Fiscal and Monetary Interactions in the Eurozone with Real Time Data

This version Sept 08:

Preliminary and Incomplete

Do not cite

John Lewis

De Nederlandsche Bank

Andrew Hughes Hallett

GeorgeMasonUniversity & CEPR

Abstract

Using real time data, we estimate reaction fiscal and monetary reaction functions for the eurozone both individually and as a system. We find that monetary policy does respond to fiscal stance, but that the reverse does not hold. A loosening of the cyclically adjusted budget balance by one percentage point prompts a monetary tightening of 37 basis points. Estimating equivalent reaction functions based on ex post data yields a different story, and disguises the central bank’s reaction to fiscal policy.

JEL Codes: E63 (Comparative or Joint Analysis of Monetary and Fiscal Policy)

E61 (Policy Designs and Consistency, Policy Coordination)

Keywords:Policy co-ordination, Fiscal Policy, Monetary Policy, Real Time Data

The authors would like to thank…. for useful comments. Part of this paper was written during John Lewis visit to the Robert Schuman Centre for Advanced Studies(European University Insitutute) under the auspices of the Pierre Warner Chair programme, which is gratefully acknowledged. The empirical section benefited from the helpful comments of Kerstin Bernoth and Steven Poelhekke. The views expressed are those of the authors and not necessarily those of the institutions they are affiliated to.
1. Introduction

In recent years, there has been a growing appreciation that analyses of policymakers behaviour need to consider the data the policymaker had at the time (real time data), as opposed to the revised data available many years hence (ex post data). As Orphanides (2001) points out, any policy rule based on ex post data could not have been implemented by the policymaker since it relies of information the policymaker did not have. Accordingly, it cannot be interpreted a description of what the policymaker was trying to do. By the same token, that suggests empirical estimates of policymakers “reaction functions”, should be formulated in terms of the real time data that the policymaker could have reacted to.

Since Orphanides’ groundbreaking article there has been a proliferation of papers examining the behaviour of monetary authorities based on real time data. In recent years, a similar literature has sprung up in the field of fiscal policy. However, to our knowledge, there are no papers which have examined the interactions betweens fiscal and monetary policy based on real time data.

Clearly Orphanides real-time critique applies equally to empirical characterisations of monetary-fiscal interactions. Any descriptive account of policymaking which aims to capture what a policymaker was trying to do, must be conditioned on data available at the time. Aside from the well known problems of real-time output gap measurement, several papers have chronicled similar real-time data issues with budget deficits and their cyclically adjusted counterparts (Jonung & Larch, 2003; Hughes Hallett et al, 2007).

The goal of this paper is to analyse fiscal-monetary interactions in the eurozone, using real time data. We do this by estimating reaction functions, where each policymaker is permitted (but not required) to respond to the other policymakers instrument, as well as economically relevant target variables such as the output gap, inflation and the debt ratio.

There exists a relatively large theoretical literature on the optimal interaction of fiscal and monetary policy, and an empirical literature (based on ex post data), which seeks to describe these interactions in practice. A natural extension to these theoretical literatures is to examine how fiscal and monetary policymakers have interacted with each other based on real time data.

On the monetary side, the literature suggests that the distinction between real time and ex post data can have important implications for characterisations of policymaker’s behaviour. Orphanides (2001) finds that a taylor rule fitted with real time data yields a better to fit to observed Fed behaviour than its ex post counterpart. In a similar paper, Orphanides (2002) also shows that during the great inflation, it was persistent mismeasurement of the output gap, rather than a change in policymakers’ preferences that was the cause of looser monetary policy. Geberding et al (2003), are able to resolve to puzzling finding that Bundesbank reaction functions appeared not to have a significant response to monetary growth. By re-estimating the same reaction functions with real time data, a strong reaction to monetary variables is uncovered. A number of papers have also attempted to estimate reaction functions for the ECB[1]. Several have also attempted to do this using real time data (Gerdesmeier and Roffia, 2004; Sauer and Sturm, 2007; Gerlach, 2007; Gorter et al 2007: Castelnuovo, 2006). A general finding of such papers is that the relatively good fit of the simple Taylor rule also carries over to eurozone data.

Papers which explicitly contrast the performance of real time versus ex post data in fitting Taylor rulesfind that the real time based rules typically tend to yield more “plausible” co-efficients. Shifting from ex post to real-time data is often crucial in pushing the coefficient on inflation above unity- and thus satisfying the “Taylor principle”.[2]

On the fiscal side, the emerging literature also finds an important distinction between the accounts of behaviour derived from real data as compared to those using ex post data. As well as mismeasurement of the state variable (the output gap), the policy instrument (budget balance, possibly cyclically adjusted) is also subject to measurement error in real time, which adds an additional dimension to the problem. Hughes Hallett et al (2007) show that cyclically adjusted budget balances are subject to considerable revisions over time.

