How do Expectations about the Macroeconomy. A ect Personal Expectations and Behavior?

How do Expectations about the Macroeconomy A ect Personal Expectations and Behavior? Christopher Roth Johannes Wohlfart August 9, 2017 Using a representative online panel from the U.S., we examine how people update their economic expectations and alter their economic behavior in response to changes in their macroeconomic expectations. We collect novel data on people s subjective expectations regarding the probability of a recession and create exogenous variation in these expectations by providing respondents with di erent professional forecasters assessments of the likelihood of a fall in real GDP. We find that respondents strongly update their beliefs about the likelihood of a recession and future unemployment in response to the forecasts. People extrapolate from their perceptions of aggregate downside risk in the economy to their personal financial prospects and adjust their financial and consumer behavior as measured in a follow-up survey. The average adjustment in personal economic expectations in response to the forecasts masks substantial heterogeneity along the skill and earnings distribution and heterogeneity by personal unemployment experiences and subjective exposure to aggregate risk. Finally, people s expectations about local unemployment, firm profits and interest rates respond to their macroeconomic outlook, but expectations about the probability of a recession do not causally a ect people s inflation expectations. JEL Classification: D12, D14, D83, D84, E32, G11 Keywords: Expectation Formation, Information, Updating, Aggregate Uncertainty, Macroeconomic Conditions. We would like to thank Roland Bénabou, Enzo Cerletti, Elisabeth Falck, Alexis Grigorie, Michalis Haliassos, Lukas Hensel, Markus Kontny, Peter Maxted, Markus Parlasca, Ricardo Perez-Truglia, Luigi Pistaferri, Carlo Pizzinelli, Simon Quinn, Timo Reinelt, Sonja Settele, and Basit Zafar for helpful comments. We thank Goethe University Frankfurt and Vereinigung von Freunden und Förderern der Goethe Universität for financial support. We received ethics approval from the University of Oxford. The online Appendix is available at https://goo.gl/ MTJ8hG and the experimental instructions are available at https://goo.gl/1c9vlk. Christopher Roth, Department of Economics, University of Oxford and CSAE, e-mail: christopher.roth@economics.ox.ac.uk Johannes Wohlfart, Department of Economics, Goethe University Frankfurt, e-mail: wohlfart@econ.unifrankfurt.de

1 Introduction Macroeconomic conditions have large consequences for the economic outcomes and financial wellbeing of individuals. Recessions, in particular, have high welfare costs and result in job loss and a reduced real income for many households. Exposure to macroeconomic risk should be reflected in people s subjective expectations. In particular, people should adjust their expectations about job security and income growth when they update their expectations about GDP growth. According to standard models, people s expectation about future GDP growth should also a ect their expectations about related macroeconomic variables. For instance, many macroeconomic models entail some form of a Phillips Curve, that is, a negative relationship between inflation and unemployment. If individuals form their expectations in line with these models, we would expect them to adjust their inflation expectations in response to a revised outlook regarding GDP growth and aggregate unemployment. Understanding how households react to changes in their macroeconomic expectations is important for several reasons. First, people s ability to self-insure against macroeconomic risk by accumulating precautionary savings depends on whether they extrapolate from aggregate to personal economic expectations. Second, the macroeconomic outlook is very salient in the news, and consumers should regularly update their expectations about future GDP growth. Changes in macroeconomic expectations could therefore help to explain fluctuations in consumption and financial behavior. Third, documenting how households respond to expectations about macroeconomic conditions is important for assessing the e ectiveness of government and central bank policies (e.g. forward guidance) that aim to influence people s behavior by shifting macroeconomic expectations. Finally, characterizing people s subjective expectations about the co-movement of di erent macroeconomic variables, such as inflation and unemployment, is important to test key assumptions of standard models. In this paper, we conduct an experiment to examine whether and how consumers perceived likelihood of an economic downturn causally a ects their beliefs about related macroeconomic variables as well as their expectations about personal outcomes and their behavior. We recruit 1,124 participants from an online panel that is representative of the U.S. population in full- 1

time employment in terms of age, gender, income and region of residence. We recontact 737 respondents in a follow-up survey two weeks after the main survey. Our experiment proceeds as follows: First, we elicit our respondents prior beliefs about the likelihood of a recession. We define a recession as a fall in real US GDP around three months after the time of the survey. Then, we provide our respondents with one of two truthful predictions about the likelihood of a recession taken from the Survey of Professional Forecasters. Respondents in the high recession treatment receive information from a pessimistic forecaster who predicts a probability of a fall in US real GDP of 35 percent. Respondents in the low recession treatment receive a prediction from a very optimistic forecaster who attaches a probability of 5 percent to a fall in US real GDP. Thereafter, we measure our respondents expectations about the evolution of other macroeconomic variables and their personal economic situation over the 12 months after the survey, as well as their financial and consumption behavior. Specifically, we measure our respondents beliefs about changes in the national unemployment rate, firm profits, inflation, interest rates, own earnings, job security and general financial prospects of their household. We conduct a follow-up survey two weeks after the information provision in order to measure people s consumer and financial behavior and to examine whether people persistently update their expectations after the administration of the treatment. The follow-up survey allows us to provide evidence in which numerical anchoring, priming, and experimenter demand e ects are less of a concern (Cavallo et al., 2017). We start by establishing a series of new stylized facts: First, respondents from our representative sample have much more pessimistic views about the likelihood of a recession than professional forecasters. Second, there is substantial disagreement among consumers about the likelihood of a recession. Third, several observable characteristics, such as political a liation, gender and personal unemployment experiences, are strongly correlated with people s beliefs about the likelihood of a recession. Fourth, expectations about the likelihood of a recession are strongly and significantly correlated with expectations about people s personal economic outlook as well as with most other macroeconomic variables in expected directions. On the other hand, expected inflation and the expected change in interest rates positively and significantly correlate 2

