- Email: [email protected]
The effect of inflation targeting on the behavior of expected inflation: evidence from an 11 country panel
Journal of Monetary Economics 49 (2002) 1521–1538
The effect of inflation targeting on the behavior of expected inflation: evidence from an 11 country panel$ David R. Johnson* Department of Economics, School of Business and Economics, Wilfrid Laurier University, 75 University Avenue West, Waterloo, Ont., Canada, N2L 3C5 Received 7 December 1999; received in revised form 12 September 2001; accepted 15 November 2001
Abstract Inflation targets were introduced in the 1990s in Australia, Canada, New Zealand, Sweden and the United Kingdom. Change in the behavior of expected inflation in the five targeting countries is measured in a panel that includes six non-targeting countries: France, Germany, Italy, the Netherlands, Japan and the United States. The level of expected inflation in targeting countries falls after the announcement of targets with controls for country effects, year effects, ongoing inflation reduction and the business cycle. Neither the variability of expected inflation nor the average absolute forecast error falls after the announcement of targets with controls for the level and variability of past inflation. Both targeting and non-targeting countries experience unexpected disinflation. r 2002 Elsevier Science B.V. All rights reserved. JEL classification: E52; E58 Keywords: Inflation targets; Expected inflation
$ I want to thank Nathan Braun for excellent research assistance and Susan Johnson, Pierre Siklos, Angelo Melino and the referees and editor for their comments. The support of an Interdisciplinary Grant from the Social Sciences and Humanities Research Council of Canada, administered by the Wilfrid Laurier University Research Office, is acknowledged. I also thank the Faculty of Economics and Politics at the University of Cambridge, England for their hospitality when this project began. *Corresponding author. Fax: +1-519-888-1015. E-mail address: [email protected] (D.R. Johnson).
0304-3932/02/$ - see front matter r 2002 Elsevier Science B.V. All rights reserved. PII: S 0 3 0 4 - 3 9 3 2 ( 0 2 ) 0 0 1 8 1 - 2
1522
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1. Introduction In the 1990s a number of countries used announcements of explicit numerical targets for inflation as a key component of monetary policy. Inflation targets, and other monetary reforms often implemented at the same time, were adopted for three reasons: to communicate more clearly the goals of monetary policy; to create accountability for the achievement of these goals and to alter expectations of inflation.1 This paper measures change after the announcement of targets in three aspects of the behavior of expected inflation: the level of expected inflation; the standard deviation of expected inflation; and the average absolute size of inflation forecast errors. The behavior of expected inflation in five countries with targets (Australia, Canada, New Zealand, Sweden and the United Kingdom) and six countries without targets (France, Germany, Italy, the Netherlands, Japan and the United States) is measured in a panel of data from 1984 to 2000. The results show that the observed reduction in expected inflation in an inflationtargeting country is not fully explained by a targeting country’s recently observed reduction in actual inflation, an ongoing increase in unemployment in that country or other factors reducing inflation throughout the world. The additional reduction in the level of expected inflation is an important success of inflation targets. Once the level of actual inflation falls and a country re-establishes a history of inflation control, uncertainty about future inflation falls, this effect is present in both targeting and non-targeting countries. But there is little evidence that the period after the announcement of targets is associated with a further reduction in uncertainty about future inflation. There is very little evidence that average absolute forecast errors were lower in targeting countries than in non-targeting countries. Both targeting and non-targeting countries experienced unexpected disinflation. There is the strong evidence noted above that targets reduced the level of expected inflation in targeting countries. Bernanke et al. (1999) note that unexpected disinflation is common at introduction of targets. These two results are reconciled if the size of the reduction in expected inflation associated with targets is sufficient only to reduce the magnitude of unexpected disinflation in targeting countries, but did not reduce forecast errors in targeting countries relative to forecast errors in the non-targeting countries. Inflation targets are a success in the sense that forecast errors in targeting countries were smaller than they would have been in the absence of targets. Two approaches have been taken in the literature to ask if target announcements alter expectations of inflation, a direct approach and an indirect approach. Bernanke et al. (1999) use the indirect approach. Following Ball (1994), they ask if sacrifice ratios for individual countries, lost output per unit of actual reduction in inflation, are lower after targets. No clear evidence of lower sacrifice ratios is found. Bernanke et al. (1999) and Groenveld (1998) also study the actual inflation process in 1
Bernanke et al. (1999) present individual case studies of the five targeting countries used in my study (as well as Israel and Spain). They document carefully the extent of other reforms associated with the announcement of targets in these countries. My methodology treats the announcement of inflation targets as the primary change in policy.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1523
individual countries using VAR models of inflation. There is no clear evidence of change in the actual inflation process after targets. Groenveld (1998) compares model-generated measures of inflation variability in three countries with targets, New Zealand, Canada and the United Kingdom, and three ‘‘partner’’ countries without targets, Australia, the United States and Germany, respectively.2 He concludes targets play a limited role. Berg and Lundkvist (1997) estimate a univariate model of inflation for Sweden and while they find some evidence of a target-induced shift in the inflation process, they report that their evidence is not robust. There are several studies that attempted to measure the direct impact of inflation targets on expected inflation. Expected inflation is measured using survey responses of professional forecasters. Bernanke et al. (1999) compare actual inflation at the end of the calendar period to average survey predictions of inflation 6 months, 12 months and 18 months earlier. Laidler and Robson (1993) and three different studies in a volume edited by Leiderman and Svensson (1995) Fischer (1995), Svensson (1995), and Bowen (1995) graph expected and actual inflation in Canada, New Zealand, Sweden and the United Kingdom around the introduction of targets. Johnson (1998) measures forecast errors before and after targets.3 All these studies note that in the early years of inflation targets, expected inflation exceeds actual inflation, and in observing unexpected disinflation, infer that inflation targets did not shift expected inflation very much. None of the studies above examine the behavior of expected inflation in a panel. There are enormous advantages in measuring the effects of inflation targets in a panel including both targeting and non-targeting countries. First, the panel of targeting countries greatly increases the number of observations of years with inflation targets and thus the precision of the estimates of the effects of targets. When an individual country is studied, the introduction of the inflation target regime is a one-time event and any given year of targeting is a one-time event. While some progress can be made in looking at inflation targeting as an event (see Johnson, 1997, 2000), the panel allows measurement of both the common effects of targets by year of targeting across countries and common effects of targets across different years in the same targeting country. The second advantage of the panel, and one of equal importance, is that in a study of an individual country it is difficult to control directly for all other factors that may influence the behavior of expected inflation and then isolate the additional effect of inflation targets. In contrast, the panel controls for worldwide events that may affect expected inflation in all the countries in the same calendar year, for example, an oil price shock. The panel compares the change in the 2
Targets begin in New Zealand in 1990 and in Australia (at the earliest) in 1993. This is the basis for Groenveld’s match of these two countries as target and non-target countries. 3 Johnson (1998) also uses the individual survey data set to create a direct measure of the credibility of targets. The specific question asked: Are individual forecasts in a year with targets drawn from a distribution centered on the midpoint of the target band? Using data ending in 1995 (with only a few years of targets), there is substantial evidence that expectations do not shift immediately to the target rate of inflation, that is, that targets were not immediately credible. The present paper focuses on a broader question: What was the effect of announcement of targets on the behavior of expected inflation?
1524
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
behavior of expected inflation in targeting and non-targeting countries in the same calendar year. The paper proceeds as follows. Section 2 describes the inflation-targeting regimes. Section 3 describes the survey data. Section 4 presents the panel analysis. There is a brief conclusion.
