## Abstract

We investigate empirically how fiscal shocks—unanticipated and exogenous changes of government consumption growth—affect the sovereign default premium. For this purpose, we assemble a new data set for 38 emerging and developed economies. It contains approximately 3,000 observations for the sovereign default premium and three alternative measures of fiscal shocks. We condition our estimates on whether shocks are positive or negative and initial conditions in terms of fiscal stress. An increase of government consumption barely affects the default premium. A reduction raises the premium if fiscal stress is severe but decreases it if initial conditions are benign.

## I. Introduction

IN the years following the global financial crisis, many European governments implemented sizable austerity measures. These included spending cuts and tax increases and were meant to confront concerns about rising levels of public debt or outright solvency issues. In fact, yields on debt issued by several European sovereigns started to take off by 2010, reflecting sizable default premiums (Krishnamurthy, Nagel, & Vissing-Jorgensen, 2018). Still, as austerity measures were implemented, default premiums kept rising. Also, along the cross section of countries, large spending cuts were associated with strong increases of the default premium.1 Against this background, we ask whether austerity causes the default premium to decline, that is, whether austerity pays off.

We find that the answer depends on the conditions under which austerity takes place. Our main result can be summarized as follows. If the default premium is high to begin with, that is, if the economy is under severe fiscal stress, a reduction of government spending induces the default premium to increase further: austerity does not pay off. In contrast, if the default premium is initially moderate, cutting government spending induces the default premium to decline: austerity pays off during benign times. This result is robust across a variety of econometric specifications, modifications of our sample, and alternative ways to measure both fiscal stress and changes in government spending.

Throughout, we focus on how financial markets respond to austerity measures and sidestep the issue of how such measures affect the actual health of government finances. While austerity has an impact on fiscal fundamentals such as the level of sovereign debt, these observed fundamentals typically fail to provide a sufficient statistic for assessing the sustainability of debt. For the ability and willingness of governments to service its debt obligations and to roll over liabilities also depends on market conditions (Calvo, 1988; Cole & Kehoe, 2000; Roch & Uhlig, 2018) and a number of country-specific, partly unobserved factors such as the ability to raise tax revenues (Bi, 2012; Lorenzoni & Werning, 2018; Trabandt & Uhlig, 2011). The same level of debt may thus have very different implications for debt sustainability at different times and in different countries. The default premium demanded by financial markets, instead, is based on a broader assessment and provides a more comprehensive picture.

To answer our research question, we assemble a new panel data set for the sovereign default premium in a large number of countries. Specifically, we construct time series for the sovereign default premium in 38 developed and emerging economies, covering the period since the early 1990s. We compute the default premium as the difference in sovereign yields vis-à-vis a riskless reference country where sovereign default can be ruled out for practical purposes. Importantly, we consider only yields on government securities issued in a common currency in order to eliminate confounding factors due to expectations of inflation and currency depreciation. For more recent observations, we also rely on data for credit default swap (CDS) spreads. In our sample we only include country-quarter observations right up to, but not including, periods of default. We analyze the variation of the default premium in some detail, both across countries and over time, and summarize it by the cumulative distribution function.

Our interest is the impact of austerity measures on the sovereign default premium. Such measures include reductions of government consumption and transfers, as well as tax increases. However, in our study, we focus on government consumption because identification is more straightforward in this case.2 We identify fiscal shocks, that is, variations of government consumption that are unanticipated and do not represent a systematic response to the state of the economy. Our identification scheme is based on a two-step approach and builds on earlier work by Ramey (2011b) and Blanchard and Perotti (2002). First, we compute three distinct measures of forecast errors of real government consumption growth in order to purify the changes of government spending growth of the anticipated component. For this purpose, we rely on data by professional forecasters: Oxford Economics and the OECD. Both provide forecasts for a large number of countries, although they do not cover all 38 countries in our sample. For this reason we also estimate a fairly conventional fiscal vector autoregression (VAR) in order to compute forecast errors of government consumption. Analyzing the distribution of forecast errors, we find that all three measures are reasonably well behaved. In our baseline, we rely on the data by Oxford Economics, since they are available at quarterly frequency and the country coverage is quite large.

In a second step, we use local projections à la Jordá (2005) and estimate the response of the default premium to the forecast error of government consumption growth. At this stage, we impose the identification assumption that unanticipated changes of government consumption are not an endogenous response to structural innovations. We assume, in other words, that the systematic component of government consumption is predetermined. This assumption is plausible because neither does government consumption automatically respond to the state of the economy nor is it, because of decision lags, adjusted instantaneously in a discretionary manner. While we cannot test our identification assumption, we provide evidence suggesting that it is indeed satisfied in our sample. Moreover, we show that while our baseline is estimated on quarterly observations, our main result also obtains for a specification that exploits monthly forecast updates.

We allow the effect of austerity to depend on the amount of fiscal stress because austerity measures can be conducted under very different circumstances. Fiscal adjustments may take place under fairly calm conditions. However, often they are implemented when the sustainability of public finances is already in doubt and the default premium is high. The specific circumstances are likely to matter for how fiscal shocks propagate through the economy because mounting evidence suggests that the transmission of fiscal policy is state dependent (Auerbach & Gorodnichenko, 2012; Corsetti, Meier, & Müller, 2012; Ilzetzki, Mendoza, & Végh, 2013). In order to measure fiscal stress, we compare the default premium in the period prior to the fiscal shock to the other realizations in the sample. In this regard, it is key that our sample covers a large variety of countries and time episodes. Formally, we run a smooth transition regression where we use the cumulative distribution function of the sovereign default premium as an indicator function for the degree of fiscal stress.

We find that how fiscal shocks affect the default premium indeed depends on initial conditions in terms of fiscal stress. According to our baseline specification, a reduction of annualized government consumption growth by 1 percentage point lowers the default premium by about 30 basis points if fiscal stress is already low. If fiscal stress is severe, the same shock raises the default premium by about 60 basis points. We define benign times as a level of the default premium sufficiently low for a reduction of government consumption to reduce the default premium. Our smooth transition model determines a threshold for benign times that differs for advanced and emerging economies—30 and 200 basis points, respectively.

We shed some light on our findings by investigating the adjustment dynamics of important macroeconomic indicators to the shock. Government consumption growth declines somewhat persistently, implying that fiscal shocks tend to lower the level of government consumption permanently. The behavior of public debt and output growth in the aftermath of the shock depends strongly on initial conditions. A fiscal shock that materializes in times of severe stress raises the debt-to-output ratio and lowers output growth significantly. If the same shock takes place in the absence of stress, the debt ratio declines and output growth is unaffected. Finally, we also assess the impact of fiscal shocks on the term structure of default premiums and find that it responds more strongly at the short end.

In terms of empirical contributions, our paper relates to a number of studies that have explored the state dependence of the fiscal transmission mechanism. In particular, we build on the smooth transition framework of Auerbach and Gorodnichenko (2012, 2013). Early work by Perotti (1999) had established that fiscal policy affects the economy differently in good times and bad. Giavazzi, Jappelli, and Pagano (2000) suggest that the size and persistence of fiscal measures also matter for their effects. Earlier studies have also focused on how financial markets respond to fiscal policy measures. Ardagna (2009), for instance, reports that interest rates tend to decline in response to large fiscal consolidations. Laubach (2009) finds that future debt and deficits tend to raise U.S. interest rates. Akitoby and Stratmann (2008) focus on how sovereign yield spreads in emerging markets react to changes of fiscal indicators.

The remainder of the paper is organized as follows. We discuss our econometric model as well as our approach to identification in section II. Section III details the construction of our data set. Here, we also establish a number of basic facts regarding the time series properties of the sovereign default premium and the series of fiscal shocks. Section IV presents the results, as well as an extensive robustness analysis. It also offers a brief discussion of our results in light of available theoretical models. Section V concludes.

## II. Model Specification and Identification

In our analysis we rely on local projections as introduced by Jordá (2005). In order to establish the causal effect of austerity measures on the sovereign default premium and other variables of interest, we project these variables on fiscal shocks, that is, variations in fiscal policy that are unanticipated and not systematically related to the state of the economy. In what follows, we first introduce our model specification and then explain how we measure fiscal shocks.

### A. Local Projections with Smooth Regime Transitions

Local projections are quite flexible in accommodating a panel structure and offer a straightforward way to condition the short-run effects of fiscal shocks on the extent of fiscal stress that we measure by the initial level of the sovereign default premium.3 In addition, local projections allow us to assess whether the sovereign default premium responds asymmetrically to positive and negative shocks. Auerbach and Gorodnichenko (2013), Owyang, Ramey, and Zubairy (2013), and Ramey and Zubairy (2018), among others, also rely on local projections while analyzing fiscal policy. Their focus, however, is on the fiscal multiplier on output and on whether it changes with the business cycle or when interest rates are close to 0.

Formally, letting $xi,t+h$ denote the response of a particular variable in country $i$ at time $t+h$ to a fiscal shock $ɛi,tg$ at time $t$, we consider the following linear model as our point of departure:
$xi,t+h=αi,h+ηt,h+ψhɛi,tg+ui,t+h,$
(1)
where $αi,h$ and $ηt,h$ capture country and time fixed effects, respectively. At each horizon $h$, the response of the dependent variable to the shock is given by the coefficient $ψh$. An inclusion of further controls is not necessary to the extent that the fiscal shock is identified correctly.4 We discuss how we measure fiscal shocks below. The error term $ui,t+h$ is assumed to have a 0 mean and strictly positive variance.
Our preferred specification is a version of model (1) which conditions the response of $xi,t+h$ on the extent of fiscal stress. For this purpose, we use a smooth transition model:
$xi,t+h=αi,h+ηt,h+ψm,hF(ri,t-1)ɛi,tg+ψn,h1-F(ri,t-1)ɛi,tg+ui,t+h.$
(2)

Here, the response of the dependent variable to the shock is allowed to differ at each horizon $h$ across regimes $m$ (maximum stress) and $n$ (no stress), with the $ψ$-coefficients indexed accordingly. For each country-time observation, the indicator function $0≤F(ri,t-1)≤1$ determines the weight of each regime. This indicator function depends on $ri,t-1$, that is, the sovereign default premium in country $i$ at the end of the previous period. As a result, the weights are predetermined with respect to the fiscal shock.

