Uncertainty matters: Evidence from close elections

Journal Pre-proof Uncertainty matters: Evidence from close elections

Chris Redl PII:

S0022-1996(20)30015-5

DOI:

https://doi.org/10.1016/j.jinteco.2020.103296

Reference:

INEC 103296

To appear in:

Journal of International Economics

Received date:

26 August 2019

Revised date:

10 December 2019

Accepted date:

17 January 2020

Please cite this article as: C. Redl, Uncertainty matters: Evidence from close elections, Journal of International Economics(2020), https://doi.org/10.1016/j.jinteco.2020.103296

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

© 2020 Published by Elsevier.

Journal Pre-proof

of

This work was completed while the author was an employee of the Bank of England. The views expressed in this paper are those of the author and do not necessarily reflect those of the Bank of England. I would like to thank Alex Tuckett, Roland Meeks, Ethan Illzetzki, Andreas Dibiasi, Matteo Iacoviello, Steffen Elstner, Ambrogio Cesa-Biachi, Pietro Veronesi, Barbara Rossi and seminar participants at Queen Mary University of London, Stanford SITE 2018, Lancaster University, Young Finance Scholars Conference 2019, 19th IWH-CIREQ-GW Macroeconometric Workshop: Uncertainty, Expectations and Macroeconomic Modelling, MMF 50th Annual Conference, Twenty-Seventh Annual Symposium of The Society for Nonlinear Dynamics and Econometrics and the NBER International Seminar on Macroeconomics 2019 for helpful comments. I would especially like to thank Juan Antolin-Diaz for sharing his code for narrative sign restrictions and John Fernald for extensive comments and helpful suggestions. Manveer Sohki provided excellent research assistance. Any errors are my own.

-p

ro

Uncertainty Matters: Evidence from Close Elections

re

Chris Redl*

Jo

ur

na

lP

International Monetary Fund, 700 19th Street NW, Washingto, DC, United States (e-mail: [email protected])

Journal Pre-proof ABSTRACT This paper uses a data rich environment to produce direct econometric estimates of macroeconomic and financial uncertainty for 11 advanced nations. These indices exhibit significant independent variation from popular proxies and provide a refinement of the influential work of [Jurado et~al., 2015] that results in improved real-time performance. We use this new data in combination with narrative evidence to jointly identify macro uncertainty and financial shocks. Macro uncertainty shocks are identified with close elections and financial shocks with financial stress during financial crises. We find that macro uncertainty shocks matter for the majority of

of

countries and that the real effects of macro uncertainty shocks are generally larger conditioning on

ro

close elections. We find that macro uncertainty shocks are more important, on average over our sample, in explaining business cycle fluctuations in real variables than financial shocks. These

-p

results are robust to controlling for news shocks and global uncertainty as well as a variety of shocks considered to be important drivers of the business cycle.

re

Keywords: economic uncertainty, business cycles, elections.

Jo

ur

na

lP

JEL classification numbers: D80, E32, D72

Journal Pre-proof

1.

Introduction

The global financial crisis (GFC) has renewed interest in two drivers of the business cycle: financial shocks and uncertainty shocks. For example, Stock and Watson [66] find that shocks to credit spreads and uncertainty accounted for two thirds of the movements in US GDP growth from 2008-2012. However, the GFC was associated with large increases in uncertainty and a significant deterioration of financial conditions. Thus in samples where this episode dominates, it can be

of

difficult to separate the effect of one shock from the other. Indeed, while there is a broad consensus that independent financial shocks can produce a recession, there is significant debate as to whether

ro

uncertainty shocks that act independently of a financial channel, have significant business cycle

-p

effects (Caldara et al. [20],Ludvigson et al. [50] and Born et al. [17]).

re

Figure 1: Uncertainty rises with elections, credit spreads do not Macro uncertainty is calculated as per section 2. BBD uncertainty is based on Baker et al. [8] and

lP

taken from policyuncertainty.com.

na

Close elections dates are as in table 4.

This paper addresses this question by looking for events that are associated with a rise

ur

uncertainty about the real economy but that are not driven by large financial shocks. I identify 29 closely contested general elections across the G10 as such periods. During these events there is a

Jo

rise in measures of uncertainty that relate to the real economy but they are not typically associated with financial stress as indicated by credit spreads - see figure 1. This study makes two contributions to this debate: improved measurement, and a novel identification procedure using narrative information to jointly identify both macro uncertainty and financial shocks. On measurement, this debate has focused almost exclusively on US data

1

where the

global financial crisis originated and where its effects were particularly pronounced. Here I extend this to 11 advanced nations increasing the potential coverage of non- financial events that cause uncertainty. I contribute to the measurement of uncertainty in the cross country context by 1

Popescu and Smets [60] is an early exception studying this question using German data and in a co mpanion paper

Redl [62] studies the case of the UK in detail.

Journal Pre-proof producing separate measures of macro and financial uncertainty following an adjusted version of the approach of Jurado et al. [42], hereafter JLN. This adjusted approach makes this uncertainty measure more useful for policy in that it has significantly improved real time performance and can be used as an input to forecasting. To identify macro uncertainty shocks, I impose a small number of relatively uncontroversial sign restrictions and then sharpen this identification by imposing additional narrative sign restrictions (as developed by Antolín-Díaz and Rubio-Ramírez [2]) requiring that macro uncertainty shocks take place during close elections. The latter are natural candidates for exogenous variation in uncertainty shocks since they portend potentially large

of

policy shifts and their timing is largely unrelated to the business cycle. In particular, these are

ro

periods when macro uncertainty is less likely to act through a financial channel compared to events where large economic shocks take place. Similarly, financial shocks are identified with sign

-p

restrictions that only impose that credit spreads and financial uncertainty are positive with identification sharpened by imposing that financial shocks are positive during periods of peak

re

financial stress as determined by the financial distress index of Romer and Romer [63]. Thus I

lP

jointly identify orthogonal macro uncertainty and financial shocks and allow them to compete in explaining real variables.

I find that macro uncertainty shocks matter for the majority of countries studied where the

na

transmission is consistent with firms pausing investment and reducing hours worked, precipitating a decline in GDP. I find evidence that the response of consumption to uncertainty shocks is

ur

stronger when conditioning on close elections, although this varies by country. Identifying macro

Jo

uncertainty shocks with close elections increases the share of GDP fluctuations they explain to one of the most important sources at the 1-2 year horizon, and roughly doubles their role in explaining variation in hours from around 10% to over 20%. Interestingly, financial shocks are found to be less important, on average, than macro uncertainty shocks in driving real variables, despite being identified with periods of peak financial stress. A key challenge to empirical studies of uncertainty shocks is to control for the fact that uncertainty is likely to rise at times when negative first moment shocks hit or when mean expectations deteriorate (Haddow et al. [36]). I address this concern by identifying a news shock using 1- year ahead professional forecasts for GDP growth and imposing that these forecasts don’t deteriorate on impact when the uncertainty shock hits. Cross-country studies have provided evidence that the real effects of domestic uncertainty shocks may be driven by increases in global uncertainty (Berger et al. [12], Mumtaz and Theodoridis [53], and Ozturk

Journal Pre-proof and Sheng [55]). I add a measure of global uncertainty to the model and identify a global uncertainty shock. In general, I find our results are robust to including a global uncertainty and news shock in the model, however, consistent with these studies I find that rises in global uncertainty as well as a deterioration in mean expectations are an important part of the transmission of domestic uncertainty shocks in some countries. A wide range of approaches to measuring uncertainty have been pursued in the literature. Closely related to the JLN approach is the use of realised and implied stock market volatilties as initiated by Bloom [15] and used by Caggiano et al. [19], Basu and Bundick [10] for the US and

of

Carriere-Swallow and Cespedes [23], Cesa-Bianchi et al. [25] and Berger et al. [12] in a

ro

cross-country context. These proxies are focused exclusively on financial measures and don’t control for a deterioration in mean expectations as volatilties rise. The JLN approach improves on

-p

this by using a broad information set of macro and financial variables, separately measures macro and financial variables and uses a forecasting model to remove mean expectations.

re

The influential work of Baker et al. [8] employed a news count of articles discussing policy

lP

uncertainty which has been extended to cover monetary policy (Husted et al. [37]), geopolitical risk (Caldara and Iacoviello [21]) as well as a wide range of countries 2 . Text based measures are valuable for capturing uncertainty not reflected in data however, they can be very volatile and

na

potentially provide a misleading signal when their correlations with the data change (see Forbes [32] for the case of the UK during the Brexit vote). The JLN approach aggregates the uncertainty

ur

in a large number of series and is thus less susceptible to unusual changes in just one series. This is

Jo

also an advantage over other econometric estimates of uncertainty that use estimated volatility in a small number of macro variables such as Fernandez-Villaverde et al. [31], Fernandez-Villaverde et al. [30], and Croce et al. [27].

A number of studies have shown a reduced role for uncertainty shocks once financial shocks are appropriately controlled for. Popescu and Smets [60] show that the real effects of financial stress are much larger and more persistent than those of uncertainty with lower inflation, GDP, and higher unemployment. Caldara et al. [20] find that both financial and uncertainty shocks matter for real fluctuations but that uncertainty shocks matter significantly more when they 2

These include Dendy et al. [29] and Haddow et al. [36] for the UK, Popescu and Smets [60] for Germany, Zalla

[68] for Ireland, Kok et al. [46] for the Netherlands, Arbatli et al. [3] for Japan, Armelius et al. [5] for Sweden, Larsen [47] for Norway, and Redl [61] for South Africa.