2. Inflation targets Table 1 lists the month and year of the announcement of targets in five countries. In these five countries (and the six comparison countries) detailed individual forecast data exist (see Section 3 below). Five aspects of Table 1 require further comment. First, the panel uses annual calendar-year data. The first ‘‘target year’’ in each country is the announcement year as listed in Table 1 except in the United Kingdom when the first ‘‘target year’’ is defined as 1993 rather than 1992 (the announcement is in October 1992). In all countries most or all forecasts made in a ‘‘target year’’ are made with the knowledge targets have been announced. Second, the announcement date in Australia is listed as March 1993. The word ‘‘target’’ does not appear in Australian central bank documents until October 1995. However in March 1993, the Governor of the Reserve Bank of Australia stated a specific desired numerical outcome for the inflation rate. Third, Table 1 reveals that in three of five countries the inflation target is very cautious. Only in the United Kingdom and New Zealand is the target rate of inflation lower than the rate of inflation observed in the last full Table 1 The adoption of targets in the countries studied Country
Month and year of target announcement
Inflation rate in previous calendar yeara
Inflation target
Australia Canada New Zealand Sweden United Kingdom
March 1993 February 1991 March 1990 January 1993 October 1992
1 3 5.5 2.3 5.6
2–3 percent 1–3 percent by 1995 0–2 percent 1–3 percent by 1995 1–2.5 percent by 1997
0.9%c
Below 2%
The European Central Bank Announcementb Germany, France, January 1999 Italy, Netherlands a
Percent change in the GDP deflator for Canada, percent change in the consumer price index for all other countries. b These four countries formally join the European Central Bank (the ECB) on this date. The ECB has both an inflation target and uses monetary growth as a guide to monetary policy. See the text for further discussion. c In European Central Bank (1999a), 0.9% is reported as the harmonized index of consumer prices (HICP) inflation rate for the 12 months ending November 1998. The ECB inflation target is specified as a percent change in the HICP.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1525
calendar year before the announcement. It may be difficult to see much impact of targets on expected inflation in the announcement year. Studying a longer period of monetary policy after the introduction of targets is useful. Fourth, Table 1 shows that in 3 of 5 countries, a time period was specified between announcement of the target and its eventual attainment, another reason to study a long period after the announcement. Fifth and finally, Table 1 notes that four countries in the study, Germany, France, Italy and the Netherlands formally joined the European Central Bank (ECB) on January 1, 1999. As described in European Central Bank (1999b), the ECB in October 1998, announced that price stability, its mandate, would be quantified and would be achieved using a strategy with ‘‘two pillars.’’ One ‘‘pillar,’’ price stability, was quantified in January 1999 as inflation below 2%, a specific numerical target for inflation. However, the ECB has also stressed that the conduct of monetary policy would also be based on a ‘‘prominent role for money,’’ that is, ‘‘a quantitative reference for monetary growth,’’ the second ‘‘pillar.’’ There is agreement that the ECB monetary regime is not an inflation-targeting regime, see Feldstein (2000) or Svensson (2000). ECB countries are not treated as inflation-targeting countries in this study.
3. The measurement of the behavior of expected inflation The data used in this study are found in Consensus Forecasts and its predecessor, Economic forecasts: a monthly worldwide survey.4 Each forecaster reports an expected rate of inflation for the current calendar year and an expected rate of inflation in the next calendar year. The month of the survey response is known. Forecasters drop in and out of the survey and many new forecasters are added over time. The data are not a monthly panel of forecasters. There are far more responses for the United States, the United Kingdom, and Japan than for other countries. The number of individual responses in a given month or even year in any country is highly variable before 1996.5 Survey responses are spread evenly over the year, the average month of response is either June or July in all countries. The survey data, with individual responses from each country within a calendaryear, are used to create three calendar-year measures related to the behavior of expected inflation. The first measure is the average next-year forecast, averaged over forecasters within a calendar year. This measure and the corresponding actual inflation rate are plotted in Fig. 1 for all countries. The vertical lines in Fig. 1 delimit 4 The data for New Zealand and Australia after 1996 are found in Asia Pacific Consensus Forecasts. This publication succeeded, for these two countries, Consensus Forecasts. The survey for the United States is not from Consensus Forecasts. Rather it is the Survey of Professional Forecasters (website www.phil.frb.org) maintained by the Federal Reserve Bank of Philadelphia. This is a quarterly survey with a structure similar to that Consensus Forecasts. I thank Dean Croushore for his help using this survey. 5 The data are described in more detail in Johnson (1998). One other small note is that until 1995 the reported measure of expected inflation in Canada is a response by forecasters to a question about the expected rate of increase in the GDP deflator and after 1995, expected inflation rate is measured using the consumer price index. Over the long run the consumer price index and the GDP deflator are closely linked.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Fig. 1. Actual and expected inflation in the next calendar year.
1526
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1527
years after the announcement of targets. In 9 of the 11 countries, excluding the Netherlands and Germany, average expected inflation and average actual inflation fall after 1990. Disinflations are larger in the targeting countries. The second measure is the standard deviation of the individual forecasts within a calendar year, a measure of forecast variability. This measure is plotted in Fig. 2 for both next-year and current-year forecasts. Variation across individual forecasts within a calendar year has three sources. First, in the same month different forecasters may interpret the same public information differently and/or may have different private information. Second, as time passes, a forecaster may revise his forecast, either using new information or changing his mind about the same information. Third, the data include different forecasters in different months of the same year. The standard deviation of current-year forecasts in Fig. 2 is generally smaller than the standard deviation of next-year forecasts in the same country in the same year. This should be the case. By the end of a calendar year, forecasters know most of the actual values of the price indexes needed for the current-year forecast. A monthly price index is usually announced with a one month lag.6 Forecasters should begin to agree on the current-year inflation forecast by year end. This same effect is present but much weaker for next-year forecasts. This is an important reason to concentrate on the properties of next-year forecasts in both standard deviations and levels. Fig. 2 shows that, in targeting countries, the standard deviation of next-year inflation forecasts falls after targets are introduced in Canada, New Zealand, the United Kingdom and Sweden but seems to rise in Australia. In the six non-targeting countries the standard deviation of next-year forecasts also seems to decline over time. The panel analysis measures any additional effect of the announcement of targets in reducing the standard deviation of next-year forecasts among forecasters. Fig. 3 presents the average next-year forecast errors in all years in all countries. A positive value of the forecast error indicates unexpected disinflation. Fig. 3 shows that in every case, the forecast error on the next-year forecast made in the year of the announcement of targets was positive. If you consider forecast errors for the first 5 years after the target announcement in each targeting country, 6 of 25 forecast errors are negative and 19 are positive. Target announcements did not prevent unexpected disinflation. However, it is not enough to look at forecast errors and draw the conclusion that the announcement of targets did not change the level of expected inflation. In countries with or without targets, there was an ongoing recession associated with disinflation. And in all countries, as Fig. 1 shows, as actual inflation fell, expected inflation did not catch up immediately (the dotted line is generally above the solid line as actual inflation falls). Thus, the disinflations in non-targeting countries were also unexpected. The third measure of the behavior of expected inflation studied is the average absolute next-year forecast error, averaged across individual forecasts with a year. Each of the three measures of the behavior of expected inflation is used to ask whether that behavior changed after the announcement of targets in targeting 6
The situation is slightly more complicated when the price index is quarterly and/or is a GDP deflator. This is true in Australia, New Zealand and in part of the Canadian sample.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Fig. 2. The variability of forecasts.
1528
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Fig. 3. Next year forecast errors.
1529
1530
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
countries relative to non-targeting countries after controlling for both countryspecific effects and worldwide effects.
4. Modeling the behavior of expected inflation 4.1. The framework The measure of the behavior of expected inflation in country i at time t; denoted Mi;t ; is modeled as a function of a vector of k country-related factors, Xi;t (made specific in Section 4.2); a variable Ci equal to 1 if the observation is from country i; a vector of dummy variables Yt with element equal to 1 if the observation is from year ECB t; and a dummy variable Ti;t that equals one if country i is a member of the European Central Bank in year t: Each observation from country i at time t is modeled as ECB þ bi Xi;t þ dTi;t þ eit : Mi;t ¼ bi;0 Ci þ yYt þ dECB Ti;t
ð1Þ
In (1) the coefficient bi;0 on the constant term Ci is country-specific and represents the effect of factors within that country that do not vary over the sample. The vector of coefficients y on the vector Yt represent worldwide effects on the measure of expected inflation in all countries that occur in each calendar year t: The single coefficient dECB is the effect on the measure of expected inflation M when a country joins the European Central Bank. As discussed in Section 2, the European Central Bank is not considered to be an inflation targeting bank.7 In (1) the vector of coefficients bi is specific to each country, that is, the effects of the Xi;t variables specific to each country differ by country. Finally, there are the dummy variables that capture the presence or absence of targets in that country. These variables are quite complicated. If targets have been announced in country i as of year t; then for that observation the element of the vector Ti;t is equal to one, otherwise it is zero. The elements of the vector of coefficients d measure the effects of targets on expected inflation. The d are the coefficients of central interest. The dummy variables that capture the effect of target announcements are organized in three ways in Table 2. One vector of dummy variables, called the ‘‘year-based’’ dummies, take values of one in all targeting countries in a given year after the targeting announcement. They j are denoted Ti;t ¼ 1 where j runs from 1 to 5 years after the target announcement and then 6,7 and 8, or 9,10 and 11 for the sixth, seventh or eighth, ninth, tenth or eleventh year of targets in country i: The coefficients on these seven dummy variables measure the average effect of targets across all targeting countries in the same year of targets, the year of targets being defined relative to the announcement year. For 1 example, the coefficient on dummy variable Ti;t measures the average effect on measure M over all targeting countries in the first year after targets were announced, the impact 7 The coefficients on the target-based dummies are not sensitive to treating the ECB as an inflationtargeting bank for the first 2 years of its existence. This is a very short additional period of targets if these countries are counted as targeting countries.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1531