Projection (2) directly captures the dynamic effects of a fiscal innovation conditional on the circumstances under which it occurs. Formally, the response in period $t+h$ to a government consumption impulse in period $t$, $ɛi,tg$, conditional on the economy experiencing a particular state today, indexed by $ri,t-1$, is given by the regression coefficients on $ɛi,tg$ in equation (2):
$∂xi,t+h∂ɛi,tg|ri,t-1=ψm,hFri,t-1+ψn,h1-Fri,t-1.$
(3)
This expression illustrates that computing impulse responses based on a single-equation approach do not require us to make additional assumptions on the economy staying in a particular regime (see also the discussion in Ramey & Zubairy, 2018). Rather, the local projection at time $t$ directly provides us with a measure of the average response of an economy in state $ri,t-1$ going forward. Note also that equation (3) is just a linear combination of regression coefficients. We can thus rely on a Wald-type test to assess whether responses at a particular horizon are significantly different from each other as a result of different initial conditions.
Our framework also allows us to investigate the effect of positive as opposed to negative shocks by splitting the shock series according to their sign and including both as distinct regressors (Kilian & Vigfusson, 2011). In this case, we estimate the following model:
$xi,t+h=αi,h+ηt,h+ψm,h+F(ri,t-1)ɛi,tg++ψm,h-F(ri,t-1)ɛi,tg-+ψn,h+1-F(ri,t-1)ɛi,tg++ψn,h-1-F(ri,t-1)ɛi,tg-+ui,t+h,$
(4)

where a positive shock is constructed as $ɛi,tg+=ɛi,tg$ if $ɛi,tg≥0$ and 0 otherwise, and similarly for negative shocks. The rest of the notation follows model (2).

The specification of the transition function $F(·)$ involves two steps: the choice of the indicator and the specification of the mapping of the indicator into weights from 0 to 1. First, we choose the (lagged) default premium as our indicator because we think that it provides a natural benchmark to measure the extent of fiscal stress for a given country-time observation. As a market-based measure, it is more comprehensive than specific fiscal indicators such as the debt-to-GDP ratio. Second, we specify the indicator function $F(ri,t-1)$ on the basis of the empirical cumulative density function (CDF) in our sample. As a result, we do not have to postulate a specific parametric indicator function as Auerbach and Gorodnichenko (2012) do in order to capture the state of the business cycle. Moreover, extreme values are given by observations that actually materialize in sample.5 Formally, we have
$F(ri,t-1)=1N∑j=1N1rj≤ri,t-1,$
(5)
where $N$ is the number of country-time observations in our sample, $1$ is an indicator function, and $j$ indexes all country-time observations, separately for the group of advanced and emerging economies (baseline). $F(ri,t-1)$ equals 1 if the premium is at the maximum of the sample: a situation of maximum stress. $F(ri,t-1)$ equals 0 if the premium is at the minimum: a situation of no stress. Of course, actual economies hardly ever operate in either of these two polar regimes. This notion is captured in the estimation, as, for each observation, the impact of the regressors is a weighted average of the dynamics in the two regimes. Put differently, regime transition is smooth as in Auerbach and Gorodnichenko (2012). Also, all observations contribute to identifying the dynamics that govern in the polar regimes.

### B. Fiscal Shocks

In our empirical analysis, we focus on government consumption to construct a measure of fiscal shocks. Of course, actual austerity programs may have sizable tax and transfer components, but identification is less thorny in the case of government spending. Under our identification assumption (discussed below), a fiscal shock is an unanticipated innovation to the rate of change in government spending. The innovation represents fiscal “news” and can be captured by the forecast error (Ramey, 2011b). Formally, we have
$ɛi,tg=Δgi,t-Et-1Δgi,t,$
(6)
where $Δgi,t$ is the realization of government spending growth and $Et-1Δgi,t$ is the previous period's forecast. It captures predictable changes due to, for instance, systematic responses to the state of the economy.

In our analysis, we pursue three alternative approaches to compute $ɛi,tg$. They differ in the source or method for obtaining the government spending forecast. The first two approaches closely follow Ramey (2011b) and use professional forecasts as a direct measure for $Et-1Δgi,t$. Ramey (2011b) in her analysis of U.S. time series relies on the Survey of Professional Forecasters, which is unavailable for our sample. We consider two different sets of professional forecasts instead.

In our baseline specification, we use proprietary data on quarter-on-quarter growth rate projections provided by Oxford Economics, a large forecasting firm that serves 1,500 clients, among them international corporations, financial institutions, government organizations, and universities. It employs some 200 economists and analysts. The reason we focus on growth rates rather than levels is that there are irregular base-year changes for the countries in our sample that would show up as structural breaks if we were considering levels.

Our second set of professional forecasts is compiled twice a year by the OECD and disseminated in its Economic Outlook. Forecasts are prepared at the end of an observation period, in June and December each year, and tend to perform quite well (Auerbach & Gorodnichenko, 2012). While the coverage of some countries starts earlier than in the Oxford Economics sample, observations are available only at lower frequency, that is, semiannually.

Our third specification employs a panel VAR model to forecast government consumption growth and compute the forecast error. Formally, let $Xi,t$ denote a vector of endogenous variables, which includes government spending growth, output growth, and the default premium. We estimate the following panel VAR,
$Xi,t=αi+ηt+A(L)Xi,t-1+νi,t,$
(7)
where $A(L)$ is a lag polynomial and $νi,t$ is a vector of reduced-form disturbances with covariance matrix $E(νi,tνi,t')=Ω$. In our analysis, we allow for four lags since the model is estimated on quarterly data. Assuming a lower Cholesky factorization $L$ of $Ω$, and that government consumption growth is ordered on top in the vector $Xi,t$, the structural shock $ɛi,tg$ equals the (scaled) first element of the reduced-form disturbance vector $νi,t$: $ɛi,tg=L-1νi,t$.6

This forecasting model allows for the broadest coverage in terms of countries, time periods, and frequency as it relies on high-quality quarterly national accounts data (see Ilzetzki et al., 2013) but not on professional forecasts. However, influential contributions by Ramey (2011b) and Leeper, Walker, and Yang (2013) have highlighted the potential limitation of VAR models in accounting for fiscal foresight. Intuitively, fiscal measures that are not predictable by the econometrician on the basis of conventionally observable time series may still be known in advance, for instance, because it takes time to pass legislation. Hence, all things equal, we prefer to rely on professional forecasts that encompass the broadest possible information set to a finite-order VAR that may not span the relevant state space entirely. But given that professional forecasts are available only for a restricted sample in terms of coverage and frequency, we think that the VAR still serves as an important reference point. We report results for all three forecasting models.

In terms of identification, we assume for all three specifications of the forecasting model that the forecast error of government spending growth is not systematically caused by the contemporaneous state of the economy. Hence, it represents a genuine fiscal shock. This identifying assumption—that the systematic component of government consumption is predetermined—goes back to Blanchard and Perotti (2002) but is also implicit in Ramey (2011b) when she considers news shocks based on professional forecasts (see also Ramey & Zubairy, 2018). To see this, assume, to the contrary, that government spending is adjusted instantaneously to other structural innovations in a systematic way. A positive technology shock may, for example, drive up output, and if fiscal policy can immediately react to output, resulting in an unforeseen change of government consumption. In this scenario, a technology shock would give rise to “fiscal news.” However, as Blanchard and Perotti (2002) originally argued, government consumption is unlikely to respond automatically to the cycle and b) to be adjusted instantaneously in a discretionary manner by policymakers. In this context, it is important to note that our measure of government consumption is derived from national accounts and therefore accrual based,7 and, unlike transfers, it is not composed of cyclical items. Discretionary changes of government spending in turn are subject to decision lags that prevent policymakers from responding instantaneously to contemporaneous developments in the economy.

Anecdotal evidence suggests that this holds true also in times of fiscal stress. Still, we cannot rule out that policy measures—while debated for some time—are sometimes spurred by contemporaneous financial market developments.8 Against this background, we exploit the specific panel structure of our data set and assess whether policymakers adjust government consumption systematically in response to contemporaneous movements of the sovereign default premium. Importantly, we find that while the common component of the sovereign default premium induces the country-specific component of the premium to move on impact, it does not affect government consumption significantly (see section IVA). Hence, we maintain our identification assumption with some confidence.

Before we turn to the data, we briefly explain why the narrative approach to identify fiscal shocks is not suited to analyze the issue at hand. Following the work of Romer and Romer (2010) for the United States, Devries et al. (2011) constructed a data set of fiscal measures for a sample of OECD countries (see also Guajardo, Leigh, & Pescatori, 2011). These fiscal policy measures are identified as being orthogonal to the business cycle on narrative grounds. A large number of these measures are taken in order to rein in public debt or budget deficits, which, in turn, move systematically with the sovereign default premium. These “shocks” are therefore likely to be endogenous with respect to the default premium and may not be used to identify the causal effect of fiscal policy on the former.

## III. Data

Our analysis is based on a new data set. In addition to standard time series data, it contains observations for the sovereign default premium and for fiscal shocks for up to 38 emerging and advanced economies.

### A. The Sovereign Default Premium

In what follows, we detail the construction of the sovereign default premium. Since our focus is on how fiscal shocks affect the default premium, we consider only those countries for which quarterly observations on government consumption are available (see below). As we stressed in section I, we construct a mostly spread-based measure using yields for securities issued in common currency. To the extent that financial markets are sufficiently integrated, we thus eliminate fluctuations in yield spreads due to inflation expectations and the risk premiums associated with them. In addition to a default risk premium, if duration differs or drifts, yield spreads may still reflect a term premium (Broner, Lorenzoni, & Schmukler, 2013). We try to minimize the term premium by constructing the yield spread on the basis of yields for bonds with a comparable maturity and coupon.9 As a result, yield spreads should primarily reflect the probability and expected extent of a sovereign default—as assessed by market participants.10

We obtain our default-risk measure based on four distinct sources or strategies. First, for a subset of (formerly) emerging markets, we directly rely on J.P. Morgan's Emerging Markets Bond Index (EMBI) spreads. Second, we add to these observations data for euro-area countries based on the long-term interest rate for convergence purposes. Third, we make use of the issuance of foreign-currency government bonds in many advanced economies during the 1990s and 2000s in order to extend our sample to non-euro-area countries and the pre-euro period. Finally, in the more recent part of the sample, a direct measure of default risk has become available in the form of CDS spreads.11

The use of CDS spreads also allows us to include the benchmark countries United States (EMBI) and Germany (long-term convergence yields) in the sample. In order to get an absolute measure of default risk for the other countries, we add the CDS spread of the respective benchmark countries to the relative country spread. For the period before CDS data are available, we add the value of the average CDS spread of the period prior to the default of Lehman Brothers.12 Online appendix A.1.6 illustrates the construction of the default premium measure by means of an example. In our sample, we only include country-quarter observations right up to, but not including, periods of actual default (see the note to table A.1 in the online appendix for the excluded episodes). During default episodes, trading in secondary markets typically collapses, and the information content of observed interest rates is limited. Excluding default episodes does not result in sample-selection bias, however, because we conduct inference about the effects of austerity when a country has not yet defaulted and austerity still is a choice.