Journal Pre-proof coincide with a tightening of credit spreads. In a forecasting exercise comparing a wide range uncertainty proxies for the U.S., they find that only the JLN a nd a proxy based on forecaster disagreement are informative about real industrial production 12 months ahead once current financial conditions are accounted for. This highlights the measurement contribution of this paper in extending the JLN indices to the G10. This paper is closely related to Ludvigson et al. [50]. They identify financial and macro uncertainty shocks and their impact on industrial production using two sets of shock-based constraints. Firstly, narrative event constraints, requiring financia l uncertainty shocks are at least 4

of

standard deviations in October 1987 (Black Monday) and at some period during the 2007-2009

ro

financial crisis, while macro uncertainty shock are no larger than 2 standard deviations. Secondly, correlation constraints which impose that the identified uncertainty shocks are negatively

-p

correlated with an external variable, aggregate stock market returns, but that correlation is larger (in absolute value) for financial uncertainty. They find that macro uncertainty is a fully

re

endogenous response to real shocks that cause business cycles but that financial uncertainty shocks

lP

have negative effects on real variables. Here I pursue a related but alternative strategy, while Ludvigson et al. [50] employ narrative restrictions on events where financial uncertainty should play a larger role than macro uncertainty (financial crises), I use events where macro uncertainty

na

should play a larger role than financial uncertainty (close elections). Ludvigson et al. [50] employ correlation with an external variable (stock returns) whereas we employ sign restrictions on the

ur

response of variables to the shocks. While the former is a novel and an appealing approach is it

Jo

more challenging to use for a larger model (I have up to 11 variables rather than the 3 used in Ludvigson et al. [50]) where finding the appropriate external variables is not straightforward. Moreover, given the wide range of macro and financial variables that uncertainty indices may respond to, it is important to control for a braod range of macro and financial variables. Our identification relies on the positive link between macro uncertainty and close elections. Kelly et al. [44] present a model where firm profitability depends on government policies and agents learn about the impact of those policies from political news. Elections create uncertainty by resetting agents beliefs about government policy. They show that this model predicts a positive relationship between option prices and elections, and in their empirical work find evidence of a 5% premium on options that cover political events (national elections and global summits) relative to those that do not. Azzimonti [6] develops a model where the quality of

Journal Pre-proof government policies influence the probability of a recession. Partisan conflict lowers the quality of those policies promoting tail risk that reduces investment spending. Agents rely on signals to learn the degree of partisan conflict where elections generate a spike in uncertainty about partisanship through resetting agents priors. A number of papers provide empirical support to this link. Li and Born [49] find that realised US stock market volatility rises prior to the election date if there is no clear leader in election polls. Bialkowski et al. [13] find that realised stock market volatility is 23% higher within a two month window around elections using data on 27 OECD countries. They find evidence that a

of

small margin of victory is a significant determinant of that rise in volatility. Goodell and Vähämaa

ro

[35] find similar evidence of increased implied volatility around elections using the VIX. These results suggest that rises in financial uncertainty may play an important role in the transmission of

-p

macro uncertainty shocks around elections - I find some evidence for this link but do not find that it is always and everywhere the case. Gao and Qi [34] provide evidence that municipal bond rates

re

rise around gubernatorial elections in the US while Jens [39] documents falls in corporate

lP

investment around these elections. Julio and Yook [40] and Canes-Wrone and Park [22] document uncertainty induced declines in investment around general elections across a variety of developed and developing countries. Julio and Yook [41] use election timing as a source of fluctuations in

na

political uncertainty, documenting a significant drop in FDI flows to receipt countries from the US around domestic elections. They find this effect is more pronounced for closer elections. Larsen

ur

[47] develops topic-specific measures of uncertainty using text mining too ls on a corpus of articles

Jo

from the major Norwegian business daily. He shows that uncertainty relating to elections is one of the most important types of uncertainty driving investment. The remainder of this paper is structured as follows: section 2 outlines the econometric framework used to measure macro and financial uncertainty; section 3 describes the data set used in estimation; section 4 describes the estimates of uncertainty I find; section 5 describes the macroeconomic impact of uncertainty shocks; and section 6 concludes.

2.

Measuring Uncertainty: Econometric Framework

I measure uncertainty following JLN, the reader is directed to their paper for full details of that approach. Uncertainty is defined as the conditional variance of the unforecastable component

Journal Pre-proof common to a large set of macro or financial variables. The approach proceeds as follows: (1) collect a large set of macro and financial data; (2) summarise the data using principal component factors; (3) use the factors in a VAR model to make fo recasts of each series in the dataset; (4) estimate the latent, time varying, conditional variance of the unforecastable component of each series by feeding the forecast errors from the VAR model to a stochastic volatility model; (5) estimate uncertainty as the average of the estimated stochastic volatilities of the series of interest 3 (e.g. macro variables for macro uncertainty, financial variables for financial uncertainty). This methodology ensures that measured uncertainty captures when the economy has become less

of

predictable (rather than just more volatile) and also reduces dependencies on one (or a small number of) observable series.

ro

Following Ludvigson et al. [50], let y jt  Yt = ( y1t , y2t ,..., yNt ) be one of the N variables

-p

for a given country dataset. A forecast, E  y jt 1 | I t  , is taken from a factor augmented forecasting

re

model:

ˆ   W ( L) W  v y y jt 1 =  jy ( L) y jt   Fj ( L)Fˆt   Gj ( L)G t j t jt 1

(1)

lP

Where  jk ( L) for k = { y, F , G,W } are finite order lag polynomials. The factors, Fˆ t , are

na

drawn from the information set of agents, I t , comprised of the full data set of macro and financial

ˆ is drawn in the same way except that the squares of the original data variables for that country. G t

ur

are used to capture potential non- linearities, and Wt is the square of the first component of Fˆ t ,

Jo

ˆ and W are permitted capturing non- linearity in this factor. The prediction error for y jt 1 , Fˆ t , G t t to have time- varying volatility4 . The forecasting model can be cast as FAVAR in first order

3

To estimate the expected forecast error variance for each series requires using a forecasting function based on the

set-up assumed in the stochastic volatility model. In the case of JLN this corresponds to an AR(1) and so is not simp ly an average of the stochastic volatility estimates. However, in the stochastic volatility model used here I assume a random walk process for the volatilities which means the expectation at any future horizon is just the estimate of volatility today and taking simple averages is a correct description. 4

JLN allow for stochastic volatility in both the estimates of the factors used to augment the VAR and the variables

included in the VA R. Th is results in four sources of time variation in the forecast errors due to the stochastic volatility of the VAR shocks, the factors, the covariance between these two, and an autoregressive term due to persistence in the volatility of the VA R shocks. Without stochastic volatility the forecast error would not vary with

t but only with h

Journal Pre-proof ˆ  , W) , Y = ( y , y ,..., y  companion form with Zt = (Fˆt, G jt jt jt 1 jt  q 1 ) , and t t  t   Z  =  Y jt   j

0   t 1    C  Yj   Y jt 1  

Where  Z an autoregressive term for

t

Z t Y jt

  

t

= (Zt ,..., Zt q 1 ) :

(2)

. The mean squared forecast error varies over

time due to the fact that shocks to y jt 1 and Z t have time varying variance. The forecast error  variance of y jt 1 , denoted by  jt = Et  

Y jt 1



Y jt 1

  5, is the conditional volatility of the purely :





2 = 1j  jt 1 j = E  y jt 1  E  y jt 1 | I t  | I t   

(3)

-p

jt

jt

ro

time t . The definition of uncertainty for y jt is denoted

of

unforecastable component of the future value of the series, conditional on all information known at

re

This procedure results in an uncertainty measure for each series in Yt . To arrive at an aggregate measure of uncertainty in that category I use simple averages of those indices (as in t

=

1 N



N j =1

jt

. To form the macro uncertainty index I use only macro variables in the

lP

JLN):

average, similarly for the financial uncertainty index.

Y jt 1

, are taken to be noisy measures of the true underlying stochastic volatility

ur

predictions errors,

na

Clearly, a key input to this uncertainty measure is the expected forecast error variance. The

process  jt and are fed to a stochastic volatility model to infer the latter. A stochastic volatility

Jo

model allows for shocks to the second moment of a variable to be independent of the first moment ensuring that these estimates capture a mean preserving increase in volatility rather than a rise in volatility that accompanies a deterioration in the mean (as is often seen in survey forecasts used widely in uncertainty proxies). JLN use the STOCHVOL package in R which implements the highly efficient stochastic volatility smoother model of Frühwirth-Schnatter [43]. This method uses the entire data series to estimate the latent stochastic volatility at date t . This means that future forecast errors are used to infer the level of uncertainty today. This has a number of

. See JLN, p.1188. 5

JLN consider horizons greater than 1 period. However when standardised these estimates are very similar and so

here I proceed with outlining the simpler case of a one period ahead based measure.

Journal Pre-proof drawbacks. Firstly, the uncertainty estimates are arguably inappropriate for forecasting as they would entail look ahead bias. Secondly, and as demonstrated below, the two-sided nature of the filter leads to large real time revisions making these estimates unappealing for policy use. Finally, it is unappealing from a theoretical perspective that an agents estimate of uncertainty today is informed by information outside the agents information set. To address these concerns I employ a one-sided stochastic volatility filter that only employs current and past values of forecast errors to infer the latent state of volatility. I use a minor alteration of the Jacquier et al. [38] algorithm, with

Data

ro

3.

of

the details outlined in the appendix.