with the perceived likelihood of an economic downturn, contrary to what many models would predict. The correlations between recession expectations and expectations about other macroeconomic variables and personal outcomes cannot be given a causal interpretation. First, unobserved omitted variables, such as general optimism, may lead people to answer to the di erent questions in similar ways. Second, the correlations could be driven by reverse causality, such as extrapolation from the personal situation to macroeconomic expectations (Kuchler and Zafar, 2016). Third, measurement error in expectations about the likelihood of a recession may bias the coe cient estimates towards zero. We circumvent these issues by instrumenting the expectations about the likelihood of a recession using random assignment to the predictions of the di erent professional forecasters. Respondents who receive a more pessimistic forecast about the likelihood of a recession develop more pessimistic views on changes in future unemployment at the national and county level and think that firm profits will be lower. A negative macroeconomic outlook has a negative causal e ect on people s perceived financial prospects and people s perceived job security. We do not find a significant e ect on people s earnings growth conditional on keeping their jobs. Moreover, we find no evidence that changes in beliefs about the likelihood of a recession causally a ect people s inflation expectations. Most of our results on people s updating of expectations decrease in size, but remain economically and statistically significant in the follow-up survey two weeks after the main experiment. This suggests that our findings reflect genuine information updating and that concerns related to numerical anchoring are less important. Next, we consider behavioral outcomes: We find that a more pessimistic macroeconomic outlook leads people to report a significantly lower planned consumption growth in the main survey. We also find suggestive evidence of changes in actual spending after the treatment using data from our follow-up survey. Furthermore, we find strong and sizable evidence that respondents in the low recession treatment actively increase their self-reported stockholdings compared to individuals in the high recession treatment during the time between the main intervention and the follow-up survey. Subsequently, we characterize whether and how di erent demographic groups di er in the way 3

they update their economic expectations. We find that expectations about an economic downturn have di erent e ects at the top and the bottom of the earnings and skill distributions. People with higher earnings significantly reduce their earnings expectations conditional on keeping their jobs. Perceived aggregate risk has a significantly negative e ect on perceived job security among individuals with lower earnings or low education. Moreover, we find that older individuals report significantly lower earnings expectations upon receiving a more pessimistic forecast. We also demonstrate that the negative e ect of recession expectations on perceived job security is driven by people who have been unemployed in the past and people from counties with higher unemployment rates. Moreover, we collect novel data on people s subjective beliefs on how firms adjust their hiring, firing and profits in response to aggregate conditions and show that people think that firms insure workers wage risk substantially, in line with recent empirical evidence (Fagereng et al., 2017b). Moreover, respondents who think that their employer s firing is more dependent on the macroeconomy or who think that their employer does not provide them with insurance report a higher job loss probability as a result of increased pessimism about the macroeconomic environment. Our findings contribute to a growing literature that uses survey data to study the formation of people s macroeconomic expectations and the e ect of these expectations on individual behavior (Manski, 2004; Coibion and Gorodnichenko, 2012, 2015a; Kumar et al., 2015; Roth and Wohlfart, 2017; Manski, 2017; Coibion et al., 2017; Beshears et al., 2013; Bordalo et al., 2017; Kuhnen and Miu, 2017). Some of these papers conduct information experiments to study the role of perceived past realization such as past inflation (Armantier et al., 2016; Cavallo et al., 2017; Coibion et al., 2015) or past house price growth (Armona et al., 2016). 1 Moreover, the literature has established that inflation and house price expectations are an important driver of individual behavior (Armantier et al., 2015; D Acunto et al., 2016; Bailey et al., 2017a). Our paper is the first to exogenously shift people s expectations about future GDP growth and to document whether people extrapolate from expectations about aggregate macroeconomic conditions to their personal economic outlook and whether these expectations a ect consumer 1 Other studies use observational data to examine how macroeconomic expectations are influenced by social interactions (Bailey et al., 2017b; Arrondel et al., 2017), macroeconomic and local experiences (Malmendier and Nagel, 2011, 2015; Malmendier and Shen, 2016; Kuchler and Zafar, 2016) or trust in the central bank(christelis et al., 2016). 4