Table A.1 provides basic descriptive statistics for the default premium, $rt$. The total number of observations in our sample is 3,013—1,648 for developed economies and 1,365 for emerging economies.13 The default premium is measured at the end of the quarter in percentage points and varies considerably across our sample.14 In a couple of euro-area countries, the lowest realizations of the default premium are slightly negative.15 For the group of developed economies (see table A.1 for the classification), we observe the highest premiums in Portugal (12 percentage points) and Greece (24 percentage points). For emerging economies, the highest values are reached in Brazil (24 percentage points), Ecuador (21 percentage points), and Argentina (20 percentage points).16

Compared to these values, most realizations of the default premium in our sample are small. This is apparent from the CDF plotted in figure 1 for the entire sample (solid line), but also for the set of developed (dash-dotted line) and emerging economies (dashed line) in isolation. In each case, the mass of observations is very much concentrated on the left. For the full sample, about 50% of the observations for the default premium are below 1.15 percentage points. Still, there are considerable differences across the two country groups: 99.4% of observations are below 10 percentage points in the sample of developed economies. The corresponding number is only 95% in the sample of emerging-market economies.
Figure 1.

Sovereign Default Premium: Empirical Cumulative Density Function (CDF)

The horizontal axis measures the default premium in percentage points. The vertical axis measures the fraction of observations for which the default premium is at most the value on the horizontal axis. The solid line displays CDF for full the sample, the dashed-dotted line for developed economies only, and the dashed line for emerging economies only.

Figure 1.

Sovereign Default Premium: Empirical Cumulative Density Function (CDF)

The horizontal axis measures the default premium in percentage points. The vertical axis measures the fraction of observations for which the default premium is at most the value on the horizontal axis. The solid line displays CDF for full the sample, the dashed-dotted line for developed economies only, and the dashed line for emerging economies only.

As explained in section II, the empirical CDF of the default premium provides a natural benchmark to measure the extent of fiscal stress for a given time-country observation. Importantly, it is a purely empirical benchmark and allows us to refrain from making specific and perhaps arbitrary assumptions. Also, as a market-based measure, it is arguably more comprehensive than specific fiscal indicators such as the debt-to-GDP ratio. As a practical matter, we specify the function $F(·)$ in projections (2) and (4) on the basis of the country-group-specific CDFs shown in figure 1. Our indicator for stress takes on the highest value of 1 when a country at a given point in time experiences the highest default premium observed in our sample. In contrast, it takes a value of 0 when the country experiences the lowest premium observed in our sample.

Figures A.3 to A.5 in the online appendix show the fiscal stress indicator over time for the countries in our sample. They also show the empirical CDF of the smoothed output gap measure of Auerbach and Gorodnichenko (2012). While there is some correlation between fiscal stress and the state of the cycle, there are long periods during which the measures move in opposite directions.

In table A.1 in the online appendix, we also report the correlation of the sovereign default premium with output growth and the growth of government consumption, based on national accounts data. It turns out that the default premium is countercyclical in all countries, although sometimes the correlation is negligible. In contrast, the within-country correlation of the default premium and government consumption growth varies across countries. It is negative for most countries, but often weakly so.

### B. Fiscal Data

We now turn to the second original contribution in terms of data collection: the construction of fiscal shocks. These shocks capture unanticipated and nonsystematic innovations in government consumption growth. As we explained in section II, we construct three alternative measures.

For the first measure (baseline), we use Oxford Economics as a data source. Oxford Economics provides data on quarterly real government consumption. The data come in monthly vintages and are unbalanced, with some missing vintages in between. The earliest available vintage is November 1996. The latest available vintage used in our analysis is December 2017. Nowcasts and forecasts are produced by Oxford Economics, while the back data are provided to Oxford Economics by Haver Analytics, which in turn usually obtains its data from national statistical offices. The data come with a flag indicating whether the most recent vintage is seasonally adjusted. For the countries where this is not the case (Argentina, Malaysia, Thailand, and Turkey), we use Tramo/SEATS for seasonal adjustment. For the other countries, we visually inspect the vintages for the presence of a seasonal pattern and remove the seasonal figure in all vintages preceding the move to seasonal adjustment at the source.17 A few countries move between exhibiting seasonal patterns and being seasonally adjusted several times. Because there is no consistent way to identify the break points, we remove these countries from our sample. The first two columns of table A.2 in the online appendix report the resulting sample.

Oxford Economics updates its quarterly forecasts for a number of macroeconomic aggregates, including real government consumption on a monthly basis. To compute the forecast error for a given quarter $t$, we compute the difference between the real-time value of government spending or GDP growth and the value forecast for that quarter a quarter earlier. As Oxford Economics forecasts are usually made every month, we use a geometric average over the available monthly values in a given quarter $t$ to arrive at quarterly forecasts for spending growth next period $EtΔgi,t+1$ and real-time realizations of spending growth $Δgi,t$. We also exploit monthly observation directly in the context of our sensitivity analysis.

Our second measure of fiscal shocks relies on data provided by the OECD in its Economic Outlook, which is published in June and December every year. Before 1996S2, the forecast target period referred to half-years. Since 1996S2, the OECD forecast target has referred to quarterly values. We sum the quarterly-level observations for each half-year period in order to obtain a sample for which the frequency at which forecasts are made and the forecast target frequency align. This allows us to obtain observations at semiannual frequency and compute the growth-rate forecasts and real-time growth-rate realizations, as well as the associated forecast errors.

Our third measure is based on an estimated VAR model that we rely on to forecast government consumption growth. For a long time, studies of the fiscal transmission mechanism have been limited to a small set of countries because high-quality quarterly data for government consumption were not available.18 Rather, quarterly data were often derived from indirect sources using time disaggregation or interpolation. In a recent contribution, Ilzetzki et al. (2013) collected quarterly data based on direct sources for government consumption for 44 countries. To estimate our VAR model, we collect quarterly data for government consumption expenditure based on national accounts and nonfinancial accounts of the government along the lines of Ilzetzki et al. (2013). Our sample differs to some extent, depending on the countries for which we are able to compute the sovereign default premium. We also extend their sample to include more recent observations and additional countries for which we were able to confirm with statistical agencies the availability of government consumption data based on direct sources.19 The full sample coverage is shown in table A.2 in the online appendix. Our earliest observation for which we obtain data on the default premium and on government consumption is 1991Q1, namely, for Denmark and Italy. Our sample runs up to 2017Q4.

Table 1 displays descriptive statistics for the three forecast errors. All three forecasts perform reasonably well but are not perfect. Over the full sample, the average forecast errors are close to 0 for all three models. In the VAR model, this is by construction. On an individual country basis, both the OECD and Oxford Economics produce forecasts with a relatively low root mean squared error (RMSE). The VAR forecasts exhibit the largest RMSE, but for a somewhat more challenging sample of countries. The OECD and VAR forecasts are unbiased, having only one country, each with a significant mean forecast error.20 Oxford Economics performs a bit worse in this regard. Regardless of the forecasting model, about 20% of the countries in our sample have autocorrelated forecast errors, as measured by a Ljung and Box (1978) Q-statistic test with eight lags. Our conjecture is that this result is partially driven by the depth of the Great Recession that caught most observers by surprise. This autocorrelation suggests that forecasts could be improved by including lags of previous forecast errors. In a robustness check below, we find results similar to the baseline when we do so.

Table 1.
Forecast Errors of Government Consumption Growth: Descriptive Statistics
OEOECDVAR
Countries 23 21 38
1,696 887 2,832
Mean −0.016 −0.039
RMSE 0.616 0.402 2.947
Number of significant constants
Number of significant LBQ tests
Wald $F$-statistic 44.41 0.99 1790.86
OEOECDVAR
Countries 23 21 38
1,696 887 2,832
Mean −0.016 −0.039
RMSE 0.616 0.402 2.947
Number of significant constants
Number of significant LBQ tests
Wald $F$-statistic 44.41 0.99 1790.86

Forecast errors measured in percentage points. Semiannual OECD forecast errors are rescaled to reflect errors in the growth rate at a quarterly level. “Number of significant constants” refers to the number of countries where the mean forecast error is significantly different from 0. “Number of significant LBQ tests” refers to the number of countries where the Ljung and Box (1978) Q-statistic with eight lags rejects the null hypothesis of the forecast error being white noise. The significance level for both tests is 5%. Kleibergen and Paap (2006) $rk$-Wald $F$-statistic computed using Stata's xtivreg2 in a first-stage regression of government consumption growth on the respective forecast error. Robust covariance estimator clustered at country and quarter level.

Figure 2 shows the kernel density estimates of the forecast errors. In line with table 1, it illustrates that the forecast errors based on Oxford Economics and OECD forecasts are considerably less dispersed than the VAR-based forecast errors. Forecast errors based on Oxford Economics forecasts, however, exhibit somewhat fatter tails than those based on the OECD.
Figure 2.

Distribution of Forecast Errors

The lines correspond to kernel density smoother estimates for forecast errors based on Oxford Economics.

Figure 2.

Distribution of Forecast Errors

The lines correspond to kernel density smoother estimates for forecast errors based on Oxford Economics.

In the last row of table 1, we report a measure of the predictive power of our identified shocks in the form of an $F$-statistic along the lines of the tests conducted in Ramey (2011b) and Ramey and Zubairy (2018).21 Reassuringly, our baseline forecast errors based on the Oxford Economics forecasts, as well as the VAR-based forecast errors, are comfortably above the rule-of-thumb threshold of 10 proposed by Staiger and Stock (1997).22 The semiannual OECD forecast errors, instead, do not seem to be a good predictor for semiannual government consumption growth. Given their frequent use in the literature, we nevertheless report the results based on this forecasting model.

Finally, our data set also includes data on government debt for a subset of the countries in our sample. Here we draw on two sources. First, Eurostat provides data on gross quarterly government debt of the general government (gov_10q_ggdebt) for EU countries. Second, the Worldbank's Quarterly Public Sector Debt (QPSD) database provides data on gross general debt for some additional countries in our sample. If the latter database does not provide data for the general government but the central government, we use that series instead.

## IV. Results

In what follows we present the main results of our analysis. We first focus on the instantaneous response of the default premium to fiscal shocks. Then we provide additional evidence on the adjustment dynamics and the transmission mechanism. Throughout, we highlight the importance of initial conditions under which austerity takes place: the extent to which there is fiscal stress.

