics in Panel B are for the same firms during the same periods in 2020 and 2019, many are similar. An exception is the average ABS_FE, which is larger during the 2020 lockdown periods, providing initial descriptive evidence on forecast quality deterioration. We also note that the number of 1-year ahead earnings forecasts during the 2020 lockdown periods nearly doubles that for the same firms during the 2019 pseudo-lockdown periods, at 25,495 versus 13,267. This provides suggestive evidence of a significant increase in analyst effort (Jacob et al., 1999).

Table 1.

Descriptive Statistics

Panel A: Full sample
Variable	N	Mean	p50	SD	p25	p75
Dependent variable and main variables of interest
ABS_FE	166,302	3.388	0.632	10.115	0.182	2.105
LOCKDOWN	166,302	0.153	0.000	0.360	0.000	0.000
ANALYST_LOCKDOWN	131,830	0.149	0.000	0.357	0.000	0.000
Control variables: Firm characteristics
ABS(UNEX_EARN)	166,302	0.059	0.023	0.092	0.006	0.067
MGT_EPS_GUIDE	166,302	0.153	0.000	0.361	0.000	0.000
EA_SPEED	166,302	−42.673	−42.000	14.936	−55.000	−29.000
#ANALYSTS	166,302	15.985	15.000	9.418	8.000	21.000
SIZE	166,302	37,824	6,282	104,025	1,612	23,434
PY_ROA	166,302	−0.010	0.026	0.187	−0.001	0.070
LEVERAGE	166,302	0.270	0.246	0.222	0.075	0.398
AGE	166,302	3.055	3.135	0.769	2.398	3.611
LN_PRICE	166,302	3.718	3.782	1.271	2.945	4.538
Control variables: Analyst characteristics
BSIZE	166,302	59.557	50.000	47.107	21.000	90.000
#FIRMS	166,302	20.109	19.000	8.830	14.000	25.000
#INDUSTRIES	166,302	3.620	3.000	2.423	2.000	5.000
FIRM_EXP	166,302	5.442	4.000	5.144	2.000	8.000
FCST_FREQ	166,302	6.649	6.000	3.190	5.000	8.000
Control variables: Forecast characteristics
FCST_HORIZON	166,302	185.897	176.000	94.415	90.000	262.000
DAYS_LASTFCST	166,302	7.283	1.000	16.298	0.000	6.000

Panel B: Sample during lockdown periods in 2020 and pseudo-lockdown periods in 2019
Variable	Lockdown periods in 2020 (n = 25,495)		Pseudo-lockdown periods in 2019 (n = 13,267)
Variable	Mean	p50	Mean	p50
Dependent variable and main variable of interest
ABS_FE	3.735	1.039	2.354	0.459
LOCKDOWN	1.000	1.000	0.000	0.000
Control variables: Firm characteristics
ABS(UNEX_EARN)	0.067	0.030	0.051	0.017
MGT_EPS_GUIDE	0.103	0.000	0.190	0.000
EA_SPEED	−42.766	−41.000	−40.852	−37.000
#ANALYSTS	16.443	15.000	18.575	17.000
SIZE	40.034	5.679	48.252	7.031
PY_ROA	−0.026	0.023	−0.001	0.031
LEVERAGE	0.274	0.245	0.240	0.215
AGE	3.024	3.135	3.049	3.135
LN_PRICE	3.867	3.920	3.855	3.951
Control variables: Firm characteristics
BSIZE	59.541	55.000	60.735	50.000
#FIRMS	20.184	19.000	19.991	19.000
#INDUSTRIES	3.613	3.000	3.519	3.000
FIRM_EXP	5.617	4.000	5.237	4.000
FCST_FREQ	6.623	6.000	6.193	6.000
Control variables: Firm characteristics
FCST_HORIZON	238.166	246.000	234.466	245.000
DAYS_LASTFCST	5.471	1.000	5.206	0.000

Note. Panel A presents descriptive statistics of the main variables for the full sample of 1-year ahead earnings forecasts issued between 2019 and 2020. Panel B is for forecasts issued during the lockdown (pseudo-lockdown) periods in 2020 (2019) for firms whose headquarters are in the lockdown states. Pseudo-lockdown periods are the same periods in 2019 that correspond to the lockdown periods in 2020.

Next, to illustrate the effect of the change in communication modality on forecast errors, in Figure 1 we show the mean forecast errors for each week during the lockdown period in 2020. Because lockdowns occurred at different times for different states, we capture the first 10 weeks of lockdown rather than anchor on the same date, such that t = 1 is for the first week of a state’s lockdown and t = 10 for the 10th week. We then compare forecast errors in the 10 weeks with those for the same firms in the same periods in 2019.

Figure 1.

Mean Forecast Errors by Week into a Lockdown Period in 2020 Compared to a Pseudo-Lockdown Period in 2019.

Consistent with the literature (e.g., Clement & Tse, 2003), Figure 1 indicates that absolute forecast errors become smaller as forecast horizon shortens. The errors in the first week are similar between 2019 and 2020. However, as the forecast horizon shortens, analyst forecast errors decrease at a slower pace in the lockdown period in 2020 than those in the same weeks in 2019. The stock price-scaled difference in forecast errors is 0.31% in the first week, and ranges from 0.73% to 1.76% during the second to the 10th week. This result is consistent with inferior forecast quality during the lockdown period. Considering that analysts issue forecasts more frequently during the lockdown periods as indicated in Table 1, this suggests increased barriers to learn about earnings information as one potential explanation for the difference in forecast errors. Below we describe results from multivariate analyses.

Regression Results

Main Results and Analyses on Variation With Analyst and Firm Characteristics

Column (1) of Table 2 presents results of estimating regression (1) with firm, year, and week fixed effects, and clustering standard errors by state. The positive and statistically significant coefficient on LOCKDOWN supports larger forecast errors for firms while their headquarter states are subject to lockdown orders than those for forecasts issued for other firms in the same periods or those issued in other periods. The difference at 0.274 is approximately 8% of the mean forecast error of 3.388.⁷ The results support our hypothesis that reduced in-person communication in the lockdown periods is related to lower information quality as reflected by lower forecast accuracy. Among the control variables, absolute forecast errors are positively correlated with changes in business fundamentals (ABS(UNEX_EARN)), firm performance (PY_ROA), leverage (LEVERAGE), age (AGE), and forecast horizon (FCST_HORIZON), and negatively correlated with stock price (LN_PRICE) and forecast frequency (FCST_FREQ). In untabulated analyses that include only firm fixed effects, or firm, year, week, and analyst fixed effects, we document consistent results, with the coefficients on LOCKDOWN at 0.381 and 0.260, respectively.

