Measuring Global and Country-Specific Uncertainty
  • 1 0000000404811396 Monetary Fund
  • | 2 0000000404811396 Monetary Fund

Contributor Notes

Authors’ E-Mail Addresses:;

Motivated by the literature on the capital asset pricing model, we decompose the uncertainty of a typical forecaster into common and idiosyncratic uncertainty. Using individual survey data from the Consensus Forecasts over the period of 1989-2014, we develop monthly measures of macroeconomic uncertainty covering 45 countries and construct a measure of global uncertainty as the weighted average of country-specific uncertainties. Our measure captures perceived uncertainty of market participants and derives from two components that are shown to exhibit strikingly different behavior. Common uncertainty shocks produce the large and persistent negative response in real economic activity, whereas the contributions of idiosyncratic uncertainty shocks are negligible.


Motivated by the literature on the capital asset pricing model, we decompose the uncertainty of a typical forecaster into common and idiosyncratic uncertainty. Using individual survey data from the Consensus Forecasts over the period of 1989-2014, we develop monthly measures of macroeconomic uncertainty covering 45 countries and construct a measure of global uncertainty as the weighted average of country-specific uncertainties. Our measure captures perceived uncertainty of market participants and derives from two components that are shown to exhibit strikingly different behavior. Common uncertainty shocks produce the large and persistent negative response in real economic activity, whereas the contributions of idiosyncratic uncertainty shocks are negligible.

1. Introduction

Heightened economic uncertainty, at both national and global levels, greatly contributed to the 2007–09 recession and shaped the speed of the subsequent recovery. Eight years after the end of the recession, there is still no sign of a complete global recovery. Advanced economies are uncertain about the effects of monetary policy normalization and emerging market economies are uncertain about the growth challenges ahead. Surrounded with unprecedentedly high uncertainty, economists face great challenges in understanding the origins of economic uncertainty and analyzing its causal impacts on real economy, e.g. Stock and Watson (2012).

Since there is no objective measure of uncertainty, economists have used numerous different proxies. A ubiquitous proxy is the implied or realized volatility in stock markets, such as VIX, e.g. Bloom (2009). However, the volatility in Wall Street might not reflect uncertainty in Main Street. For instance, changes in the VIX might be due to leverage or financial stress, despite low levels of economic uncertainty; see Bekaert et al. (2013). Jurado, et al. (2015) develop an alternative measure of economic uncertainty: the common variation in uncertainty across hundreds of economic series. Their measure reflects uncertainty around objective statistical forecasts, rather than perceived uncertainty by market participants. Moreover, as they focus on common, not idiosyncratic, uncertainty, there is no role for private information and heterogeneous agent models. A third leading proxy is based on the frequency of references to policy-related uncertainty in the newspapers, e.g. Baker, et al. (2016). But, like all measurements of this type, this news-based uncertainty measure puts a high bar for the attentiveness of reporters and editors, who might miss uncertainty events if they neglect to write a story on the subject. The fourth proxy for uncertainty is cross-sectional disagreement of economic agents, calculated as the dispersion in directional or point forecasts, e.g. Bachmann et al. (2013). When disagreement is taken to indicate uncertainty, the underlying assumption is that this inter-personal dispersion measure is an acceptable proxy for the average dispersion of intra-personal uncertainty. As shown by Lahiri and Sheng (2010), however, disagreement is only a part of uncertainty and misses an important component: the volatility of aggregate shocks.

To address some of the limitations in the existing measures, we develop a comprehensive measure of economic uncertainty by incorporating rich information reflected in the surveys of professional forecasters. Similar to Jo and Sekkel (2015), Rossi and Sekhposyan (2015) and Scotti (2016), our measure is based on subjective forecasts of market participants and reflects their perceived uncertainty. In contrast to these three papers, our uncertainty measure includes two components: common uncertainty as emphasized in Jurado et al. (2015) and idiosyncratic uncertainty as documented in the macroeconomics literature. Our decomposition of uncertainty of a typical forecaster into common and idiosyncratic parts is similar to Campbell et al. (2001) that decompose the volatility of a typical stock into market and firm-level volatility. We estimate the common component as the perceived variability of future aggregate shocks and idiosyncratic component as the disagreement among professional forecasters across three different layers. First, we estimate the variable-specific uncertainty for eight nominal and real economic indicators. Second, we measure the country-specific uncertainty as the weighted average of standardized components of variable-specific uncertainty measures. Finally, we propose an index of global uncertainty, which is a rather new concept in the literature.1 Constructed from a large set of countries, corresponding to more than 90 percent of the world economy, this global measure is more comprehensive than the previously proposed measures, e.g. Berger and Herz (2014).