### A. The Impact of Fiscal Shocks on the Default Premium

In table 2 we report our estimates for the response of the sovereign default premium to fiscal shocks. The different columns in the table show results for the alternative models that we use to compute the forecast error of government consumption growth (see equation [6]). The different panels, in turn, refer to different specifications of the projections (1) to (4). In each case, we focus on the instantaneous response of the premium ($h=0$) and normalize coefficients so that they reflect the response of the default premium to an unanticipated reduction of government consumption growth by 1 percentage point (annualized). Standard errors are given in parentheses. They are robust with respect to heteroskedasticity, as well as serial and cross-sectional correlation (Driscoll & Kraay, 1998).

Table 2.
Instantaneous Response of Default Premium (Basis Points) to Reduction of Government Consumption Growth
Forecasting model
Oxford EconomicsOECDVARVAR (Maximum Sample)
Unconditional projection
All 6.02* 8.28 0.74 0.35
(3.70) (6.25) (0.47) (0.47)
Cuts only 15.61** 20.79** 1.83* 2.43**
(7.78) (10.32) (1.01) (1.05)
Maximum stress
All 20.86 26.99*** 1.52 0.58
(12.99) (9.03) (2.10) (1.08)
Cuts only 58.81*** 55.27*** 15.91*** 10.40***
(17.34) (9.99) (5.25) (3.21)
No fiscal stress
All −6.99 −20.18*** 0.25 −0.04
(4.98) (6.17) (0.77) (0.66)
Cuts only −33.40*** −39.00*** −7.66*** −9.67***
(9.85) (14.27) (2.63) (2.83)
Diff. max stress—no stress
All 27.85 47.17*** 1.27 0.62
Cuts only 92.21*** 94.27*** 23.57*** 20.07***
Countries 23 21 23 38
Observations 1,515 708 1,443 2,689
Forecasting model
Oxford EconomicsOECDVARVAR (Maximum Sample)
Unconditional projection
All 6.02* 8.28 0.74 0.35
(3.70) (6.25) (0.47) (0.47)
Cuts only 15.61** 20.79** 1.83* 2.43**
(7.78) (10.32) (1.01) (1.05)
Maximum stress
All 20.86 26.99*** 1.52 0.58
(12.99) (9.03) (2.10) (1.08)
Cuts only 58.81*** 55.27*** 15.91*** 10.40***
(17.34) (9.99) (5.25) (3.21)
No fiscal stress
All −6.99 −20.18*** 0.25 −0.04
(4.98) (6.17) (0.77) (0.66)
Cuts only −33.40*** −39.00*** −7.66*** −9.67***
(9.85) (14.27) (2.63) (2.83)
Diff. max stress—no stress
All 27.85 47.17*** 1.27 0.62
Cuts only 92.21*** 94.27*** 23.57*** 20.07***
Countries 23 21 23 38
Observations 1,515 708 1,443 2,689

Response to unanticipated reduction of government spending growth by 1 percentage point (annualized). Estimates based on projections (1) to (4), with $h=0$. Driscoll and Kraay (1998) standard errors in parentheses. Significant at $***$1%, $**$5%, and $*$10%. Forecast error for spending growth computed using different forecasting models: Oxford Economics, OECD, and VAR model; the right-most column uses the largest possible sample; the second-to-right-most column uses only the sample for which forecasts by Oxford Economics are available.

Consider first the results for our baseline specification of the forecasting model, shown in the leftmost column. Here, we compute fiscal shocks as the difference of actual government consumption growth (as perceived in real time) and the forecast of Oxford Economics. In this case, our sample covers 23 countries and consists of 1,515 observations. In the top panel, to set the stage, we show results for the unconditional linear projection (1), that is, we do not condition on initial conditions in terms of fiscal stress. We find that an unanticipated and exogenous reduction of government consumption raises the default premium significantly. Quantitatively, the effect is rather small: the default premium increases by 6 basis points. Once we estimate a version of model (4) and allow for different effects of spending cuts and hikes (but without accounting for fiscal stress), the effects of cuts are larger—about 16 basis points. There is no significant effect of spending increases (not shown).

In the second and third panels, we report results for the polar regimes of maximum fiscal stress and no fiscal stress, respectively, which we obtain as a result of estimating the conditional model, equation (2). Recall that we condition the response of the premium on the extent of fiscal stress to begin with (that is, in the previous quarter) relative to the empirical distribution in our sample. In the baseline, our measure of stress is computed on the basis of distinct distributions for advanced and emerging economies. It turns out that results are robust to modifying this assumption, as we discuss below.

For regime maximum stress (regime m), results are rather stark. Lumping together cuts and hikes in one time series does not yield significant results. However, if we allow for a differential effect of hikes and cuts as in model (4), we find that spending cuts induce the default premium to rise strongly. Severe fiscal stress alters the quantitative impact of cuts dramatically. We now find an increase of the premium by almost 60 basis points.

For regime no stress (regime n), we again find no significant effects if we do not distinguish between cuts and increases of government consumption. Spending cuts per se, however, tend to lower the default premium if stress is absent and quite strongly so. This result requires some qualification. Formally, we define no stress as a situation where the default premium is at its lowest point in the sample. Strictly speaking, a further reduction of the default premium is hard to achieve.23

The strong reduction of the premium in regime n reflects a more fundamental issue: the estimated projection (2) extrapolates from low-stress observations to the pure regime of no stress. Put differently, our finding suggests that reducing government consumption may actually bring down the default premium even if there is some fiscal stress. In what follows, we refer to such a moderate level of fiscal stress as “benign times,” defined as the region of the default premium for which a reduction of government consumption does not induce the default premium to rise. As we combine our estimates for regimes $m$ and $n$, we obtain the threshold value of the indicator function and, in turn, the implied threshold for the default premium. For advanced economies, benign times correspond to values of the default premium of up to approximately 30 basis points. The corresponding value for emerging economies is 200 basis points.

In sum, we find that whether austerity pays off depends on the initial conditions in terms of fiscal stress. If stress is at its maximum (historically speaking), a reduction of government consumption does not pay off. It raises the default premium strongly. If the reduction takes place while stress is low, that is, during benign times, it tends to reduce the default premium.

The last panel of table 2 underscores once more the importance of conditioning the response of the premium on initial conditions in terms of fiscal stress. It shows the difference in the response of the premium across regimes. It amounts to almost 1 percentage point (cuts only). A Wald-type test shows that the difference is also statistically significant.

Our result is robust to changes in the way we compute fiscal shocks. Specifically, in the second column, we report results that obtain once we measure the forecast error of government consumption growth on the basis of OECD forecasts. Recall that in this case, forecasts for government consumption are compiled twice a year. Also, the country coverage is smaller than in the baseline. The sample consists of twenty countries and 708 semiannual observations. However, the results, shown in the second column of table 2 are similar to the baseline.24 Importantly, this holds for all model specifications. We find once more that whether fiscal shocks raise or lower the default premium depends on the extent of fiscal stress.

The third forecasting model we consider is a conventional VAR model with four lags of government spending growth, real GDP growth, and the default premium. A potential downside of the VAR model is that it may fail to capture fiscal news that are known to market participants (but not the econometrician) and hence provide an incorrect measure of fiscal surprises. The country coverage of professional forecasters is limited. And indeed the sample for which we are able to compute quarterly VAR-based forecasts is considerably larger: it comprises 38 countries and 2,689 country-time observations. We compute the unanticipated innovations of government spending growth on the basis of the VAR model and use these fiscal shocks to estimate projections (1) to (4). The two right-most columns of table 2 report the results. In the first of those columns, we show VAR-based estimates for a restricted sample, namely, the sample for which forecasts of government spending growth by Oxford Economics are available. In the second column, we estimate the model on the largest possible sample.

We find that results are very similar across samples. Qualitatively, they are also similar to those of our baseline specification for the unconditional model, as well as for benign times and times of severe fiscal stress. We find that reducing quarterly government spending growth raises the default premium if fiscal stress is very high to begin with. As before, spending cuts in the absence of stress tend to lower the default premium. However, compared to the forecasting models that employ professional forecasts, the effects are estimated to be quite a bit weaker. This may be due to the fact that the VAR does not capture all fiscal measures that are foreseen by market participants (Ramey, 2011b; Leeper et al., 2013).25 That said, we note that our main result also obtains once we consider a much larger group of countries and rely on a VAR to compute forecast errors: the effects of spending cuts on the default premium depend on initial conditions.

### B. Robustness

To the extent that our findings are surprising, notably for regime m, it may raise doubts about the identification assumption that we entertain throughout: that the systematic component of government consumption is predetermined. A possible objection to our analysis is that results might be driven by reverse causality: as the sovereign default premium rises, governments may immediately cut government consumption, for instance, in order to calm financial markets. The panel structure of our data set allows us to assess this objection formally. We first extract a common factor in the default premium along the cross-sectional dimension of our panel by means of a principal component analysis (see Longstaff et al., 2011, for a similar approach), separately for developed and emerging economies. Then we project the default premium and the forecast error on this common factor, as well as on its lags.26 We consider the impact period as well as the impact in the next period ($h=0,1$). Identification rests on the assumption that the common factor—variations of which may, for instance, reflect changes in the stochastic discount factor of global investors—is not contemporaneously affected by country-specific developments.27

We show results in table 3. As before, we condition the effects on the extent of fiscal stress and normalize coefficients so that they capture the response to an increase of the common factor of the default premium by 1 percentage point. In the upper panel, we show the response of the default premium measured in basis points. We observe that the “local” default premium increases significantly. Upon closer inspection, we find that this result is driven by regime m (right column). In contrast, there is no significant response of the forecast error for government consumption growth (bottom panel), regardless of whether fiscal stress is severe. While not a formal test of our identification assumption, this evidence is highly suggestive.28 It shows that in response to a higher default premium caused by global developments, there is no systematic, unforeseen contemporaneous fiscal adjustment even if there is severe stress to begin with.

Table 3.
Responses to Global Default Premium Component
Linear ModelNo Fiscal StressMaximum Stress
Default premium in individual country
Horizon 0 1.662 −1.848 3.536*
(1.355) (2.292) (1.988)
Horizon 1 0.969** −2.005*** 2.496**
(0.428) (0.422) (1.072)
Forecast error of government consumption growth
Horizon 0 0.001 0.005 −0.007
(0.003) (0.003) (0.006)
Horizon 1 0.001 0.001 0.004
(0.002) (0.004) (0.006)
Linear ModelNo Fiscal StressMaximum Stress
Default premium in individual country
Horizon 0 1.662 −1.848 3.536*
(1.355) (2.292) (1.988)
Horizon 1 0.969** −2.005*** 2.496**
(0.428) (0.422) (1.072)
Forecast error of government consumption growth
Horizon 0 0.001 0.005 −0.007
(0.003) (0.003) (0.006)
Horizon 1 0.001 0.001 0.004
(0.002) (0.004) (0.006)

Estimates based on forecast error for government consumption growth by Oxford Economics and common component of default premium (first principal component). Driscoll and Kraay (1998) standard errors in parentheses. Significant at $***$1%, $**$5%, and $*$10%.