Table 2.

Absolute Forecast Errors in the Lockdown Periods Versus Other Periods and Variation with Analyst and Firm Characteristics

Variables	(1)	(2)	(3)	(4)	(5)	(6)
LOCKDOWN	0.2736**	0.4845***	−0.0091	1.9117***	0.3841**	1.5360***
LOCKDOWN	(2.16)	(2.90)	(−0.05)	(4.06)	(2.43)	(3.48)
LOCKDOWNFIRM_EXP*		−0.0374***				−0.0066
LOCKDOWNFIRM_EXP*		(−3.70)				(−0.76)
LOCKDOWN#INDUSTRIES*			0.0788**			0.0890**
LOCKDOWN#INDUSTRIES*			(2.19)			(2.61)
LOCKDOWNAGE*				−0.5401***		−0.4853***
LOCKDOWNAGE*				(−4.08)		(−4.46)
LOCKDOWNSIZE*					−0.0027**	−0.0018*
LOCKDOWNSIZE*					(−2.19)	(−1.66)
Firm characteristics:
ABS(UNEX_EARN)	17.1264***	17.1475***	17.1110***	17.1519***	17.0993***	17.1177***
ABS(UNEX_EARN)	(5.13)	(5.14)	(5.12)	(5.16)	(5.12)	(5.15)
MGT_EPS_GUIDE	−0.0638	−0.0685	−0.0555	−0.0766	−0.0606	−0.0646
MGT_EPS_GUIDE	(−0.37)	(−0.40)	(−0.32)	(−0.44)	(−0.35)	(−0.37)
EA_SPEED	−0.0245	−0.0245	−0.0245	−0.0246	−0.0247	−0.0247
EA_SPEED	(−0.66)	(−0.66)	(−0.66)	(−0.66)	(−0.66)	(−0.66)
#ANALYSTS	0.0928	0.0920	0.0939	0.0899	0.0916	0.0906
#ANALYSTS	(1.46)	(1.45)	(1.48)	(1.41)	(1.45)	(1.42)
SIZE	0.0151	0.0154	0.0153	0.0155	0.0180	0.0176
SIZE	(1.24)	(1.26)	(1.25)	(1.27)	(1.44)	(1.40)
PY_ROA	4.4839**	4.4825**	4.4923**	4.4376**	4.4697**	4.4421**
PY_ROA	(2.58)	(2.58)	(2.58)	(2.57)	(2.58)	(2.57)
LEVERAGE	3.7849**	3.7635**	3.7844**	3.6462**	3.7201**	3.6131**
LEVERAGE	(2.17)	(2.17)	(2.17)	(2.08)	(2.12)	(2.06)
AGE	4.5748**	4.2960**	4.6323**	3.2643*	4.3796**	3.2847*
AGE	(2.48)	(2.37)	(2.52)	(1.79)	(2.43)	(1.81)
LN_PRICE	−8.6099***	−8.6162***	−8.6206***	−8.6312***	−8.6302***	−8.6556***
LN_PRICE	(−7.06)	(−7.07)	(−7.10)	(−7.13)	(−7.10)	(−7.18)
Analyst characteristics:
BSIZE	0.0006	0.0006	0.0005	0.0006	0.0006	0.0006
BSIZE	(1.30)	(1.32)	(1.28)	(1.32)	(1.32)	(1.32)
#FIRMS	−0.0002	−0.0002	−0.0001	−0.0002	−0.0002	−0.0001
#FIRMS	(−0.07)	(−0.06)	(−0.04)	(−0.06)	(−0.07)	(−0.03)
#INDUSTRIES	0.0142	0.0142	0.0017	0.0143	0.0141	0.0001
#INDUSTRIES	(1.07)	(1.06)	(0.13)	(1.07)	(1.06)	(0.01)
FIRM_EXP	−0.0050	0.0005	−0.0051	−0.0050	−0.0050	−0.0040
FIRM_EXP	(−1.50)	(0.15)	(−1.50)	(−1.49)	(−1.49)	(−1.09)
FCST_FREQ	−0.0396***	−0.0404***	−0.0396***	−0.0404***	−0.0397***	−0.0406***
FCST_FREQ	(−3.23)	(−3.29)	(−3.22)	(−3.31)	(−3.24)	(−3.32)
Forecast characteristics:
FCST_HORIZON	0.0483***	0.0483***	0.0485***	0.0478**	0.0481**	0.0479**
FCST_HORIZON	(2.69)	(2.70)	(2.69)	(2.67)	(2.67)	(2.66)
DAYS_LASTFCST	−0.0004	−0.0004	−0.0003	−0.0004	−0.0003	−0.0003
DAYS_LASTFCST	(−0.17)	(−0.17)	(−0.15)	(−0.17)	(−0.12)	(−0.12)
Observations	166,302	166,302	166,302	166,302	166,302	166,302
Adjusted R²	0.661	0.661	0.661	0.661	0.661	0.661
Firm, week, and year fixed effects	Yes	Yes	Yes	Yes	Yes	Yes

Note. Column (1) reports the results of estimating the regression below for the average impact of lockdown on analyst forecast errors based on 166,302 1-year ahead analyst earnings forecasts issued in 2019 and 2020 for firms with December 31 fiscal year ends and one or more analyst coverage. Only forecasts issued by analysts who cover a firm in both years are retained. ABS_FE_ijd = α₀ + α₁LOCKDOWN_jd + $\sum_{k = 1}^{n} β_{k} \cdot {C o n t r o l}_{k}$ + Fixed effects for firm, year, and week + ε (1). Columns (2)–(6) examine the variation of the impact with analyst (FIRM_EXP and #INDUSTRIES) and firm (AGE and SIZE) characteristics. All variables are defined in Appendix A. Robust t-statistics are in parentheses. Standard errors are clustered by state. ***, **, * indicate p-values at 0.01, 0.05, and 0.10, respectively.

We further examine variation in the main results with analyst and firm characteristics to enhance our argument for reduced in-person communication as the underlying mechanism. Regarding analyst characteristics, we expect analysts with shorter firm-specific experience (FIRM_EXP) or covering more industries (#INDUSTRIES) to face more challenges when receiving reduced in-person communication. The former group of analysts has not developed in-depth knowledge of the firms and management and is less likely to offset the loss of in-person communication with alternative sources. The latter group is more constrained in time and effort as their attention is dispersed across diverse industries. We augment regression (1) with the two variables and their interactions with LOCKDOWN in Columns (2) and (3) of Table 2. The significantly negative coefficient on LOCKDOWN*FIRM_EXP and positive coefficient on LOCKDOWN*#INDUSTRIES support our expectations. Taken together, these results are consistent with reduced in-person communication being related to lower analyst forecast accuracy through more demand for analyst’s effort and attention.