Table 2 The dummy variables whose coefficients measure the effect of targets Variable
Year and country where dummy ¼ 1 (inclusive): all other values zero
Year-based dummies (seven separate variables) First year of targets: Ti;j1 ¼ 1 Australia (1993), Canada (1991), New Zealand (1990), Sweden (1993), United Kingdom (1993) Second year of targets: Ti;j2 ¼ 1 Australia (1994), Canada (1992), New Zealand (1991), Sweden (1994), United Kingdom (1994) Third year of targets: Ti;j3 ¼ 1 Australia (1995), Canada (1993), New Zealand (1992), Sweden (1995), United Kingdom (1995) Fourth year of targets: Ti;j4 ¼ 1 Australia (1996), Canada (1994), New Zealand (1993), Sweden (1996), United Kingdom (1996) Fifth year of targets: Ti;j5 ¼ 1 Australia (1997), Canada (1995), New Zealand (1994), Sweden (1997), United Kingdom (1997) Sixth, seventh or eighth year of targets: Australia (1998, 1999, 2000), Canada (1996, 1997, 1998), Ti;j628 ¼ 1 New Zealand (1995, 1996, 1997), Sweden (1998, 1999, 2000), United Kingdom (1998, 1999, 2000) Ninth, tenth or eleventh year of targets: Canada (1999, 2000), New Zealand (1998, 1999, 2000) Ti;j9211 ¼ 1 Country-based dummies (five separate variables) Australia: Ti;jAustralia ¼ 1 Australia 1993–2000 Canada: Ti;jCanada ¼ 1 Canada 1991–2000 New Zealand: Ti;jNZ ¼ 1 New Zealand 1990–2000 Sweden: Ti;jSweden ¼ 1 Sweden 1993–2000 United Kingdom 1993–2000 United Kingdom: Ti;jUK ¼ 1 The all targets dummy: Ti;jall ¼ 1
Australia 1993–2000 Canada 1991–2000 New Zealand 1990–2000 Sweden 1993–2000 United Kingdom 1993–2000
628 effect of the target announcement. The coefficient on dummy variable Ti;t similarly measures the average effect on measure M over all targeting countries in the sixth, seventh and eighth years after targets were announced, the longer-term effect of targets. The other vector of dummy variables related to targets are ‘‘country-based.’’ Here Country the variable Ti;t takes a value of 1 in all years after the specified country Australia announces targets and is zero otherwise. For example Ti;t equals one in each calendar year from 1993 to 2000 for the observations from Australia and is zero for all other observations. The coefficient on this dummy variable measures the average effect of targets on measure M over all years of targets in Australia. Five such coefficients are estimated, one for each targeting country. Imposing either the constraints that the effects of targets are equal in all years after the announcement year or the constraints that the effects of targets are equal in all targeting countries is equivalent to estimating the model with a third dummy all variable. The third target-related dummy variable Ti;t equals unity for any observation in any country when targets have been announced in any country i at
1532
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
all time t and is zero otherwise. For the ‘‘country-based’’ vector of dummies, using the Ti;t dummy in (1) imposes four constraints that the effects of targets are equal in all targeting countries. For the ‘‘year-based’’ vector of dummies, using this dummy in (1) imposes six constraints that the effect of targets in the seven designated targeting periods are equal. These constraints are always rejected, an interesting finding in itself. The effects of targets are not statistically equivalent across years of targeting or across all countries. The single coefficient on the Ti;t dummy variable is useful only as a summary statistic of the average effect of targets over all years and all targeting countries. The 11 countries and 17 years (1984–2000) of data provide up to 185 observations, there are no data in Australia for 1984 and there are no next-year forecast responses in the Netherlands in 1986, only the current-year forecast question was answered. The set of 11 equations, one per country defined using (1), is a system of seemingly unrelated regression (SUR) equations, see Greene (2000), Chapter 15 for further discussion. The most general specification of the error term in such a system is described by
Eðeit ; eit Þ ¼ s2i
ð2Þ
Eðeit ; ejt Þ ¼ s2ij
ð3Þ
and
and, in some country equations Eðeit ; eit1 Þ ¼ ri :
ð4Þ
Expression (2) represents error terms that are heteroscedastic across countries. Expression (3) represents error terms in different countries that are correlated in a given calendar year. This correlation could be the result of year-specific worldwide effects on expected inflation in all countries, for example an oil shock. This correlation could also result when linkages or other similarities occur between specific country pairs. Finally (4) represents the possibility that error terms in (1) within a country may be auto-correlated. Diagnostic tests reveal that the error terms in (1) are typically heteroscedastic and are usually correlated across years. Some but not all estimated values of ri are large, there are some country-specific dynamics. Results are presented with and without the correction for serial correlation. The AR(1) specification estimates up to 11 additional AR(1) coefficients and drops the first observation for each country. Greene (2000) reports that the efficiency gain from the AR(1) correction may be outweighed by the loss of the one observation in a relatively small sample.8 Results from specifications with and without AR(1) terms are presented for this reason.9 8 I do not report the individual country AR(1) coefficients in the results below. The model was first estimated with all 11 AR(1) coefficients, then AR(1) coefficients that were insignificant were dropped. Results are presented in Table 3 with and without the AR(1) correction. 9 Results from two other econometric specifications common in the panel literature were also estimated and are presented in the working paper version of this paper available at http://www.wlu.ca/~wwwsbe/ sbe2000/html/economics/working papers2001.html. The time-series cross-section (TSCS) specification imposes the restrictions that the slope coefficients (but not the constant terms) are equal across countries and that the common year effects (measured by the vector of coefficients y) are zero on expression (1). Correlations across countries in the same year as in (3) are considered sufficient to capture common year
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1533
4.2. The effects of inflation targets on the behavior of expected inflation Measures of the effect of targets on the level of expected inflation are presented in the left-hand 2 columns of Table 3. There are two types of country-specific control variables (the Xi;t ) in the equations measuring the effect of targets on the level of expected inflation. The business cycle variable is the rate of unemployment observed in the 12 months closest to and preceding January of the current calendar year minus the rate of unemployment observed in the past 60 months closest to and preceding January of the current calendar year. A value of 1 indicates observed unemployment in that country is 1 percentage point higher in the past 12 months than the average of the preceding 60 months.10 The element of the bi vector that is the coefficient on this variable, while it varies from country to country, is usually negative and significant, an increase in unemployment relative to its immediate past reduces the level of expected inflation. The other four elements of the Xi;t vector of variables are four lagged inflation rates measured over 12 months intervals using the forecasters’ knowledge as of January 1 of the current calendar year.11 Countries with a history of high inflation are expected to have higher expected inflation. The most important results in Table 3 are that inflation target announcements reduce, with both economic and statistical significance, the average level of expected inflation in the next calendar year after controlling for both individual country and worldwide effects. If the purpose of announcing inflation targets is to reduce the level of expected inflation, targets are a success in most countries in most years. The seven coefficients on the year-based dummies (the year of targeting relative to the announcement year) present an interesting pattern in both the AR(1) and non-AR(1) estimates. The coefficients on the different years of targeting are not equal. The effect of targets seems to peak later in the targeting period, the reduction in next-year (footnote continued) effects. These extremely strong restrictions require every slope coefficient to be the same in every country and every year-effect coefficient in the vector y to be zero. These restrictions are rejected. A third specification, the fixed effects (FE) specification restores non-zero coefficients on the y vector. The TSCS and FE specifications use the additional assumptions of equal slope coefficients across countries to more precisely identify the effects of targets. The countries are assumed to be similar enough (much as individuals or firms in other panel data sets are assumed to be similar enough) to allow a panel estimator of the characteristics of the ‘‘average’’ country. The TSCS and FE estimates of the coefficients on the dummy variables associated with targets are very similar to the SUR estimates. In earlier versions of the paper, the results are also robust to dropping a country (the Netherlands) and to reducing the sample, dropping the years 1984, 1998, 1999 and 2000. The results presented in Table 3 are robust to a variety of samples and estimation procedures. 10 This variable represents the forecaster’s knowledge of the state of the business cycle as the beginning of the year. This measure of the state of the business cycle is roughly comparable across the different countries with very different average unemployment rates. This is useful in the TSCS and FE specifications discussed above. 11 Models with both three and four lags of actual inflation were estimated. The results are not sensitive to this choice. It is difficult to choose between these two specifications, the lagged inflation measures are collinear. When the fourth lag of inflation is included, its coefficient is typically positive and statistically significant but the coefficients on either the second or third lag dwindle while with three lags of known inflation, all three coefficients tend to be significant.
1.22(0.26)n
1.16(0.26)n
1.38(0.23)n
1.25(0.19)n
0.41(0.25)
1.73(0.39)n
1.12(0.20)n
3.79(0.71)n
2.57(1.56)
1.16(0.26)n
2.42(0.20)n
3 Year 3: Ti;t
4 Year 4: Ti;t
5 Year 5: Ti;t
628 Years 6, 7 or 8: Ti;t
9211 Years 9, 10 or 11: Ti;t
Australia Australia: Ti;t
Canada Canada: Ti;t
NZ New Zealand: Ti;t
Sweden Sweden: Ti;t
UK United Kingdom: Ti;t
all All targets: Ti;t
2.55(0.25)n
0.32(0.32)
4.33(1.7)n
8.10(0.16)n
1.07(0.26)n
2.35(1.0)n
1.01(0.43)n
3.02(0.38)n
3.56(0.43)n
3.01(0.46)n
2.21(0.44)n
1.85(0.39)n
0.90(0.29)
0.07(0.02)n
0.13(0.19)
0.29(0.19)
0.84(0.14)n
0.03(0.02)
0.46(0.24)
0.03(0.03)
0.00(0.03)
0.02(0.03)
0.02(0.04)
0.07(0.03)n
0.01(0.03)
0.36(0.02)
Indicates coefficient on the dummy variable is statistically significant at 5 percent.