-p

For each country, the forecasts above are formed on the basis of two monthly data sets, one capturing macroeconomic series and one capturing financial variables. The data sources are

re

described in full in the appendix. The data generally covers early 1990s to 2018 6 . The original JLN work employed a monthly model and I do the same here to capture higher frequency changes in

lP

forecast errors which may not be captured in a quarterly model7 . The macro series range in number from 50 (US) to 15 (Switzerland), and broadly cover the labour market (unemployment,

na

employment, wages, vacancies), retail sales, industrial production, new orders, inflation, trade (exports, imports and their prices), vehicle sales as well as business consumer confidence and a

ur

composite leading indicator. The financial series are fewer in number and range from 27 (the UK) to 8 (Spain), and broadly cover exchange rates, money supply, credit extension, foreign reserves,

Jo

interest rates (interbank rates, government bond yields) and share price indices. The original JLN measure of financial data captures only asset returns whereas here it is defined more broadly to include credit extension - which is important in models featuring financial frictions. For each country, the macro and financial data sets are combined to form the information set in the forecasting model from which the forecasting factors are drawn. The forecasting model

6

Regularly undated versions of the uncertainty indices are available from the authors website.

7

Experiments with a quarterly dataset for the US, covering similar series to those used for the other countries here,

showed that a quarterly model does well in capturing macro uncertainty but less well in capturing financial uncertainty when compared to the orig inal JLN indices (experiment used a two -sided filter to be co mparable). However, the JLN financial data focuses exclusively on asset returns where I take this measure to be broa der, see above.

Journal Pre-proof uses a large set of potential predictors in the factors, Ft , and Wt (which is comprised of squares of the first principal component in Ft ), and G t a further set of factors drawn from the squares of

ˆ , are chose n the original data set. From the potential factors, Ft and G t , a subset, Fˆ t and G t based on the information criterion in Bai and Ng [7].The set of predictors,

Fˆ , Gˆ , W  , are t

t

t

selected for inclusion in the forecasting model based on their incremental predictive power using a t-test (with the threshold set at t = 2.575 , corresponding to a 99% confidence interval) for each

Estimates of Uncertainty

ro

4.

of

y jt 8 .

-p

Figure 2 compares the estimates for macro uncertainty across countries. The GFC is largest uncertainty event for most countries but there remains significant idiosyncratic variation. For

re

example, in March 2011 a 4 standard deviation rise in uncertainty took place in Japan as the 9.0

lP

magnitude Tohoku earthquake hit the east coast. Italy experienced a significant rise in macro uncertainty during 1992 as the Amato government cut pension and benefit entitlements (Miniaci

na

and Weber [51]). The UK experienced high uncertainty around 2002 linked to poor performance in the manufacturing sector (Redl [62]). Recently, the slowdown in global trade has had a significant

ur

impact on Swiss export performance leading to a large uptick in uncertainty at the end of 2018.

Jo

Figure 2: Macro Uncertainty

Figure 3: Financial Uncertainty

Similar patterns are present in the financial uncertainty measures - see figure 3. Switzerland experienced very high financial uncertainty around the announcement of the Swiss Franc- Euro exchange rate floor in September 2011 and the ending of the floor in January 2015. The Netherlands experience a significant increase in financial uncertainty in 2001 as share prices collapses following the dot-com bust in the US. Germany experienced high financial uncertainty as interest rates rose and credit growth declined sharply in the early 1990s. 8

The equations each contain four lags of their own series.

Journal Pre-proof

SWITZER

NETHER

AIN DEN

LAND

LANDS

66

55

58

27

44

46

60

48

48

32

28

38

53

of

Table 1: Correlations across Macro and Financial Uncertainty measures U

JAP

GER

ITA

U

FRA

CAN

SP

S

AN

MAN

LY

K

NCE

ADA

57

53

57

5

50

SWE

46

14

39

46

58

61

54

30

42

A USA

5 JAPAN

20

49

48

5 0

GERMAN

52

34

54

3

ITALY

47

56

57

4

UK

60

24

26

47

FRANCE

57

54

65

72

46 3

CANADA

60

45

38

56

64

54

36

26

59

56

45

32

38

53

48

64

17

37

57

45

38

33

43

lP

7

45

re

9

4

ro

8

-p

Y

36

0

45

27

60

80

4

na

SPAIN

68

43

62

51

47

54

43

54

42

59

45

59

44

4

45

LAND NETHER

53

37

33

55

38

52

ur

SWITZER

56

45

Jo

SWEDEN

56

LANDS

55

4 1 4

21

1 66

3

48

6

Note: Below main diagonal are financial uncertainty correlations (%), above main diagonal are macro uncertainty correlations.

Table (1) presents the cross-correlations in macro and financial uncertainty across countries. There are higher levels of correlation for financial compared to macro uncertainty, as one might expect given open capital accounts. The G7 have stronger links on both measures however Japan’s financial uncertainty is more weakly correlated with uncertainty in the rest of the

Journal Pre-proof countries. The Netherlands and Switzerland are an outliers in terms of the independence of their experience of macro uncertainty relative to the other nations, although financial uncertainty is more closely aligned with other European nations.

Figure 4: Comparing Macro Uncertainty Indices to Baker, Bloom & Davis (2016) EPU All indices are standardised. BBD style news index for Switzerland is provided by KOF Swiss Economic

Institute,

available

at

of

https://www.kof.ethz.ch/en/forecasts-and-indicators/indicators/kof- uncertainty- indicator.html.

ro

The macro uncertainty measures exhibit significant independent variation from the news based indices of Baker et al. [8] labeled as BBD in figure (4). The indices computed in this paper

-p

show less short term volatility and greater persistence for uncertainty spikes, and register larger increases in uncertainty around the GFC. The JLN based indices also do not accord with the recent

re

increases in EPU seen in the UK, Germany, France and Canada. This may be due to coverage of

lP

political events that have not resulted in greater inability to forecast the path of real macro variables. Given the lack of correlation between the JLN and BBD indices the question of which to use naturally arises. I investigate the usefulness of each in the context of a simple forecasting

na

exercise following Caldara et al. [20]. I estimate the following equation:

ur

4 y  ln  t  h  =  h  h xt   ht    h , j  ln yt  j   t h j =1  yt 1 

(4)

Jo

Where y is GDP, xt is a measure of uncertainty,  t is a set of controls that includes credit spreads, global uncertainty and consensus 1 year ahead GDP growth forecasts. The time period is quarterly. This tests how reliable a predictor of economic activity each domestic uncertainty index is once I control for financial conditions, global uncertainty and, news about future growth, given by h . The JLN based indices outperform the BBD indices both in terms of statistical significance and scale of impact, with the macro index performing best (table 2). Interestingly, there is relatively little overlap in the countries or horizons where the different indices perform well suggesting that they may fruitfully employed in combination 9 . The joint 9

What drives the differences in forecasting performance across the different uncertainty indices, e.g. structural

differences across countries, or the types of events that drive uncertainty is a valuable area for future research.

Journal Pre-proof monitoring of the indices may be useful for narrative purposes for example, when a rise in the BBD indices is subsequently followed by a rise in the JLN based measures this maybe a useful cross-check that the news-based uncertainty measure is leading to a deterioration in forecastability in the economy, with larger real effects10 . Table 2: Forecast performance of uncertainty indices for GDP ( h ) Macro Uncertainty Horizo

4

8

12

Financial Uncertainty 16

4

8

12

ers):

**

-0.48 -0.54

-0.7

***

***

8**

0.11

0.12

-0.0

-0.04

0.52

0.14

-0.01

6

-0.81 -0.44 -0.82 -1.29 ***

***

ur

*

-1.08 -1.28 -1.40 -1.24

France

Canad

-0.11

***

0.05

-0.3

0.30

0.53

0.73

0.62

0.98

0.29

0.44

-0.18 -0.20

-0.3

**

-0.93 -1.85 -1.36 **

-0.42 -0.09

***

-0.84 -0.42

0.10

*** 0.49

1 -0.14 -0.09

0.20

0

-0.7

***

***

1*

0.21

0.15

-0.0

Jo

***

-0.60 -1.39 -1.45 -1.24 **

***

**

*

Spain

0.20

0.88

1.49

0.19

0.33

-1.03 -2.10 -1.00 -0.07 -0.23 -1.03 *

***

*

-0.48 -0.79 -0.33 -0.07 -0.50 -0.53

-1.4 2 -0.5 5

0.36

0.39

1.04

-0.29

0.26

0.44

* -0.6

0.72 **

5

a

10

-0.1

na

ny

UK

16

*

lP

8

Italy

12

-1.29 -1.10 -0.60 -0.16 -0.00

re

*

Germa

8

-p

-0.90 -1.26 -2.28 -2.82

Japan

4

ro

(Quart

US

16

of

n

Baker et al. [8] Uncertainty

4

-1.04 -1.15 -1.13 -0.15

The improvement of model fit based on estimating (4) with and without

-0.0

0.27

0.81

1.50

xt aligns with these results: adding

macro uncertainty adds between 2.4 - 4.6 percentage points to r-squared, financial uncertainty 1.5 - 3.3 and Baker et al. [8] uncertainty 1.0 - 2.4. See online appendix for full details on the marginal fit of these models when an uncertainty proxy is included.