From the perspective of firms, for those with rich information available, when in-person contact is limited, the marginal reduction in information communicated is relatively small. Thus, we expect a more limited decrease in forecast quality during lockdowns for such firms. We proxy for firm information environments with firm age and size. Per columns (4) and (5) of Table 2, the coefficients on LOCKDOWN are significantly positive and the coefficients on its interaction terms are significantly negative. We thus infer that higher absolute forecast errors in the lockdown periods are driven by younger and smaller firms, which have more scarce overall information. Column (6) presents results when including all four interaction terms simultaneously. Except the interaction term on FIRM_EXP, which loses significance and magnitude, all interaction terms maintain a comparable magnitude and significance.

We infer from the overall results in Table 2 that reduced in-person communication negatively impacts forecast accuracy, especially for analysts with shorter firm-specific experience or covering more industries and for younger or smaller firms.

Analysis of Corporate Events During the Lockdown and Pseudo-Lockdown Period

Table 2 supports larger forecast errors for firms during the periods when they are subject to lockdown orders. We argue that reduced in-person communication from these firms explains the results and document variation in the results with analyst and firm characteristics that supports this argument. This section describes additional analyses to further validate this mechanism.

We first re-examine our primary analysis using a subsample of firms in lockdown states during the lockdown periods in 2020 and the pseudo-lockdown period in 2019, analogous to our descriptive tests reported in Figure 1 and Table 1. This test allows for a sharper analysis of forecasts that are for the same firms and same time periods across two adjacent years, mitigating the effect on forecast errors associated with differences in firms, states, or forecast horizons. A limitation with this analysis is that we lose observations for firms that are never subject to lockdowns (e.g., those in Iowa or Utah) and those made outside of the (pseudo) lockdown period in (2019) 2020.

Table 3, column (1) reports the results of estimating equation (1). Because the value of LOCKDOWN varies solely with year in this analysis, we remove year fixed effects. Additionally, because the comparison is within the same states, we cluster standard errors by firms instead of states. We observe significantly higher absolute forecast errors during the lockdown periods, with a difference of 1.29 percentage points. This result is consistent with the descriptive analyses reported in Figure 1 and Table 1.

Table 3.

Forecast Errors Around Events in Lockdown Versus Pseudo-lockdown Periods

Variables	(1)	(2)	(3)	(4)
	ABS_FE	ABS_FE	ABS_FE	ABS_FE
	[All forecasts]	[In-person to virtual events]	[In-person to virtual events (2020 events ≥2019 events)]	[Virtual to virtual events]
LOCKDOWN	1.2899***	0.8871**	1.6109*	0.4915
LOCKDOWN	(4.92)	(2.09)	(1.89)	(0.88)
Firm characteristics
ABS(UNEX_EARN)	15.6941***	12.4198***	11.5482**	3.7245
ABS(UNEX_EARN)	(4.02)	(2.84)	(2.68)	(1.34)
MGT_EPS_GUIDE	−0.0088	0.1853	0.4202	0.0109
MGT_EPS_GUIDE	(−0.06)	(1.32)	(0.89)	(0.08)
EA_SPEED	0.0388	0.0171	0.1363**	−0.0433
EA_SPEED	(0.96)	(0.44)	(2.47)	(−0.77)
#ANALYSTS	0.0898	−0.0801	−0.2588***	−0.0822
#ANALYSTS	(0.81)	(−1.21)	(−3.03)	(−0.85)
SIZE	0.0099	0.0008	−0.0039	−0.0103
SIZE	(1.34)	(0.30)	(−1.35)	(−0.46)
PY_ROA	7.3352***	8.0191**	−3.0001	−5.6454*
PY_ROA	(2.64)	(2.67)	(−0.56)	(−1.75)
LEVERAGE	2.1318	−0.5046	1.5952	0.8662
LEVERAGE	(1.01)	(−0.29)	(0.63)	(0.33)
AGE	7.8711**	4.6448	7.1504	0.7140
AGE	(1.96)	(0.98)	(1.12)	(0.07)
LN_PRICE	−7.9565***	−3.0352***	−4.2778***	−0.3069
LN_PRICE	(−5.63)	(−4.06)	(−3.64)	(−0.42)
Analyst characteristics
BSIZE	0.0005	0.0004	−0.0000	−0.0006
BSIZE	(0.95)	(0.73)	(−0.04)	(−0.60)
#FIRMS	−0.0033	0.0005	0.0176	−0.0063
#FIRMS	(−0.88)	(0.12)	(1.52)	(−1.35)
#INDUSTRIES	0.0358**	0.0227	−0.0513*	−0.0019
#INDUSTRIES	(2.14)	(1.36)	(−1.83)	(−0.09)
FIRM_EXP	−0.0010	0.0019	−0.0064	−0.0028
FIRM_EXP	(−0.24)	(0.34)	(−0.94)	(−0.43)
FCST_FREQ	−0.0095	−0.0152	−0.0060	−0.0206
FCST_FREQ	(−0.67)	(−0.98)	(−0.31)	(−1.35)
Forecast characteristics
FCST_HORIZON	0.0259	−0.0448	−0.2815**	0.0126
FCST_HORIZON	(1.14)	(−0.85)	(−2.29)	(0.37)
DAYS_LASTFCST	0.0067*	0.0051	−0.0000	0.0075
DAYS_LASTFCST	(1.84)	(0.79)	(−0.00)	(0.69)
Observations	38,614	8,046	2,407	4,227
Adjusted R²	0.730	0.813	0.797	0.892
Fixed effects	Firm, week	Firm, week	Firm, week	Firm, week

Note. Column (1) examines forecasts issued during the 2020 lockdown periods for firms headquartered in the lockdown states and forecasts for the same firms during the 2019 pseudo-lockdown periods. In this subsample, LOCKDOWN is year-specific and renders year fixed effects unnecessary. Column (2) results are solely within firms that switch from in-person events in the 2019 pseudo-lockdown periods to virtual events in the 2020 lockdown periods. Column (3) limits the sample in column (2) to firms with the same number of, or more, events in the 2020 lockdown period compared to the 2019 pseudo-lockdown period. Column (4) presents a falsification test examining forecasts for firms involved in virtual events in both the 2019 pseudo-lockdown periods and the 2020 lockdown periods. All variables are defined in Appendix A. Robust t-statistics are in parentheses. Standard errors are clustered by firm. ***, **, * indicate p-values at 0.01, 0.05, and 0.10, respectively.