and financial behavior. 2 We thereby relate to recent work by Mian et al. (2017) whodocument that Democrats and Republicans update their economic expectations in opposite directions following presidential elections. However, they find no evidence that shifts in economic expectations due to partisan bias a ect consumer spending. Moreover, our paper is the first to examine whether people s inflation and interest rate expectations causally respond to changes in expectations about future unemployment. We thereby contribute to a small literature that tries to measure people s subjective models of the economy and that studies whether subjective models are in line with economic theory and historical facts. Specifically, there was substantial disinflation during most recessions in the past (Coibion and Gorodnichenko, 2015b). Similarly, a standard Taylor rule would predict that interest rates decrease in response to higher unemployment. We find no evidence that our estimates of a subjective Phillips Curve or a subjective Taylor rule are in line with the theoretical predictions. 3 We thereby contribute to existing work by Carvalho and Nechio (2014) and Dräger et al. (2016) who use observational data to examine how beliefs about unemployment correlate with beliefs about interest rates and inflation. As opposed to these previous papers, we deal with omitted variable bias and reverse causality by exogenously shifting expectations about future unemployment. Our paper also contributes to a small literature examining the role of professional forecasters for household s expectation formation about the economy (Carroll, 2003; Armantier et al., 2016). Carroll (2003) outlines a model in which households expectations derive from news reports based on the views of professional forecasters. Armantier et al. (2016) provide experimental evidence that people s inflation expectations respond to professional forecasts of future inflation. We contribute to this literature by showing that consumers significantly update their expectations upon receiving professional forecasts about the likelihood of a recession. This updating occurs even though people receive a forecast from only one forecasting institution, which is a relatively noisy signal about the likely future path of the economy. 2 We thereby also contribute to a literature in labor economics on the determinants of subjective job security (Campbell et al., 2007; Dickerson and Green, 2012; Geishecker et al., 2012). This literature finds that individual job loss expectations strongly predict actual transitions into unemployment. 3 If anything, we find e ects of an expected economic downturn going into the opposite direction. 5

We also contribute to a recent literature on workers exposure to aggregate and firm-level risk. Guvenen et al. (2017) use administrative tax data to show that there is large variation across industries and demographic groups in the dependence of individual earnings growth on GDP growth. Fagereng et al. (2017b) use administrative data from Norway to show that firms provide their workers with substantial insurance against firm-level shocks. We contribute to this literature by showing that these empirical facts are reflected in workers subjective beliefs. We also show that respondents think that their employer provides them with substantial insurance, and that they expect part of an economic downturn to be absorbed by firm profits. Our paper is structured as follows: in section 2 we describe the design of the main experiment and the follow-up survey. In section 3, we provide details on the data collection and characterize our sample. In section 4, we present evidence on prior beliefs about the likelihood of a recession and on belief updating in response to the professional forecasts. In section 5, we show results on the causal e ect of expectations about a recession on other macroeconomic variables, personal expectations and behavior. In Section 6, we explore further heterogeneity by unemployment experiences and exposure to aggregate risk. Section 7 concludes. 2 Experimental design 2.1 Baseline experiment First, we ask subjects to complete a questionnaire on demographics, which includes questions on gender, age, income, education, and region of residence. Subsequently, we give our respondents a brief introduction on how to probabilistically express expectations about future outcomes, and also explain to them several relevant economic concepts, such as recession and GDP. Then, we ask our respondents to estimate the likelihood that there will be a fall in US real GDP in the fourth quarter of 2017 compared to the third quarter of 2017. 4 The survey was conducted in the summer of 2017, so this corresponds to a fall in real GDP three to six months after the survey. Thereafter, we ask our respondents how confident they are in their estimate. The Federal Reserve Bank of Philadelphia regularly collects and publishes predictions by 4 We refer to these beliefs as recession expectations throughout the paper. 6

professional forecasters about about a range of macroeconomic variables in their Survey of Professional Forecasters (SPF) (Croushore, 1993). The SPF is conducted in the middle of each calendar quarter, and forecasters have to estimate the likelihood of a decline in real GDP in the quarter of the survey and for each of the four following quarters. The average probability assigned to a drop in GDP in the quarter after the survey has become known as the anxious index and has had high predictive power for actual recessions in the past. In our survey we randomly assign our respondents to receive one of two forecasts taken from the micro data of the wave of the SPF conducted in the second quarter of 2017, the most recent wave of the SPF available at the time of our survey. To make the forecast more meaningful to respondents, we tell them that it is from a financial services provider that regularly participates in a survey of professional forecasters conducted by the Federal Reserve Bank of Philadelphia. In the high recession treatment, respondents receive a forecast from the most pessimistic panelist in the SPF, who assigns a probability of 35 percent to a fall in US real GDP in the fourth quarter compared to the third quarter of 2017. In the low recession treatment, respondents receive information from one of the most optimistic forecasters, who expects a fall in US real GDP with a probability of 5 percent. In order to make the treatment more meaningful to our respondents, we provide them with a figure that contrasts their prior belief with the prediction from the professional forecaster. 5 Thereafter, all respondents are asked to estimate the likelihood that the unemployment rate in the US will increase over the 12 months after the survey, as well as blocks of questions on personal economic expectations and on a range of other macroeconomic variables. We randomize the order in which the two blocks of questions are shown to our respondents. While we elicit most expectations probabilistically, we also include some qualitative questions with categorical answer options. 6 In the question block on personal economic prospects, we first ask our respondents whether they think that their family will be better or worse o 12 months after the survey. Then, we elicit 5 The professional forecasts correspond to SPF panelists beliefs about a drop in real GDP two quarters after this wave of the SPF was conducted. 6 The question framing we use to elicit people s expectations closely follows the New York Fed s Survey of Consumer Expectations (SCE). Fed economists have experimented with di erent ways to elicit people s expectations and have settled on this question framing after extensive testing (Armantier et al., 2017). More generally, we mostly follow the guidelines on the measurement of expectations as laid out by Manski (2017). 7