Journal Pre-proof 2* Swede n

-0.56 -1.23 -1.00 -0.68 *

Switze

*

-0.0

*

-0.25 -0.54 -0.28 -0.12

lands

1.53

1.75

***

***

0.67

1.18 *

-0.4 1*

0.08

0.31

8

rland Nether

-0.46 -0.47 -0.07 -0.02

0.70

-0.69 -0.52 -0.20 -0.00 ***

-0.2

0.07

0.30

*

***

-0.83 -0.72 -1.17 -0.75 -1.51 -2.03

4

0.82

*

**

***

***

***

-1.5 3*

* significant at 10%, ** signifcant at 5%, *** significant at 1% for t-test based on Newey and West

ro

of

[54] standard errors

I have pursued a one-sided filter approach to extracting the stochastic volatility in place of

-p

the two-sided stochastic volatility smoother approach used in JLN. This leads to greater real time stability. I perform a pseudo real time test of the two approaches. I calculate the index up to 2005

re

and then re-estimate the model as I roll forward one quarter at a time using the latest data and

lP

collect the revisions that take place at each date for the next 8 quarters 11 . The results for the UK macro uncertainty index are shown in figure 5. The smoother based estimate is, unsurprisingly,

na

less volatile than the filtered estimate however the revisions are significantly larger: the probability that the initial estimate is revised by more than 0.25 standard deviations in the subsequent 8

ur

quarters is 21.2% for the smoothed estimate but only 4.1% for the filtered estimate (see figure 6). This revisions noise, which is completely unrelated to data revisions since the data is held constant

Jo

in this exercise, would entail significant news to a policy makers forecasting model were that model to incorporate uncertainty. The real time stability of the filtered estimates makes this measure more appropriate for policy use.

Figure 5: Pseudo-real time performance: smoother v. filter approach to stochastic volatility for macro uncertainty for the UK

Figure 6: Distribution of pseudo-real time revisions based on different stochastic volatility model 11

The implementation of the two-sided filter employs the Matlab code of Chan and Hsiao [26] rather than the

two-sided filter of the R STOCHVOL package for speed and ease of use in the Matlab environment. Testing showed that both two-sided filters resulted in near identical aggeregate uncertainty indices.

Journal Pre-proof Probability density function from a non-parametric Epanechnikov kernel density estimate of the pseudo real time revisions in the 8 quarters after the initial estimate.

5.

Macroeconomic Impact of Uncertainty Shocks

We identify the following structural VAR model following Antolín- Díaz and Rubio-Ramírez [2]. The model can be written as: p

A0 yt = Aj yt  j  c   t

(5)

of

j =1

Where yt is a nx1 vector of endogenous variables, p is the lag length, A j is an nxn

ro

matrix of parameters and c is a vector of nx1 parameters,  t is a nx1 vector of exogenous

-p

structural shocks which is Guassian conditional on past information with mean zero and covariance matrix given by the nxn identity matrix: E ( t  t) = I n . The matrix A0 captures the

p

re

impact of shocks to  t on yt , to recover this we must first estimate the reduced form model:

lP

yt =  B j yt  j  b  ut

(6)

j =1

na

Where B j = A01 Aj , ut = A01 t , b = A01c and E (ut ut) = u . In order to identify the structural model we need to impose restrictions on the matrix A0 . A common approach is to

ur

assume a recursive relationship between  t and yt by using A0 = chol (u ) , where chol ()

we define

Jo

refers to the Cholesky decomposition and A0 A0 = u . We instead employ sign restrictions where

A0 = QA0 where

A0 = chol (u ) and Q is a

nxn orthogonal matrix i.e.

QQ = QQ = I n . The sign restrictions are imposed sampling Q from a standard normal random matrix but only retaining draws which entail a A0 that correspond to our assumed traditional sign restrictions. To identify the structural shocks we follow Antolín-Díaz and Rubio-Ramírez [2]. Their approach to implementing zero and sign restriction follows Rubio-Ramírez et al. [64] complemented with the recent insights in Arias et al. [4]. Let B = [ B1 , B2 ,..., Bp , b] . We will impose sign restrictions on the IRFs. Let L = [ L0 , L1 ,..., L ] represent the stacked IRFs at each

Journal Pre-proof horizon12 . The algorithm proceeds as follows: 1. Draw ( B, u ) from the posterior distribution of the reduced form parameters. 2. Draw an orthogonal matrix Q from a QR decomposition of a nxn matrix whose columns are independent random draws from a N (0, I k ) 13 . 3. Let L0 be the Cholesky decomposition of u , then L0Q will be a draw from the posterior distribution of A01 before the sign restrictions are imposed. 4. Compute L with L0Q replacing L0 . Retain the draw if L accords with the sign

of

restrictions for all variables at all horizons.

ro

5. Repeat steps 2- 4 M times.

This reduced form VAR is estimated with Bayesian methods using a normal inverse

-p

Wishart Prior 14 . We estimate the above model for each country. The variables included in the

re

matrix yt are a measure of short term interest rates typically the policy rate, Consumer Price Index, hours or if unavailable employment, investment, consumption, GDP, credit spreads and a

lP

measure of uncertainty15 . The monthly uncertainty indices are averaged to get quarterly figures. All variables are de-trended using a cubic trend except for credit spreads, bank rate and the

information criteria.

na

uncertainty measures. The model includes 2 lags following the Schwartz and Hannan-Quinn

ur

We identify 2 shocks to the system using traditional sign restrictions on the response of variables to the shock - see table 3. In addition we also require that the identified series of shocks is

12

Jo

positive on particular dates. For macro uncertainty, the identified shocks must be positive in the The impulse response functions (IRFs) at horizon

 A1 A01  J  = [ I n ,0,...,0] and F =   Ap 1 A01  1  Ap A0 13

are defined as

Lh = A01 ( J F h J ) , where

0   In   0 

In 0 0

In the case of zero and sign restrictions the

h

Q matrix is drawn recursively so that each column of Q is

consistent with the zero restrictions as outlined by Arias et al. [4] 14

The Normal inverse Wishart prior assumes a normal prior for the VAR coefficients and a inverse Wishart prior

for the covariance matrix, see Blake and Mumtaz [14]. 15

See online appendix for full data description.

Journal Pre-proof months preceding the close election event where as for the financial shock, identified shocks must be positive on the dates of the largest change in the Romer and Romer [63] financial distress index, see tables 4 and 5. This final step of the algorithm simply computes the time series of the structural shocks identified above and keeps the draw if the shock has the correct sign on the required date 16 .

5.1. Results The majority of empirical studies of macro uncertainty employ recursive identification schemes (for example, Baker et al. [8], Leduc and Liu [48]). However, recursive ordering imposes a rigid

of

structure on the response of the VAR system requiring that the timing of each variable to a shock is

ro

known. This challenge is especially relevant when a ttempting to separate fast-moving financial and uncertainty shocks that, on average, may move together. I identify macro uncertainty shocks

-p

using small number of traditional sign restrictions and then sharpen this identification by imposing

re

additional narrative sign restrictions requiring that macro uncertainty shocks take place during close elections (see table 3and table 4). Similarly, I identify financial shocks requiring only that

lP

financial uncertainty and credit spreads rise and then require that the s hocks identified from these weak assumptions is positive across periods of peak financial distress as identified by Romer and

na

Romer [63]. The traditional sign restrictions impose only that investment falls following a macro uncertainty shock in line with a large number of empirical and theoretical results (see for example,

This procedure is computationally intensive since the narrative restrictions reduce the number of admissible

Jo

16

ur

Bloom [15], Basu and Bundick [10], Baker et al. [8], Fernandez-Villaverde et al. [30]) 17 . We apply

draws substantially. In the baseline case this ranges from around 7% to as low as 0.5%, depending on the country. We use 100 000 draws for each country to ensure that sufficient draws are retained for computing statistics of interest. Having a low nu mber of ad missible draws has been criticised by Fry and Pagan [33] as a sign that the restrictions are implausible. I share the view Kilian and Lütkepohl [45] that this is not evidence against a SVAR model’s restrictions. They highlight that the fraction of admissible models is indicative of how informative the identifying restrictions are and that a high number would suggest insufficiently identified mod els and thus be unattractive. Moreover, as they point out, in terms of fit with the data all the rotations fit the data equally well i.e. since each rotation remains a decomposition of the reduced form covariance matrix of the model. However, it remains th e case that sigificant computational resources can be required to use narrative restrictions especially if a large number of shocks are considered. 17

The response of inflation is less clear, theoretical models focusing on a precautionary demand channel ind icate

that inflation should fall (Leduc and Liu [48], Basu and Bundick[10]) but others find evidence that uncertainty can

Journal Pre-proof minimal use of traditional sign restrictions in order to put more weight on the narrative restrictions in identifying the shocks18 . It is worth noting that there is also evidence that the relationship between uncertainty and investment may be state dependent where idiosyncratic uncertainty present in firms investing in new technologies is associated with large potential productivity ga ins (Pastor and Veronesi [58]) or where the uncertainty may be around a switch of regime from normal times towards high growth (David and Veronesi [28]) - both are cases of good uncertainty. Segal et al. [65] provide empirical evidence that good uncertainty is associated with positive fixed investment in the US 19 .

of

An alternative mechanism by which uncertainty can lead to increased growth are the so called

ro

Oi-Hartmann-Abel effect: firms that have relatively mild adjustment costs can expand in response to positive outcomes and self- insure through contraction during bad times. However, as noted by