1.37(0.26)n
2 Year 2: Ti;t
n
1.14(0.22)
1 Year 1: Ti;t n
0.02(0.03)
0.15(0.11)
0.16(0.19)
0.02(0.26)
0.04(0.03)
0.43(0.24)
0.15(0.07)n
0.14(0.06)n
0.12(0.07)
0.20(0.07)n
0.06(0.07)
0.00(0.06)
0.04(0.05)
No AR(1)
AR(1) n
AR(1)
No AR(1)
On the standard deviation of expected inflation
On the level of expected inflation
n
Coefficients on target dummies (standard errors)
Table 3 The effect of inflation targets on the behavior of expected inflation in the next calendar year
0.21(0.14)
0.61(0.43)
0.27(0.54)
0.28(0.30)
0.19(0.18)
0.86(0.89)
0.08(0.21)
0.58(0.19)n
0.46(0.26)
0.04(0.21)
.13(0.18)
0.20(0.15)
0.00(0.14)
AR(1)
0.19(0.13)
0.11(0.05)n
0.62(0.50)
0.30(0.34)
0.21(0.18)
0.83(0.91)
0.04(0.20)
0.64(0.19)n
0.54(0.26)n
0.09(0.21)
0.05(0.18)
0.24(0.14)
0.02(0.13)
No AR(1)
On the absolute average error in next-year forecasts
1534 D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1535
expected inflation is smallest in the first and last years of targets and highest in the middle years. This pattern is more pronounced in the specification without the AR(1) correction.12 Since years 9, 10 and 11 of targeting occur only in Canada and New Zealand, the coefficient on this element of the ‘‘year-based’’ vector of dummy variables is estimated using data from only 2 and not 5 targeting countries and some caution is needed in its interpretation. However, a slow effect of targets on the level of expected inflation makes good economic sense. Targets were designed to alter long-term patterns of expected inflation in relatively high inflation countries. These countries experienced significant recessions to reduce inflation. Targets then seem to prevent the increase in expected inflation to ‘‘normal’’ (for that country) levels which would have been associated with economic recovery. There is variation in the estimates of the average effect of inflation targets on the level of expected inflation across countries. The estimates of the coefficient on the NZ dummy variable measuring the effect of targets in New Zealand ðTi;t Þ vary from 3:79 to 8:10; meaning that in the period after the announcement of targets, the level of expected inflation in New Zealand was reduced by 3.79–8.1 percentage points. The 8 percentage point effect in New Zealand seems implausible although Fig. 1 presents an extremely rapid decline in expected inflation in New Zealand. Very large changes in the level of expected inflation in New Zealand could have been associated with the large scale fiscal, trade and monetary reforms in New Zealand that coincided with targets. These reforms were much broader than in other countries and could account for some of the strength of the New Zealand effect. The Swedish case also generates a relatively large ‘‘country-based’’ coefficient on dummy Sweden variable Ti;t (2:57 or 4:33), albeit with a larger standard error than the analogous coefficient in New Zealand.13 all The single coefficient on the dummy variable, Ti;t ; is a summary measure of the average effect of targets on the level of expected inflation in all the years in all the targeting countries. Constraints are imposed that the effects of targets are equal across years and countries, constraints we can reject.14 The summary coefficient suggests a substantial reduction in the level of expected inflation after the announcement of targets, a reduction not explained by either country-specific effects or worldwide effects. Table 3 provides strong evidence inflation targets directly reduced the level of expected inflation. The other two measures of the behavior of expected inflation studied are the standard deviation of next-year forecasts across forecasters within a given calendar 12
There is a similar coefficient pattern in both the TSCS and FE specifications as well. Some of the coefficient variation and the larger standard errors seem to be associated with the most complex SUR specification (1) rather than the TSCS or FE specifications described in footnote 9. There are some reasons to prefer the estimates from the TSCS or FE models. In these specifications the elements of the bi coefficient vector are constrained to be equal across the countries. Without the AR(1) correction there are additional observations. Alternative estimation procedures yield tighter estimates of the effects of inflation targets. The effects of targets remains strongest in New Zealand and are generally comparable in Australia, Canada, Sweden and the United Kingdom. 14 The test statistics on the null hypotheses of equal country or year effects are not reported. The coefficient values as reported in Table 3 makes it clear these equality constraints are rejected. 13
1536
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
year, plotted as the dotted line in Fig. 2, and the average absolute forecast error plotted in Fig. 3. The panel specifications with these dependent variables have two country-specific independent variables: the lagged level of average inflation calculated over the last 5 years as observed at the beginning of the calendar year, and the average standard deviation of inflation observed over the previous 5 years.15 For both dependent variables, the coefficient on the lagged level of inflation is usually positive and significant. A higher past level of inflation in a country is associated both with more diverse inflation forecasts and larger average next-year forecast errors. Past variability of inflation, given the level of past inflation, is usually marginally significant in the models of inflation variability but not usually significant in the models of average absolute forecast errors. After the effects of higher past inflation or higher past inflation variability are taken into account, the announcement and implementation of inflation targets has little additional effect on the standard deviation of inflation forecasts. The middle panel of Table 3 presents these results. Most of the coefficients on target-related dummy variables do not differ significantly from zero. One oddity is that in first year of inflation targets in the AR(1) specification, there is even some weak evidence of a significant increase in next-year inflation uncertainty, not the result desired by monetary authorities at the announcement of targets. By the fourth and fifth years of targets the signs on ‘‘year-based’’ target dummies are negative but not generally significant. When the coefficients on ‘‘country-based’’ target dummies are examined, in New Zealand targets significantly increase the dispersion of forecasts in the AR(1) specification but not in the specification without the AR(1) correction. The New Zealand result is obtained with quite limited data, prior to targets New Zealand has only three forecasts in 1984 and 1988, four forecasts in 1985, five forecasts in 1986 and six forecasts in both 1987 and 1989. New Zealand and the Netherlands have the fewest number of individual forecasts. However, Table 3 suggests that targeting, beyond an important indirect effect through successful reduction in the level of past inflation, has no additional effect on the standard deviation of expected inflation. The last investigation in this study compares the behavior of average absolute forecast errors in targeting and non-targeting countries. These results are presented in the right-hand side panel of Table 3. To this point in the history of inflation targets, there is very little evidence that targeting regimes experience lower average absolute value next-year forecast errors relative to the non-targeting countries. Figs. 1 and 3 leave little doubt that the reduction in inflation in targeting countries was unexpected. This is the standard result of previous studies. Italy, Japan, France and the United States, non-targeting countries, also experienced unexpected disinflation. The dependent variable in (1) is the average absolute next-year forecast error across forecasters. The country-specific control variables are the average level and standard deviation of inflation over the past 5 years. Higher lagged inflation is related to higher average forecast errors in some countries. But the effects are weak and unsystematic. Thus, the reduction in the level of inflation associated with targets 15
These models were also estimated with similar results when lagged inflation or its standard deviation is calculated over the previous 10 years of observed inflation.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1537
indirectly reduces average forecast errors in some countries by lowering past inflation levels. Any additional effects of the target announcement are quite weak. The only observable direct effect of targets on absolute forecast errors is that the coefficients on target-year dummies do become negative later in the targeting process. By years 5–8 after the announcement of targets, there may be some small all effect of targets in reducing average forecast errors. The summary coefficient on Ti;t is negative but not significant. Although targeting does not directly reduce forecast errors in targeting countries relative to those in non-targeting countries, recall the strong evidence that targets reduced the level of expected inflation in targeting countries after the announcements. Fig. 1 shows that actual inflation fell more rapidly in targeting countries than in non-targeting countries. Targeting countries, through the effect of targets on the level of expected inflation, avoided even larger forecast errors than would have occurred in the absence of targets. For a larger disinflation, targeting countries had similar forecast errors. This is a success of targets.
5. Conclusions One purpose in introducing inflation targets was to alter the behavior of expected inflation. The evidence is very strong that the period after the announcement of inflation targets is associated with a large reduction in the level of expected inflation. That reduction is of both economic and statistical significance, at least one to two percentage points in most years after targets. That reduction took place in all 5 countries with inflation targets. This is an important success of inflation targets. Targets were most successful in New Zealand where other reforms may have played a role in reducing expected inflation. The actual rate of inflation falls in both targeting and non-targeting countries. With lower actual inflation, the dispersion of inflation forecasts falls in all countries. But there is little or no additional reduction in the dispersion of inflation forecasts associated with the period after the announcement of inflation targets. Targets are not yet a success in this dimension. Finally, in the disinflations of the 1990s both targeting and non-targeting countries experience large forecast errors. Inflation targets did not reduce absolute average forecast errors in targeting countries relative to those in non-targeting countries. However in targeting countries expected inflation falls more rapidly with inflation targets. These events are reconciled if inflation targets prevented forecast errors in targeting countries from being larger than they would have been without targets. Inflation targets allowed a larger disinflation with smaller forecast errors to take place in targeting countries. This is also an important success of inflation targets.