More generally, the assumption that the systematic component of government consumption is predetermined and may not be adjusted to the state of the economy within the period is less restrictive the shorter the period under consideration. Hence, we find it reassuring that our main result obtains also as we consider monthly updates of Oxford Economics' forecasts. Specifically, we construct a monthly measure of fiscal news for the current quarter: the change of the forecast for the current quarter in a given month compared to the last month when a forecast for this quarter was made.29 Column 1 of table 4 shows the results. The structure of the table mimics table 2, except that we use the baseline measure for the forecast error by Oxford Economics throughout. The results for monthly news are qualitatively similar to our baseline for quarterly news. Quantitatively, the effects are smaller but still significant.

Table 4.
Instantaneous Response of Default Premium (Basis Points) to Fiscal Shock: Alternative Specifications
Specification
(1)(2)(3)(4)(5)(6)(7)(8)
Unconditional projection
All 2.07** 2.57* 5.50* 3.09*** 6.30 7.41** 5.77* 6.81
(1.01) (1.58) (3.10) (1.18) (3.95) (3.31) (3.39) (4.38)
Cuts only 2.34 8.25*** 13.28** 5.50** 16.62** 15.16** 14.12** 16.99**
(1.76) (3.28) (6.89) (2.38) (8.38) (7.39) (7.18) (7.84)
Maximum stress
All 6.08** 7.24 18.02** 8.49** 19.69* 17.24* 18.12 5.56
(3.14) (5.01) (7.59) (3.58) (11.96) (10.02) (11.69) (8.11)
Cuts only 13.13*** 27.44*** 54.40*** 13.75*** 47.15*** 41.42*** 46.95*** 85.28***
(4.73) (5.74) (9.89) (24.74) (16.55) (14.39) (16.28) (12.11)
No fiscal stress
All −0.81 −1.80 −5.31** −0.97 −6.72 −3.70 −5.73 0.11
(1.25) (2.23) (2.73) (1.43) (5.64) (4.08) (5.07) (22.86)
Cuts only −6.93*** −15.56*** −32.22*** −2.89*** −24.62** −20.64*** −25.39*** −169.47**
(1.75) (4.71) (5.22) (3.50) (10.13) (8.12) (9.09) (88.15)
Diff. max stress–no stress
All 6.89* 9.04 23.33** 9.45** 26.41 20.95 23.85 5.46
Cuts only 20.06*** 43.00*** 86.62*** 16.64* 71.77*** 62.06*** 72.34*** 254.75***
Countries 23 23 23 23 23 23 23 23
Observations 4,153 1,515 1,362 1,226 1,445 1,504 1,515 1,515
Specification
(1)(2)(3)(4)(5)(6)(7)(8)
Unconditional projection
All 2.07** 2.57* 5.50* 3.09*** 6.30 7.41** 5.77* 6.81
(1.01) (1.58) (3.10) (1.18) (3.95) (3.31) (3.39) (4.38)
Cuts only 2.34 8.25*** 13.28** 5.50** 16.62** 15.16** 14.12** 16.99**
(1.76) (3.28) (6.89) (2.38) (8.38) (7.39) (7.18) (7.84)
Maximum stress
All 6.08** 7.24 18.02** 8.49** 19.69* 17.24* 18.12 5.56
(3.14) (5.01) (7.59) (3.58) (11.96) (10.02) (11.69) (8.11)
Cuts only 13.13*** 27.44*** 54.40*** 13.75*** 47.15*** 41.42*** 46.95*** 85.28***
(4.73) (5.74) (9.89) (24.74) (16.55) (14.39) (16.28) (12.11)
No fiscal stress
All −0.81 −1.80 −5.31** −0.97 −6.72 −3.70 −5.73 0.11
(1.25) (2.23) (2.73) (1.43) (5.64) (4.08) (5.07) (22.86)
Cuts only −6.93*** −15.56*** −32.22*** −2.89*** −24.62** −20.64*** −25.39*** −169.47**
(1.75) (4.71) (5.22) (3.50) (10.13) (8.12) (9.09) (88.15)
Diff. max stress–no stress
All 6.89* 9.04 23.33** 9.45** 26.41 20.95 23.85 5.46
Cuts only 20.06*** 43.00*** 86.62*** 16.64* 71.77*** 62.06*** 72.34*** 254.75***
Countries 23 23 23 23 23 23 23 23
Observations 4,153 1,515 1,362 1,226 1,445 1,504 1,515 1,515

Response to unanticipated reduction of government spending growth by 1 percentage point (annualized). Estimates based on projection (2), with $h=0$ and using Oxford Economics forecasts. Driscoll and Kraay (1998) standard errors in parentheses. Significant at $***$1%, $**$5%, and $*$10%. 1 $=$ monthly model using fiscal news about the current quarter, 2 $=$ spreads relative to country-group common factor as dependent variable, 3 $=$ four lags of the forecast error included, 4 $=$ four lags of the default premium, government spending growth, and GDP growth included, 5 $=$ controlling for global factors instead of time-fixed effects (see text), 6 $=$ fiscal deficit and trade balance (as shares of GDP) included, 7 $=$ ICRG political risk index included, 8 $=$ pooled mean group estimator.

Next, we turn to further robustness tests. In column 2 of table 4, we report results for a specification where, instead of the default premium, we consider the difference between a country's default premium and the country-group common factor as the outcome variable. In this case, just as with a time fixed effect, the common factor captures changes in global risk valuation, but since each country has a different loading, we are able to capture different country betas. Also in this case, we find results are fairly similar to the baseline.

Our main result is also robust to further variations of our econometric specification, as we illustrate in columns 3 to 8 of table 4. Here, we report the results for six alternative specifications. First, given the evidence of autocorrelation in the government spending shocks, we include four lags of the forecast error. Second, we include four lags of government spending growth, GDP growth, and the default premium as additional regressors. Third, in light of evidence on the determinants of spread movements provided by Juvenal and Wiseman (2015), instead of time-fixed effects, we include the following global factors as contemporaneous regressors: a euro high-yield index, the ECB shadow rate, the Fed Funds shadow rate, and the VIX. We also include a global commodities index. Fourth, we include four lags of the fiscal-deficit-to-GDP ratio and the trade-balance-to-GDP ratio to control for fiscal and external sustainability.30 Fifth, we include the PRS Group's International Country Risk Guide (ICRG) political risk index (the sum of all components in their Table 3B). Finally, we consider the pooled mean group estimator to account for the fact that our panel of countries is quite heterogeneous (Pesaran & Smith, 1995). Again, our main result obtains for all specifications.

Recall that we measure initial conditions on the basis of the empirical CDF of the default premium. Our results already discussed are based on distinct CDFs for emerging and advanced economies (“country-group specific CDF”) and the end-of-quarter observation of the premium. Table A.3 in the online appendix shows results for six different specifications for the indicator function, based on the average (rather than the end-of-quarter) premium, a joint empirical CDF for all countries, and a parametric specification of the indicator function. We focus on our baseline measure of the forecast error by Oxford Economics and find that our results are robust: in all instances, a cut of government consumption lowers the default premium in benign times but raises it significantly if fiscal stress is high.31

In the online appendix, we explore more systematically to what extent our results also obtain for specific country groups—even though we stress that the mean group estimator allows for slope heterogeneity across countries and delivers results that are similar to the baseline (see column 8 in table 4). By and large, the results in table A.4 are similar to the baseline.

In fact, countries issue government debt with different maturities and, hence, correspondingly, there is a term structure of spreads. In our baseline, because of data limitations, we consider only one default premium for each country and investigate how it responds to fiscal shocks. For a limited number of time-country observations, we are able to collect spreads for CDS of different maturities. The resulting sample starts after 2004 and is therefore dominated by the financial and euro-area sovereign debt crises. With this limited sample at hand, we reestimate our baseline model for CDS spreads of different maturities.32 Specifically, we report in table 5 the impact response of a government spending shock in our baseline model on one-year to ten-year CDS spreads.33 Here, we focus on a sample where we also have a bond-based spread measure available, the response of which is reported in the right-most column for comparison. We find that the response of the ten-year CDS spread is very similar to the bond-based measure. This is reassuring, for the latter is dominated by bonds with a maturity of close to ten years (see section III). More interesting, we find that the short end of the spread curve responds more strongly to the shock. This implies that fiscal shocks alter the outlook in terms of default for the short run more strongly than for the long run.34

Table 5.
Instantaneous Response of Default Premium (Basis Points) to Fiscal Shock: CDS with Different Maturities Versus Bond Based
CDS
12345710Spreads
Unconditional projection
All 12.32** 11.77** 10.72** 9.77** 8.81* 7.95* 7.49* 5.79
(6.13) (5.90) (5.50) (5.05) (4.68) (4.21) (3.91) (3.60)
Cuts only 20.44** 20.18** 18.83** 17.36** 15.74** 15.37** 14.16** 13.37**
(10.12) (10.04) (9.33) (8.43) (7.83) (7.01) (6.12) (5.74)
Maximum stress
All 49.99** 49.97** 43.47** 39.33** 35.41** 31.86** 29.01** 23.22**
(21.36) (20.15) (18.26) (16.41) (15.02) (13.47) (12.13) (11.78)
Cuts only 79.74*** 78.94*** 73.76*** 67.89*** 62.23*** 58.66*** 53.67*** 50.87***
(25.51) (25.05) (22.98) (20.51) (18.95) (16.54) (14.63) (12.66)
No fiscal stress
All −21.42** −20.66** −18.63** −16.71** −15.01** −13.46** −11.80** −9.91*
(10.60) (10.17) (9.20) (8.25) (7.48) (6.79) (6.05) (5.86)
Cuts only −43.67*** −43.57*** −41.00*** −37.80*** −35.13** −32.17*** −29.26*** −28.42***
(16.83) (17.23) (16.43) (15.02) (14.03) (12.75) (11.68) (11.24)
Diff. max stress–no stress
All 71.41** 68.63** 62.09** 56.04** 50.42** 45.33** 40.81** 33.13*
Cuts only 123.41*** 122.51*** 114.75*** 105.69*** 97.36*** 90.83** 82.92** 79.30***
Countries 15 15 15 15 15 15 15 15
Observations 676 676 676 676 676 676 676 676
CDS
12345710Spreads
Unconditional projection
All 12.32** 11.77** 10.72** 9.77** 8.81* 7.95* 7.49* 5.79
(6.13) (5.90) (5.50) (5.05) (4.68) (4.21) (3.91) (3.60)
Cuts only 20.44** 20.18** 18.83** 17.36** 15.74** 15.37** 14.16** 13.37**
(10.12) (10.04) (9.33) (8.43) (7.83) (7.01) (6.12) (5.74)
Maximum stress
All 49.99** 49.97** 43.47** 39.33** 35.41** 31.86** 29.01** 23.22**
(21.36) (20.15) (18.26) (16.41) (15.02) (13.47) (12.13) (11.78)
Cuts only 79.74*** 78.94*** 73.76*** 67.89*** 62.23*** 58.66*** 53.67*** 50.87***
(25.51) (25.05) (22.98) (20.51) (18.95) (16.54) (14.63) (12.66)
No fiscal stress
All −21.42** −20.66** −18.63** −16.71** −15.01** −13.46** −11.80** −9.91*
(10.60) (10.17) (9.20) (8.25) (7.48) (6.79) (6.05) (5.86)
Cuts only −43.67*** −43.57*** −41.00*** −37.80*** −35.13** −32.17*** −29.26*** −28.42***
(16.83) (17.23) (16.43) (15.02) (14.03) (12.75) (11.68) (11.24)
Diff. max stress–no stress
All 71.41** 68.63** 62.09** 56.04** 50.42** 45.33** 40.81** 33.13*
Cuts only 123.41*** 122.51*** 114.75*** 105.69*** 97.36*** 90.83** 82.92** 79.30***
Countries 15 15 15 15 15 15 15 15
Observations 676 676 676 676 676 676 676 676

