key: cord-0986645-w6yhtcxn
authors: Dietrich, Alexander M.; Kuester, Keith; Müller, Gernot J.; Schoenle, Raphael
title: News and uncertainty about COVID-19: Survey evidence and short-run economic impact()
date: 2022-02-09
journal: J Monet Econ
DOI: 10.1016/j.jmoneco.2022.02.004
sha: 854fa1163368d863db0aa7158cb6eed0f7eccb7b
doc_id: 986645
cord_uid: w6yhtcxn

A tailor-made survey documents consumers’ perceptions of the US economy’s response to a large shock: the advent of the COVID-19 pandemic. The survey ran at a daily frequency between March 2020 and July 2021. Consumer’s perceptions regarding output and inflation react rapidly. Uncertainty is pervasive. A business-cycle model calibrated to the consumers’ views provides an interpretation. The rise in household uncertainty accounts for two-thirds of the fall in output. Different perceptions about monetary policy can explain why consumers and professional forecasters agree on the recessionary impact, but have sharply divergent views about inflation.

Expectations are central to economic decision making and so is the expectation formation process. The 2 full information rational expectations hypothesis (FIRE) provides a natural benchmark according to which 3 people adjust their expectations adequately and immediately in the face of new information. Survey evidence, 4 by contrast, suggests that expectations tend to adjust only sluggishly to macroeconomic shocks. This 5 holds not only for professional forecasters but also for policymakers, firms, and households (Coibion and 6 Gorodnichenko, 2012). 1 In normal times, the response of households tends to be even more sluggish than 7 the response of professional forecasts (Carroll, 2003; Carroll et al., 2020) . There is, however, evidence that 8 expectations adjust more quickly in times of high uncertainty, in response to large shocks, and as media 9 coverage intensifies (Coibion and Gorodnichenko, 2015; Baker et al., 2020b; Larsen et al., 2021) . All of these 10 conditions are met in the context of the COVID-19 pandemic: it offers a natural experiment to study the 11 expectation formation process in some detail. 12 In order to do so, we exploit a special resource: a daily survey of consumer expectations that we have 13 been running since the start of the pandemic. The survey asks a representative sample of consumers in 14 real time-that is, starting with the onset of the pandemic-how they expect the COVID-19 shock to affect 15 income and inflation over a 12-month horizon. We find that consumer expectations respond very rapidly 16 and that uncertainty about the economic effects of the shock is pervasive-and much more so than what 17 comparable measures for professional forecasters suggest. Our survey is unique in that it directly elicits the shift in consumers' conditional expectations that the COVID-19 shock brings about, in real time. Translated the majority of respondents expect the pandemic to raise prices while lowering GDP. In this, consumers' 23 inflationary views stand in sharp contrast to the views held by professional forecasters. Observation 4: the 24 pandemic strongly raises consumer uncertainty about future inflation. Once more, the impact on uncertainty 25 is much more pronounced for consumers than for professional forecasters. 26 The second part of the paper rationalizes these observations based on a FIRE model. Here we make 27 only a first pass at the data, but FIRE seems justified from an ex ante point of view in light of the distinct 28 features of the pandemic-the massive increase in uncertainty, the large shock, and the intensive news 29 coverage. The model is a simplified version of the representative-household New Keynesian framework with 30 demand uncertainty shocks developed by Basu and Bundick (2017). 2 Next to such demand uncertainty 31 shocks, business cycles arise in the model because of level shocks to demand preferences and to supply. The 32 latter set of shocks also features news about future supply conditions, since anticipation of future supply 33 disruptions seems essential for pandemics. We solve the model controlling for the effective lower bound on 34 interest rates. 35 In the model, no single shock in isolation is sufficient to generate the identified moments. Hence, we 36 devise a COVID-19 scenario based on several large shocks. So as to replicate the response of uncertainty in 37 the survey, for example, the COVID-19 scenario sees an increase in the volatility of demand shocks by 17.5 38 standard deviations. These shocks reduce the natural rate of interest by 15 percentage points (annualized) 39 and make the effective lower bound bind. Next, comparably large adverse news shocks about productivity 40 are required to match households' stagflationary views. The representative-household FIRE approach, thus, 41 provides a nuanced interpretation of the survey facts. policies, the outcomes differ. In the counterfactual, the outcomes look much more similar to the professional 1 forecasters' view of the pandemic: the pandemic's effect on uncertainty about output and inflation falls by 2 half. In addition, the inflation response switches sign. The reason for this is that tracking the natural rate of 3 interest prevents two effects that are inflationary: it avoids the future accommodation built into the baseline 4 policy and the uncertainty about marginal costs (Fernndez-Villaverde et al., 2015). 5 Thus, the same model-with the same size and timing of shocks-can replicate both the consumers' and 6 the professional forecasters' views on the impact of the pandemic, the only difference being the perceived 7 monetary policy response. This suggests an important policy implication: communicating effectively with the 8 broader public (and not only professional forecasters) about monetary policy and the state of the economy (as 9 captured by the natural rate) could itself dampen economic uncertainty and the fallout after large unexpected 10 shocks. 11 There is a different reading of the model-based exercise, of course: namely, that it points toward important 12 gaps in modeling and survey methodology that future work should address. The shocks that are needed to 13 make the model replicate the survey evidence are very large. Depending on one's view, this may adequately 14 reflect the depth of the recession or cast doubt on the representative-household FIRE model. By its very 15 nature this approach also cannot account for the heterogeneity of households and of households' expectations, 16 both of which could make activity more exposed to shocks. Another weakness is that we cannot account for Relative to these papers, our survey makes three contributions. First, we identify expectations conditional 27 on an exceptionally large shock. Second, we do so in real time at high frequency. Complementary work by 28 Andre et al. (2021) also studies expectations conditional on shocks, but they consider hypothetical shocks 29 rather than the exceptional event that is the focus of our paper. Relative to other existing surveys, such as 30 the Survey of Consumer Expectations and its analysis in regard to COVID-19, for example, by Armantier Third, we use the survey responses as identified moments in order to calibrate a business-cycle model. This, 33 in turn, allows us to analyze the role of expectations in transmitting the shock. Our analysis also relates to 34 work by Bloom (2009) and many others who have stressed the role of uncertainty as a potential source and 35 amplification channel of the business cycle, a view recently supported by direct survey evidence (Coibion 36 et al., 2021). 37 The remainder of the paper is structured as follows. We introduce our survey in the next section and 38 present the main results of the survey in Section 3. Section 4 introduces our business-cycle model, which 39 allows us to develop a structural scenario for the expected impact of the COVID-19 shock. A final section 40 concludes. 41 