-p

Bloom [16], this is more likely to be important in the meduim or long run than in the short run when the cost to adjust output is higher. The measurement of uncertainty pursued here aggregates

re

uncertainty in a large panel of data and is thus less likely to reflect idiosyncratic uncertainty and we

lP

the find that the sample average correlation of investment growth and uncertainty is negative for all countries. However, 8- year rolling correlations do reveal significant time variation in the relationship where pre-crisis the correlation is weaker and turns positive just prior to the crisis in a

na

few countries, where this is more pronounced for financial uncertainty than macro uncertainty20 . Whilst good uncertainty is an interesting and understudied area of the literature, we proceed here

ur

with the more representative assumption, at least in our data, that macro uncertainty shocks are

Jo

negatively related to investment. The traditio nal sign restricitons imposed in table (3) may thus put greater evidence on the post-crisis period and may ignore aggregate implicaitons of the create an upward pricing bias in firms price setting decision (Fernandez-Villaverde et al. [30]) similarly there is empirical evidence that this can go either way (for inflationary see Popescu and Smets [60], Redl [61, 62]; for dis-inflationary see Leduc and Liu [48], Basu and Bundick [10]). Hence we remain agnostic on the response of inflation. 18

Robustness checks below impose stronger traditional sign restrictions that ensure separation of the shocks

regardless of the ability of the narrative estimates to separate uncertainty and financial shocks. 19

A related strand of the finance literature has provided theoretical and empirical evidence that the value of equity

can be positively related to volatility of earnings, see Abel and Eberly [1], Pastor and Veronesi [56] and Pastor and Veronesi [57]. 20

See appendix for details.

Journal Pre-proof mechanisms outlined by Pastor and Veronesi [58], David and Veronesi [28] or Oi-Hartmann-Abel effects.

Table 3: Sign restrictions on response of variables to shocks Variable / Shock

Macro

Financial

Uncertainty Short term interest rate CPI

-

ro

Investment

of

Hours or Employment

Consumption

-p

GDP

Macro Uncertainty

+

lP

Financial Uncertainty

+

+

re

Credit Spreads

na

No entry represents an unrestricted response.

In addition to the above standard sign restrictions we impose narrative sign restrictions on

ur

the macro uncertainty shocks which require a positive shock takes place around close general elections, following the framework Antolín-Díaz and Rubio-Ramírez [2]. Table 4 outlines which

Jo

general elections we have selected as close and presents some ex-post evidence that these were close elections. This includes the results of the election in terms of popular vote which would represent a broad measure of the voter disagreement in the country. However, what matters for the ability of politicians to affect the business environment is the split in the legislature, this is provided in the percentage of seats. On average, these metrics are both very close for the selected elections. A new ruling party may bring more potential changes in economic policies and thus more uncertainty. This takes place in about half the elections here (16/29). Further narrative evidence around these events is outlined in the appendix. A key source of ex-ante macro uncertainty around close elections is the difference in the policy plans of the leading parties. To measure this we construct an economic policy analogue to the RILE measure of left-right sentiment in party manifestos used widely in political science

Journal Pre-proof (Budge et al. [18]). This measure uses the Manifesto Project Database (Volkens et al. [67]) which employs human coders to assign codes to each sentence (or part sentence) in each manifesto which express a positive or negative sentiment in a variety of categories: external relations, freedom and democracy, the political system, the economy, etc. The database then expresses these coded sentences as a proportion of all coded sentences in the manifesto. For example, in the US presidential election of 2008 (with Barrack Obama as candidate), in the Democratic party manifesto 2.91% of all coded sentences expressed support for market regulation (a subsection of the economy section of the codes) whereas for the Republican party this was 1.19%. The original

of

right- left position or RILE measure adds code scores relating to left leaning sentences and

ro

subtracts the right leaning ones. This is done for a selection of codes across all topics in the database. For our purposes I focus on the economy topic to measure left-right position in terms of

-p

market policies. I add all the codes that express support for free market policies and subtract all the codes expressing support for greater intervention in the free market, within the economy

re

modules 21 . I label this EconRILE. Thus a positive value suggests the party promotes policies that

lP

are pro- free market and a negative value indicates greater focus on market intervention. I present the gap between the two leading parties EconRILE measures as an indicator of the difference in their planned policies, as the more pro- free market party less the more interventionist or socialist

na

party. The greater this gap (in absolute value) the larger the disagreement in policy a nd the more plausible it is that a close election should cause greater macroeconomic uncertainty. If there was

ur

little disagreement this value would be close to zero, however it is typical for their to be significant

Jo

differences between parties based on this gap. Following the above above approach to identifying macro uncertainty shocks I impose minimal sign restrictions on the financial shock - only that financial uncertainty and credit spreads are positive after the shocks hits. Additional narrative sign restrictions are used based on the financial distress index of Romer and Romer [63]. The latter produce a real time financial distress index based on the authors reading of semi-annual OECD Economic Outlook publications. I use the dates of peak distress as episodes where financial shocks are of primary importance in the business cycle. It is unclear which dates are most relevant for this exercise, where the index reaches its peak level may indicate that financial shocks hit some time ago, whereas peak changes in the index may do a better job of indicating when those shocks hit but may also ignore times 21

More details are provided in the online appendix

Journal Pre-proof when the level has been slowly building. To balance a weight on both the level and the change I use the interaction of the change and the level (product of the change and the level of the index) to choose the dates to impose narrative restrictions.

Table 4: Close Election Events Country

Elections

Narrative

Winner

sign date

restriction

(runner-up)

22/9/2002

Jo

Germany

18/9/2005 Italy †

21/4/1996

9/4/2006

24/2/2013

UK

9/4/1992

Q3

Q1

Q1

Q1

Q1

Party

(average

47.9

50.4

53.2

(Kerry)

(48.3)

(46.7)

Trump

46.1

56.5

(Clinton)

(48.2)

(42.2)

Mori

28.3

48.5

(Hatoyama)

(25.2)

(26.5)

Koizumi

35

49.4

(Kan)

(37.4)

(36.9)

Schröder

38.5

41.6

(Stoiber)

(38.5)

(41.1)

Merkel

35.2

36.8

(Schröder)

(34.2)

(36.2)

Prodi

42.6

52.0

(Berlusconi)

(40.3)

(38.3)

Prodi

49.4

53.5

(Berlusconi)

(50.0)

(46.2)

Bersani

30.6

49.5

(Berlusconi)

(30.0)

(25.6)

Major

41.9

51.6

re

lP

Q4

Q3

Gap 

50.7

Bush

ur

9/11/2003

Q2

Ruling

(49.4)

Q3

25/6/2000



(48.4)

na

Japan

Bush (Gore)

Q3

8/11/2016

Popular

-p

2/112004

EconRILE

of

Q3

New

ro

7/11/2000

% Seats

vote

date US

%

gap) Yes

4.8 (6.3)

No

9.5 (6.3)

Yes

-

No

3.1 (5.5)

No

12.1 (5.5)

No

3.9 (4.1)

Yes

6.4 (4.1)

Yes

-

Yes

15.1 (9.0)

Yes

19.2 (10.8)

No

13.9 (9.8)

Journal Pre-proof

7/5/2015

France *

Q1

7/5/1995

(41.6)

Cameron

36.1

47.1

(Brown)

(29.0)

(39.7)

Cameron

36.8

50.8

(Miliband)

(30.4)

(35.7)

Chirac

52.6

Pres.

(Jospin)

(47.4)

election

Sarkozy

53.1

Pres.

(Royal)

(46.9)

election

Hollande

51.6

Pres.

Q2

6/5/2007

Q2

6/5/2012

Q2

(Sarkozy)

17/9/2006

19/9/2010

Switzerland

19/10/2003

23/10/2011

(29.6)

(32.1)

Harper

36.3

40.3

(Martin)

(30.2)

(33.4)

Aznar

38.8

48.0

(González)

(37.6)

(39.8)

Zapatero

42.6

34.2

(Rajoy)

(37.7)

(44.5)

Zapatero

43.9

46.1

(Rajoy)

(39.9)

(45.7)

Rajoy

28.7

44.3

(Sánchez)

(22.0)

(24.6)

Reinfeldt

26.3

27.8

(Persson) 

(35.0)

(37.2)

Sahlin

30.7

32.1

(Reinfeldt)

(30.1)

(30.7)

Maurer

26.7

26.0

(C.Brunner)

(23.3)

(25.1)

T.Brunner

26.6

24.0

re

Q1

ur

Jo

Sweden

(Harper)

na

Q1

9/3/2008

20/12/2015

54.5

Martin

Q1

3/3/1996

14/3/2004

36.7

Q2

23/1/2006 Spain ‡

election

-p

28/6/2004

(48.4)

lP

Canada

Q1

Q4

Q3

Q3

Q4

Q4

of

Q1

(34.4)

ro

6/5/2010

(Kinnock)

Yes

0.4 (9.8)

No

6.4 (9.8)

Yes

-

Yes

7.6 (-)

Yes

12.1 (-)

No

-4.4 (0.41)

Yes

3.8 (0.41)

Yes

1.9 (5.5)

Yes

6.5 (5.5)

No

5.2 (5.5)

No

6.0 (5.5)

No

9.6 (16.2)

No

3.1 (16.2)

Yes

16.1 (14.3)

No

24.8

Journal Pre-proof

Netherlands

15/5/2002

Q4

9/6/2010

(Levrat)

(18.7)

(23.2)

Balkenende

27.9

28.7

(Fortuyn)

(17.0)

(17.3)

Rutte

20.5

20.7

(Cohen)

(19.6)

(20.0)

Rutte

26.6

27.3

(Samsom)

(24.8)

(25.3)

Q2

12/9/2012

Q3

(14.3) Yes

1.1 (4.6)

Yes

11.3 (4.6)

No

14.8 (4.6)

 All elections require approximately 50% of seats to form a government. * Second round

of

run-off. † Italy popular votes data taken as an average of popular vote from the chamber of

ro

Deputies and the Italian Senate . ‡ Spain data used for both congress of deputies (350 seats) and the Senate (266 seats however only 208 seats were up for election).