Appendix A. Data All forecasts of inflation except for the United States were drawn from economic forecasts: a monthly worldwide survey, its successors Consensus Forecasts and
1538
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Asia-Pacific consensus forecasts between 1984 and 1998. Forecasts for the United States are the ASA-NBER forecasts available at the Reserve Bank of Philadelphia website: www.phil.frb.org. Actual inflation rates and actual unemployment rates for all countries are drawn from the OECD database and updated from various national websites to February 2000. Unemployment rates are seasonally adjusted. These are not the OECD standardized unemployment rates. They are the unemployment rates reported by the national statistical agencies to the OECD. In New Zealand unemployment rates from a labour force survey are available only after December 1985. Before 1985, a proxy for the unemployment rate is constructed by dividing the number of registered unemployed by employment and unemployment.
References Ball, L., 1994. What determines the sacrifice ratio? In: Mankiw, N.G. (Ed.), Monetary Policy. University of Chicago Press, Chicago, pp. 155–182. Berg, C., Lundkvist, P., 1997. Has the inflation process changed? Sveriges Riksbank Quarterly Review 5, 5–25. Bernanke, B.S., Laubach, T., Mishkin, F.S., Posen, A., 1999. Inflation Targeting: Lessons from the International Experience. Princeton University Press, Princeton, NJ. Bowen, A., 1995. British experience with inflation targetry. In: Leiderman, L., Svensson, L.E.O. (Eds.), Inflation Targets. Centre for Economic Policy Research, London, pp. 53–68. European Central Bank, 1999a. Editorial, European Central Bank Monthly Bulletin, January 1999, pp. 9–10. European Central Bank, 1999b. The stability-oriented monetary policy strategy of the Eurosystem. European Central Bank Monthly Bulletin, January 1999, pp. 39–50. Feldstein, M., 2000. The European Central Bank and the Euro: the First Year, National Bureau of Economic Research Working Paper, 7517. Fischer, A., 1995. New Zealand’s experience with inflation targets. In: Leiderman, L., Svensson, L.E.O. (Eds.), Inflation Targets. Centre for Economic Policy Research, London, pp. 32–52. Greene, W.H., 2000. Econometric Analysis, 3rd Edition. MacMillan, New York. Groenveld, J.M., 1998. Inflation patterns and monetary policy: lessons for the European Central Bank. Edward Elgar, Cheltenham. Johnson, D.R., 1997. Expected inflation in Canada 1988–1995: an evaluation of Bank of Canada credibility and the effect of inflation targets. Canadian Public Policy 23, 223–258. Johnson, D.R., 1998. The credibility of monetary policy: international evidence based on surveys of expected inflation. In: Macklem, T. (Ed.), Price Stability, Inflation Targets and Monetary Policy. Proceedings of a Conference held by the Bank of Canada, May 1997. Bank of Canada, Ottawa, pp. 361–395. Johnson, D.R., 2000. The effect of targeting inflation on the level of expected inflation in five countries. Wilfrid Laurier University Working Paper 2000–01. Laidler, D.E.W., Robson, W.B.P., 1993. The Great Canadian Disinflation: The Economics and Politics of Monetary Policy in Canada 1988–1993. C.D. Howe Institute, Toronto. Leiderman, L., Svensson, L.E.O., 1995. Inflation Targets. Centre for Economic Policy Research, London. Svensson, L.E.O., 1995. The Swedish experience of an inflation target. In: Leiderman, L., Svensson, L.E.O. (Eds.), Inflation Targets. Centre for Economic Policy Research, London, pp. 69–89. Svensson, L.E.O., 2000. The first year of the Eurosystem: inflation targeting or not? National Bureau of Economic Research Working Paper No. 7598.
The effect of inflation targeting on the behavior of expected inflation: evidence from an 11 country panel$ David R. Johnson* Department of Economics, School of Business and Economics, Wilfrid Laurier University, 75 University Avenue West, Waterloo, Ont., Canada, N2L 3C5 Received 7 December 1999; received in revised form 12 September 2001; accepted 15 November 2001
Abstract Inflation targets were introduced in the 1990s in Australia, Canada, New Zealand, Sweden and the United Kingdom. Change in the behavior of expected inflation in the five targeting countries is measured in a panel that includes six non-targeting countries: France, Germany, Italy, the Netherlands, Japan and the United States. The level of expected inflation in targeting countries falls after the announcement of targets with controls for country effects, year effects, ongoing inflation reduction and the business cycle. Neither the variability of expected inflation nor the average absolute forecast error falls after the announcement of targets with controls for the level and variability of past inflation. Both targeting and non-targeting countries experience unexpected disinflation. r 2002 Elsevier Science B.V. All rights reserved. JEL classification: E52; E58 Keywords: Inflation targets; Expected inflation
$ I want to thank Nathan Braun for excellent research assistance and Susan Johnson, Pierre Siklos, Angelo Melino and the referees and editor for their comments. The support of an Interdisciplinary Grant from the Social Sciences and Humanities Research Council of Canada, administered by the Wilfrid Laurier University Research Office, is acknowledged. I also thank the Faculty of Economics and Politics at the University of Cambridge, England for their hospitality when this project began. *Corresponding author. Fax: +1-519-888-1015. E-mail address: [email protected] (D.R. Johnson).
0304-3932/02/$ - see front matter r 2002 Elsevier Science B.V. All rights reserved. PII: S 0 3 0 4 - 3 9 3 2 ( 0 2 ) 0 0 1 8 1 - 2
1522
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1. Introduction In the 1990s a number of countries used announcements of explicit numerical targets for inflation as a key component of monetary policy. Inflation targets, and other monetary reforms often implemented at the same time, were adopted for three reasons: to communicate more clearly the goals of monetary policy; to create accountability for the achievement of these goals and to alter expectations of inflation.1 This paper measures change after the announcement of targets in three aspects of the behavior of expected inflation: the level of expected inflation; the standard deviation of expected inflation; and the average absolute size of inflation forecast errors. The behavior of expected inflation in five countries with targets (Australia, Canada, New Zealand, Sweden and the United Kingdom) and six countries without targets (France, Germany, Italy, the Netherlands, Japan and the United States) is measured in a panel of data from 1984 to 2000. The results show that the observed reduction in expected inflation in an inflationtargeting country is not fully explained by a targeting country’s recently observed reduction in actual inflation, an ongoing increase in unemployment in that country or other factors reducing inflation throughout the world. The additional reduction in the level of expected inflation is an important success of inflation targets. Once the level of actual inflation falls and a country re-establishes a history of inflation control, uncertainty about future inflation falls, this effect is present in both targeting and non-targeting countries. But there is little evidence that the period after the announcement of targets is associated with a further reduction in uncertainty about future inflation. There is very little evidence that average absolute forecast errors were lower in targeting countries than in non-targeting countries. Both targeting and non-targeting countries experienced unexpected disinflation. There is the strong evidence noted above that targets reduced the level of expected inflation in targeting countries. Bernanke et al. (1999) note that unexpected disinflation is common at introduction of targets. These two results are reconciled if the size of the reduction in expected inflation associated with targets is sufficient only to reduce the magnitude of unexpected disinflation in targeting countries, but did not reduce forecast errors in targeting countries relative to forecast errors in the non-targeting countries. Inflation targets are a success in the sense that forecast errors in targeting countries were smaller than they would have been in the absence of targets. Two approaches have been taken in the literature to ask if target announcements alter expectations of inflation, a direct approach and an indirect approach. Bernanke et al. (1999) use the indirect approach. Following Ball (1994), they ask if sacrifice ratios for individual countries, lost output per unit of actual reduction in inflation, are lower after targets. No clear evidence of lower sacrifice ratios is found. Bernanke et al. (1999) and Groenveld (1998) also study the actual inflation process in 1
Bernanke et al. (1999) present individual case studies of the five targeting countries used in my study (as well as Israel and Spain). They document carefully the extent of other reforms associated with the announcement of targets in these countries. My methodology treats the announcement of inflation targets as the primary change in policy.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1523
individual countries using VAR models of inflation. There is no clear evidence of change in the actual inflation process after targets. Groenveld (1998) compares model-generated measures of inflation variability in three countries with targets, New Zealand, Canada and the United Kingdom, and three ‘‘partner’’ countries without targets, Australia, the United States and Germany, respectively.2 He concludes targets play a limited role. Berg and Lundkvist (1997) estimate a univariate model of inflation for Sweden and while they find some evidence of a target-induced shift in the inflation process, they report that their evidence is not robust. There are several studies that attempted to measure the direct impact of inflation targets on expected inflation. Expected inflation is measured using survey responses of professional forecasters. Bernanke et al. (1999) compare actual inflation at the end of the calendar period to average survey predictions of inflation 6 months, 12 months and 18 months earlier. Laidler and Robson (1993) and three different studies in a volume edited by Leiderman and Svensson (1995) Fischer (1995), Svensson (1995), and Bowen (1995) graph expected and actual inflation in Canada, New Zealand, Sweden and the United Kingdom around the introduction of targets. Johnson (1998) measures forecast errors before and after targets.3 All these studies note that in the early years of inflation targets, expected inflation exceeds actual inflation, and in observing unexpected disinflation, infer that inflation targets did not shift expected inflation very much. None of the studies above examine the behavior of expected inflation in a panel. There are enormous advantages in measuring the effects of inflation targets in a panel including both targeting and non-targeting countries. First, the panel of targeting countries greatly increases the number of observations of years with inflation targets and thus the precision of the estimates of the effects of targets. When an individual country is studied, the introduction of the inflation target regime is a one-time event and any given year of targeting is a one-time event. While some progress can be made in looking at inflation targeting as an event (see Johnson, 1997, 2000), the panel allows measurement of both the common effects of targets by year of targeting across countries and common effects of targets across different years in the same targeting country. The second advantage of the panel, and one of equal importance, is that in a study of an individual country it is difficult to control directly for all other factors that may influence the behavior of expected inflation and then isolate the additional effect of inflation targets. In contrast, the panel controls for worldwide events that may affect expected inflation in all the countries in the same calendar year, for example, an oil price shock. The panel compares the change in the 2
Targets begin in New Zealand in 1990 and in Australia (at the earliest) in 1993. This is the basis for Groenveld’s match of these two countries as target and non-target countries. 3 Johnson (1998) also uses the individual survey data set to create a direct measure of the credibility of targets. The specific question asked: Are individual forecasts in a year with targets drawn from a distribution centered on the midpoint of the target band? Using data ending in 1995 (with only a few years of targets), there is substantial evidence that expectations do not shift immediately to the target rate of inflation, that is, that targets were not immediately credible. The present paper focuses on a broader question: What was the effect of announcement of targets on the behavior of expected inflation?