The survey that we run is unique in two ways. First, our survey systematically introduces questions that 43 elicit conditional expectations on prices, quantities and behavioral variables. Namely, we ask respondents 44 to assess the impact of COVID-19 on their outlook for the economy. In doing so, our work presents the 45 empirical counterpart of the hypothetical "vignettes" in Andre et al. (2021) . These conditional expectations 46 correspond closely to how shocks move expectations in the context of models. As such, the questions allow 47 for a tighter identification of the impact of specific shocks than would eliciting conventional unconditional 48 expectations. We find that conditional and unconditional expectations differ for GDP and personal household 49 income but are similar for inflation. At the same time, disagreement is relatively similar across all variables (see Supplementary Material C.8). 1 Second, the high-frequency approach is a distinct feature of our survey. It is rooted in a daily sample 2 of respondents, an approach that presents a large option value for policymaking in practice and real time. 3 However, we do not exploit the high-frequency feature further. We contracted with Qualtrics Research Services to provide us with a survey of 60,003 nationally repre-6 sentative respondents for 16 months between March 10, 2020, and July 11, 2021. The survey was run with a 7 daily sampling size of at least 100 respondents. Over the course of one month the number of survey responses 8 (above 3000) compares favorably to that of existing consumer surveys. Balancing a more granular view on 9 the expectations process with a larger, less noisy sample size, we mainly report 11-day moving averages 10 below. The survey required all respondents to be US residents and speak English as their primary language. 11 Other than this, our sample was taken to be representative of the US population. 12 In terms of demographics, respondents had to be male or female with 50% probability. Moreover, approx-13 imately one-third of respondents were targeted to be between 18 and 34 years of age, another third between 14 ages 35 and 55, and a final third older than age 55. We also required a distribution across US regions in 15 proportion to population size, drawing 20% of our sample from the Midwest, 20% from the Northeast, 40% 16 from the South and 20% from the West. 17 The survey includes filters to eliminate respondents who enter gibberish for at least one response or who 18 complete the survey in less (more) than five (30) minutes. We also employ CAPTCHA tests to reduce the 19 possibility that bots participate in the survey. 22 Respondents match the US population demographics along key dimensions. To improve the fit further, we 23 additionally compute a survey weight for each respondent. To do so, we apply iterative proportional fitting 24 to create respondent weights after completion of the survey (raking, see, for example, Bishop et al. (1975) 25 or Idel (2016) ). This allows us to calculate statistics that are exactly representative of the US population 26 also according to ethnicity, income and education, that is, the variables in the right column of Table 1 . forecast, we directly elicit consumers' assessments of the "impact of the coronavirus" or changes in economic 31 aggregates "because of the coronavirus." Otherwise, we stick to the wording of the SCE as closely as possible.