people s density forecast about their earnings growth conditional on working at the same place they currently work. 7 We ask our respondents to assign probabilities to ten brackets of earnings growth over the next 12 months, which are mutually exclusive and collectively exhaustive. Respondents could not continue to the next screen if the entered probabilities did not sum up to 100 percent. The elicitation of a subjective probability distribution allows us to measure both mean expected earnings growth and uncertainty about earnings growth. 8 Thereafter, respondents estimate their subjective probability of job loss and their subjective probability of finding a new job within three months in case they lose their job over the next 12 months. In the next block of questions, we elicit density forecasts of inflation over the next 12 months using the same methodology as for earnings expectations. In addition, we ask respondents to estimate the likelihood that interest rates on savings accounts will be higher in 12 months than they are at the time of the survey. Thereafter, we ask our respondents some questions related to their consumption behavior. First, we ask them whether they think it is a good time to buy major durable goods. Second, our respondents are asked how they plan to adjust their consumption expenditures on food at home, food away from home and leisure activities during the four weeks after the survey compared to the four weeks prior to the survey. Then, our respondents answer a question on how firm profits will change over the next 12 months. Subsequently, they estimate the percent chance that unemployment in their county of residence will increase over the next 12 months. After that we re-elicit beliefs about the likelihood of a fall in real US GDP in the fourth quarter of 2017 compared to the third quarter of 2017. Finally, our respondents complete a series of additional questions on the combined dollar value of their spending on food at home, food away from home, clothing and leisure activities over the seven days before the survey, the industry they work in, and their tenure at their employer, as well as a set of questions measuring their financial literacy. 9 Moreover, we ask them a series of 7 In contrast to the question in the SCE, we do not elicit earnings expectations conditional on working in the same job and for the same number of hours. Instead, we allow for changes in hours worked as well as for job promotions or demotions at their workplace as this provides us with additional variation. 8 In addition, means of density forecasts are straightforward to interpret, while point forecasts could capture mean, median, mode or some other moment of our respondents subjective probability distributions (Manski, 2017). 9 We use the three questions on interest compounding, inflation and risk diversification that have now become standard to measure financial literacy (Lusardi and Mitchell, 2014). 8

questions on their assets, their political a liation as well as their zipcode of residence. 2.2 Follow-up survey Two weeks after the main study our survey provider re-invited our respondents to a five-minute follow-up survey. In the follow-up survey, we re-elicited some of the key outcome questions from the main survey, such as the likelihood of an increase in national- and county-level unemployment, expectations about firm profits, as well as personal economic expectations such as subjective job security and earnings expectations. We re-elicit our respondents estimated likelihood of a fall in real GDP in the fourth quarter of 2017 compared to the third quarter of 2017. Moreover, we collect data on our respondents consumer and financial behavior in the time between the main intervention and the follow-up survey. First, we ask our respondents about their spending on food at home, food away from home, clothing and leisure activities over the seven days before the follow-up survey. Second, we ask them whether they bought any major durable goods and whether they actively increased or decreased their stockholdings during the 14 days prior to the follow-up. Finally, we elicit our respondents beliefs about their employers exposure to aggregate risk, their personal unemployment history, as well as their beliefs about the most likely causes of a potential recession. Figures 1 and 2 show detailed timelines of the experiment and the relevant reference periods for behavioral outcomes and expectations. 3 Data 3.1 Survey administration We collect a sample of 1,124 respondents that is representative of the US population in full-time employment in terms of gender, age, region and total household income through the market research company Research Now. We only invite people in full-time employment. The data for the main survey were collected in summer 2017. We conducted the follow-up survey approximately two weeks after the main survey was administered. We managed to recontact 737 respondents, which corresponds to a recontact rate of 65 percent. 9