1524
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
behavior of expected inflation in targeting and non-targeting countries in the same calendar year. The paper proceeds as follows. Section 2 describes the inflation-targeting regimes. Section 3 describes the survey data. Section 4 presents the panel analysis. There is a brief conclusion.
2. Inflation targets Table 1 lists the month and year of the announcement of targets in five countries. In these five countries (and the six comparison countries) detailed individual forecast data exist (see Section 3 below). Five aspects of Table 1 require further comment. First, the panel uses annual calendar-year data. The first ‘‘target year’’ in each country is the announcement year as listed in Table 1 except in the United Kingdom when the first ‘‘target year’’ is defined as 1993 rather than 1992 (the announcement is in October 1992). In all countries most or all forecasts made in a ‘‘target year’’ are made with the knowledge targets have been announced. Second, the announcement date in Australia is listed as March 1993. The word ‘‘target’’ does not appear in Australian central bank documents until October 1995. However in March 1993, the Governor of the Reserve Bank of Australia stated a specific desired numerical outcome for the inflation rate. Third, Table 1 reveals that in three of five countries the inflation target is very cautious. Only in the United Kingdom and New Zealand is the target rate of inflation lower than the rate of inflation observed in the last full Table 1 The adoption of targets in the countries studied Country
Month and year of target announcement
Inflation rate in previous calendar yeara
Inflation target
Australia Canada New Zealand Sweden United Kingdom
March 1993 February 1991 March 1990 January 1993 October 1992
1 3 5.5 2.3 5.6
2–3 percent 1–3 percent by 1995 0–2 percent 1–3 percent by 1995 1–2.5 percent by 1997
0.9%c
Below 2%
The European Central Bank Announcementb Germany, France, January 1999 Italy, Netherlands a
Percent change in the GDP deflator for Canada, percent change in the consumer price index for all other countries. b These four countries formally join the European Central Bank (the ECB) on this date. The ECB has both an inflation target and uses monetary growth as a guide to monetary policy. See the text for further discussion. c In European Central Bank (1999a), 0.9% is reported as the harmonized index of consumer prices (HICP) inflation rate for the 12 months ending November 1998. The ECB inflation target is specified as a percent change in the HICP.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1525
calendar year before the announcement. It may be difficult to see much impact of targets on expected inflation in the announcement year. Studying a longer period of monetary policy after the introduction of targets is useful. Fourth, Table 1 shows that in 3 of 5 countries, a time period was specified between announcement of the target and its eventual attainment, another reason to study a long period after the announcement. Fifth and finally, Table 1 notes that four countries in the study, Germany, France, Italy and the Netherlands formally joined the European Central Bank (ECB) on January 1, 1999. As described in European Central Bank (1999b), the ECB in October 1998, announced that price stability, its mandate, would be quantified and would be achieved using a strategy with ‘‘two pillars.’’ One ‘‘pillar,’’ price stability, was quantified in January 1999 as inflation below 2%, a specific numerical target for inflation. However, the ECB has also stressed that the conduct of monetary policy would also be based on a ‘‘prominent role for money,’’ that is, ‘‘a quantitative reference for monetary growth,’’ the second ‘‘pillar.’’ There is agreement that the ECB monetary regime is not an inflation-targeting regime, see Feldstein (2000) or Svensson (2000). ECB countries are not treated as inflation-targeting countries in this study.
3. The measurement of the behavior of expected inflation The data used in this study are found in Consensus Forecasts and its predecessor, Economic forecasts: a monthly worldwide survey.4 Each forecaster reports an expected rate of inflation for the current calendar year and an expected rate of inflation in the next calendar year. The month of the survey response is known. Forecasters drop in and out of the survey and many new forecasters are added over time. The data are not a monthly panel of forecasters. There are far more responses for the United States, the United Kingdom, and Japan than for other countries. The number of individual responses in a given month or even year in any country is highly variable before 1996.5 Survey responses are spread evenly over the year, the average month of response is either June or July in all countries. The survey data, with individual responses from each country within a calendaryear, are used to create three calendar-year measures related to the behavior of expected inflation. The first measure is the average next-year forecast, averaged over forecasters within a calendar year. This measure and the corresponding actual inflation rate are plotted in Fig. 1 for all countries. The vertical lines in Fig. 1 delimit 4 The data for New Zealand and Australia after 1996 are found in Asia Pacific Consensus Forecasts. This publication succeeded, for these two countries, Consensus Forecasts. The survey for the United States is not from Consensus Forecasts. Rather it is the Survey of Professional Forecasters (website www.phil.frb.org) maintained by the Federal Reserve Bank of Philadelphia. This is a quarterly survey with a structure similar to that Consensus Forecasts. I thank Dean Croushore for his help using this survey. 5 The data are described in more detail in Johnson (1998). One other small note is that until 1995 the reported measure of expected inflation in Canada is a response by forecasters to a question about the expected rate of increase in the GDP deflator and after 1995, expected inflation rate is measured using the consumer price index. Over the long run the consumer price index and the GDP deflator are closely linked.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Fig. 1. Actual and expected inflation in the next calendar year.
1526
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1527
years after the announcement of targets. In 9 of the 11 countries, excluding the Netherlands and Germany, average expected inflation and average actual inflation fall after 1990. Disinflations are larger in the targeting countries. The second measure is the standard deviation of the individual forecasts within a calendar year, a measure of forecast variability. This measure is plotted in Fig. 2 for both next-year and current-year forecasts. Variation across individual forecasts within a calendar year has three sources. First, in the same month different forecasters may interpret the same public information differently and/or may have different private information. Second, as time passes, a forecaster may revise his forecast, either using new information or changing his mind about the same information. Third, the data include different forecasters in different months of the same year. The standard deviation of current-year forecasts in Fig. 2 is generally smaller than the standard deviation of next-year forecasts in the same country in the same year. This should be the case. By the end of a calendar year, forecasters know most of the actual values of the price indexes needed for the current-year forecast. A monthly price index is usually announced with a one month lag.6 Forecasters should begin to agree on the current-year inflation forecast by year end. This same effect is present but much weaker for next-year forecasts. This is an important reason to concentrate on the properties of next-year forecasts in both standard deviations and levels. Fig. 2 shows that, in targeting countries, the standard deviation of next-year inflation forecasts falls after targets are introduced in Canada, New Zealand, the United Kingdom and Sweden but seems to rise in Australia. In the six non-targeting countries the standard deviation of next-year forecasts also seems to decline over time. The panel analysis measures any additional effect of the announcement of targets in reducing the standard deviation of next-year forecasts among forecasters. Fig. 3 presents the average next-year forecast errors in all years in all countries. A positive value of the forecast error indicates unexpected disinflation. Fig. 3 shows that in every case, the forecast error on the next-year forecast made in the year of the announcement of targets was positive. If you consider forecast errors for the first 5 years after the target announcement in each targeting country, 6 of 25 forecast errors are negative and 19 are positive. Target announcements did not prevent unexpected disinflation. However, it is not enough to look at forecast errors and draw the conclusion that the announcement of targets did not change the level of expected inflation. In countries with or without targets, there was an ongoing recession associated with disinflation. And in all countries, as Fig. 1 shows, as actual inflation fell, expected inflation did not catch up immediately (the dotted line is generally above the solid line as actual inflation falls). Thus, the disinflations in non-targeting countries were also unexpected. The third measure of the behavior of expected inflation studied is the average absolute next-year forecast error, averaged across individual forecasts with a year. Each of the three measures of the behavior of expected inflation is used to ask whether that behavior changed after the announcement of targets in targeting 6
The situation is slightly more complicated when the price index is quarterly and/or is a GDP deflator. This is true in Australia, New Zealand and in part of the Canadian sample.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Fig. 2. The variability of forecasts.