While we keep the way of measuring inflation the same as in the SCE, we elicit responses for two different 33 measures of income. On the one hand, we follow the SCE by asking for the "total income of all members 34 of your household (including you)." On the other hand, we are interested in GDP as a measure of income, in terms of GDP over the next 12 months, averaged across respondents daily (jagged black solid line), the 5 11-day moving average thereof (red dotted line), and a measure of the impact of the pandemic as viewed 6 by the professional forecasters that contribute to the Blue Chip forecasts (blue dashed line). 4 In addition, 7 panel (a) also reports how much actual GDP over the next 12 months has deviated from the pre-pandemic 8 trend (triangles). Panel (b) refers to personal household income. 9 What is most striking, perhaps, is the speed with which consumer expectations react: by late March/early 10 April 2020 the average expected GDP impact (across households) is close to -15%. The maximum effect in 11 terms of the moving average is -18% and observed on April 01, 2020. For personal household income, we 12 observe the maximum drop in conditional income expectations on March 24; it is -13%. What is striking, too, is that consumers' expectations initially react by more than professional fore-21 casters'. 5 This is noteworthy because it seems to run counter to the received idea of sluggish responses of 22 household expectations and a one-way information flow from professional forecasts to households (Carroll, 23 2003). Instead, household expectations and the Blue Chip forecasts converge to the middle ground: by 24 May/June 2020 they are very well aligned at approximately a -7% impact on GDP, and remain surprisingly 25 aligned all the way until the end of our sample period. 26 Our measures of income-GDP and household income-require different levels of abstraction of house- ity that, below, leads us to rely on personal household income expectations for our model-based calibration 33 targets for the impact on output.

The COVID-19 shock also triggered a massive increase in uncertainty. This increase has been documented uncertainty also shows up on the consumer side. 6 In particular, regarding our survey we make 38 Observation 2 (Income Uncertainty). Consumers' uncertainty about the impact of the pandemic in terms 39 of GDP rises fast, faster and by much more than professional forecasters'. First, consumer disagreement in our data leads disagreement of professional forecasters, suggestive of a 2 real-time information content of the daily consumer survey. Second, consumer disagreement rises by an 3 order of magnitude more than disagreement of professional forecasters (recall that in panel (a) we measure 4 disagreement against the left axis for our survey and against the right axis for the Blue Chip survey). This 5 finding is, perhaps, not entirely unexpected, given that our survey's respondents, consumers, are considerably 6 more heterogeneous than the respondents in the Blue Chip survey. Consumer uncertainty about the economic 7 effects on income is high from the start of the survey/pandemic. 8 [FIGURE 2 ABOUT HERE] 9 Later, in Section 4 we will use a representative-household model to take a first pass at the role that 10 consumer uncertainty had in shaping the recession. By its nature, this model does not feature disagreement. 11 Therefore, to calibrate the model, we rely on a second measure of household uncertainty about income: un- The Supplementary Material revisits the conceptual concern of how well households understand the In addition to income expectations, our survey asks respondents about the likely impact of the pandemic 27 on inflation. We summarize the results with 28 Observation 3 (Inflation Expectations). On average, consumers expect inflation to rise strongly in response 29 to the COVID-19 shock, in contrast to professional forecasters, who expect a deflationary effect. Moreover, 30 most consumers expect an inflationary impact, independently of whether they expect economic activity to 31 contract or rise in response to the pandemic. consumers' uncertainty regarding the inflationary impact of the pandemic is also pronounced if we turn to 41 the measure of subjective uncertainty. 9 42 8 Our survey documents that households disagree about the impact of the COVID-19 shock on inflation. That households hold heterogeneous inflation expectations more generally is well documented, for example, by Mankiw et al. (2004) . 9 The survey also includes questions on savings and purchasing behavior and plans in response to COVID-19, the expected duration of the pandemic, and whether respondents have hoarded food and medical supplies. Economic expectations elicited within our survey vary in a meaningful way with the behavioral adjustments and financial decisions of survey participants. We also document demographic and socio-economic heterogeneity in expectations. Supplementary Material C provides these findings.