Switzerland data used for

-p

both National Council (200 seats) and council of States (46 seats).  Reinfeldt received fewer

re

votes but lead government by forming a coalition with smaller parties.  Author calculations using Manifesto Project Database, gap defined as more free market party less more socialist party. The

lP

average gap is the EconRILE gap between leading parties in postwar data (where available).

na

The index is semi-annual so I impose that the identified financial shock must be positive on both quarters corresponding to those dates - see table 5. These dates tend to emphasize the

ur

financial crisis, the exceptions being Japan’s experience of the Asian crisis in 1998H1, the Netherlands exposure to the Euro crisis in 2012H1 and Sweden’s housing crash in 1992H2.

Jo

The results for the macro uncertainty shock are summarised in figure 7 using the pooled mean group estimator of Pesaran et al. [59], which is simply an average of the impulse response functions for each country. The macro uncertainty shock is normalised to be the same size under both identification schemes (1 standard deviation). Two key results are present: (1) macro uncertainty shocks have significant effects on the cyclical component of GDP and (2) conditioning on electoral uncertainty implies larger real effects of uncertainty shocks. The average peak response is around -1% for GDP (quarterly annualised) where the effects tend to be expressed via a larger drop in investment and employment or hours as britishemphasised by Baker et al. [8]. There is some evidence of a greater response of consumption, concentrated in Japan, Italy, the UK and France. There is also a rise in financial uncertainty and a moderate ris e in credit spreads, although neither is significant at the aggregate level. The individual country responses for

Journal Pre-proof de-trended GDP show larger responses conditioning on close elections for the US, Japan, Italy, the UK, France and the Netherlands - figure 7.

(2) Max

(3) Interaction: Change (1)

Level

Change

* Level (2)

US

2007H2

2008H2

2008H2

Japan

1991H2

1998H2

1998H1

Germany

1974H2

2008H2

2008H2

Italy

2008H1

2012H2

UK

2007H2

2008H2

France

2008H2

2008H2

Canada

2007H2

2008H2

2007H2

Spain

2008H1

2012H2

2008H1

Sweden

1992H2

1993H1

1992H2

Switzerland

2007H2

2008H1

2008H1

Netherlands

2012H1

2009H1

2012H1

-p

ro

2008H2 2007H2 2008H2

na

lP

Country

of

(1) Max

re

Table 5: Romer and Romer [63] Financial Distress Index Peak Dates

Where H refers to a semi-annual period e.g. 2008H2 is 2008Q1 and 2008Q2.The dates above are

ur

the peak readings of the Romer & Romer financial conditions semi- annual indices for the (1) level of the index, (2) change in the level and the (3) interaction of change and level. The narrative

Jo

restrictions to identify a financial shock use the interaction (3) dates for both quarters of the period above.

Figure 7: Macro Uncertainty Shock: Impulse Response Functions of Mean Group Estimates for all Countries Responses in blue (with 68% credible set in Grey) are results with baseline sign restrictions given in table (3). Responses in red show the effect of adding narrative information contained in tables (4) and (5). All IRFs are mean group estimates across all countries. Macro uncertainty shock normalised to 1 standard deviation under both identification schemes.

I have argued above that tight elections would plausibly lead to higher macro uncertainty in

Journal Pre-proof the months leading to an election. However, it is equally plausible that the major change in uncertainty may be driven by the resolution of uncertainty shortly after the election is completed. This assumption is compared to the baseline in figure (7), where the narrative restriction requires that the identified macro uncertainty shock is negative in the quarte r after the election has taken place. The results are similar on aggregate, however, under this assumption financial uncertainty is less elevated compared to the baseline case 22 . See figure 8 for some of the country level variaiton this alternative assumption implies.

of

I investigate the the transmission of macro uncertainty shocks identified with close

ro

elections with special attention on the financial channel in figure 9. This shows the countries where the impact of conditioning on close elections is associated with a larger response of GDP. There is

-p

some evidence that credit spreads rise in the UK and France but are not particularly elevated for the other countries. Credit spreads do not appear to be driving the larger response on aggregate, in

re

line with the anecdotal evidence in figure 1. However, financial uncertainty does play an important

lP

part in the transmission mechanism with significant rises in all countries except Italy and France. I can shut down the role of financial uncertainty in the transmission mechanism by imposing a traditional sign restriction that identifies macro uncertainty shocks where there is no rise in

na

financial uncertainty in the first two quarters - this is shown with the dotted black lines. Making this strict assumption leads to smaller response of GDP to the macro uncertainty shock for Japan,

ur

the Netherlands and the UK. Thus, financial uncertainty is an important part of the transmission of

Jo

uncertainty shocks that are associated with close elections.

Figure 8: GDP Response to Macro Uncertainty Shock

Responses in blue (with 68% credible set in Grey) are results with baseline sign restrictions given in table (3). Responses in red show the effect of adding narrative information contained in tables

22

An additional robustness check on the impact of the timing restriction on investment is computed in the

appendix. This assumes that investment falls only 4 quarters after the uncertainty shock hits - the results are very similar to the baseline assumption that investment falls contemporaneously. In experiments where I assume that uncertainty falls in the quarter prior to a close election, the effect of the IRFs is roughly symmetrical where the reponses are smaller by approximately the same amount they are larger when assuming that uncertainty rises prior to close elections.

Journal Pre-proof (4) and (5). Macro uncertainty shock normalised to 1 standard deviation under both identification schemes.

Figure 9: GDP Response to Macro Uncertainty Shock & the Financial Channel Responses in blue (with 68% credible set in Grey) are results with baseline sign restrictions given in table (3). Responses in red show the effect of adding narrative information contained in tables (4) and (5). Black dashed line illustrates the median response when traditional sign restrictions require that financial uncertainty falls following a macro uncertainty shock. Macro uncertainty

ro

of

shock normalised to 1 standard deviation under both identification schemes.

Historical decompositions illustrating the contribution of macro uncertainty and financial

-p

shocks to de-trended GDP based on the baseline model are provided in figure 10. As in the above analysis, there is a heterogeneous impact across countries with macro uncertainty shocks being

re

less important in Sweden, Switzerland, Canada and the US. Macro uncertainty shocks played a

lP

relatively minor role in the US business cycle with the exception of the period around the 2016 election. In a number of countries macro uncertainty has been an important driver of both rises and declines in GDP. This is most pronounced in the expansion prior to the global financial crisis of

na

2008 when macro uncertainty tended to be unusually low, especially in the UK, Japan, France and the Netherlands. In countries where there has been a protracted weakness in GDP following the

ur

crisis, macro uncertainty shocks have been an important drag on growth, in particular, the UK,

Jo

France and the Netherlands. These countries may have suffered from repeat uncertainty shocks relating to the Euro crisis around 2011/12.

Figure 10: Historical Decomposition for detrended GDP Figure shows the historical decomposition for model identified with standard sign restriction in table 6 and narrative information relating to macro uncertainty and financial shocks.

5.2. Robustness Table 6: Sign restrictions on response of variables to shocks

Journal Pre-proof Variable / Shock

Macro Uncertainty

Financial

News Shock

Global Uncertainty

0, + (Q4-Q6) *

-

Short term interest rate CPI Hours or Employment Investment

-

Consumption 0, + (Q4-Q6) * +

Macro Uncertainty

+ +

-p

Financial

ro

Credit Spreads

of

GDP

0

forecasts Global Uncertainty

0

lP

CE GDP growth

re

Uncertainty

+

+

na

No entry represents an unrestricted response. * Zero response on impact but a positive response for

ur

2 quarters a year after the shock hits.

I check the robustness of the above results by adding two additional shocks to the 2 shock

Jo

system presented in table 3:23 news shocks and global uncertainty shocks. A significant challenge to using uncertainty indices in policy is that positive uncertainty shocks (second moment) are typically correlated with negative confidence shocks (first moment), as highlighted by Haddow et al. [36]. The news shock will capture changes in confidence that may be incorrectly attributed to uncertainty shocks. Moreover, a number of studies have highlighted that uncertainty shocks are closely related to news shocks (Berger et al. [11], Cascaldi- Garcia and Galvao [24]) and thus there is a risk of conflating the two without separately identifying news in the model. News shocks are

23

Co mputational constraints make it difficult to include a larger set of shocks as in table 6 alongside the news and

global uncertainty shocks since as this must be calculated for 11 countries with 2 sets of narrative sign restrictions as well as the dynamic sign restrictions used for the news shock as this requires a very large number of draws.

