News-Driven Uncertainty Fluctuations

Abstract

We investigate the channels through which news influences the subjective beliefs of economic agents, with a particular focus on their subjective uncertainty. The main insight of the article is that news that is more at odds with agents’ prior beliefs generates an increase in uncertainty; news that is more consistent with their prior beliefs generates a decrease in uncertainty. We illustrate this insight theoretically and then estimate the model empirically using data on U.S. output and professional forecasts to provide novel measures of news shocks and uncertainty. We then estimate impulse responses from the identified shocks to show that news shocks can affect macroeconomic variables in ways that resemble the effects of uncertainty shocks. Our results suggest that controlling for news can potentially diminish the estimated effects of uncertainty shocks on real variables, particularly at longer horizons.

Keyword:

• Bayesian learning
• Discrete environment
• News shocks
• Survey forecasts
• Uncertainty

Supplementary Materials

The supplementary materials comprise the Online Appendix, code and data files.

Acknowledgments

We thank Jaroslav Borovicka, Christophe Chamley, Nicholas Kozeniauskas, Andre Kurmann, Minchul Shin, Jonathan Wright, and the seminar and conference participants at Boston College, Johns Hopkins University, University of Washington, University of Wisconsin-Madison, the 2017 Society for Economic Measurement Conference, the 2017 NBER Time-Series Conference, the 2017 NBER DSGE Meeting, the 2018 Zurich Asset Pricing Conference, the 2018 Society for Nonlinear Dynamics and Econometrics Conference, the 2018 BC-BU Green Line Macro Meeting, the 2018 SFS Cavalcade Meeting, the 2018 Barcelona GSE Summer Forum for helpful comments. The views expressed in this article are those of the authors and do not necessarily represent the views of the Federal Reserve Bank of Boston or the Federal Reserve System.

Disclosure Statement

The authors report there are no competing interests to declare.

Notes

1 This index is based on responses to the following question in the University of Michigan’s Survey of Consumers: “During the last few months, have you heard of any favorable or unfavorable changes in business conditions?” The index is constructed as the percent replying favorable minus the percent replying unfavorable plus 100. We interpret the negative of this index as a measure of bad news.

2 A reduced-form interpretation is applicable to our news shock because it either indicates enhancement or deterioration of the future economy, potentially proxying various forms of news.

3 Veronesi (1999) shows a similar relationship between dividend growth and uncertainty in a model where agents learn about a hidden dividend growth state by observing realized dividend growth. Here, we consider a shock that explicitly gives information about future states and that can generate uncertainty fluctuations that are independent of the current state. See the literature review below for a more extensive discussion of the differences between this article and Veronesi (1999).

4 Note that the label switching problem arises for $\chi <\frac{1}{K}$. Therefore, we restrict to $\frac{1}{K}\le \chi \le 1$.

5 The assumption that a wrong signal is equally likely for all states is for simplicity. In reality, it is plausible to think that a signal increases the likelihood of “close” states. Another form of extension where multiple signals with different informativeness can be considered.

6 The expressions for the numerator of (3) are provided in (1) and (2) and the calculation of the denominator is provided in the Appendix, supplementary materials.

7 If the support of possible ${\mu }_{S_t}$ changes as the number of states increases, the role of news on uncertainty becomes more important. Instead, we take a more conservative assumption.

8 Specifically, when $1/K<\chi <1$, the distribution is not only shifted from $S_{t+1}=1$ to $S_{t+1}=K$, but some of it is also shifted to $S_{t+1}=2,\dots ,K-1$. However, the differences this creates in forecast uncertainty as K increases seem to be second-order.

9 We acknowledge that fixing the remain probability q while changing the number of states K impacts the distance between states as well as the switching probabilities to other states. However, our main intuition carries over when we relax this assumption.

10 Note that we could also use mean forecast, $E(y_{t+1}|I_t)={\sum }_{i=1}^2{\mu }_{i}p(S_{t+1}=i|I_t)$, which is a function of $p(S_{t+1}=i|I_t)$. Instead, we rely on a direct measure of $p(S_{t+1}=i|I_t)$.

11 The choice of the probit linking function is just for convenience.

12 The evolution of parameter estimates is provided in Figure A-4, supplementary materials. The credible intervals at time 0 correspond to the 90% prior intervals. As more information from observed data is incorporated into the posterior distributions over time, the 90% credible intervals shrink. Those at time T are posterior credible intervals one would obtain from the entire time series of data.

13 Figure A-5, supplementary materials provides the probability of each regime in the benchmark estimation. Note that the economy spends the most time in the “good state, good news” regime. However, the probability of “good state, bad news” is nonnegligible. This is the regime in which the economy is currently in the good state, but news suggests that the economy will enter the bad state in the following period. On the other hand, the economy assigns very small probabilities to the “bad state, good news” and “bad state, bad news” regimes. Nevertheless, in periods when the probability of being in a bad state is high, the conditional probability of receiving good news can be substantial.

14 The correlation between the posterior median forecast uncertainty and the macroeconomic uncertainty in Jurado, Ludvigson, and Ng (2015) is 0.60. We find a similar correlation between the posterior median forecast uncertainty and the estimated GARCH(1,1) process of 0.57.

15 Unlike in Table 1, we do not find that current GDP growth has predictive power, but this is not surprising as our model assumed, for simplicity, that growth is iid conditional on states.

16 Our results are robust to a frequentist OLS estimation of the VAR.

17 With the addition of the S&P 500 index, Y₁ is the set of variables used in the 8-variable VARs in Bloom (2009) and Jurado, Ludvigson, and Ng (2015). However, we exclude the S&P 500 from the benchmark “no news” specification as it’s a very forward-looking variable that has been often used in VARs to identify news shocks (see Beaudry and Portier 2006; Barsky and Sims 2011; Barsky, Basu, and Lee 2014).

18 Note that this GARCH estimate differs from the stochastic volatility models estimated in Jurado, Ludvigson, and Ng (2015) and Ludvigson, Ma, and Ng (2021) by not allowing for an independent driver of uncertainty. However, recall that our estimates of ex-ante uncertainty are from a structural estimation of a model where uncertainty moves only endogenously with news and growth state shocks. Therefore, the GARCH estimate is more consistent with our structural estimates of ex-ante uncertainty.