1528
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Fig. 3. Next year forecast errors.
1529
1530
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
countries relative to non-targeting countries after controlling for both countryspecific effects and worldwide effects.
4. Modeling the behavior of expected inflation 4.1. The framework The measure of the behavior of expected inflation in country i at time t; denoted Mi;t ; is modeled as a function of a vector of k country-related factors, Xi;t (made specific in Section 4.2); a variable Ci equal to 1 if the observation is from country i; a vector of dummy variables Yt with element equal to 1 if the observation is from year ECB t; and a dummy variable Ti;t that equals one if country i is a member of the European Central Bank in year t: Each observation from country i at time t is modeled as ECB þ bi Xi;t þ dTi;t þ eit : Mi;t ¼ bi;0 Ci þ yYt þ dECB Ti;t
ð1Þ
In (1) the coefficient bi;0 on the constant term Ci is country-specific and represents the effect of factors within that country that do not vary over the sample. The vector of coefficients y on the vector Yt represent worldwide effects on the measure of expected inflation in all countries that occur in each calendar year t: The single coefficient dECB is the effect on the measure of expected inflation M when a country joins the European Central Bank. As discussed in Section 2, the European Central Bank is not considered to be an inflation targeting bank.7 In (1) the vector of coefficients bi is specific to each country, that is, the effects of the Xi;t variables specific to each country differ by country. Finally, there are the dummy variables that capture the presence or absence of targets in that country. These variables are quite complicated. If targets have been announced in country i as of year t; then for that observation the element of the vector Ti;t is equal to one, otherwise it is zero. The elements of the vector of coefficients d measure the effects of targets on expected inflation. The d are the coefficients of central interest. The dummy variables that capture the effect of target announcements are organized in three ways in Table 2. One vector of dummy variables, called the ‘‘year-based’’ dummies, take values of one in all targeting countries in a given year after the targeting announcement. They j are denoted Ti;t ¼ 1 where j runs from 1 to 5 years after the target announcement and then 6,7 and 8, or 9,10 and 11 for the sixth, seventh or eighth, ninth, tenth or eleventh year of targets in country i: The coefficients on these seven dummy variables measure the average effect of targets across all targeting countries in the same year of targets, the year of targets being defined relative to the announcement year. For 1 example, the coefficient on dummy variable Ti;t measures the average effect on measure M over all targeting countries in the first year after targets were announced, the impact 7 The coefficients on the target-based dummies are not sensitive to treating the ECB as an inflationtargeting bank for the first 2 years of its existence. This is a very short additional period of targets if these countries are counted as targeting countries.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1531
Table 2 The dummy variables whose coefficients measure the effect of targets Variable
Year and country where dummy ¼ 1 (inclusive): all other values zero
Year-based dummies (seven separate variables) First year of targets: Ti;j1 ¼ 1 Australia (1993), Canada (1991), New Zealand (1990), Sweden (1993), United Kingdom (1993) Second year of targets: Ti;j2 ¼ 1 Australia (1994), Canada (1992), New Zealand (1991), Sweden (1994), United Kingdom (1994) Third year of targets: Ti;j3 ¼ 1 Australia (1995), Canada (1993), New Zealand (1992), Sweden (1995), United Kingdom (1995) Fourth year of targets: Ti;j4 ¼ 1 Australia (1996), Canada (1994), New Zealand (1993), Sweden (1996), United Kingdom (1996) Fifth year of targets: Ti;j5 ¼ 1 Australia (1997), Canada (1995), New Zealand (1994), Sweden (1997), United Kingdom (1997) Sixth, seventh or eighth year of targets: Australia (1998, 1999, 2000), Canada (1996, 1997, 1998), Ti;j628 ¼ 1 New Zealand (1995, 1996, 1997), Sweden (1998, 1999, 2000), United Kingdom (1998, 1999, 2000) Ninth, tenth or eleventh year of targets: Canada (1999, 2000), New Zealand (1998, 1999, 2000) Ti;j9211 ¼ 1 Country-based dummies (five separate variables) Australia: Ti;jAustralia ¼ 1 Australia 1993–2000 Canada: Ti;jCanada ¼ 1 Canada 1991–2000 New Zealand: Ti;jNZ ¼ 1 New Zealand 1990–2000 Sweden: Ti;jSweden ¼ 1 Sweden 1993–2000 United Kingdom 1993–2000 United Kingdom: Ti;jUK ¼ 1 The all targets dummy: Ti;jall ¼ 1
Australia 1993–2000 Canada 1991–2000 New Zealand 1990–2000 Sweden 1993–2000 United Kingdom 1993–2000
628 effect of the target announcement. The coefficient on dummy variable Ti;t similarly measures the average effect on measure M over all targeting countries in the sixth, seventh and eighth years after targets were announced, the longer-term effect of targets. The other vector of dummy variables related to targets are ‘‘country-based.’’ Here Country the variable Ti;t takes a value of 1 in all years after the specified country Australia announces targets and is zero otherwise. For example Ti;t equals one in each calendar year from 1993 to 2000 for the observations from Australia and is zero for all other observations. The coefficient on this dummy variable measures the average effect of targets on measure M over all years of targets in Australia. Five such coefficients are estimated, one for each targeting country. Imposing either the constraints that the effects of targets are equal in all years after the announcement year or the constraints that the effects of targets are equal in all targeting countries is equivalent to estimating the model with a third dummy all variable. The third target-related dummy variable Ti;t equals unity for any observation in any country when targets have been announced in any country i at
1532
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
all time t and is zero otherwise. For the ‘‘country-based’’ vector of dummies, using the Ti;t dummy in (1) imposes four constraints that the effects of targets are equal in all targeting countries. For the ‘‘year-based’’ vector of dummies, using this dummy in (1) imposes six constraints that the effect of targets in the seven designated targeting periods are equal. These constraints are always rejected, an interesting finding in itself. The effects of targets are not statistically equivalent across years of targeting or across all countries. The single coefficient on the Ti;t dummy variable is useful only as a summary statistic of the average effect of targets over all years and all targeting countries. The 11 countries and 17 years (1984–2000) of data provide up to 185 observations, there are no data in Australia for 1984 and there are no next-year forecast responses in the Netherlands in 1986, only the current-year forecast question was answered. The set of 11 equations, one per country defined using (1), is a system of seemingly unrelated regression (SUR) equations, see Greene (2000), Chapter 15 for further discussion. The most general specification of the error term in such a system is described by
Eðeit ; eit Þ ¼ s2i
ð2Þ
Eðeit ; ejt Þ ¼ s2ij
ð3Þ
and
and, in some country equations Eðeit ; eit1 Þ ¼ ri :
ð4Þ
Expression (2) represents error terms that are heteroscedastic across countries. Expression (3) represents error terms in different countries that are correlated in a given calendar year. This correlation could be the result of year-specific worldwide effects on expected inflation in all countries, for example an oil shock. This correlation could also result when linkages or other similarities occur between specific country pairs. Finally (4) represents the possibility that error terms in (1) within a country may be auto-correlated. Diagnostic tests reveal that the error terms in (1) are typically heteroscedastic and are usually correlated across years. Some but not all estimated values of ri are large, there are some country-specific dynamics. Results are presented with and without the correction for serial correlation. The AR(1) specification estimates up to 11 additional AR(1) coefficients and drops the first observation for each country. Greene (2000) reports that the efficiency gain from the AR(1) correction may be outweighed by the loss of the one observation in a relatively small sample.8 Results from specifications with and without AR(1) terms are presented for this reason.9 8 I do not report the individual country AR(1) coefficients in the results below. The model was first estimated with all 11 AR(1) coefficients, then AR(1) coefficients that were insignificant were dropped. Results are presented in Table 3 with and without the AR(1) correction. 9 Results from two other econometric specifications common in the panel literature were also estimated and are presented in the working paper version of this paper available at http://www.wlu.ca/~wwwsbe/ sbe2000/html/economics/working papers2001.html. The time-series cross-section (TSCS) specification imposes the restrictions that the slope coefficients (but not the constant terms) are equal across countries and that the common year effects (measured by the vector of coefficients y) are zero on expression (1). Correlations across countries in the same year as in (3) are considered sufficient to capture common year
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1533