Journal Pre-proof identified as shocks that increase Consensus Economics 1- year ahead GDP growth forecasts on impact, have no contemporaneous effect on GDP and investment but see a rise in those variables for two quarters a year after the shock hits. These assumptions are in line with the theoretical and empirical results in Barsky and Sims [9]. To control for a deterioration in mean expectations driving the results of the domestic uncertainty shock I impose that the identified domestic macro uncertainty shock does not cause a drop in growth expectations on impact. It may be that the effects of domestic macro uncertainty shocks are not due to domestic developments but rather through correlation with global uncertainty shocks as highlighted by

of

Cesa-Bianchi et al. [25], Mumtaz and Theodoridis [52] and Berger et al. [12]. I employ a measure

ro

of global uncertainty developed in Redl [62] to test this hypothesis 24 . That measure of global uncertainty applies the JLN methodology to a wide set of global macro and financial variables.

-p

The index uses global macro and financial data covering stock market returns, sovereign bonds yields, exchange rates, commodity prices, trade volumes, retail sales, consumer and business

re

confidence from emerging and advanced economies. I allow the global uncertainty shock to

lP

compete with the domestic macro uncertainty shock by identifying it in a similar way, assuming that it also drags on investment but imposing zero correlation between the shocks on impact. While this strong assumption will not in general be true, domestic macro uncertainty shocks may often

na

reflect global shocks, however here I am interested in identifying only those shocks in sample that were not (contemporaneously) driven by global shocks.

ur

Including news and global uncertainty shocks in the model is broadly consistent with the

Jo

main results, see figure 11. Growth expectations tend to deteriorate in Italy and France following a macro uncertainty shock and global uncertainty tends to be higher for all countries in this sub-sample see figure 12. These amplification channels are particularly important in the UK and Japan, where the response of GDP is less pronounced, although it remains statistically signifcant. Additional robustness checks replace the replace the JLN based measures of macro uncertainty with the news based measure produced by Baker et al. [8], labeled BBD in figure 11. The transmission mechanism is broadly in line with the 2 shock model presented above, with significant declines in inputs to production (investment and hours) resulting in lower GDP.

24

That paper used global variables excluding the UK as it focused exclusively on the UK. In constructing t his

global index I use all global data including the UK.

Journal Pre-proof Figure 11: Robustness: Mean Group Estimates for Different Models Responses in blue (with 68% credible set in Grey) . Responses in red show the effect of adding narrative information. All IRFs are mean group estimates across all countries.

Figure 12: GDP Response to Macro Uncertainty Shock: role of GDP growth News and Global Uncertainty Responses in red (with 68% credible set) are results with sign restrictions given in table (6) and narrative restrictions described in tables (4) and (5). Black dashed line shows the median impulse

of

response function under the baseline 2 shock model. Macro uncertainty shocks normalised to 1

ro

standard deviation under both identification schemes.

-p

Variance Decompositions

Macro

Financia

Shock

Uncertaint

l

interest rate

Employment Investment

Jo

CPI

Monetar Cost y Policy

Supply

ur

Short term

Hours or

r

na

y

Labou

lP

Variable /

re

Table 7: Sign restrictions on response of variables to shocks

Pus

Preferenc

Generi

Generi

e

c

c

Supply

Deman

h

d +

-

-

+

+

-

-

+

-

+

+

+

+

-

Consumptio n GDP

-

Credit

+

Spreads Macro Uncertainty

+

Journal Pre-proof Financial

+

Uncertainty No entry represents an unrestricted response.

In order to asses the role of uncertainty shocks in driving the business cycle I augment the 2 shock VAR system described above to include 6 additional shocks and compute variance decompositions using the mean group IRFs for all countries. The additional s hocks are identified using the sign restrictions in table 7. The macro uncertainty and financial shocks are identified as

of

before. The aggregate IRFs are shown in figure 11.

ro

The variance decompositions are computed with and without the narrative sign restrictions in order to asses how this information alters the role of macro uncertainty and financial shocks in

-p

the model - see figure 13. Using only traditional sign restrictions, I see that preference and supply shocks are the dominant explanations for moveme nts in de-trended GDP while labour supply

re

shocks are most important for explaining the variation in hours. Macro uncertainty shocks play a

lP

moderate role in GDP fluctuations in the medium term but financial shocks play a relatively small role overall. The addition of narrative information raises the importance of both macro uncertainty

na

and financial shocks - right panel of figure 13. At the horizon of 2-3 years, macro uncertainty shocks are now on par with supply shocks in explaining GDP variations and their role is around

ur

twice as important for hours: rising from around 10% to 20% and the most important shock aside from labour supply shocks. Financial shocks are more important for both GDP and hours variation

Jo

when narrative information is included but are less relevant than macro uncertainty shocks25 .

Figure 13: Variance Decomposition for GDP and Hours

Figure shows the forecast error variance decomposition for model identified with standard sign restriction in table 7 without narrative information (LHS) and with narrative information (RHS) relating to macro uncertainty and financial shocks.

6.

Conclusion 25

The full variance decompositions for each shock are available in the appendix and show that financial shocks

are important for explaining movements in financial uncertainty and credit spreads.

Journal Pre-proof This paper uses a data rich environment to produce new econometric measures of macroeconomic and financial uncertainty for 11 advanced nations. These new mac ro uncertainty measures show significant independent variation from other popular proxies such as those of Baker et al. [8], with improved readiness for policy use through better real- time performance and a construction appropriate for an input for forecasting. Moreover, these new measures of financial uncertainty go beyond narrow measures of share price or interest rate implied volatility to capture credit extension and the external environment. I apply these measures to study the impact of macro uncertainty shocks controlling for

of

financial shocks identified using narrative information on financial crises. I find that real macro

ro

uncertainty shocks matter for the majority of countries. I further isolate the macro uncertainty channel by employing narrative information from closely contested elections. I find that this

-p

induces a larger real effect of macro uncertainty shocks. I find that the transmission mechanism of these shocks does not rely on a rise in credit spreads, however there is evidence that increases in

re

financial and global uncertainty can be an important part of the transmission mechanism for some

lP

countries. These results are robust to controlling for news shocks as well as a variety of shocks considered to be important drivers of the business cycle. I find that identifying macro uncertainty shocks with close elections raises their importance

na

as a source of fluctuations in GDP especially at the horizon of 1-2 years where they explain almost 20% of the forecast error variance of GDP. Interestingly, I find that these shocks are more

ur

important for real variables than financial shocks, which explain only around 10% of the variation

Jo

in GDP. Similarly, identifying macro uncertainty shocks with close elections approximately doubles their importance in explaining labour market fluctuations (hours) from around 10% to slightly more than 20%.

Future research could employ the JLN approach to estimate macro economic uncertainty in developing countries where digital newspaper archives are rare and the narrative signs approach used to isolate macro uncertainty shocks using elections or other political events.

References [1]

Abel, A. B. and Eberly, J. C. (1994). A unified model of investment under uncertainty. The American Economic Review, 84(5):1369–1384.

Journal Pre-proof [2]

Antolín-Díaz, J. and Rubio-Ramírez, J. F. (2018). Narrative sign restrictions for svars. American Economic Review, 108(10):2802–29.

[3]

Arbatli, E. C., Davis, S. J., Ito, A., Miake, N., and Saito, I. (2017). Policy uncertainty in japan. Working Paper 23411, National Bureau of Economic Research.

[4]

Arias, J. E., Rubio-Ramírez, J. F., and Waggoner, D. F. (2018). Inference based on structural vector autoregressions identified with sign and zero restrictions: Theory and applications. Econometrica, 86(2):685–720.

[5]

Armelius, H., Hull, I., and Stenbacka Kohler, H. (2017). The timing of uncertainty shocks

Azzimonti, M. (2017). Partisan conflict, news and investors’ expectations. Technical

ro

[6]

of

in a small open economy. Economics Letters, 155(C):31–34.

report.

Bai, J. and Ng, S. (2002). Determining the number of factors in approximate factor models.

-p

[7]

Econometrica, 70(1):191–221.

Baker, S. R., Bloom, N., and Davis, S. J. (2016). Measuring economic policy uncertainty*.

re

[8]

[9]

lP

The Quarterly Journal of Economics, 131(4):1593. Barsky, R. B. and Sims, E. R. (2011). News shocks and business cycles. Journal of Monetary Economics, 58(3):273–289.

Basu, S. and Bundick, B. (2017). Uncertainty shocks in a model of effective demand.

na

[10]

Econometrica, 85(3):937–958.

Berger, D., Dew-Becker, I., and Giglio, S. (2017). Uncertainty shocks as second-moment

ur

[11]

[12]

Jo

news shocks. Working Paper 23796, National Bureau of Economic Research. Berger, T., Grabert, S., and Kempa, B. (2016). Global and country-specific output growth uncertainty and macroeconomic performance. Oxford Bulletin of Economics and Statistics, 78(5):694–716. [13]

Bialkowski, J., Gottschalk, K., and Wisniewski, T. P. (2008). Stock market volatility around national elections. Journal of Banking Finance, 32(9):1941–1953.

[14]

Blake, A. P. and Mumtaz, H. (2012). Applied Bayesian econometrics for central bankers. Number 4 in Technical Books. Centre for Central Banking Studies, Bank of England.

[15]

Bloom, N. (2009). The impact of uncertainty shocks. Econometrica, 77(3):623–685.

[16]

Bloom, N. (2014). Fluctuations in Uncertainty. Journal of Economic Perspectives, 28(2):153–176.

Journal Pre-proof [17]

Born, B., Breuer, S., and Elstner, S. (2018). Uncertainty and the great recession. Oxford Bulletin of Economics and Statistics, 80(5):951–971.