The analysis in column (1) that is solely within the same firms during the same weeks in 2020 versus 2019 provides more support for reduced in-person communication as a driver for higher forecast errors during the lockdown periods are likely driven by. To further validate this mechanism, we next limit our analysis to only analyst forecasts issued subsequent to corporate events and examine variation in forecast accuracy with the event modality. We use Capital IQ Key Development data, along with manual work, to collect dates and modality of three events, including analyst/investor days, company conference presentations, and special shareholders meetings. Appendix B provides details of the collection procedures. We select these events because they are mostly held in-person before the COVID pandemic and analysts actively participate in them or pay attention to them. In untabulated statistics, we observe a significant decrease in the proportion of in-person events, from 98.8% in 2019 to 48.6% in 2020, and a slight increase in the number of events, from 23,371 to 23,731.

In Column (2) of Table 3, we retain only forecasts issued within 1 month after corporate events for firms that switch between in-person events in the 2019 pseudo-lockdown period to virtual events in the 2020 lockdown period. This results in a sample of 8,046 analyst forecasts, of which 5,618 are in 2019 and 2,428 in 2020. Consistent with results in column (1), the coefficient on LOCKDOWN at 0.887 indicates significantly higher errors of forecasts issued subsequent to virtual events than errors of those issued after in-person events in 2019.

We note that the three types of events we focus on are either voluntarily organized by management (analyst-investor days and special or extraordinary shareholder meetings) or organized by third parties (company conference presentation), the latter accounting for more than 98% of our sample events. Because management has little to no discretion over these events – unlike shareholder annual meetings – information communicated therein is less likely to reflect management manipulation, such as allocating less time to audience questions (Schwartz-Ziv, 2021). We therefore infer that the decline in forecast accuracy following such events, as shown in column (2) of Table 3, is more likely driven by a change in communication modality than by management manipulation.

An alternative explanation for the larger forecast errors in the lockdown period is that firms participate in or hold fewer information events, as reflected in fewer events during the lockdown period than the pseudo-lockdown period. In other words, a reduction in overall information available to the analysts, rather than reduction in mere in-person communication potentially explains our main findings. To mitigate this concern, in Table 3, column (3), we further limit our sample to only firms that have a greater or equal number of events during the lockdown period compared to the pseudo-lockdown period. Based on this subsample of 2,407 forecasts, we continue to find significantly larger forecast errors following virtual events in the 2020 lockdown periods than those following in-person events in the 2019 pseudo-lockdown periods. We infer that having fewer information events is unlikely to explain the main results. Per columns (1) through (3), the coefficients on LOCKDOWN range from 38% (0.887/2.354) to 68% (1.611/2.354) of the mean forecast error in the pseudo-lockdown periods. In Section 6, we further examine whether outliers contribute to the large magnitude of the coefficient on LOCKDOWN.

Results in column (3) of Table 3 align more closely with the lack of tacit information as an explanation for the main results than with the management manipulation explanation. If management manipulates how virtual communication is conducted, for example, by shortening time allocated to the audience, it becomes increasingly difficult to sustain such behaviors when more virtual events are subsequently held. This is because the audience could repeatedly request additional time for previously unanswered questions. Under this view, more virtual events mitigate the reduction in information virtually conveyed and lessen the negative effects of virtual communication on forecast accuracy. Conversely, multiple virtual communications could lead to an accumulation of tacit information, meaning that the audience does not necessarily receive more and higher-quality information over more virtual events. Therefore, the result of worse forecast quality in column (3) is more consistent with the lack of tacit information than management manipulation as an explanation.⁸

Finally, in Table 3, column (4), we perform a falsification test that examines forecasts that are issued in 1 month after virtual events in the 2019 pseudo- or 2020 actual lockdown periods for the same firms. For these forecasts, we anticipate that the lockdown has a weaker effect on the reduction of in-person communication.⁹ Among the 4,227 forecasts, 3,952 (275) are in the lockdown (pseudo-lockdown) period, issued after 2,034 (23) virtual events in the corresponding period. In the lockdown (pseudo-lockdown) period, 14, 1997, and 23 (11, 11, and 1) events are analyst/investor days, company conference presentations, and special/extraordinary shareholder meetings, respectively.¹⁰ The coefficient on LOCKDOWN in column (4) is insignificant and approximately half of the magnitude reported in column (2).

We also note that uncertainty associated with lockdowns affects all subsamples in Table 3. The lower and insignificant coefficient on LOCKDOWN for firms that maintain virtual communication is inconsistent with greater forecast errors being driven by higher uncertainty in the lockdown periods. However, we acknowledge that the small number of virtual events in 2019 may reduce the power of the test and contribute to the insignificant coefficient in column (4). We note this limitation and refrain from drawing a definitive conclusion based on this result.¹¹

Collectively, Table 3 shows reduced forecast accuracy when firms switch away from in-person events irrespective of whether more virtual events are held, and insignificant accuracy change when maintaining the virtual modality. These results strengthen our inference that reduced in-person communication from firms is related to information quality deterioration.

Additional Analyses and Robustness Tests

Lockdown Status Based on Analyst Location

This section examines variation in forecast errors among analysts covering the same firms in the same time periods. To more clearly capture variation in lockdown status among analysts, we focus on forecasts issued by analysts located in states different from the headquarter state of the covered firm. As detailed in Appendix B, we manually collect analyst office location from I/B/E/S and FINRA, yielding a sample of 131,830 observations. Table 4 documents that analysts located in states under lockdown produce significantly larger forecast errors than other analysts covering the same firm-year in the same week. The magnitude of the difference at 0.1724, or 5.09% of the mean analyst forecast error, is comparable to 0.2736 documented in column (1) of Table 2. The results based on analysts’ lockdown status triangulate our inference from the main analyses that forecast accuracy deteriorates with reduced in-person communication. Similar to Table 2, in untabulated analyses, we also include only firm fixed effects or additionally add analyst fixed effect and document similar results, with the coefficients on ANALYST_LOCKDOWN at 0.3245 and 0.1380, respectively, and both statistically significant.¹²

Table 4.