Our main findings are summarized as follows. All uncertainty measures are countercyclical and at all layers, combined uncertainty is more countercyclical than its common or idiosyncratic component. A comparison of our country-specific uncertainty measures with alternative leading measures from the literature for a subset of countries shows that our measures have fewer peaks, all around the recessions, and have persistent and heightened uncertainty during the recession episodes. Shocks to our measures of uncertainty are associated with large and persistent drops in real activity at both national and global levels. Further investigation shows that common uncertainty shocks produce large and persistent responses in real activity, whereas the contributions of idiosyncratic uncertainty shocks are negligible.

The rest of the paper is organized as follows. Section 2 details the methodology on measuring uncertainty. Section 3 introduces the data used in this paper. Section 4 describes the properties of economic uncertainty measures. Section 5 presents the dynamic relationship between uncertainty and economic activity and Section 6 concludes. The online appendix includes detailed information on the dataset, alternative measures of country-specific uncertainty using principal component analysis and regional measures of uncertainty.

2. Methodology: Estimating Uncertainty

2.1 Uncertainty Decomposition

Our decomposition of the uncertainty of a typical forecaster is motivated by the literature on the capital asset pricing model (CAPM) that decomposes the return volatility of a typical stock into market volatility and firm-specific volatility. We start off by presenting the traditional CAPM decomposition that requires estimation of firm-specific betas and then move to the approach in Campbell et al. (2001) that does not require any information about individual betas on the aggregate level.

Let eit be individual i’s forecast error at time t. Then, consensus forecast error, et, is defined as the weighted average of individual forecast errors:


where wit is the weight of individual forecast error in consensus forecast error. Parallel to the CAPM literature that connects firm-specific return to market return, we specify the relationship between individual and consensus forecast errors as follows


where βi measures individual i’s tendency to respond to common shocks, as proxied by consensus forecast error et. Beta is important since it captures the risk arising from exposure to general economic conditions as opposed to idiosyncratic factors. The βi. below 1 indicates that an individual forecast error is not highly correlated with consensus forecast error. In equation (2), ɛit is orthogonal by construction to et. Equations (1) and (2) together impose the following restriction Σi=1Nwitβi=1, which is the standard assumption in the CAPM literature that the weighted sums of the different betas equal unity. Equation (2) permits a simple variance decomposition in which the covariance term is zero:


In equation (3), Var(et) measures the common volatility and Var(εit) captures the idiosyncratic volatility. The problem with this decomposition, however, is that it requires knowledge of individual-specific betas that are difficult to estimate and introduce another layer of uncertainty in parameter estimation. To avoid this problem, we follow the approach in Campbell et al. (2001) that does not require any information about individual betas on the aggregate level. To fix ideas, let uit denote the difference between eit and et :


Plugging equation (4) into equation (2) and re-arranging yields


The apparent drawback of equation (4) is that uit and et are not orthogonal, and so we cannot ignore their covariance. Taking the variance on both sides of equation (4), we have


where the second equality follows from equation (5). Again, taking into account the covariance term introduces the individual forecaster beta into the variance decomposition.

Note, however, that although the variance of an individual forecast error contains the covariance term, the weighted average of variances across forecasters is free of the covariance term and individual betas:


The covariance term from equation (6) aggregates out due to the standard restriction Σi=1Nwitβi=1. The weighted average Σi=1NwitVar(eit) can be interpreted as the volatility of a “typical” forecaster, selected randomly from among all forecasters with probability equal to its weight wit, e.g. Giordani and Söderlind (2003). Equation (7) states that the volatility of a typical forecaster can be decomposed into two parts: volatility that is common to all forecasters and volatility that arises from the heterogeneity of individual forecasters.

The observed disagreement among forecasts (or forecast errors) can be expressed as


The sample variance dt is a random variable prior to observing the forecasts. Taking expectations, we get an expression for the non-random disagreement, denoted by Dt, as


where the last equality holds since E(etɛit) = 0 and E(et) = 0 by assumption. Taking the variance on both sides of equation (5), we have


Plugging equation (10) into equation (9) yields


Combining equation (11) with equation (7), we get


Equation (12) decomposes the uncertainty of a typical forecaster into common and idiosyncratic uncertainty. The first component is the empirical variance of the consensus forecast, which is conventionally the common uncertainty in the literature; see Clements (2014). The second component is the forecast disagreement and captures idiosyncratic uncertainty2 Finally, we need to point out that our uncertainty decomposition is similar to the decomposition as in Lahiri and Sheng (2010) under a panel data framework.