3.2 Representativeness In Table 1 we provide summary statistics for our sample for a large set of variables. Moreover, Table A1 in the online Appendix 10 displays the distribution of a range of individual characteristics among respondents in full-time employment in the 2015 American Community Survey (ACS) and in our data. 11 As can be seen, we match the distributions of gender, age, region and total household income very precisely. In addition, the composition of our sample is quite close to the composition of the population in full-time employment along non-targeted dimensions, such as industry and hours worked. The main di erence is that our sample is more highly educated and has higher labor earnings on average than the US population in full-time employment. Around 80 percent of the respondents indicate that they are the main earner in their household. In Table A2 we provide evidence on the integrity of the randomization. As can be seen, our sample is well-balanced for a set of key demographics and pre-treatment beliefs about the likelihood of a recession. Moreover, we observe no di erential attrition in our main survey across treatment arms, and response to the follow-up survey is not related to treatment status in the main experiment. Table A3 shows that the sample of individuals that responded to the follow-up is balanced across the two treatment arms in terms of key covariates. 3.3 Data quality We provide evidence that our expectation data on earnings and inflation is of high quality by comparing it to a panel survey by the New York Fed launched as a predecessor of the Survey of Consumer Expectations (SCE) (Armantier et al., 2013). For example, for inflation expectations we find that in our sample 80 percent of respondents assign positive probability to more than one bin (89.4 percent in the Fed survey) and the average number of bins with positive probability is 4.24 (3.83). While a larger fraction of our respondents assigns positive probability to noncontiguous bins (6.9 percent vs 0.9 percent), this still accounts for a very small share of our respondents. In line with the findings in Armantier et al. (2016), we find that females and respondents with low education expect significantly higher inflation. 10 The online Appendix is available at https://goo.gl/mtj8hg. 11 In the ACS we classify as full-time employed individuals who report working at least 30 hours per week. 10

In addition, survey respondents sometimes express mental overload when answering probabilistic questions by simply entering 0 percent, 50 percent or 100 percent (Manski, 2017). Only 0.4 percent, 6.5 percent and 0.3 percent of our respondents enter a prior probability of a fall in real GDP of 0 percent, 50 percent and 100 percent, respectively. 12 We also provide evidence for the quality of our data on non-durable spending in the week prior to the main survey. In Figure A.7 we show that log spending is correlated with age and log income in expected directions. As a further check of data quality we examine whether people s self-reported zip code in the follow-up matches the zip-code they report in the main study. More than 93 percent of respondents enter matching zipcodes which reassures us of the high data quality in our sample. 3.4 Definition of variables In what follows we define some of the variables we use in our analysis. We calculate a measure of composite job insecurity as the perceived probability of losing the main job within the next 12 months and not finding a new job within the following three months: Pr(unemployment) = Pr(job loss) (1 Pr(job finding job loss)) For each respondent we calculate the mean and standard deviation of expected inflation and expected earnings growth using the mid-points of the bins to which the respondent assigned probabilities. 13 Moreover, we create an unweighted index of people s planned change in nondurable consumption from the four weeks prior to the main survey to the four weeks after the survey, using their qualitative spending plans for food at home, food away from home and leisure activities. Finally, we create a measure of people s actual changes in spending on food at home, food away from home, clothing and leisure during the seven days before the survey based on their responses in the main survey and in the follow-up survey. 14 The questions on expected 12 In Figures A.8 to A.13 in the online Appendix, we display the distributions of expectations for future unemployment and inflation, inflation uncertainty, mean expected earnings, earnings uncertainty as well as subjective job finding and job loss probabilities. 13 Following the question framing of the Survey of Consumer Expectations (Armantier et al., 2017), we elicit probabilities over eight closed bins between -12 percent and 12 percent and two open bins for outcomes outside this range. For the open bins we assign -14 percent and 14 percent, respectively. 14 We take the di erence in log spending from the follow-up and the baseline survey, so this variable measures the percent change in spending. We deal with outliers by setting spending growth to missing for respondents in the top and bottom two percent of observed spending growth. We obtain qualitatively similar results if we instead 11

firm profits, the financial situation of the household or the change in stockholdings between main survey and follow-up were elicited on 5- and 7-point scales. We code these variables such that higher values refer to increase or improve and lower values refer to decrease or worsen. All of these qualitative outcome variables are normalized using the mean and standard deviation separately for the main survey and the follow-up survey. 15 4 Results: Prior beliefs and updating 4.1 Prior beliefs: Stylized facts Our paper provides novel data on people s expectations about the likelihood of a recession for a population of full-time employed Americans. First, we characterize our respondents prior expectations and compare them to the expectations of professional forecasters in the SPF. Figure 3 highlights that there is a large dispersion in beliefs about the likelihood of a recession among consumers. We find a substantial proportion of respondents who think that the probability of a recession is below 20 percent, but also a large fraction of respondents who think that the probability of a recession is higher than 50 percent. Figure A.1 shows the distribution of predictions from the SPF, which range from four professional forecasters giving an estimate of a 5 percent chance of a recession to one forecaster assigning a 35 percent chance. Respondents in our sample have a much more pessimistic macroeconomic outlook than professional forecasters. The median respondent in our sample thinks that the likelihood of a recession is 40 percent, 16 while the median professional forecaster assigns a probability of just 15 percent. Indeed, the most pessimistic professional forecast of 35 percent is below the median forecast in the online panel. 4.2 Correlates of recession expectations Next, we examine how recession expectations correlate with a range of individual characteristics. Table A4 shows that neither education nor age are related to people s recession expectations, but use one or five percent as cuto, or if we winsorize the variable instead. 15 For all remaining quantative measures we do not normalize outcome variables. 16 The modal estimate is 50 percent, selected by 6.5 percent of respondents. 12