Lockdown Status Based on Analyst Location

Variables	(1)
Variables	ABS_FE
ANALYST_LOCKDOWN	0.1724**
ANALYST_LOCKDOWN	(2.17)
Firm characteristics
ABS(UNEX_EARN)	17.2419***
ABS(UNEX_EARN)	(8.82)
MGT_EPS_GUIDE	−0.1082*
MGT_EPS_GUIDE	(−1.75)
EA_SPEED	−0.0190
EA_SPEED	(−1.40)
#ANALYSTS	0.1194**
#ANALYSTS	(2.70)
SIZE	0.0141***
SIZE	(4.54)
PY_ROA	5.1657***
PY_ROA	(4.88)
LEVERAGE	3.4444***
LEVERAGE	(3.63)
AGE	3.8810
AGE	(1.58)
LN_PRICE	−8.6594***
LN_PRICE	(−13.63)
Analyst characteristics
BSIZE	0.0005*
BSIZE	(1.73)
#FIRMS	0.0016
#FIRMS	(0.45)
#INDUSTRIES	0.0160
#INDUSTRIES	(1.57)
FIRM_EXP	−0.0045
FIRM_EXP	(−1.56)
FCST_FREQ	−0.0449***
FCST_FREQ	(−4.80)
Forecast characteristics
FCST_HORIZON	0.0461***
FCST_HORIZON	(6.34)
DAYS_LASTFCST	−0.0021
DAYS_LASTFCST	(−1.48)
Observations	131,830
Adjusted R²	0.661
Fixed effects	Firm, week, year

Note. This table presents results based on forecasts issued by analysts located in states different from the headquarter state of the covered firm. In this table, lockdown periods are defined based on analyst office locations (ANALYST_LOCKDOWN). All variables are defined in Appendix A. Robust t-statistics are in parentheses. Standard errors are clustered by state. ***, **, * indicate p-values at 0.01, 0.05, and 0.10, respectively.

Robust Regressions to Mitigate Effects of Outliers

To examine whether influential observations contribute to the large magnitude of the coefficients on LOCKDOWN in Tables 2 and 3, we conduct robust regressions in Table 5.¹³ Leone et al. (2019) demonstrate that robust regression outperforms winsorization, truncation, and influence diagnostics for identifying and mitigating the effect of influential observations.

Table 5.

Robust Regressions to Mitigate the Effect of Outliers

Variables	(1)	(2)
Variables	ABS_FE	ABS_FE
LOCKDOWN	0.0835***	0.1643***
LOCKDOWN	(9.91)	(2.76)
Firm characteristics
ABS(UNEX_EARN)	2.7923***	2.3763***
ABS(UNEX_EARN)	(56.87)	(24.21)
MGT_EPS_GUIDE	−0.0056	−0.0218
MGT_EPS_GUIDE	(−0.59)	(−1.18)
EA_SPEED	−0.0041***	−0.0050***
EA_SPEED	(−5.06)	(−2.96)
#ANALYSTS	−0.0001	0.0006
#ANALYSTS	(−0.64)	(1.46)
SIZE	0.0012**	0.0010
SIZE	(1.98)	(1.00)
PY_ROA	−0.0011***	−0.0008
PY_ROA	(−3.09)	(−1.19)
LEVERAGE	−0.0012	−0.0031
LEVERAGE	(−0.72)	(−0.98)
AGE	−0.0005	0.0017
AGE	(−0.90)	(1.61)
LN_PRICE	−0.0050***	−0.0038*
LN_PRICE	(−4.99)	(−1.83)
Analyst characteristics
BSIZE	−0.0407***	−0.0378***
BSIZE	(−22.55)	(−10.47)
#FIRMS	0.0000***	0.0000**
#FIRMS	(3.91)	(2.17)
#INDUSTRIES	−0.3973***	−0.4664***
#INDUSTRIES	(−9.35)	(−5.72)
FIRM_EXP	0.2561***	−0.0227
FIRM_EXP	(5.51)	(−0.25)
FCST_FREQ	0.0113	−0.4380***
FCST_FREQ	(0.19)	(−3.96)
Forecast characteristics
FCST_HORIZON	0.0005***	0.0000
FCST_HORIZON	(3.02)	(0.06)
DAYS_LASTFCST	−0.7239***	−0.5105***
DAYS_LASTFCST	(−80.74)	(−27.60)
Observations	166,302	38,614
Adjusted R²	0.205	0.138
Sample	Full	Pseudo-lockdown and lockdown periods
Fixed effects	Firm (Demeaned), week, year	Firm (Demeaned), week

Note. This table presents results from robust regression analyses that assign lower weights to outliers in estimating the minimum aggregated residuals based on the total sample and the subsample from the lockdown and pseudo-lockdown periods. All variables are defined in Appendix A. Robust t-statistics are in parentheses. Standard errors are clustered by state. ***, **, * indicate p-values at 0.01, 0.05, and 0.10, respectively.

With the robust regression we continue to document significantly larger analyst forecast errors during lockdown periods, although with attenuated coefficients on LOCKDOWN. For example, the coefficient drops to 2.5% (0.084/3.388) for the main sample in Table 2, and to 7.0% (0.164/2.354) for the pseudo-lockdown subsample in Table 3. We infer that influential observations amplify, but do not manufacture, the effects we document.

As the robust regression downweights more influential observations, the attenuation of the robust regression coefficients relative to those from the OLS regression is expected and interpretable. The influential observations are likely those most affected by COVID-19 lockdowns, which caused severe and abrupt disruptions to firm operations and information environments. Large forecast errors among the most affected firms reflect real economic conditions, not data artifacts. Leone et al. (2019) emphasize that influential observations driven by real economic events represent legitimate data values, and that downweighting them risks attenuating the estimated effect precisely where treatment was most consequential. The OLS estimation captures the average effects on analyst forecast quality across the full distribution of treated firms, which aligns with our research question more appropriately than the effect from a reweighted distribution. We therefore treat Table 5 as confirming the robustness of our findings rather than as grounds to revise the primary estimates downward.

Additional Analyses of Analyst Effort and Forecast Dispersion

This section describes two additional analyses that examine the effect of reduced in-person information on analyst effort and forecast dispersion. First, we expect analysts to increase their effort to offset limitations to in-person dissemination of information and the associated challenges in interpreting information communicated through virtual channels. We proxy analyst effort input with the number of forecasts issued for each firm in a month (Jacob et al., 1999). We then estimate equation (1) with the natural logarithm of forecast frequency (LN(NUM_FCSTS)) as the dependent variable.

Second, we argue that reduced in-person communication could lead to higher forecast dispersion, an alternative proxy for information quality (De Franco et al., 2011; Gurun & Butler, 2012). This could arise in two ways. The first is through more difficulty in interpreting nonverbal cues and differential influence of this challenge on different analysts’ forecasting ability, which we document in Section 5 and Table 2. This influence likely leads to more divergent opinions among analysts, and thus, higher forecast dispersion (Diether et al., 2002; Thomas, 2002). In the second way, forecasts dispersion could rise when analysts place more weight on their private information and less on public information when the latter has less clarity (De Franco et al., 2011).