of the COVID-19 pandemic. What remains to be understood are the potential mechanisms that are behind 1 the survey responses and their implications. Toward this end, we now put forward a business-cycle model 2 for which we devise a specific COVID-19 scenario. 3 The model assumes rational expectations and full information (FIRE) even though evidence suggests 4 that this assumption is generally too restrictive Gorodnichenko, 2012, 2015) . Against this 5 background, we understand our modelling exercise as a first pass in accounting for the response of consumer 6 expectations to the COVID-19 shock. This seems reasonable because of three distinct features of this very 7 special episode. First, the increase in uncertainty was massive and Coibion and Gorodnichenko (2015) find 8 that information rigidities decline precisely in times of increased macroeconomic uncertainty. Specifically, 9 their estimate of information rigidities for the volatile 1970s and early 1980s suggests that the distance to 10 FIRE was smaller. Second, the media focus on COVID-19 was also exceptional. This matters since Larsen follows the exposition in BB (and their notation) closely. 26 There is a representative household that has Epstein-Zin preferences over current and future consumption, 27 C t , and hours worked, N t . The household faces competitive labor, goods, and financial markets. Let 

E t is the expectation operator, a t is a preference shifter ("demand shock") and η ∈ (0, 1). The household 32 purchases consumption at nominal price P t per unit. In addition, the household can buy infinitely lived shares 33 S t+1 at price P E t or a real one-period pure discount bond bearing real gross interest R R t . The household 34 funds these expenditures through labor income (with W t marking the nominal wage rate) and past savings. 10 Here and in the following, to preserve on notation, we use t interchangeably as an indicator of time, a summary measure of the information set in period t, or to mark states in period t.

subject to the production function

Above, M t,t+s is the stochastic discount factor arising from the household problem. It prices in period t 6 claims in t + s. θ µ > 1 is the elasticity of demand, α ∈ [0, 1), Φ is a fixed cost measured in terms of goods 7 used in the production process, and φ p > 0 indexes price adjustment costs. Capital is in fixed supply and 8 does not depreciate. A bar on top of a variable marks the variable's steady-state value. So, for example, 9 the presence of Π above reflects that prices are indexed to steady-state inflation, where inflation is given by 10 Π t := P t /P t−1 .

As in BB, each intermediate goods firm is assumed to issue real bonds in proportion to the capital stock, 12 B t (i) = νK, with ν ∈ [0, 1). Total cash flows of the firm are divided between dividends to equity holders and 13 interest paid to bond holders, so that dividends are given by

The financing 14 structure of the firm is without consequence; it only serves to introduce the return to equity as an observable 15 variable. 16 The monetary policy instrument is the gross nominal interest rate R t on one-period risk-free nominal 17 bonds that are in zero net supply. Let R tar t denote the target interest rate and R the effective lower bound on 18 gross interest rates, such that R t = max[R tar t , R]. For the target interest rate itself we assume a conventional

Taylor rule:

where ρ Π > 1 and ρ y ≥ 0 determine the responses to inflation and the output gap, respectively. The output 21 gap, Y n t , is defined as the gap between actual output and its natural level. The equilibrium condition for 22 nominal bonds is given via the conventional consumption Euler equation

t 's mark iid, zero-mean, standard-normal innovations. Following BB, there are both first-moment 24 shocks to demand and "uncertainty" shocks to demand, namely:

Productivity is a convolute of two components. The first is a front-loaded productivity component as in 26 BB. The second component allows a gradual build-up of productivity (the news shock). Namely, productivity

log(X t ) = ρ X,1 log(X t−1 ) + ρ X,2 log(X t−2 ) + σ X X t . We include a news component X t about future productivity since we consider the news important for tracing 29 some of the features of the COVID-19 crisis, the anticipation inherent in the survey, in particular. In each 30 case, the shock processes' parameters are restricted such that all the shocks are stationary. 31 In equilibrium, all intermediate goods firms choose the same price. Hence, they all have the same level of production, the same demand for inputs, and the same financing structure. Goods market clearing implies Y t (i) = Y t and

Labor-market clearing implies N t (i) = N t . Next, the bond and equity markets clear, so that D E t (i) = D E t 32 and B t (i) = B t .