[18]

Budge, I., Klingemann, H.-D., Volkens, A., Bara, J., and Tanenbaum, E. (2001). Mapping Policy Preferences. Estimates for Parties, Electors, and Governments 1945-1998. Oxford University Press, Oxford.

[19]

Caggiano, G., Castelnuovo, E., and Groshenny, N. (2014). Uncertainty shocks and unemployment dynamics in u.s. recessions. Journal of Monetary Economics, 67(Supplement C):78 – 92. Caldara, D., Fuentes-Albero, C., Gilchrist, S., and Zakrajsek, E. (2016). The

of

[20]

88:185 – 207. SI: The Post-Crisis Slump.

Caldara, D. and Iacoviello, M. (2018). Measuring Geopolitical Risk. International Finance

-p

[21]

ro

macroeconomic impact of financial and uncertainty shocks. European Economic Review,

Discussion Papers 1222, Board of Governors of the Federal Reserve System (U.S.). Canes-Wrone, B. and Park, J.-K. (2014). Elections, uncertainty and irreversible

re

[22]

[23]

lP

investment. British Journal of Political Science, 44(1):83 to 106. Carriere-Swallow, Y. and Cespedes, L. F. (2013). The impact of uncertainty shocks in emerging economies. Journal of International Economics, 90(2):316 – 325. Cascaldi-Garcia, D. and Galvao, A. B. (2018). News and Uncertainty Shocks. International

na

[24]

Finance Discussion Papers 1240, Board of Governors of the Federal Reserve System

Cesa-Bianchi, A., Pesaran, M. H., and Rebucci, A. (2014). Uncertainty and Economic

Jo

[25]

ur

(U.S.).

Activity: A Global Perspective. CESifo Working Paper Series 4736, CESifo Group Munich. [26]

Chan, J. C. and Hsiao, C. Y. (2014). Estimation of Stochastic Volatility Models with Heavy Tails and Serial Dependence, chapter 6, pages 155–176. John Wiley & Sons, Ltd.

[27]

Croce, M., Colacito, R., Liu, Y., and Shaliastovich, I. (2017). Volatility risk pass-through. Technical report.

[28]

David, A. and Veronesi, P. (2013). What ties return volatilities to price valuations and fundamentals? Journal of Political Economy, 121(4):682–746.

[29]

Dendy, C., Mumtaz, H., and Silver, L. (2013). An uncertainty index for the uk 1986-2012. mimeo.

Journal Pre-proof [30]

Fernandez-Villaverde, J., Guerron-Quintana, P., Kuester, K., and Rubio-Ramírez, J. (2015). Fiscal volatility shocks and economic activity. American Economic Review, 105(11):3352–84.

[31]

Fernandez-Villaverde, J., Guerron-Quintana, P., Rubio-Ramirez, J. F., and Uribe, M. (2011). Risk matters: The real effects of volatility shocks. American Economic Review, 101(6):2530–61.

[32]

Forbes, K. (2016). Uncertainty about uncertainty. Technical report, Speech to be given by Kristin Forbes, External MPC Member, at J.P. Morgan Cazenove Best of British

Fry, R. and Pagan, A. (2011). Sign restrictions in structural vector autoregressions: A

ro

[33]

of

Conference, London.

critical review. Journal of Economic Literature, 49(4):938–60. Gao, P. and Qi, Y. (2013). Political uncertainty and public financing costs: Evidence from

-p

[34]

u.s. gubernatorial elections and municipal bond market. Technical report. Goodell, J. W. and Vähämaa, S. (2013). Us presidential elections and implied volatility:

re

[35]

[36]

lP

The role of political uncertainty. Journal of Banking Finance, 37(3):1108 – 1117. Haddow, A., Hare, C., Hooley, J., and Shakir, T. (2013). Macroeconomic uncertainty: what is it, how can we measure it and why does it matter? Bank of England Quarterly

[37]

na

Bulletin, 53(2):100–109.

Husted, L. F., Rogers, J. H., and Sun, B. (2017). Monetary Policy Uncertainty.

[38]

Jo

System (U.S.).

ur

International Finance Discussion Papers 1215, Board of Governors of the Federal Reserve

Jacquier, E., Polson, N. G., and Rossi, P. E. (1994). Bayesian analysis of stochastic volatility models. Journal of Business & Economic Statistics, 12(4):371–389.

[39]

Jens, C. E. (2017). Political uncertainty and investment: Causal evidence from u.s. gubernatorial elections. Journal of Financial Economics, 124(3):563 – 579.

[40]

Julio, B. and Yook, Y. (2012). Political uncertainty and corporate investment cycles. The Journal of Finance, 67(1):45–83.

[41]

Julio, B. and Yook, Y. (2016). Policy uncertainty, irreversibility, and cross-border flows of capital. Journal of International Economics, 103(Supplement C):13 – 26.

[42]

Jurado, K., Ludvigson, S. C., and Ng, S. (2015). Measuring uncertainty. American Economic Review, 105(3):1177–1216.

Journal Pre-proof [43]

Kastner, G. and Frühwirth-Schnatter, S. (2014). Ancillarity-sufficiency interweaving strategy (asis) for boosting mcmc estimation of stochastic volatility models. Computational Statistics & Data Analysis, 76:408 – 423. CFEnetwork: The Annals of Computational and Financial Econometrics.

[44]

Kelly, B., Pastor, L., and Veronesi, P. (2016). The price of political uncertainty: Theory and evidence from the option market. The Journal of Finance, 71(5):2417–2480.

[45]

Kilian, L. and Lütkepohl, H. (2017). Structural Vector Autoregressive Analysis. Themes in Modern Econometrics. Cambridge University Press. Kok, S., Kroese, L., and Parlevliet, J. (2015). Beleidsonzekerheid in Nederland.

of

[46]

ro

Economisch Statische Berichtenl, 100.

Larsen, V. H. (2017). Components of uncertainty. Working Paper 2017/5, Norges Bank.

[48]

Leduc, S. and Liu, Z. (2012). Uncertainty shocks are aggregate demand shocks. Working

-p

[47]

Paper Series 2012-10, Federal Reserve Bank of San Francisco. Li, J. and Born, J. A. (2006). Presidential election uncertainty and common stock returns in

re

[49]

[50]

lP

the united states. Journal of Financial Research, 29(4):609–622. Ludvigson, S. C., Ma, S., and Ng, S. (2018). Uncertainty and business cycles: Exogenous impulse or endogenous response? Mimeo. Miniaci, R. and Weber, G. (1999). The Italian recession of 1993: Aggregate implications

na

[51]

of microeconomic evidence. The Review of Economics and Statistics, 81(2):237–249. Mumtaz, H. and Theodoridis, K. (2015). The international transmission of volatility

ur

[52]

Jo

shocks: An empirical analysis. Journal of the European Economic Association, 13(3):512–533. [53]

Mumtaz, H. and Theodoridis, K. (2017). Common and country specific economic uncertainty. Journal of International Economics, 105:205 – 216.

[54]

Newey, W. K. and West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica, 55(3):703–708.

[55]

Ozturk, E. O. and Sheng, X. S. (2017). Measuring global and country-specific uncertainty. Journal of International Money and Finance.

[56]

Pastor, L. and Veronesi, P. (2003). Stock valuation and learning about profitability. The Journal of Finance, 58(5):1749–1789.

[57]

Pastor, L. and Veronesi, P. (2005). Rational ipo waves. The Journal of Finance,

Journal Pre-proof 60(4):1713–1757. [58]

Pastor, L. and Veronesi, P. (2009). Technological revolutions and stock prices. The American Economic Review, 99(4):1451–1483.

[59]

Pesaran, M. H., Shin, Y., and Smith, R. P. (1999). Pooled mean group estimation of dynamic heterogeneous panels. Journal of the American Statistical Association, 94(446):621–634.

[60]

Popescu, A. and Smets, F. R. (2010). Uncertainty, Risk-taking, and the Business Cycle in Germany. CESifo Economic Studies, 56(4):596–626. Redl, C. (2015). Macroeconomic uncertainty in south africa. Technical report, Economic

[62]

Redl, C. (2017). The impact of uncertainty shocks in the United Kingdom. Bank of

-p

England working papers 695, Bank of England. [63]

ro

Research Southern Africa Working Paper Series.

of

[61]

Romer, C. D. and Romer, D. H. (2017). New evidence on the aftermath of financial crises

Rubio-Ramírez, J. F., Waggoner, D. F., and Zha, T. (2010). Structural Vector

lP

[64]

re

in advanced countries. American Economic Review, 107(10):3072–3118.

Autoregressions: Theory of Identification and Algorithms for Inference. The Review of Economic Studies, 77(2):665–696.

Segal, G., Shaliastovich, I., and Yaron, A. (2015). Good and bad uncertainty:

na

[65]

Macroeconomic and financial market implications. Journal of Financial Economics,

Stock, J. H. and Watson, M. W. (2012). Disentangling the channels of the 2007-2009

Jo

[66]

ur

117(2):369 – 397.

recession. Working Paper 18094, National Bureau of Economic Research. [67]

Volkens, A., Lehmann, P., Matthiess, T., Merz, N., Regel, S., and Wessels, B. (2017). The manifesto data collection. manifesto project (mrg/cmp/marpor). version 2017a.

[68]

Zalla, R. (2017). Economic policy uncertainty in Ireland. Atlantic Economic Journal, 45.

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13