that females are significantly more pessimistic about the macroeconomic outlook than men. We find that more financially literate respondents report lower recession expectations, even though this correlation is noisily measured. Interestingly, we find that there are large di erences according to people s political a liation. Democrats are much more pessimistic compared to Independents, while Republicans are much more optimistic. This is in line with recent evidence on partisan bias in economic expectations (Mian et al., 2017). In addition, we examine whether people who personally have been unemployed in the past and people with higher experienced aggregate unemployment rates (Malmendier and Shen, 2016) display di erent beliefs. Interestingly, we find that people who have been personally unemployed in the past are significantly more pessimistic about aggregate economic conditions. This is in line with Kuchler and Zafar (2016) who find that individuals who lose their jobs become significantly less optimistic about the aggregate economy. We find no e ect of experienced aggregate unemployment rates. We also analyze whether di erences in objective and subjective exposure to macroeconomic risk a ect beliefs about the likelihood of a recession. Guvenen et al. (2017) estimate GDP betas capturing the correlation of individual earnings growth with GDP growth for di erent groups defined by gender, age, earnings and industry. We assign to each of our respondents the respective GDP beta and use it as an objective measure of exposure to macroeconomic fluctuations. We also use our respondents assessment of whether their employer s firing decisions depend on the aggregate economy as a subjective measure of exposure to aggregate risk. We find no evidence that objective exposure to GDP fluctuations a ects people s beliefs about the likelihood of a recession. However, our measure of subjective exposure to aggregate risk positively correlates with expectations about the likelihood of a recession. Finally, we find no significant relationship between the level of our respondents priors and their confidence in the priors. 4.3 Updating of recession expectations: Non-parametric analysis We examine non-parametrically how our respondents update their recession expectations upon receiving di erent professional forecasts. Figure 3 shows that the information provision strongly shifts expectations towards the professional forecast in both treatment arms. In Figure 4 we 13

plot updating of beliefs, defined as the di erence between posterior and prior, against people s perception gap, which we define as the di erence between the professional forecast and people s prior belief: 8 >< 35 prior i if highrecession i =1 perceptiongap i = >: 5 prior i if highrecession i =0 where highrecession i takes value one for respondents who were given the prediction that the likelihood of a recession is 35 percent, while it takes value zero for respondents who received the prediction that the likelihood of a recession is 5 percent. The figure shows that people strongly update their prior beliefs in the direction of the professional forecasts. Figure 5 displays scatter plots of prior and posterior beliefs. Observations along the red horizontal lines indicate full updating of beliefs towards the professional forecast, while respondents along the 45 degree line do not update at all. Bayesian updating implies that posterior beliefs should lie somewhere between the two lines. Table A5 displays the number of respondents behaving in a Bayesian and in a non-bayesian manner in the two treatment arms. We find that 80 percent of respondents in the low recession treatment and 67.3 percent of respondents in the high recession treatment behave in a Bayesian manner. 17 11.5 percent of respondents in the low recession treatment and 19.5 percent of respondents in the high recession treatment do not update their beliefs at all, while 68.6 percent (47.8 percent) of respondents either fully or partially update their beliefs towards the signal. Our evidence is in line with previous results on the non-belief in the law of large numbers as we show that people respond strongly to just one professional forecast and strongly and persistently update their beliefs. 18 4.4 Updating of recession expectations: Estimating learning rates We adopt the approach in Armantier et al. (2016) and Bottan and Perez-Truglia (2017) to quantify the degree of learning from information by estimating the slope between updating and 17 Some of the apparent non-bayesian updating may be due to rounding errors, due to respondents accidently entering numbers that slightly deviate from their true beliefs, or due to ambiguity. 18 For example, previous evidence shows that people consider averages estimated with samples sizes of 10 and 1,000 as equally precisely estimated (Benjamin et al., 2016). 14

the perception gap. Specifically, we estimate the following equation: updating i = 1 perceptiongap i + " i Table A6 shows our estimates of the Bayesian learning rates. We find learning rates of about 50 percent using the full sample. We find larger learning rates for respondents with weaker priors in line with a simple Bayesian learning model. This provides strong evidence that our respondents expectations about the likelihood of a recession are a ected by the professional forecasts. 19 4.5 Persistence of updating Following Cavallo et al. (2017) we employ a follow-up survey in order to disentangle genuine from spurious learning. Specifically, the follow-up survey alleviates two concerns: first, that our results on updating of recession expectations are driven by numerical anchoring which is a very short-lived phenomenon by definition. Second, it it less likely that respondents distort their stated expectations because of experimenter demand e ects (de Quidt et al., 2017; Zizzo, 2010; Cavallo et al., 2017). 20 In Table 8 and Figure A.2, we show that people s recession expectations still di er across the treatment groups in a two-week follow-up survey. We find that di erences in recession expectations across treatment arms in the follow-up survey make up about 40 percent of di erences in expectations across treatment arms in the main survey. The medium-term learning rate estimated using the follow-up data is slightly smaller than the short-term learning rate, but still close to 50 percent, as can be seen in Table A6 and Figure 4. 19 We also examine which covariates are correlated with people s updating of recession expectations. Table A7 in the online Appendix highlights that females and Democrats update less strongly. 20 de Quidt et al. (2017) provide evidence that people s response to explicit signals of the experimenter s expectations only very moderately change behavior in online experiments. 15