Response to unanticipated reduction of government spending growth by 1 percentage point (annualized). Estimates based on projection (2), with $h=0$ and using Oxford Economics forecasts. Driscoll and Kraay (1998) standard errors in parentheses. Significant at $***$1%, $**$5%, and $*$10%. The sample has been restricted to the country-time observations where CDS at all seven maturities as well as a bond-based spread measure are available. The reported number of observations is for the unconditional model. For the conditional model, we lose an additional two observations where we cannot construct the stress indicator. For CDS, columns represent maturities in years. For spreads, numbers are bond based.

### C. Adjustment Dynamics

So far, we have focused on the impact response of the default premium to fiscal shocks. In order to shed some light on the adjustment dynamics, we estimate the local projection (2) also for $h=1,⋯,8$. This means that we are considering the adjustment dynamics over a two-year horizon as we focus on the baseline forecasting model (Oxford Economics). In addition to the adjustment of the default premium over time, we study the response of other macroeconomic indicators to the fiscal shock.35

Figure 3 shows the estimated impulse response functions. The horizontal axis measures time in quarters. The vertical axis measures the deviation from the preshock level in percentage or basis points, depending on the variable. As before we distinguish initial conditions in terms of fiscal stress: the dashed line represents the response in times of maximum fiscal stress (with dotted lines indicating 90% confidence bounds based on Driscoll and Kraay (1998) standard errors), the solid line represents the response without stress (with shaded bands indicating 90 percent confidence bounds).
Figure 3.

Dynamic Adjustment to Fiscal Shock (Reduction of Government Consumption by 1 Percentage Point Annualized) in Times of No Fiscal Stress (Solid Line) and Maximum Stress (Dashed Line)

The horizontal axis measures quarters, and the vertical axis measures deviation from a preshock path in percentages and basis points (default premium). Shaded areas and dotted lines indicate 90% confidence bounds. Results based on Oxford Economics forecast sample where debt data are available.

Figure 3.

Dynamic Adjustment to Fiscal Shock (Reduction of Government Consumption by 1 Percentage Point Annualized) in Times of No Fiscal Stress (Solid Line) and Maximum Stress (Dashed Line)

The horizontal axis measures quarters, and the vertical axis measures deviation from a preshock path in percentages and basis points (default premium). Shaded areas and dotted lines indicate 90% confidence bounds. Results based on Oxford Economics forecast sample where debt data are available.

The upper-left panel shows the response of the default premium. For regime m, we observe that after the initial increase, the premium remains elevated. It declines only very gradually. Even toward the end of the two-year horizon under consideration, it is significantly above the preshock level. For regime $n$, we see that the decline of the premium is fairly persistent as well, but it is no longer significant at the end of the two-year horizon.

A natural question is how the default premium adjusts to the fiscal shock in the long run. However, local projections are not particularly well suited to address this question because it is costly (in terms of degrees of freedom) to expand the horizon for which impulse responses are estimated. In an earlier version of this paper, we estimated the effect of fiscal shocks on the default premium on the basis of a VAR model. The results for the short run are similar to those obtained on the basis of local projections. In particular, in the VAR, we also obtain an immediate increase of the premium in response to a reduction of government consumption provided fiscal stress is pervasive. And yet for the long run, the VAR predicts a significant decline of the default premium for regime m (Born et al., 2015).

The upper-right panel of figure 3 shows the response of real-time government consumption growth. It slows on impact and continues to be reduced relative to the preshock path for an extended period. The temporary decline of government consumption growth implies a reduction of the level of government consumption during the two-year horizon under consideration. The dynamics are fairly similar across the two regimes, even though the growth decline is somewhat stronger in the second and third quarters in regime m.

In order to understand the differential response of the premium across regimes, we turn to other macroeconomic variables. The sovereign default premium compensates investors for the probability that governments may repudiate part of their debt obligations. Recent models of sovereign default have highlighted the importance of two variables for a government's default decisions: output and the existing stock of debt. For this reason, we consider the impulse responses of both variables in the bottom panels of figure 3.

The lower-left panel displays the response of real-time output growth. Here, the dynamics are markedly different across regimes. Initially, growth slows in both regimes. However, if fiscal stress is absent, output growth rebounds quickly. After about one year, the initial effect on the output level is essentially undone. During times of maximum fiscal stress the dynamics are fundamentally different: output growth remains subdued for the entire horizon under consideration. Assuming an average government-consumption-to-GDP ratio of 20%, our estimates imply a cumulative fiscal multiplier on output for regime m of about 1 after one quarter and of about 2 after four quarters. These values are perhaps high but not unheard of (House, Proebsting, & Tesar, 2017; Ramey, 2011a).

In the lower-right panel, we show how the debt-to-GDP ratio evolves over time in response to a cut of government consumption. Here, we again observe stark differences across regimes. In the absence of stress, the reduction of government consumption entails a sharp and lasting decline of the debt ratio. Instead, if fiscal stress is maximal, the debt ratio rises in response to the shock. Put differently, we find evidence consistent with the view that austerity may be self-defeating at times. However, while Krugman (2010) outlines such a scenario for a liquidity trap, we provide evidence that the initial conditions in terms of fiscal stress are crucial in this regard. In a related recent study, Auerbach and Gorodnichenko (2017) find that fiscal stimulus in a slump can actually improve fiscal sustainability.

### D. Discussion

A complete model-based analysis of our empirical results is beyond the scope of this paper.36 Here we offer a brief discussion in light of available theoretical models. Our focus is on the behavior of the sovereign default premium. According to fundamental no-arbitrage considerations, it compensates investors for the probability of (partial) default. Models in the tradition of Eaton and Gersovitz (1981) predict that, all else equal, sovereign default is more likely the higher the level of public debt and the lower the level of output (Arellano, 2008; Hatchondo, Martinez, & Sapriza, 2010). Hence, our finding of how the default premium on the one hand and the debt ratio and output growth on the other hand co-move in response to spending cuts aligns well with theory—in both the presence and the absence of fiscal stress (see figure 3).

More specifically, a few recent studies explicitly consider exogenous variation of either taxes or government consumption in models of sovereign default. Arellano and Bai (2017) calibrate a model of optimal default to match key features of the Greek economy and study an exogenous tax increase during a debt crisis. Their main result is particularly relevant in light of our findings: austerity programs that increase distortionary taxes in times of fiscal stress can actually be self-defeating. They amplify the recession because higher taxes reduce the incentive to work. The default premium goes up as the recession deepens, reflecting stronger incentives for the government to default.

More closely related still is work by Bianchi, Ottonello, and Presno (2019). They study changes of government consumption and sovereign default in a two-sector small, open economy that operates inside a currency union in the presence of downward nominal wage rigidity. There is an exogenous endowment of tradable goods, while nontraded goods are produced using labor. The government consumes nontraded goods, which has a direct impact on production due to the nominal rigidity. Hence, an increase of government spending above the optimal level lowers unemployment. At the same time, it raises the default premium because spending is debt financed. Conversely, a cut of government consumption lowers the default premium—in line with our findings in the absence of fiscal stress (see panel 3 of table 2).

However, Bianchi et al. (2019) assume that taxes are lump sum. This rules out adverse budgetary effects from reduced production. Such an adverse budgetary effect is at the heart of the analysis by Corsetti et al. (2013). In their model, sovereign default is not an optimal decision of the government, but is likely to take place as public debt gets close to the “fiscal limit,” that is, the maximum level of debt that the government is able to service (Bi, 2012). Importantly, tax revenues move in proportion with output, and government consumption has a direct impact on output because of nominal rigidities. In this environment, a large fiscal multiplier obtains if monetary policy is constrained by the zero lower bound. A contractionary cut in government consumption then causes a decline of tax revenues that more than offsets the initial effect of reduced expenditures on the government budget. As a result, the default premium increases in response to a cut of government spending—in line with our findings for fiscal stress (see panel 2 of table 2).37

We briefly refer to work on the optimal adjustment of government consumption in the context of sovereign default (Cuadra, Sanchez, & Sapriza, 2010; Hatchondo, Martinez, & Roch, 2017). Because in our empirical analysis we identify exogenous variations of government consumption, we are silent on the optimality of such measures. In fact, a one-time cut of government consumption is unlikely to be the optimal way to implement an austerity program. In this context, results by Hatchondo et al. (2017) are remarkable, as they illustrate the benefits of credible medium-term consolidation strategies. Specifically, they study fiscal rules, which impose a ceiling on the government budget balance when either debt or the default premium is high. It turns out that such fiscal rules allow governments to forgo sharp fiscal adjustments because they provide an ex ante fiscal anchor that mitigates concerns about the sustainability of debt.

In sum, available theoretical models provide important insights into the dynamics of the sovereign default premium, both unconditionally and conditional on fiscal shocks. They also allow us to rationalize our empirical results to a considerable degree. However, to the best of our knowledge, a comprehensive investigation of how initial conditions in terms of fiscal stress affect the fiscal transmission mechanism and, eventually, the response of the default premium to austerity measures is still lacking. Our empirical results may provide a useful reference point for such an investigation.