Momigliano and Forni (2006), Cimadomo (2007) and Bernoth et al (2008) all find that in real time, governments seek to use discretionary fiscal policy in a counter-cyclical way, but that when the same reaction functions are estimated with ex post data, the result is acyclical. Von Kalckreuth and Wolff (2007) find that governments seek to adjust spending in a counter-cyclical way, and are able to do so within the space of a quarter. Beetsma and Giuliodori (2008), find evidence of eurozone fiscal authorities reacting to each others plans based on real time data, but do not test explicitly if this result also shows up in ex post data.

Balboni et al (2007) present a model of fiscal monetary interactions in real time when the policymakers themselves disagree about the output gap. However, this does not attempt an empirical characterisation of the observed interactions. For the US, Claeys (2003) estimates fiscal and monetary reaction functions which include the other policymakers instrument as an argument. The present paper is complementary to this paper, in the sense that it performs a similar exercise for the Eurozone (rather than the US), and on the basis of real time (as opposed to ex post) data.

This paper makes a number of contributions to the literature. First, it extends the existing empirical work on fiscal monetary interactions by making use of real time data. In this way, it get closer to “what policymakers were trying to do” than the work which utilises ex post data. Second, it tests whether some of the previous results on fiscal and monetary policymakers behaviour are robust to the inclusion of the “other” policymakers behaviour in the reaction function. Third, it provides specific evidence on the eurozone, in contrast to most of the other literature which tends to focus on the US.

  1. Dataset

There is no single eurozone dataset available for all our relevant variables. The Euro Area Real Time Database is the single most complete dataset, but the vintages only begin in 2001 and some only run up to 2006. For that reason, to obtain data for a longer time period it was necessary to compile out dataset independently, using data from several sources. In all cases our data is at the quarterly frequency.

The bulk of the real time time data is taken from successive editions of the OECD’s economic outlook from December 1994 (No 56) onwards. It is similar to that used by Hughes Hallett et al (2007) and Bernoth et al (2008). The variables from this source are the primary balance, total balance (and their cyclically adjustedcounterparts), government debt and the output gap.

Economic Outlook is published twice per year- one edition in June and one in edition December. The published values of the variables are all on a yearly basis[3]. To derive our quarterly data, we take the latest available vintage at the start of a given quarter, and then perform the Lisman method[4]to interpolate quarterly values for the whole time series. This procedure generates the property that the average of the four quarterly figures equals the annual figure from the official data.

Economic Outlook does not report eurozone figures for the whole period. Therefore, we construct our own euerozone data, based on a weighted average of national data. Weights are determined by the nominal GDP (in millions of euro) of each country[5]. In each case, we use a vintage of GDP which matches the vintage of the variable being measured- e.g real time budget deficits are weighted according to real time GDP, ex post budget deficits are weighted using ex post GDP and so on. Economic Outlook does not report figures across the whole period for Luxembourg, Slovenia, Malta and Cyprus and therefore, these countries are effectively assigned a weight of zero in our analysis. However, the bias from excluding these countries from the construction of our eurozone data is extremely small, since they account for around 1% of Eurozone GDP (and for most of the sample, only Luxembourg was an EMU member).

For monetary policy, the policy instrument is the ECB repo rate. Data is taken from Eurostat.. There is no distinction between real time and ex post data here, since the observation of the discount rate in real time is not subject to any measurement error.

Inflation expectations data is taken from Consensus Forecasts. This is a monthly survey of over 200 forecasters, who report inflation expectations for around 20 countries. Participants are asked to forecast year end inflation for the current year and the next year- i.e. in December of each year. To generate a forecast for inflation in the intermediate months, we follow a number of authors[6] in performing linear interpolation. This of course only provides a proxy for “true” inflation expectations, but nevertheless respects the “real-time principle” of restricting our information set to information known to policymakers at the time. The eurozone figure is obtained by taking a weighted average of the national figures using eurostat’s yearly HICP country weights[7]. Consensus Forecasts do not collect data on Luxembourg, Slovenia, Malta and Cyprus, therefore our “eurozone” figure exclude these countries. However, since they have a combined weight of around 1% in the HICP, our bias from excluding them is likely to be very small.

Data on inflation itself was taken from from Eurostat, using the year on year change in the HICP. Given that initial releases are seldom revised (Coenen et al 2003), the real time data and ex post data for current inflation are largely the same[8], although there is typically a lag of around 2 months in the reporting of inflation figures. In any case, in the bulk of the analysis, we assume monetary policy is set on a forward looking basis, and hence the inflation data that we use are forecasts, rather than contemporaneous inflation data.

Figure 1 compares the real time, ex post and 1 year forecast of the output gap. The forecast variable is lagged by one year, so that the figure reported for year X quarter Q is the forecast made at X-1:Q, for the variable at time X:Q.