On the other hand, if analysts rationally recognize the differential influence of reduced in-person contact on their peers, they may herd towards the forecasts by those less affected, a tendency that would lead to lower forecast dispersion (Liu & Natarajan, 2012; Ramnath et al., 2012). We measure forecast dispersion with the standard deviation of analyst earnings estimates in a month (SD_FCSTS) and estimate equation (1) with SD_FCSTS as the dependent variable.

As both SD_FCSTS and LN(NUM_FCSTS) require a time window to measure, we collapse our data to firm-year-month observations. Accordingly, we modify LOCKDOWN to take the value of 1 if a lockdown appears during the firm-year-month, and 0 otherwise. Among the control variables, the focus at firm-year-month level renders those related to analysts or forecasts unnecessary. We thus retain only firm-related controls and firm and year fixed effects.

Column (1) of Table 6 shows a significant increase in the number of analyst earnings forecasts for firms during the lockdown months. Based on the coefficient on LOCKDOWN in column (1), approximately 1.2 more forecasts are issued in the lockdown month, which is economically significant as compared to the average (median) NUM_FCSTS at 4.90 (3.00). The result is consistent with the descriptive statistics in Section 4 and Table 1, where we note more analyst forecast issuances in the lockdown periods than in the pseudo-lockdown periods for the same firms (25,495 versus 13,267). The result complements finding of increased analyst forecast frequency post the COVID pandemic by Bilinski (2023) and Hao et al. (2022). By focusing on the lockdown periods our results suggest that analysts significantly increase their effort when in-person information is limited.

Table 6.

Firm-Month Analysis of Forecast Frequency and Dispersion During Lockdown Periods

Variables	(1)	(2)
Variables	LN(NUM_FCSTS)	SD_FCSTS
LOCKDOWN	0.1629***	0.1313***
LOCKDOWN	(4.56)	(5.14)
Firm characteristics
ABS(UNEX_EARN)	0.0519	0.9031***
ABS(UNEX_EARN)	(1.12)	(6.25)
MGT_EPS_GUIDE	0.4200***	−0.0548***
MGT_EPS_GUIDE	(18.42)	(−5.21)
EA_SPEED	−0.0000	−0.0035**
EA_SPEED	(−0.00)	(−2.28)
#ANALYSTS	−0.0009	−0.0174***
#ANALYSTS	(−0.40)	(−4.73)
SIZE	0.0019***	0.0035
SIZE	(2.91)	(1.36)
PY_ROA	0.0253	−0.0893*
PY_ROA	(0.72)	(−1.84)
LEVERAGE	0.0777*	0.1005
LEVERAGE	(1.88)	(1.21)
AGE	−0.3744***	−0.7918**
AGE	(−4.02)	(−2.12)
LN_PRICE	0.0017	−0.0263
LN_PRICE	(0.23)	(−1.07)
Observations	33,461	24,237
Adjusted R²	0.370	0.532
Fixed effects	Firm, year	Firm, year

Note. This table examines the effect of lockdowns on monthly analyst forecast frequency (NUM_FCSTS) and monthly dispersion of 1-year ahead earnings forecasts (SD_FCSTS). As the observations are for each firm-month, analyst level control variables are omitted. All variables are defined in Appendix A. Robust t-statistics are in parentheses. Standard errors are clustered by state. ***, **, * indicate p-values at 0.01, 0.05, and 0.10, respectively.

Column (2) of Table 6 reports a significant increase in analyst forecast dispersion during the month firms are subject to the lockdowns, by 40.53% of the sample mean at 0.324 or over 100% of the median at 0.120. Taken together, Table 6 shows that when in-person information is limited analysts both increase effort and exhibit more disagreement. In other words, despite more effort, analysts issue forecasts of lower quality during the lockdown periods when receiving less in-person communication from firms.

Alternative Measures of Lockdown

While the stay-at-home orders proxy for a plausibly exogenous reduction in firms’ in-person communication to market participants, using the period right after the effective date is subject to some noise when in-person communication occurred immediately before state-mandated lockdowns. To mitigate this confounding effect, we recode LOCKDOWN as 0 for forecasts issued during the lockdown period but within 1 month of an in-person event held just before the mandate, and label this adjusted measure ADJ_LOCKDOWN.¹⁴

To identify such in-person communication, we use the corporate event data discussed in the previous section and in Appendix B. We replicate column (1) of Table 2 after replacing LOCKDOWN in regression (1) with ADJ_LOCKDOWN and document consistent results in Table OA1 in the Online Appendix. Relative to the LOCKDOWN coefficients in Table 2 and from the untabulated analyses that include various fixed effects, the ADJ_LOCKDOWN coefficients in Table OA1 are slightly larger (0.280, 0.390, and 0.267 versus 0.274, untabulated 0.381, and untabulated 0.260) and more statistically significant in two out of the three analyses. This suggests that the modified measure contains less measurement noise and better captures the effect of limited in-person information. The results strengthen our confidence that forecasts issued during the lockdown without information from in-person communication are indeed less accurate.

Control for Management Guidance Issuance

Hope et al. (2023) document 272 management guidance withdrawals in March 2020 because of economic uncertainty and increased analyst forecast dispersion following the withdrawal. We have controlled for management guidance issuance in the main analyses. To specifically address the possibility that lower management disclosure explains our findings, we separately examine firms that issued guidance during the 2019 pseudo-lockdown period but discontinued it during the 2020 lockdown period (STOP_GUIDE). Per results in Table OA2 in the Online Appendix, for these firms, the coefficients on LOCKDOWN are not significantly different from those for other firms, suggesting that changes in management earnings guidance disclosures are unlikely to fully explain our results.

Results Among Essential Business Industries

This section restricts the sample to firms in industries designated as essential businesses under federal critical infrastructure guideline.¹⁵ Although lockdowns required all firms to hold corporate events virtually, they likely had a smaller impact on the business fundamentals of essential firms. Thus, changes in analyst forecast accuracy for these firms are less likely to be attributable to shifts in business fundamentals. We classify grocery (SIC 54), banks (60), utilities (49), hardware (52), communications (48), and defense (97) as essential industries and exclude those directly affected by the pandemic, including healthcare, gas stations, and transportation. Panel A of Table OA3 in the Online Appendix indicates that the magnitude of the coefficients on LOCKDOWN is comparable to those described in Section 5 based on the full sample. Panel B shows that the results remain robust to addressing outliers with robust regressions. These results further support the role of communication modality in driving the forecast errors.

Conclusion

We document a significant impact of in-person communication on analyst forecast accuracy. Focusing on firms whose headquarters undergo state-mandated lockdowns in early 2020 due to the COVID-19 pandemic, a period with little or no in-person communication from firms, we find that forecast errors for firms under lockdowns are significantly higher than in other periods or for other firms in the same periods. The increase amounts to about 8% of the mean absolute forecast error, suggesting that reduced in-person communication impairs the quality of information available to analysts. The effect is stronger for analysts with less firm-specific experience and broader industry coverage, as well as for smaller or younger firms. These are analysts that presumably face more challenges in filling the information gap caused by reduction of in-person communication, and firms that tend to have weaker information environment.