## V. Conclusion

In this paper, we make two distinct contributions. First, we set up a new data set. It comprises observations for 38 emerging and advanced economies since 1991, notably for the sovereign default premium and three alternative measures of fiscal shocks. We provide a detailed description of the data and establish a number of basic facts. Second, we assess how the default premium responds to a cut of government consumption. In doing so, we account for initial conditions in terms of fiscal stress. We find that the premium rises in response to a cut of government consumption if stress is very severe but declines if stress is low. This result is robust across a variety of alternative specifications and shock measures. It also holds for specific country groups in our sample.

Our results have important implications for policy. If a government is confronted with a situation of severe fiscal stress, a reduction of government consumption is not rewarded by financial markets. In fact, financial markets may appear schizophrenic about austerity in that they demand austerity measures as public debt builds up, but fail to reward them as austerity slows output growth (Blanchard, 2011; Cotarelli & Jaramillo, 2013). Under these circumstances, a commitment to a credible medium-term strategy might be more promising, along the lines of our discussion in section IVD. Moreover, to the extent that fiscal stress builds up gradually over time, our results suggest that austerity can actually pay off if implemented in a sufficiently timely manner. A main conclusion of our analysis is thus that delaying fiscal consolidations can be particularly harmful.

## Notes

1

See figure A.2 in the online appendix.

2

Some authors find that whether austerity is tax based or spending based is crucial for its impact on the economy (for recent contributions, see Alesina & Ardagna, 2013; Alesina, Barbiero, et al., 2015; Alesina, Favero, & Giavazzi, 2015).

3

In an earlier working paper version, we also consider VAR models. The results are very similar to those obtained with local projections (see Born, Müller, & Pfeifer, 2015).

4

Still, in a robustness check, we also include a vector of controls, $Xi,t-1$, which features lags of GDP growth, government spending growth, and the default premium.

5

We check the robustness of our results by using a parametric logistic transition function and find that they are robust. See table A.3 in the online appendix.

6

The estimated shocks $ɛ^i,tg$ in this third specification are generated regressors in the second stage. However, as Pagan (1984) showed, the standard errors on the generated regressors are asymptotically valid under the null hypothesis that the coefficient is 0. See also Coibion and Gorodnichenko (2015), note 18, on this point.

7

Hence, a change in cash flows that results from, say, a government deferring payments during a crisis is not recorded as a change of government consumption in our data set.

8

See the discussion in online appendix A.2.2.

9

We focus on long-term rates whenever possible. As they are closely linked to the average of expected future short-term rates, they are a more appropriate measure of governments' refinancing costs than short-term rates. We investigate the issue of different maturities in section A.4.2 in the online appendix.

10

In principle, spreads may also reflect a liquidity premium, an issue we ignore in what follows because we consider government debt traded in mature markets. See online appendix A.1.4 for a more detailed discussion.

11

See online appendix A.1.1 for details.

12

Before the Lehman Brothers default, German and U.S. CDS were below 8 basis points and thus virtually 0. After Lehman, they peak at about 70 basis points and slowly return to about 15 basis points.

13

In order to construct the CDF, we rely on the broadest possible sample, that is, we do not restrict the default premiums to country-quarter observations where we have also have government spending forecast errors available (which would also differ across the three forecasting approaches).

14

Results are unchanged when using the average over the quarter.

15

The reason is that the long-term convergence yields are sometimes slightly lower than the German ones. This is presumably due to their construction not controlling for different bond duration characteristics and small maturity differences.

16

During default episodes, spreads in secondary markets can achieve even higher values. In the case of Argentina, the peak spread was 70 percentage points.

17

This is the case for Finland before 1999Q2, Ireland before 2004Q1, and Sweden before 1998Q2.

18

Some studies have resorted to annual data (Beetsma, Giuliodori, & Klaasen, 2006, 2008; Bénétrix & Lane, 2013). In this case identification assumptions tend to be more restrictive. However, Born and Müller (2012) consider both quarterly and annual data for four OECD countries. They find that the estimated effects of government spending shocks barely differ.

19

For several European countries, we also include earlier observations for the 1990s whenever we are able to compute a default premium (see online appendix A.2.1.

20

We run a regression of the forecast error on a constant and evaluate whether the coefficient is statistically significantly different from 0.

21

Technically, given our panel structure with potentially non-i.i.d. errors, we follow the suggestion in Baum, Schaffer, and Stillman (2007) and check the predictive power of our identified shocks using the Kleibergen and Paap (2006) $rk$ Wald $F$-statistic. It is computed in a first-stage panel fixed-effects regression of the government consumption growth variable on the respective shock measure. Computing naive $F$-statistics in our pooled sample yields very similar values.

22

The Montiel Olea and Pflueger (2013) threshold for the 5% critical value for testing the null hypothesis that the 2SLS bias exceeds 10% of the OLS bias in our context is 23.1.

23

In principle, a further reduction is possible only to the extent that even country-time observations like those for the United States in the mid-2000s were not entirely risk free. This is unlikely to rationalize our estimate, which is quantitatively nontrivial.

24

In order to ensure that results are quantitatively comparable to the baseline, we adjust our estimates: they reflect the response of the default premium to a change of government spending equivalent to a reduction of its growth rate by 1 percentage point (annualized).

25

The larger RMSE reported in table 1 indicates that professional forecasts indeed have a more complete information set. However, because our VAR model includes an inherently forward-looking variable, the sovereign default premium, potential problems due to fiscal foresight should at least be somewhat mitigated (Sims, 2012).

26

We include four lags. In this ways, we seek to establish the effect of an innovation in the common factor on country-specific variables.

27

We exclude the United States in the second stage as in this case, the assumption is questionable.

28

It is not a genuine test because it rests on an identification assumption that is itself untested: that the common factor is contemporaneously unaffected by the country-specific developments.

29

Thus, for the first month $m$ in a quarter $q$, the news refer to the revision in the forecast for quarter $q$, made in month $m-1$, which belonged to quarter $q-1$. For the other two months in the quarter $q$, the fiscal news refer to the revision in the nowcast for quarter $q$ made in the previous month belonging to the same quarter.

30

Details on the data used can be found in online appendix A.4.1.

31

Interestingly, we find that the response of the default premium is similarly sensitive to the state of the business cycle. If we condition the effect of spending cuts on booms and recessions as in Auerbach and Gorodnichenko (2012), we find that the premium rises during recessions and tends to decline during booms. Note that the recession indicator and the indicator for fiscal stress show quite distinct dynamics. See figures A.3 to A.5 in the online appendix.

32

As a practical matter, we discard CDS spread observations that are bigger than 40%. This threshold exceeds the largest value of the bond-based spread measure by 15 percentage points. Observations above this threshold are sometimes observed for periods of extreme stress during which price movements may be somewhat erratic.

33

In online appendix A.4.2, we also discuss a related issue on the distinction between an increase of default risk (the quantity of risk) or an increase of the “risk premium” (the price of risk).

34

Our findings regarding a country-specific government consumption shock are consistent with Augustin (2018), who reports that inverted CDS spread curves coincide with periods where country-specific factors as opposed to global factors drive the CDS spreads.

35

Our sample is somewhat restricted relative to section IVA because data for government debt are available only for a subset of countries.

36

In the working paper version of the paper, we rationalize our findings on the basis of model simulations (Born et al., 2015). For this purpose, we consider a variant of the model put forward by Arellano (2008).

37

Corsetti et al. (2013) also consider a “sovereign risk channel” whereby a higher sovereign default premium raises borrowing costs in the private sector. If this channel is strong, spending cuts tend to reduce the default premium.