1 by HP-filtered data. The model-based moments are computed without any adjustments for a potential zero 25 lower bound, using third-order perturbation. In line with this, we also report the data counterparts only 26 for the period before the lower bound became binding (1984Q1 to 2008Q3) . Overall, the model appears to 27 paint a reasonable picture of the standard business cycle. 28 In all the simulations that follow, we allow the conditional mean dynamics of the nominal interest rate 29 to fall at most 1.5 percentage points (annualized) below the steady state. We do so to mimic the room for To map the survey responses into the model, we devise a COVID-19 scenario based on a range of shocks.

In the survey we do not ask respondents about specific macro shocks as, for instance, Andre et al. (2021) do. 37 Rather, we ask respondents about the impact of the pandemic. Here we therefore specify a combination of 38 shocks that is meant to rationalize the conditional expectations we obtain from the survey; that is, we target 39 moments identified by the survey. The set of shocks includes a demand uncertainty shock, TFP shocks, and 40 a level shock to demand preferences. The aim of the scenario we develop is to replicate the main patterns 41 of the survey responses, taking a representative-agent perspective. 42 A key feature that emerges is that the required shocks are large, reflecting the extent of the effects manifest 43 in the survey responses. An uncertainty shock to demand helps us to target the patterns of observation we assume a 17.5-standard-deviations pandemic rise in the volatility of the demand shock (σ a t ).

Next, a fall in TFP now or later is essential for mimicking consumers' stagflationary view of the recession, All of the COVID-19 scenario simulations assume that monetary policy is expected to stabilize economic 6 activity at its pre-pandemic (no-shock) level. That is, monetary policy is expected to follow rule (3) as in 7 normal times, but with ρ y multiplying a conventional measure of the output gap, log(Y t /Y ), rather than 8 the gap between output and flex-price output. 12 The model does not feature cost-push shocks. At the same In our view, there are two possible ways to think of our COVID-19 scenario. First, the pandemic was just exceptional, and hence, it requires exceptionally large shocks to rationalize the response of expectations to the pandemic. Second, the size of the shocks testifies to the limitations of the representative-agent FIRE model and strengthens the case to move beyond that framework.

The top row of Figure 5 shows the role that news about future productivity (X t ) plays in shaping 1 the model-based recession. The panels plot the perceived impact of COVID-19 in the baseline scenario 2 (red dashed lines) against a counterfactual that is identical except that the news component is mute (blue 3 dashed-dotted line). The right panel shows the response of inflation: the negative news shock to productivity 4 is essential for explaining the stagflationary response, observation 3 (inflationary beliefs). The reason is 5 simple. In the scenarios here, the central bank leans against lower future productivity, keeping the real 6 rate below the natural rate of interest-unless it is constrained by the effective lower bound. This policy 7 response raises future marginal costs. Forward-looking price setters respond by raising prices already at the 8 onset of the pandemic (red dashed line and corresponding bounds). Absent the news shock, instead, future 9 marginal costs do not rise, and inflation falls on impact (blue dashed dotted line). Note that this means 10 that-whenever the effective lower bound is binding-the real rate of interest is higher without the news 11 shock. This in turn explains why the response of output is of comparable magnitude both with and without 12 an initial negative news shock (see the left panel of the first row). 13 The bottom row of Figure 5 shows the role that the shock to demand uncertainty plays in the model- 14 based recession. The baseline is identical to the top row (and is shown as a red dashed line again). The 15 blue dashed-dotted line now shows the response of the economy if the shock to demand uncertainty is mute. 16 In interpreting this, it is important to note that-unless monetary policy is constrained-in the context of The rise in demand uncertainty means that the natural rate of interest falls sharply in the baseline, by 15 21 pp. annualized (the bottom left panel of Figure 6 shows the response of the natural rate). Most of this is due 22 to the shock to demand uncertainty. 13 This means that the effective lower bound on interest rates becomes 23 binding, and monetary policy cannot accommodate this shock. A deep recession ensues. Absent the shock 24 to demand uncertainty, output falls only about a third as much as in the baseline (bottom row, left panel).