Further analyses show that firms switching from in-person events in the 2019 pseudo-lockdown periods to virtual events during lockdown experience higher forecast errors, whereas firms that maintain virtual communications do not. Although low power associated with the small number of virtual events in the pseudo-lockdown periods could influence the latter result, we note that the untabulated result from examining a larger number of virtual events in the full year of 2019 remain consistent. We also find that analysts exert greater effort and exhibit higher forecast dispersion during lockdowns.

Our finding of larger forecast errors during the lockdown periods is robust to a plethora of additional tests, including considering whether analysts are located in states under lockdowns, applying robust regressions to address outliers, adjusting the lockdown measure to reduce potential measurement error, accounting for firms that withdrew management guidance, and restricting the sample to essential industries whose business fundamentals are less affected by COVID-19. Collectively, the evidence indicates that despite increased analyst effort, reduced in-person interaction lowers analyst forecast quality. Although we cannot fully rule out correlated omitted variables confounding our results, the consistency of the results supports a role for communication modality in driving forecast errors.

These findings underscore the importance of in-person communication in facilitating information flow in capital markets and highlight potential information quality deterioration upon shifting to virtual formats. Firms weighing virtual versus in-person modalities should balance the savings in cost and time and broader participation associated with virtual communication against the potential decline in perceived information quality. Such a tradeoff is especially relevant for companies in weaker information environments and for market participants facing greater challenges in interpreting more subtle information.

Supplemental Material

Supplemental Material - The Impact of Reduced In-Person Contact on Information Quality: The Case of Analyst Forecast Accuracy During COVID Lockdowns

Supplemental Material for The Impact of Reduced In-Person Contact on Information Quality: The Case of Analyst Forecast Accuracy During COVID Lockdowns by Trent Krupa, Yanhua Sunny Yang, Xiao Yu in Journal of Accounting, Auditing & Finance

Footnotes

Acknowledgements

We are grateful for the support of our respective institutions and for helpful comments from Xiao-Jun Zhang (Editor in Chief), Suresh Radhakrishnan (Associate Editor), an anonymous reviewer, Vishal Baloria, Alina Lerman, Frank Murphy, George Plesko, Michael Willenborg, Ying Zhou, and seminar participants at the University of Connecticut and the University of Massachusetts at Dartmouth. We thank Tyler Johnson, Vernan Rivera, Ashley Schmid, and Anthony Barile for their excellent research assistance.

ORCID iDs

Trent Krupa

Yanhua Sunny Yang

Xiao Yu

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Supplemental Material

Supplemental material for this article is available online.

Notes

Author Biographies

Trent Krupa is an assistant professor in accounting at the Pennsylvania State University. He previously held a faculty position at the University of Arkansas and received his PhD in accounting from the University of Connecticut. His research examines how tax law, regulation, and information shape corporate behavior and economic outcomes.

Yanhua Sunny Yang is an associate professor in accounting at the University of Connecticut. She previously held a faculty position at the University of Texas at Austin and received her PhD in accounting from the University of Colorado at Boulder. Her research interests include financial analysts, initial public offerings, management disclosure, and credit ratings.

Xiao Yu is a PhD candidate in Accounting at the University of Connecticut. Her broad research interests are in financial accounting and focuses are on understanding the production and quality of accounting information and labor economics. More specifically, she studies information discovery and dissemination by information intermediaries; factors that influence the quality of information flow to capital markets; and labor-related decisions involving both top executives and rank-and-file employees.

References

Armstrong

(2006). How to make better forecasts and decisions: Avoid face-to-face meetings. FORESIGHT: The International Journal of Applied Forecasting, 5(5), 3–15.

Baek

(2023). The effect of telecommuting on information acquisition: Evidence from the U.S. Patent Office. The Accounting Review, 98(3), 21–46. https://doi.org/10.2308/TAR-2020-0153

Barry

Brown

(1985). Differential information and security market equilibrium. Journal of Financial and Quantitative Analysis, 20(4), 407–422. https://doi.org/10.2307/2330758

Ben-Rephael

Carlin

Israelsen

(2023). Security analysis and the collection of hard and soft information. NBER working paper. https://www.nber.org/papers/w31936

Bilinski

(2023). Analyst research activity during the COVID-19 Pandemic. Abacus, 59(4), 1041–1073. https://doi.org/10.1111/abac.12291

Bloom

Liang

Roberts

Ying

Z. J.

(2015). Does working from home work? Evidence from a Chinese experiment. The Quarterly Journal of Economics, 130(1), 165–218. https://www.jstor.org/stable/26372598.

Brochet

Chychyla

Ferri

(2023). Virtual shareholder meetings. Management Science, 70(9), 5896–5930. https://doi.org/10.1287/mnsc.2023.4946

Brodsky

Tolliver

(2022). No, remote employees aren’t becoming less engaged. Harvard Business Review. https://hbr.org/2022/12/no-remote-employees-arent-becoming-less-engaged?ab=hero-subleft-2

Brown

Call

Clement

Sharp

(2015). Inside the “Black Box” of sell-side financial analysts. Journal of Accounting Research, 53(1), 1–47. https://doi.org/10.1111/1475-679X.12067

10.

Brucks

M.S.

Levav

(2022). Virtual communication curbs creative idea generation. Nature, 605(7908), 108–112. DOI:https://doi.org/10.1038/s41586-022-04643-y

11.

Bushee

Gerakos

Lee

L. F.

(2018). Corporate jets and private meetings with investors. Journal of Accounting and Economics, 65(2–3), 358–379. https://doi.org/10.1016/j.jacceco.2018.01.005

12.

Bushee

Jung

J. M.

Miller

(2011). Conference presentations and the disclosure milieu. Journal of Accounting Research, 49(5), 1163–1192. https://doi.org/10.1111/j.1475-679X.2011.00426.x

13.

Causholli

Cheng

Golshan

(2022). Does working from home impact audit quality? Evidence from non-pharmaceutical intervention during COVID-19. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4155625

14.

Clement

(1999). Analyst forecast accuracy: Do ability, resources, and portfolio complexity matter? Journal of Accounting and Economics, 27(3), 285–303. https://doi.org/10.1016/S0165-4101(99)00013-0

15.

Clement

Tse

(2003). Do investors respond to analysts’ forecast revisions as if forecast accuracy is all that matters? The Accounting Review, 78(1), 227–249. https://doi.org/10.2308/accr.2003.78.1.227

16.