5 Results: The causal e ect of recession expectations 5.1 E ects on macroeconomic expectations In this section, we examine how people s beliefs about the likelihood of a recession are related to expectations about several other macroeconomic expectations. As a first step, we examine how macroeconomic expectations, macroexp i, are correlated with our respondents posterior beliefs about the likelihood of a recession, posterior i : macroexp i = 0 + 1 posterior i + T X i + " i where X i is a vector of control variables. 21 " i is an idiosyncratic error term. 22 The OLS estimate of 1 cannot be given a causal interpretation. For example, it is possible that people who are generally more optimistic or more pessimistic respond di erently to both the question on the posterior as well as the questions related to the evolution of other macroeconomic variables. It is also conceivable that the direction of causality is running from other macroeconomic expectations to expectations about a recession. Finally, the estimate of 1 could be biased towards zero because of measurement error in the posterior belief. To deal with omitted variable bias, reverse causality and measurement error, we instrument our respondents posterior beliefs with the random assignment to the di erent professional forecasts. Specifically, we use two-stage-least squares and estimate the following equations: macroexp i = 0 + 1 \ posterior i + T X i + " i where \ posterior i = 0 + 1 highrecession i + T X i + u i Table 2 displays the results on how people s posterior beliefs about a recession are related to 21 The controls are as follows: age, age squared, a dummy for females, log income, a dummy for respondents with at least a bachelor degree, dummies for the respondent s Census region of residence, a measure of the respondent s financial literacy as well as a dummy for Republicans and a dummy for Democrats. 22 We report robust standard errors throughout the paper. 16

expectations about other macroeconomic variables. Panel A shows the results from the OLS regressions, while panel B displays the estimates using the random treatment assignment as an instrument for posterior beliefs. We find strong evidence that posterior beliefs about a recession significantly a ect people s beliefs about the probability that national unemployment will increase. We find that a one percentage point higher likelihood of a recession is associated with a 0.895 percentage point increase in the chance that national unemployment will increase. The e ect size is only 0.536 for unemployment in the respondent s county of residence, consistent with di erences across counties in the exposure to aggregate risk. Moreover, respondents with higher recession expectations are more likely to expect lower firm profits. These results are significant for both the OLS and IV estimates. Moreover, OLS and IV estimates are very similar in magnitude. During most recessions in the past there was a substantial decline in inflation (Coibion and Gorodnichenko, 2015b). Moreover, many macroeconomic models incorporate some form of a Phillips Curve, that is, a negative relationship between unemployment and inflation. However, we find no evidence that this is reflected in people s subjective economic model. Indeed, mean expected inflation (as well as inflation uncertainty) is positively correlated with the perceived downward risk in the economy. However, this relationship is not statistically significant in the IV specification. Moreover, we find a significant positive e ect of recession expectations on the probability that interest rates will increase in both the OLS and IV specifications, contrary to what a simple Taylor rule would predict. Next, we consider heterogeneous e ects of beliefs about an economic downturn on other macroeconomic expectations. For heterogeneous e ects, we focus on the IV specification and estimate the following system of equations: macroexp i = 0 + 1 \ posterior i + 2 interact i + 3 \ posterior i interact i + T X i + " i 17

where \ posterior i = 0 + 1 highrecession i + 2 highrecession i interact i + 3 interact i + T X i + v i \ posterior i interact i = 0 + 1 highrecession i + 2 highrecession i interact i + 3 interact i + T X i + w i where interact i is the interaction term of interest. The results are shown in Table 3. Most interestingly, we find that the positive e ect of recession expectations on interest rate expectations is driven by individuals who have lower education, are younger or have lower earnings. Similarly, we find a positive e ect of recession expectations on expected inflation among individuals with low education. One interpretation of this finding is that less sophisticated individuals associate an increased risk of an economic downturn with general economic pessimism, which is also reflected in a higher rate of inflation. 23 In addition, our IV specifications capture the local average treatment e ects on the population of compliers, i.e. on those individuals who change their recession expectations upon receiving the professional forecasts (Angrist and Pischke, 2008). We examine whether recession expectations exert a di erent e ect on those with a positive and those with a negative perception gap. We find that individuals who received a forecast that was more pessimistic than their prior belief report a significantly higher inflation uncertainty, while there is no such e ect for those who received a forecast that was more optimistic than their prior. In Table 8 we provide evidence that these e ects persist in the two-week follow-up. People who receive more pessimistic forecasts about the likelihood of a recession still report a significantly higher probability of an increase in national unemployment. We find that for expectations 23 Am alternative explanation is that some respondents think that the recession would be caused by a supplyside shock. Recessions caused by supply-side shocks could entail di erent co-movement of inflation and GDP growth than recessions caused by demand-side shocks. To shed light on this issue, we collect novel data on beliefs about recession causes in the follow-up survey (see Figures A.4, A.5 and A.6). The most frequently mentioned recession causes in the follow-up survey are a drop in consumer confidence and political turmoil, while supplyside factors such as an oil price increase are not mentioned as frequently. In addition, we find no heterogeneous responses in inflation expectations depending on whether respondents think that the recession will be caused by supply-side or demand-side factors. We find suggestive evidence that our results on interest increases are driven by a subsample of respondents who think that the recession will likely be caused by an interest rate hike. Results are available upon request. 18