2.2 Estimation

Based on the uncertainty decomposition in equation (12), we construct time series of the two components of uncertainty measure for each variable, each country, and finally for the world. In this subsection, we discuss how we estimate common and idiosyncratic components of variable-specific, country-specific, and global uncertainty measures.

The common uncertainty shocks have long been estimated using GARCH-type models, dating back to Engle (1982). Under such a framework, the estimates of common uncertainty depend on innovations to the raw series, denoted by Yt, and therefore cannot be separated from first-moment shocks. For this reason, we use the stochastic volatility model to estimate common uncertainty in our main analysis.3 The stochastic volatility model permits construction of a shock to the second moment that is independent of innovations to Yt. This exogeneity is consistent with the theoretical literature which presumes the existence of an uncertainty shock that independently affects real activity. Estimation of the common uncertainty using a stochastic volatility model has the following specification:


We estimate this model using Markov Chain Monte Carlo (MCMC) methods as in Kim et al. (1998). To prevent the impacts of the outliers, we use median forecast errors instead of mean forecast errors to estimate common uncertainty, σt2.

We measure forecast disagreement, Dt in equation (12), as the interquartile range of forecasts of survey respondents rather than their standard deviation in order to mitigate the effect of the outliers, as is common in the literature; e.g. Mankiw, et al. (2003) and Dovern, et al. (2012). With both common uncertainty σcjt2 and idiosyncratic uncertainty Dcjt at hand, our variable-specific uncertainty Ucjt for country c, variable j at time t can be estimated as


Since these two components of uncertainty measure have different scales, we standardize them using the min-max normalization rule. Applying this rule, both common and idiosyncratic uncertainty components are scaled between 0 and 1, and the sum of these two is bounded between 0 and 2 for all eight variables including GDP, consumption, investment, industrial production, inflation, unemployment rate, short-term and long-term interest rates. Variable-specific uncertainty estimates have two prominent features.4 First, guided by the recent empirical findings that surveys provide more accurate forecasts than models (see, for example, Ang et al. (2007) and Faust and Wright (2013)), we use surveys of professional forecasters directly rather than making objective statistical forecasts. Thus, our uncertainty estimates are less prone to measurement errors due to potentially misspecified econometric models that yield large forecast errors and inflated uncertainty estimates. Second, we use forecast errors, rather than forecasts, to remove the predictable component of the raw series and estimate common uncertainty as the conditional volatility of the purely unforecastable component of the future value of the series.

To estimate country-specific economic uncertainty, we take the weighted average of eight variable-specific uncertainty estimates as follows:


We present the results using equal weights wj=1j in the paper. As an alternative, we also estimate the country-specific uncertainty as the first principal component of eight variable-specific uncertainty series and find that the results are very similar.5 This definition emphasizes that economic uncertainty is a measure of common variation in uncertainty across many series, as also pointed out by Jurado, et al. (2015).

Unlike the variable-specific and country-specific uncertainty measures, global uncertainty receives little attention in the literature. This is possibly due to insufficient data to estimate global uncertainty. The existing global uncertainty measures are based on too few countries and tend to focus on developed economies. For instance, Hirata et al. (2013) construct a measure of global uncertainty based on stock price volatility in seven advanced economies and Berger and Herz (2014) estimate global uncertainty using nine advanced economies and two variables. To address these limitations, we use a dataset of 45 advanced and emerging market economies, covering more than 90 percent of the world economy today. For these economies, we include eight variables for each country, covering both real and nominal variables. Taking advantage of this rich dataset, we construct a measure of global uncertainty as the purchasing power parity (PPP)-weighted average of the country-specific uncertainties.6

3. Data

We use survey data of macroeconomic forecasts to compute uncertainty measures. The forecast data are from the Consensus Forecasts, publications of the Consensus Economics Inc., a private macroeconomic survey firm based in London. This survey is a comprehensive dataset with a large coverage of advanced and emerging market economies. For each country, the survey asks similar questions to a panel of 10–30 professional forecasters, all based in the home country, on the first week of each month. For some countries, the definition of variables varies slightly (i.e. manufacturing production instead of industrial production) and for others some questions are omitted because of possible data limitations. Other than these, the surveys have a near uniform design for all countries in the sample, which makes the results comparable across countries. Our study covers all 45 countries with monthly forecasts available for the annual growth rates of GDP, consumption, investment, industrial production, and levels of inflation, short-term and long-term interest rates, and the unemployment rate. These eight variables enable us to capture uncertainty both in nominal and real macroeconomic series, where inflation, short-term and long-term interest rates are in nominal and the rest are in real terms. Table A.1 in the online appendix provides detailed information on the country, time and variable coverage of the dataset.