D’Augusta

Grossetti

(2023). How did Covid-19 affect investors’ interpretation of earnings news? The role of accounting conservatism. Finance Research Letters, 52, Article 103504. https://doi.org/10.1016/j.frl.2022.103504

17.

De Franco

Kothari

S. P.

Verdi

R. S.

(2011). The benefits of financial statement comparability. Journal of Accounting Research, 49(4), 895. https://doi.org/10.1111/j.1475-679X.2011.00415.x

18.

Deloitte Global Services Limited . (2012). The leadership premium: How companies win the confidence of investors. https://www2.deloitte.com/az/en/pages/human-capital/articles/leadership-premium.html

19.

Diether

K. B.

Malloy

C. J.

Scherbina

(2002). Differences of opinion and the cross section of stock returns. Journal of Finance, 57(5), 2113–2141. https://www.jstor.org/stable/3094506.

20.

Down

A. K.

Reiter

Wahid

A. S.

(2023). The adaptability of investor relations in the face of disruption and the effectiveness of virtual investor engagement. https://ssrn.com/abstract=4175392

21.

(2023). Locked-in at home: The gender difference in analyst forecasts after the COVID-19 school closures. Journal of Accounting and Economics, 76(1), Article 101603. https://doi.org/10.1016/j.jacceco.2023.101603

22.

Durney

Kyung

Markov

Park

(2024). Forecasting earnings from home. Management Science. https://doi.org/10.1287/mnsc.2024.04637

23.

Fabrizi

Ipino

Longhin

Parbonetti

(2023). The informativeness of earnings announcements during times of global uncertainty: Evidence from the Covid-19 pandemic. Corporate Governance: An International Review, 31(5), 795–813. https://doi.org/10.1111/corg.12552

24.

Ferrazzi

Zapp

(2020). The upside of virtual board meetings. Harvard Business Review. https://hbr.org/2020/07/the-upside-of-virtual-board-meetings

25.

Gajendran

R. S.

Harrison

D. A.

(2007). The good, the bad, and the unknown about telecommuting: Meta-analysis of psychological mediators and individual consequences. Journal of Applied Psychology, 92(6), 1524–1541. https://doi.org/10.1037/0021-9010.92.6.1524

26.

Gurun

Butler

(2012). Don’t believe the hype: Local media slant, local advertising, and firm value. Journal of Finance, 67(2), 561–598. https://www.jstor.org/stable/41419705.

27.

Hao

Xue

Yau

L. N. B.

Zhang

(2022). Analyst forecasting during COVID-19 pandemic. Managerial Auditing Journal, 37(3), 380–405. https://doi.org/10.1108/MAJ-12-2021-3406

28.

Hope

O. K.

(2023). Is silence golden sometimes? Management guidance withdrawals during the COVID-19 pandemic. Review of Accounting Studies, 28, 2319–2360. https://doi.org/10.1007/s11142-022-09698-w

29.

IR Magazine . (May 18, 2023). IR magazine research report: Investor events. https://www.irmagazine.com/research-reports/investor-events-report-2023

30.

Jacob

J. T.

Lys

Neale

(1999). Expertise in forecasting performance of security analysts. Journal of Accounting and Economics, 28(1), 51–82. https://doi.org/10.1016/S0165-4101(99)00016-6

31.

Jacobsen

(2020). Statewide COVID-19 stay-at-home orders and population mobility in the United States. World Med Health Policy, 12(4), 347–356. https://doi.org/10.1002/wmh3.350

32.

Kale

(2023). Do exogenous economic crises change investors’ response to earnings announcements? A detailed review using the data from COVID-19 pandemic. The Journal of Economic Asymmetries, 28, Article e00330. https://doi.org/10.1016/j.jeca.2023.e00330

33.

Lang

Lins

Miller

(2003). ADRs, analysts, and accuracy: Does cross listing in the United States improve a firm’s information environment and increase market value? Journal of Accounting Research, 41(2), 317–345. https://doi.org/10.1111/1475-679X.00106

34.

Leone

Minutti-Meza

Wasley

(2019). Influential observations and inference in accounting research. The Accounting Review, 94(6), 337–364. https://doi.org/10.2308/accr-52396

35.

Liu

Natarajan

(2012). The effect of financial Analysts’ Strategic Behavior on Analysts’ forecast dispersion. The Accounting Review, 87(6), 2123–2149. https://doi.org/10.2308/accr-50212

36.

Liu

Myers

Wang

Ward

(2023). The impact of In-Person interaction with parties near covered firm headquarters on analyst forecast accuracy. https://doi.org/10.2139/ssrn.5406762

37.

Mikhail

Walther

Willis

(1997). Do security analysts improve their performance with experience? Journal of Accounting Research, 35(1), 131–157. https://doi.org/10.2307/2491458

38.

Ramnath

Rock

Shane

(2012). The financial analyst forecasting literature: A taxonomy with suggestions for further research. International Journal of Forecasting, 24(1), 34–75. https://doi.org/10.1016/j.ijforecast.2007.12.006

39.

Rhoads

(2010). Face-to-Face and computer-mediated communication: What does theory tell us and what have we learned so far? Journal of Planning Literature, 25(2), 111–122. https://doi.org/10.1177/0885412210382984

40.

Schwartz-Ziv

(2021). How shifting from In-Person to virtual-only shareholder meetings affects shareholders’ voice. https://ssrn.com/abstract_id=3674998

41.

Senju

Johnson

M. H.

(2009). The eye contact effect: Mechanisms and development. Trends in Cognitive Sciences, 13(3), 127–134. https://doi.org/10.1016/j.tics.2008.11.009

42.

Soltes

(2014). Private interaction between firm management and sell-side analysts. Journal of Accounting Research, 52(1), 245–272. https://doi.org/10.1111/1475-679X.12037

43.

Thomas

(2002). Firm diversification and asymmetric information: Evidence from analysts’ forecasts and earnings announcements. Journal of Financial Economics, 64(3), 373–396. https://doi.org/10.1016/S0304-405X(02)00129-0

44.

Verrecchia

(1990). Information quality and discretionary disclosure. Journal of Accounting and Economics, 12(4), 365–380. https://doi.org/10.1016/0165-4101(90)90021-U

45.

Waymire

(1985). Earnings volatility and voluntary management forecast disclosure. Journal of Accounting Research, 23(1), 268–295. https://www.jstor.org/stable/2490919.

46.

Zhang

(2006). Information uncertainty and stock returns. Journal of Finance, 61(1), 105–136. https://doi.org/10.1111/j.1540-6261.2006.00831.x

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.10 MB

0.00 MB