about national and county-level unemployment the e ect sizes in the follow-up are about 50 percent and about 42 percent of the original e ect size in the main study. We believe that this reflects a substantial degree of persistence, given that people likely received relevant signals about macroeconomic conditions between the two surveys. We also find that the e ects on expected firm profits remain similar in size and significance in the follow-up. 5.2 E ects on personal economic expectations Next, we consider whether people s expectations about macroeconomic conditions a ect their expectations about their own financial and economic outcomes. In Panel A of Table 4 we show results on personal expectations using OLS. We find that people who deem it more likely that a recession occurs are also more likely to hold pessimistic beliefs about their household s financial prospects and report lower levels of subjective job security. They also expect lower earnings growth and higher earnings uncertainty conditional on keeping their jobs. As in the previous section, we deal with the issues of omitted variable bias, reverse causality and measurement error by estimating an IV specification which relies on the random assignment to the di erent professional forecasts. 24 In the IV specifications, we also find that respondents with more pessimistic views about aggregate macroeconomic conditions expect significantly worse financial outcomes for their own household. We find a similar e ect size for expected earnings growth as in the OLS specification, but the coe cient is very noisily measured. However, we find no evidence that, on average, recession beliefs are causally related to earnings uncertainty. Our IV point estimate on perceived job security is almost identical in magnitude to the OLS estimate, and is still marginally statistically significant. The e ect size is substantial: We find that a one percentage point increase in the likelihood of a recession leads to an increase in subjective unemployment risk of 0.113 percent. 25 There is substantial heterogeneity in people s actual exposure to aggregate risk (Guvenen 24 Kuchler and Zafar (2016)provideevidencethatindividualsextrapolatefromtheirpersonaleconomicsituation to expectations about the aggregate economy. Such extrapolation would be a source of reverse causality in the OLS specifications. 25 To illustrate the magnitude of this e ect, consider moving from a situation with zero risk of a recession to asituationwherearecessionwillhappenwithcertainty. 11.3percentofourrespondentswouldneedtobecome unemployed for their expectations to be accurate on average. For comparison, the unemployment rate increased from five to around ten percent during the Great Recession 2007-9, so around 5.3 percent of the employed population lost their jobs. 19

et al., 2017), suggesting that one should expect strong heterogeneity in the e ect of aggregate expectations on people s personal economic outlook. In Table 5, we examine heterogeneous e ects of beliefs about the likelihood of a recession on personal financial expectations using the random treatment assignment. We find a stronger response in terms of overall household financial prospects for people with low education. This is reflected in a large and statistically significant decrease in perceived job security for people with lower education, while there is no such e ect for people with high education. Moreover, we find that recession expectations have di erential e ects across the earnings distribution. Individuals with high earnings expect to be a ected through lower earnings growth at their workplace, while individuals with low earnings report a significantly higher perceived risk of becoming unemployed. We also find that higher recession expectations lead to significantly lower financial prospects and expected earnings growth among older individuals, while the e ects are insignificant for younger individuals. Finally, we find that the e ects on personal expectations are mostly driven by those who underestimated the risk of a recession relative to the professional forecast they received. In Table 8 we examine whether these e ects persist in a 2-week follow-up. We find that the treatment e ects for all outcomes are not statistically distinguishable from the treatment e ects in the main experiment. However, our estimates from the follow-up are less precisely estimated and slightly smaller than those in the main study. The e ects sizes for financial prospects and perceived job insecurity are about 50 percent and about 75 percent of those in the main study. 5.3 E ects on consumption and stockholdings Above, we have shown that individuals extrapolate from their macroeconomic to their personal outlook. In this section, we examine whether updating of recession expectations leads people to adjust their consumption plans, their actual consumption growth and their stockholdings. First, we examine whether updating of recession expectations a ects our measures of planned and actual changes in non-durable spending around the main intervention. According to a standard Euler equation, an innovation to expected future economic resources should induce households to immediately adjust their consumption. We focus on non-durables as for this category consumption plausibly equals expenditure. Second, we examine whether updating of 20