Figure 1: Data Across Vintages (percentage points)

Looking at the output gap (left hand panel), it is evident that compared to the ex post data the real time figures (and the 1 year forecast) underestimated the extent of the boom in the first half of the sample, and were overly pessimistic during the recovery in the latter years of the sample. Similarly, the real time CAB figures failed to pick up the substantial fiscal loosening in the early part of the sample, and were sluggish in picking up the improvement in public finances later on. Taken together, these graphs provide prima facie evidence that empirical characterisation of monetary and fiscal policy using ex post data may differ substantially from those which use only the data available to the policymaker at the time.

  1. Empirical Estimates of Reaction Functions

Monetary Policy

To capture the behaviour of the ECB we estimate a canonical Taylor rule of the form:

/ (1)

where it is the policy rate, tis the rate of inflation[9], yis the output gap and z is a vector of additional variables. k captures the policy horizon of the central bank: k=0 means the authorities respond to contemporaneous data, k>0 implies forward looking behaviour. The parameter, , captures the degree of “gradualism” or “inertia” in monetary policy.

Fiscal Policy

In generic form, the aggregate fiscal policy of the eurozone is of the form:

/ (2)

where bal is a measure of the fiscal balance. The precise measures of fiscal balance used varies across the literature: cyclically adjusted primary balance, capb, the (unadjusted) primary balance, pb, the actual budget balance b and the cyclically adjusted balance b. The time horizon of the fiscal authorities is captured by m and their inertia by .

Alternatively, one may analyse the fiscal policy of individual countries within the eurozone:

/ (3)

3.1 Econometric Considerations

We begin by considering the persistence properties of individual variables (for a tabulation of results, see table A1 of the appendix). An Augmented Dickey- Fuller (ADF) test cannot reject non-stationarity in any of the variables. However, given the low power of the ADF test when the autoregressive parameter is close to but below one (as may be the case with these variables), we follow Claeys (2003), Österholm (2003) and others , by also undertaking a KPSS test (Kwiatwoski et al,1992). The KPSS test was carried out with a Bartlett kernel where the bandwidth parameter was automatically selected following the Newey-West procedure (Newey and West, 1994). The test was run both with and without a trend term.

For the output gap and inflation, the KPSS test cannot reject stationarity, and for the interest rate the KPSS tests rejects at the 1 and 5% levels. For the fiscal variables the picture varies across vintages- some vintages show some signs of a unit root, but other vintages of the same variable do not. But in no case do all three tests suggest non-stationarity for a given time series, and for no variable are the results consistent across the three vintages. Taken together, the results do not provide clear cut evidence of non-stationarity in the time series.

As a further check, a visual plot of all series suggests non-stationarity. Furthermore, sound theoretical arguments can be presented as to why these variables should be stationary. Therefore, we proceed to model these variables as persistent but stationary.

In common with the bulk of the literature, the reaction function equations were estimated using the (two stage) Generalised Method of Moments. This overcomes the problem of potential correlation between explanatory variables and the residual term. We reported Newey-West Heteroscedasticity and Autocorrelation corrected (HAC) standard errors. All regressions use a Bartlett kernel with a bandwidth of 3. This is consistent with the literature[10] and is also motivated by the suggestion of Green (2003) of using T1/4. In any event, the results are robust to changes in the bandwidth parameter.

As instruments, we employ one to four lags of the (real time) inflation and output gap series, and one to four lags of the year on year percentage change in the euro-dollar exchange rate. The j-statistic is reported for each regression, and in each case exogeneity is strongly supported.

Favourable results for tests of the exogeneity of instruments are a necessary condition, but it is also important that instruments are “relevant”- i.e. that the correlation between the instruments and explanatory variables is high. Stock and Yogo (2002) argue that many applications of GMM and IV suffer from the problem of weak (but nevertheless exogenous) instruments. If instruments are of low relevance, then not only do standard aysmptotics fail, but the asymptotic standard errors are increased and hence the power of hypothesis tests is reduced.

As Staiger and Stock (1997) and others have demonstrated, the weak instrument problem can be present even when first stage F-tests are significant at conventional levels.

Whilst the Hansen test can detect exogeneity, it cannot say anything about the relevance of instruments. Moreover, the Hansen test is a somewhat “permissive” criterion in the sense that it permits quite a wide range of instruments, which may yield contradictory estimation results.

To ensure that our selected instruments are relevant, we employ several criteria. We start with the weak instruments test of Stock and Yogo (2002). This is based on two definitions of weak instruments, and yields two formal tests. The first tests for the bias of the IV estimator relative to OLS; the second tests whether conventional Wald test on IV statistics has a size that could exceed a certain threshold. For the sake of brevity we report the only the outcome of the first of these tests for each regression.[11]

We also use the partial R2 measure of Shea (1997) to measure the relevance of our instruments. This test allows for possible collinearity of instruments, which may be an issue when a number of lags of the same variable are used as instruments.[12] For our instruments to be relevant, the partial R2 should be also be large. For results of this, see appendix