Note that the demand uncertainty shock is also the primary driver of the rise in consumer uncertainty itself, 26 observations 2 and 4. Absent the direct effect on uncertainty, the pandemic shock would hardly affect 27 the standard deviation of (that is, the uncertainty about) output and inflation, as can be seen by observing 28 that the bands almost coincide with the effect of the shock on the means.

In sum, the above suggests that heightened consumer uncertainty about demand itself may have been 35 shaping the uncertainty that consumers face, a role to which we turn next. In detail, under the alternative policy, the target interest rate is now governed by

so that-all else equal-the interest rate tracks the actual natural rate of interest one-to-one. 8 The top row of Figure 6 shows how the change in policy affects output and inflation, and uncertainty 9 about the two. The most important result is that the alternative policy notably reduces the uncertainty 10 bands for all variables. The bands are between one-third and one-half as wide as under the baseline policy.

It is important to note that the shocks are identical in both scenarios shown here. What differs is only the 12 policy response: the policy response and perceptions thereof matter. 13 The difference in the responses of uncertainty is most easily explained for inflation. Absent the lower 14 bound, the alternative policy would stabilize inflation almost perfectly. This means that any shock would 15 hardly affect uncertainty about inflation. With the lower bound, however, such tracking is not perfect so that 16 some uncertainty remains. The bottom-right panel of Figure 6 illustrates the constraints imposed by the explain why consumers declared that they were much more uncertain about the impact of COVID-19 than 41 professional forecasters. 42 

In this paper, we assess the response of consumer expectations to the pandemic. We do so at two levels. 

How Economic Expectations React 3 to News: Evidence from German Firms

Forecaster (Mis-)Behavior. Review of Economics and Statistics. forth-5 coming

The Demographics of Inflation Opinion Surveys

Communication and the Beliefs of Economic Agents

Economic Policy Symposium

Macroeconomic Expectations of Households and Professional Forecasters

Sticky Expectations and

Inflation with Covid Consumption Baskets

Inflation Expectations, Learning, and Supermarket 17 Prices: Evidence from Survey Experiments

The Covid-19 Crisis and Consumption

The Effect of Macroeco-21

nomic Uncertainty on Household Spending. Working Paper 28625

What Can Survey Forecasts Tell Us About Information Rigidi-23 ties?

Information Rigidity and the Expectations Formation Process: 25 A Simple Framework and New Facts

Average Inflation Targeting and 27 Household Expectations

The Cost of the Covid-19 Crisis: Lockdowns

Working Paper 27141, National Bureau of Eco-30 nomic Research

Monetary Policy Communications and their Effects 32 on Household Inflation Expectations

Gender Roles Produce Divergent Economic Expec-34 tations

Inflation 36 Expectations: Category Beliefs and Spending Plans

Fiscal Volatility 38 Shocks and Economic Activity

Uncertainty Shocks and Business Cycle Research

Coronavirus Perceptions and Economic Anxiety

Covid-19 Coronavirus and Macroeconomic Policy

Inflation Dynamics: A Structural Econometric Analysis

Doves for the Rich, Hawks for the Poor? Distribu-6 tional Consequences of Systematic Monetary Policy. Opportunity and Inclusive Growth Institute Working 7 Papers 50

Macroeconomic Implications of COVID-9 19: Can Negative Supply Shocks Cause Demand Shortages? Working Paper 26918

Exposure to the COVID-19 Stock Market Crash and its 12 Effect on Household Expectations

A Review of Matrix Scaling and Sinkhorn's Normal Form for Matrices and Positive Maps. 14 arXiv: Rings and Algebras

News-driven Inflation Expectations and Informa-16 tion Rigidities

Disagreement about Inflation Expectations

NBER Chapters

The Impact of the COVID-19 Pandemic on Business 676 Expectations

COVID-19-induced Shocks and Uncertainty

Identification in Macroeconomics. The Journal of Economic Per-680 spectives

What's News in Business Cycles