Forecasts for all variables except interest rates are fixed event forecasts. Every month, each survey participant provides forecasts for both the current and next calendar year. These fixed event forecasts get closer to the actual values when the forecasting horizon is shorter. Following Dovern, et al. (2012), we transform the fixed event forecasts of all variables into fixed horizon forecasts with the following adjustment:


where Fi,t+k|t and Fi,t+12+k|t are the two forecasts based on the information set at time t with horizons of k ∈ {1, … ,12} and k + 12 months, respectively. The average of two fixed event forecasts weighted by their share in the forecasting horizon approximates the fixed horizon forecast, Fi,t+12|t, for the next 12 months. For interest rates, survey participants provide both three-month and twelve-month ahead forecasts. To be consistent with the horizon of the forecasts for other variables, we use the twelve-month ahead forecasts for both short-term and long-term interest rates.

Turning to the actual values, monthly series are available for industrial production, inflation, unemployment, short-term and long-term interest rates. For real GDP, consumption and investment, we use quarterly series as they are not available at the monthly frequency and we use the value belonging to a quarter for each of the three months in that quarter. The main sources of actual values are Global Data Source of IMF, Haver Analytics, OECD Analytical databases and country statistical offices. To match the actual values with the fixed-horizon forecasts, we perform the appropriate data transformation.7 We explore the properties of these forecasts through the Mincer-Zarnowitz regression and find that some forecasts are biased and inefficient in incorporating new information.8 Despite these inefficiencies, we use forecast data because of the advantages of surveys over purely model-based forecasts and because these surveys reflect market participants’ perceptions of economic development in the future. These perceptions are the key to capturing how economic agents experience uncertainty in the economy.

4. Properties of Economic Uncertainty

We estimate variable-specific uncertainty (VSU) for eight indicators. For most of the economies in the sample, the VSU is countercyclical for all series. Moreover, some VSU estimates are highly correlated. Table 1 shows that, for the United States, the pairwise correlations are quite high for most of the VSU estimates. Interestingly, pairwise correlations between all VSU estimates except long-term interest rate are higher for the common than for the idiosyncratic component. For instance, the correlation between inflation and investment growth is 0.27 for idiosyncratic uncertainty, but 0.78 for common uncertainty. If one estimates uncertainty at the country level using only forecast disagreement, then there would be too many uncertainty spikes due to idiosyncratic shocks in individual series. On the other hand, if one estimates uncertainty using only the common component, then the series would be too smooth. These findings imply that the combined estimate of these two reflects the uncertainty in the entire economy better than any individual component.

Table 1.

Correlation between Variable-specific Uncertainty Measures: United States

article image
Note : Output, consumption, investment, and industrial production stand for the growth rates of these indicators. The sample is between 1989M11-2014M7 for all estimates.

For all countries, common uncertainty is less volatile and on average, higher than idiosyncratic uncertainty. There are very few peaks in common uncertainty and those peaks are usually around recessions. For instance, in the United States, the uncertainty for output, consumption, investment, unemployment rate and short-term interest rates increases during all three recession periods covered in the sample of 1989–2014.9,10 Interestingly, some regional recession episodes are associated with higher uncertainty than global recession episodes. For instance, in Indonesia and South Korea, some of the VSU peaks around the 1997 Asian financial crisis are higher than those around the recent global recession. This is consistent with the findings of Hirata, et al. (2013): since the mid-1980s the importance of regional factors has increased and global factors play a lesser role in explaining international business cycles.

Turning to the country-specific uncertainty (CSU), Figure 1 plots the uncertainty estimates for 45 advanced and emerging market economies. The CSU is strongly countercyclical. Almost in all countries, the CSU peaked around 2009, even though the country itself did not experience any recession, e.g. China and Australia. For some emerging market economies, the uncertainty was higher during earlier recessions than the latest global recession. For instance, the largest uncertainty peak for Argentina is around 2001–2002 when there was a deep financial crisis in the country, whereas for Hong Kong it is around 1997–1998 Asian financial crisis.

Figure 1.
Figure 1.
Figure 1.
Figure 1.