## REFERENCES

Akitoby
,
Bernardin
, and
Thomas
Stratmann
, “
Fiscal Policy and Financial Markets,
Economic Journal
118
(
2008
),
1971
1985
.
Alesina
,
Alberto
, and
Silvia
Ardagna
, “The Design of Fiscal Adjustments,” (pp.
19
68
), in
Jeffrey R.
Brown
, ed.,
Tax Policy and the Economy
(
Chicago
:
University of Chicago Press
,
2013
).
Alesina
,
Alberto
,
Omar
Barbiero
,
Carlo
Favero
,
Francesco
Giavazzi
, and
Matteo
Paradisi
, “
Austerity in 2009–2013,
Economic Policy
30
(
2015
),
383
437
.
Alesina
,
Alberto
,
Carlo
Favero
, and
Francesco
Giavazzi
, “
The Output Effect of Fiscal Consolidation Plans,
Journal of International Economics
96
(
2015
),
S19
S42
.
Ardagna
,
Silvia
, “
Financial Markets' Behavior around Episodes of Large Changes in the Fiscal Stance,
European Economic Review
53
(
2009
),
37
55
.
Arellano
,
Cristina
, “
Default Risk and Income Fluctuations in Emerging Economies,
American Economic Review
98
(
2008
),
690
712
.
Arellano
,
Cristina
, and
Yan
Bai
, “
Fiscal Austerity during Debt Crises,
Economic Theory
64
(
2017
),
657
673
.
Auerbach
,
Alan J.
, and
Yuriy
Gorodnichenko
, “
Measuring the Output Responses to Fiscal Policy,
American Economic Journal: Economic Policy
4
(
2012
),
1
27
.
Auerbach
,
Alan J.
, and
Yuriy
Gorodnichenko
“Fiscal Multipliers in Recession and Expansion,” (pp.
63
98
), in
Alberto
Alesina
and
Francesco
Giavazzi
, eds.,
Fiscal Policy after the Financial Crisis
(
Chicago
:
University of Chicago Press
,
2013
).
Auerbach
,
Alan J.
, and
Yuriy
Gorodnichenko
Fiscal Stimulus and Fiscal Sustainability
,” NBER working paper
23789
(
2017
).
Augustin
,
Patrick
, “
The Term Structure of CDS Spreads and Sovereign Credit Risk,
Journal of Monetary Economics
96
(
2017
),
53
76
.
Baum
,
Christopher F.
,
Mark E.
Schaffer
, and
Steven
Stillman
, “
Enhanced Routines for Instrumental Variables/Generalized Method of Moments Estimation and Testing,
Stata Journal
7
(
2007
),
465
506
.
Beetsma
,
Roel
,
Massimo
Giuliodori
, and
Franc
Klaasen
, “
Trade Spill-Overs of Fiscal Policy in the European Union: A Panel Analysis,
Economic Policy
48
(
2006
),
640
687
.
Beetsma
,
Roel
,
Massimo
Giuliodori
, and
Franc
Klaasen
The Effects of Public Spending Shocks on Trade Balances in the European Union,
Journal of the European Economic Association
6
(
2008
),
414
423
.
Bénétrix
,
Agustín S.
, and
Philip R.
Lane
, “
Fiscal Shocks and the Real Exchange Rate,
International Journal of Central Banking
9
:
3
(
2013
),
6
37
.
Bi
,
Huixin
, “
Sovereign Default Risk Premia, Fiscal Limits, and Fiscal Policy,
European Economic Review
56
(
2012
),
389
410
.
Bianchi
,
Javier
,
Pablo
Ottonello
, and
Ignacio
Presno
, “
Fiscal Stimulus Under Sovereign Risk
,” NBER working paper
26307
(
2019
).
Blanchard
,
Olivier J.
, “
2011 in Review: Four Hard Truths,
iMF
(blog),
December 21
,
2011
, http://blog-imfdirect.imf.org/2011/12/21/2011-in-review-four-hard-truths/.
Blanchard
,
Olivier J.
, and
Roberto
Perotti
, “
An Empirical Characterization of the Dynamic Effects of Changes in Government Spending and Taxes on Output,
Quarterly Journal of Economics
117
(
2002
),
1329
1368
.
Born
,
Benjamin
, and
Gernot J.
Müller
, “
Government Spending in Quarterly and Annual Time Series,
Journal of Money, Credit and Banking
44
(
2012
),
507
517
.
Born
,
Benjamin
,
Gernot J.
Müller
, and
Johannes
Pfeifer
, “
Does Austerity Pay Off?
” CEPR discussion paper
10425
(
2015
).
Broner
,
Fernando A.
,
Guido
Lorenzoni
, and
Sergio L.
Schmukler
, “
Why Do Emerging Economies Borrow Short Term?
Journal of the European Economic Association
11
:
S1
(
2013
),
67
100
.
Calvo
,
Guillermo A.
, “
Servicing the Public Debt: The Role of Expectations,
American Economic Review
78
(
1988
),
647
661
.
Coibion
,
Olivier
, and
Yuriy
Gorodnichenko
, “
Information Rigidity and the Expectations Formation Process: A Simple Framework and New Facts,
American Economic Review
105
(
2015
),
2644
2678
.
Cole
,
Harold L.
, and
Timothy J.
Kehoe
, “
Self-Fulfilling Debt Crises,
Review of Economic Studies
67
(
2000
),
91
116
.
Corsetti
,
Giancarlo
,
Keith
Kuester
,
André
Meier
, and
Gernot J.
Müller
, “
Sovereign Risk, Fiscal Policy, and Macroeconomic Stability,
Economic Journal
123
(
2013
),
F99
F132
.
Corsetti
,
Giancarlo
,
André
Meier
, and
Gernot J.
Müller
, “
What Determines Government Spending Multipliers?
Economic Policy
72
(
2012
),
521
565
.
Cotarelli
,
Carlo
, and
Laura
Jaramillo
, “
Walking Hand in Hand: Fiscal Policy and Growth in Advanced Economies,
Review of Economics and Institutions
4
:
2
(
2013
),
1
25
.
Cuadra
,
Gabriel
,
Juan M.
Sanchez
, and
Horacio
Sapriza
, “
Fiscal Policy and Default Risk in Emerging Markets,
Review of Economic Dynamics
13
(
2010
),
452
469
.
Devries
,
Pete
,
Jaime
Guajardo
,
Daniel
Leigh
, and
Andrea
Pescatori
, “
A New Action-Based Dataset of Fiscal Consolidation
,” IMF working paper
11/128
(
2011
).
Driscoll
,
John C.
, and
Aart C.
Kraay
, “
Consistent Covariance Matrix Estimation with Spatially Dependent Panel Data,
this review
80
(
1998
),
549
560
.
Eaton
,
Jonathan
, and
Mark
Gersovitz
, “
Debt with Potential Repudiation: Theoretical and Empirical Analysis,
Review of Economic Studies
48
(
1981
),
289
309
.
Giavazzi
,
Francesco
,
Tullio
Jappelli
, and
Marco
Pagano
, “
Searching for Non-Linear Effects of Fiscal Policy: Evidence from Industrial and Developing Countries,
European Economic Review
44
(
2000
),
1259
1289
.
Guajardo
,
Jaime
,
Daniel
Leigh
, and
Andrea
Pescatori
, “
Expansionary Austerity: New International Evidence
,” IMF working papers
11/158
(
2011
).
Hatchondo
,
Juan Carlos
,
Leonardo
Martinez
, and
Francisco
Roch
, “
Fiscal Rules and the Sovereign Default Premium,
University of Western Ontario Mimeograph
(
2017
).
Hatchondo
,
Juan Carlos
,
Leonardo
Martinez
, and
Francisco
Roch
, and
Horacio
Sapriza
, “
Quantitative Properties of Sovereign Default Models: Solution Methods Matter,
Review of Economic Dynamics
13
(
2010
),
919
933
.
House
,
Christopher L.
,
Christian
Proebsting
, and
Linda L.
Tesar
, “
Austerity in the Aftermath of the Great Recession
,” NBER working paper
23147
(
2017
).
Ilzetzki
,
Ethan
,
Enrique G.
Mendoza
, and
Carlos A.
Végh
, “
How Big (Small?) Are Fiscal Multipliers?
Journal of Monetary Economics
60
(
2013
),
239
254
.
Jordá,
Òscar
, “
Estimation and Inference of Impulse Responses by Local Projections,
American Economic Review
95
(
2005
),
161
182
.
Juvenal
,
Luciana
, and
Kevin
Wiseman
, “
Portugal's Regained Market Access: Opportunities and Risks,
Portugal—Selected Issues, IMF Country Report 15/127
(
2015
).
Kilian
,
Lutz
, and
Robert J.
Vigfusson
, “
Are the Responses of the U.S. Economy Asymmetric in Energy Price Increases and Decreases?
Quantitative Economics
2
(
2011
),
419
453
.
Kleibergen
,
Frank
, and
Richard
Paap
, “
Generalized Reduced Rank Tests Using the Singular Value Decomposition,
Journal of Econometrics
133
(
2006
),
97
126
.
Krishnamurthy
,
Arvind
,
Stefan
Nagel
, and
Annette
Vissing-Jorgensen
, “
ECB Policies Involving Government Bond Purchases: Impact and Channels,
Review of Finance
22
(
2018
),
1
44
.
Krugman
,
Paul
, “
Self-Defeating Austerity,
New York Times
,
July 7
,
2010
. https://krugman.blogs.nytimes.com/2010/07/07/self-defeating-austerity/.
Laubach
,
Thomas
, “
New Evidence on the Interest Rate Effects of Budget Deficits and Debt,
Journal of the European Economic Association
7
(
2009
),
858
885
.
Leeper
,
Eric M.
,
Todd B.
Walker
, and
Shu-Chun Susan
Yang
, “
Fiscal Foresight and Information Flows,
Econometrica
81
(
2013
),
1115
1145
.
Ljung
,
Greta M.
, and
George E. P.
Box
, “
On a Measure of Lack of Fit in Time Series Models,
Biometrika
65
(
1978
),
297
303
.
Longstaff
,
Francis A.
,
Jun
Pan
,
Lasse H.
Pedersen
, and
Kenneth J.
Singleton
, “
How Sovereign is Sovereign Credit Risk?
American Economic Journal: Macroeconomics
3
(
2011
),
75
103
.
Lorenzoni
,
Guido
, and
Iván
Werning
, “
Slow Moving Debt Crisis,
Northwestern University mimeograph
(
2018
).
Montiel Olea
,
José Luis
, and
Carolin
Pflueger
, “
A Robust Test for Weak Instruments,
Journal of Business and Economic Statistics
31
(
2013
),
358
369
.
Owyang
,
Michael T.
,
Valerie A.
Ramey
, and
Sarah
Zubairy
, “
Are Government Spending Multipliers Greater during Periods of Slack? Evidence from Twentieth-Century Historical Data,
American Economic Review, Papers and Proceedings
103
(
2013
),
129
134
.
Pagan
,
Adrian
, “
Econometric Issues in the Analysis of Regressions with Generated Regressors,
International Economic Review
25
(
1984
),
221
247
.
Perotti,
Roberto
, “
Fiscal Policy in Good Times and Bad,
Quarterly Journal of Economics
114
(
1999
),
1399
1436
.
Pesaran
,
M. Hashem
, and
Ron
Smith
, “
Estimating Long-Run Relationships from Dynamic Heterogeneous Panels,
Journal of Econometrics
68
(
1995
),
79
113
.
Ramey
,
Valerie A.
, “
Can Government Purchases Stimulate the Economy?
Journal of Economic Literature
49
(
2011a
),
673
685
.
Ramey
,
Valerie A.
Identifying Government Spending Shocks: It's All in the Timing,
Quarterly Journal of Economics
126
:
1
(
2011b
),
1
50
.
Ramey
,
Valerie A.
, and
Sarah
Zubairy
, “
Government Spending Multipliers in Good Times and in Bad: Evidence from US Historical Data,
Journal of Political Economy
126
(
2018
),
850
901
.
Roch
,
Francisco
, and
Harald
Uhlig
, “
The Dynamics of Sovereign Debt Crises and Bailouts,
Journal of International Economics
114
(
2018
),
1
13
.
Romer
,
Christina D.
, and
David H.
Romer
, “
The Macroeconomic Effects of Tax Changes: Estimates Based on a New Measure of Fiscal Shocks,
American Economic Review
100
(
2010
),
763
801
.
Sims
,
Eric
, “
News, Non-Invertibility and Structural VARs,
Advances in Econometrics
28
(
2012
),
81
135
.
Staiger
,
Douglas
, and
James H.
Stock
, “
Instrumental Variables Regression with Weak Instruments,
Econometrica
65
(
1997
),
557
586
.
Trabandt
,
Mathias
, and
Harald
Uhlig
, “
The Laffer Curve Revisited,
Journal of Monetary Economics
58
(
2011
),
305
327
.

## Author notes

We thank Yuriy Gorodnichenko (the editor), as well as three anonymous referees. We also thank our discussants Nicola Fuchs-Schündeln, Alessandro Gioffré, Klemens Hauzenberger, Josef Hollmayr, and Tomasz Wieladek, as well as Kerstin Bernoth, Florian Kirsch, Helmut Lütkepohl, Enrique Mendoza, Valerie Ramey, Almuth Scholl, and various seminar audiences, for very useful comments and discussions. Andreas Born, Marc Faupel, Alexander Scheer, Diana Schüler, and Susanne Wellmann have provided excellent research assistance. We gratefully acknowledge research support from the Research Center SAFE, funded by the State of Hessen initiative for research LOEWE. G.M. also thanks the German Science Foundation (DFG) for financial support under the Priority Program 1578. The usual disclaimer applies.

A supplemental appendix is available online at http://www.mitpressjournals.org/doi/suppl/10.1162/rest_a_00844.