Country-Specific Uncertainty

Citation: IMF Working Papers 2017, 219; 10.5089/9781484316597.001.A001

Note : Country-specific uncertainty is the sum of idiosyncratic and common uncertainty. Gray bars indicate the period of recessions as identified in Claessens, et. al. (2016, forthcoming).

The uncertainty at the national level influences the variable-specific uncertainty. To explore this impact, Table 2 presents the proportion of variable-specific uncertainty that is explained by the country-specific counterpart. For the entire sample, on average, the explanatory power of the CSU for the variable-specific uncertainty is almost the same during recessions (R2 = 0.585) and expansions (R2 = 0.576). For the advanced economies, however, it is higher during recessions (R2 = 0.51) than expansions (R2 = 0.46). Shorter time coverage of the emerging market economies makes it difficult to compare the explanatory power at different phases of the business cycle. For eight out of fifteen advanced economies, the CSU explains output growth uncertainty the most. Furthermore, the explanatory power varies over business cycles. For instance, in the United Kingdom, the CSU explains investment growth uncertainty the best during recessions but the least during expansions. In Japan, the variable that the CSU explains the most is inflation uncertainty during recessions but output growth uncertainty during expansions. For emerging market economies, the evidence is rather mixed. For instance, R2 is highest for industrial production uncertainty in China, Poland, and Czech Republic; for consumption uncertainty in Argentina, Brazil, Colombia, Peru, South Korea, Philippines, Lithuania, and Romania; for investment uncertainty in Bulgaria, Croatia and Russia. Taken together, we see that country-specific uncertainty accounts for a large fraction of the variation in the variable-specific uncertainty. But there is a large amount of idiosyncratic variation in uncertainty across variables, as evident from many R2 statistics that are much lower than one.

Table 2.

R-square : Variable -specific Uncertainty on Country-specific Uncertainty

article image
article image
article image
Note: Each cell presents the R-square of the regressions of respective variable-specific uncertainty on country-specific uncertainty measures. Recession episodes are from Claessens, Kose, Ozturk, Terrones (2016, forthcoming). The last column presents the average of the R-square in each economy. Numbers in red are the smallest values and numbers in green are the largest values in the row they stand.

Our country-specific uncertainty measure complements the uncertainty estimate proposed by Jurado, et al. (2015) in two dimensions. First, we use surveys of professional forecasters available for many countries and focus on market participants’ perceived uncertainty; whereas they generate forecasts from augmented autoregressive models and measure uncertainty only for the U.S. around objective statistical forecasts. Second, they measure macroeconomic uncertainty as the common factor of all uncertainty estimates of hundreds of variables. In contrast, our uncertainty measure captures both common and idiosyncratic uncertainties that have different effects on economic activity as we show in the next section.

With national uncertainty at hand, we estimate global uncertainty as the weighted average of country-specific uncertainties in Figure 2. Global uncertainty is strongly countercyclical and rises during the global recessions of 1991 and 2009, identified by Kose and Terrones (2015). The country-specific uncertainty is potentially influenced by global uncertainty because of large trade and financial interconnectedness among economies. Table 3 shows the proportion of the variation in the country-specific uncertainty that is explained by global uncertainty. In some of the Asian economies, global uncertainty explains only a small fraction of the country-specific uncertainty. For instance, R2 is 0.435 in Hong Kong and 0.079 in Thailand. On the other hand, in some of the Eastern European economies, global uncertainty explains a very large fraction of the country-specific uncertainty, e.g. R2 = 0.925 in Lithuania, 0.904 in Latvia and 0.886 in Bulgaria. In addition, global uncertainty amplifies the country-specific uncertainty for almost half of the sample, where the coefficient is significantly larger than 1. This amplification is less evident for its common component than idiosyncratic component.11 Finally, global uncertainty has the largest explanatory power relative to its two components. Parallel to other layers of uncertainty, the sum of both components better reflects worldwide uncertainty than any individual component.12

Figure 2.
Figure 2.

Global Uncertainty

Citation: IMF Working Papers 2017, 219; 10.5089/9781484316597.001.A001

Note : Each line presents the PPP-weighted average of the respective measure for 46 economies. Gray bars present the global recession episodes identified by Kose and Terrones (2015).
Table 3.

R-square: Country-specific Uncertainty on Global Uncertainty

article image
Note : Economies are sorted with respect to their estimated coefficients in uncertainty (total). Each result is based on bivariate regressions of country-specific uncertainty on global uncertainty. *** indicates significance at 1 percent level.