4.2. The effects of inflation targets on the behavior of expected inflation Measures of the effect of targets on the level of expected inflation are presented in the left-hand 2 columns of Table 3. There are two types of country-specific control variables (the Xi;t ) in the equations measuring the effect of targets on the level of expected inflation. The business cycle variable is the rate of unemployment observed in the 12 months closest to and preceding January of the current calendar year minus the rate of unemployment observed in the past 60 months closest to and preceding January of the current calendar year. A value of 1 indicates observed unemployment in that country is 1 percentage point higher in the past 12 months than the average of the preceding 60 months.10 The element of the bi vector that is the coefficient on this variable, while it varies from country to country, is usually negative and significant, an increase in unemployment relative to its immediate past reduces the level of expected inflation. The other four elements of the Xi;t vector of variables are four lagged inflation rates measured over 12 months intervals using the forecasters’ knowledge as of January 1 of the current calendar year.11 Countries with a history of high inflation are expected to have higher expected inflation. The most important results in Table 3 are that inflation target announcements reduce, with both economic and statistical significance, the average level of expected inflation in the next calendar year after controlling for both individual country and worldwide effects. If the purpose of announcing inflation targets is to reduce the level of expected inflation, targets are a success in most countries in most years. The seven coefficients on the year-based dummies (the year of targeting relative to the announcement year) present an interesting pattern in both the AR(1) and non-AR(1) estimates. The coefficients on the different years of targeting are not equal. The effect of targets seems to peak later in the targeting period, the reduction in next-year (footnote continued) effects. These extremely strong restrictions require every slope coefficient to be the same in every country and every year-effect coefficient in the vector y to be zero. These restrictions are rejected. A third specification, the fixed effects (FE) specification restores non-zero coefficients on the y vector. The TSCS and FE specifications use the additional assumptions of equal slope coefficients across countries to more precisely identify the effects of targets. The countries are assumed to be similar enough (much as individuals or firms in other panel data sets are assumed to be similar enough) to allow a panel estimator of the characteristics of the ‘‘average’’ country. The TSCS and FE estimates of the coefficients on the dummy variables associated with targets are very similar to the SUR estimates. In earlier versions of the paper, the results are also robust to dropping a country (the Netherlands) and to reducing the sample, dropping the years 1984, 1998, 1999 and 2000. The results presented in Table 3 are robust to a variety of samples and estimation procedures. 10 This variable represents the forecaster’s knowledge of the state of the business cycle as the beginning of the year. This measure of the state of the business cycle is roughly comparable across the different countries with very different average unemployment rates. This is useful in the TSCS and FE specifications discussed above. 11 Models with both three and four lags of actual inflation were estimated. The results are not sensitive to this choice. It is difficult to choose between these two specifications, the lagged inflation measures are collinear. When the fourth lag of inflation is included, its coefficient is typically positive and statistically significant but the coefficients on either the second or third lag dwindle while with three lags of known inflation, all three coefficients tend to be significant.
1.22(0.26)n
1.16(0.26)n
1.38(0.23)n
1.25(0.19)n
0.41(0.25)
1.73(0.39)n
1.12(0.20)n
3.79(0.71)n
2.57(1.56)
1.16(0.26)n
2.42(0.20)n
3 Year 3: Ti;t
4 Year 4: Ti;t
5 Year 5: Ti;t
628 Years 6, 7 or 8: Ti;t
9211 Years 9, 10 or 11: Ti;t
Australia Australia: Ti;t
Canada Canada: Ti;t
NZ New Zealand: Ti;t
Sweden Sweden: Ti;t
UK United Kingdom: Ti;t
all All targets: Ti;t
2.55(0.25)n
0.32(0.32)
4.33(1.7)n
8.10(0.16)n
1.07(0.26)n
2.35(1.0)n
1.01(0.43)n
3.02(0.38)n
3.56(0.43)n
3.01(0.46)n
2.21(0.44)n
1.85(0.39)n
0.90(0.29)
0.07(0.02)n
0.13(0.19)
0.29(0.19)
0.84(0.14)n
0.03(0.02)
0.46(0.24)
0.03(0.03)
0.00(0.03)
0.02(0.03)
0.02(0.04)
0.07(0.03)n
0.01(0.03)
0.36(0.02)
Indicates coefficient on the dummy variable is statistically significant at 5 percent.
1.37(0.26)n
2 Year 2: Ti;t
n
1.14(0.22)
1 Year 1: Ti;t n
0.02(0.03)
0.15(0.11)
0.16(0.19)
0.02(0.26)
0.04(0.03)
0.43(0.24)
0.15(0.07)n
0.14(0.06)n
0.12(0.07)
0.20(0.07)n
0.06(0.07)
0.00(0.06)
0.04(0.05)
No AR(1)
AR(1) n
AR(1)
No AR(1)
On the standard deviation of expected inflation
On the level of expected inflation
n
Coefficients on target dummies (standard errors)
Table 3 The effect of inflation targets on the behavior of expected inflation in the next calendar year
0.21(0.14)
0.61(0.43)
0.27(0.54)
0.28(0.30)
0.19(0.18)
0.86(0.89)
0.08(0.21)
0.58(0.19)n
0.46(0.26)
0.04(0.21)
.13(0.18)
0.20(0.15)
0.00(0.14)
AR(1)
0.19(0.13)
0.11(0.05)n
0.62(0.50)
0.30(0.34)
0.21(0.18)
0.83(0.91)
0.04(0.20)
0.64(0.19)n
0.54(0.26)n
0.09(0.21)
0.05(0.18)
0.24(0.14)
0.02(0.13)
No AR(1)
On the absolute average error in next-year forecasts
1534 D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1535
expected inflation is smallest in the first and last years of targets and highest in the middle years. This pattern is more pronounced in the specification without the AR(1) correction.12 Since years 9, 10 and 11 of targeting occur only in Canada and New Zealand, the coefficient on this element of the ‘‘year-based’’ vector of dummy variables is estimated using data from only 2 and not 5 targeting countries and some caution is needed in its interpretation. However, a slow effect of targets on the level of expected inflation makes good economic sense. Targets were designed to alter long-term patterns of expected inflation in relatively high inflation countries. These countries experienced significant recessions to reduce inflation. Targets then seem to prevent the increase in expected inflation to ‘‘normal’’ (for that country) levels which would have been associated with economic recovery. There is variation in the estimates of the average effect of inflation targets on the level of expected inflation across countries. The estimates of the coefficient on the NZ dummy variable measuring the effect of targets in New Zealand ðTi;t Þ vary from 3:79 to 8:10; meaning that in the period after the announcement of targets, the level of expected inflation in New Zealand was reduced by 3.79–8.1 percentage points. The 8 percentage point effect in New Zealand seems implausible although Fig. 1 presents an extremely rapid decline in expected inflation in New Zealand. Very large changes in the level of expected inflation in New Zealand could have been associated with the large scale fiscal, trade and monetary reforms in New Zealand that coincided with targets. These reforms were much broader than in other countries and could account for some of the strength of the New Zealand effect. The Swedish case also generates a relatively large ‘‘country-based’’ coefficient on dummy Sweden variable Ti;t (2:57 or 4:33), albeit with a larger standard error than the analogous coefficient in New Zealand.13 all The single coefficient on the dummy variable, Ti;t ; is a summary measure of the average effect of targets on the level of expected inflation in all the years in all the targeting countries. Constraints are imposed that the effects of targets are equal across years and countries, constraints we can reject.14 The summary coefficient suggests a substantial reduction in the level of expected inflation after the announcement of targets, a reduction not explained by either country-specific effects or worldwide effects. Table 3 provides strong evidence inflation targets directly reduced the level of expected inflation. The other two measures of the behavior of expected inflation studied are the standard deviation of next-year forecasts across forecasters within a given calendar 12
There is a similar coefficient pattern in both the TSCS and FE specifications as well. Some of the coefficient variation and the larger standard errors seem to be associated with the most complex SUR specification (1) rather than the TSCS or FE specifications described in footnote 9. There are some reasons to prefer the estimates from the TSCS or FE models. In these specifications the elements of the bi coefficient vector are constrained to be equal across the countries. Without the AR(1) correction there are additional observations. Alternative estimation procedures yield tighter estimates of the effects of inflation targets. The effects of targets remains strongest in New Zealand and are generally comparable in Australia, Canada, Sweden and the United Kingdom. 14 The test statistics on the null hypotheses of equal country or year effects are not reported. The coefficient values as reported in Table 3 makes it clear these equality constraints are rejected. 13
1536
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
year, plotted as the dotted line in Fig. 2, and the average absolute forecast error plotted in Fig. 3. The panel specifications with these dependent variables have two country-specific independent variables: the lagged level of average inflation calculated over the last 5 years as observed at the beginning of the calendar year, and the average standard deviation of inflation observed over the previous 5 years.15 For both dependent variables, the coefficient on the lagged level of inflation is usually positive and significant. A higher past level of inflation in a country is associated both with more diverse inflation forecasts and larger average next-year forecast errors. Past variability of inflation, given the level of past inflation, is usually marginally significant in the models of inflation variability but not usually significant in the models of average absolute forecast errors. After the effects of higher past inflation or higher past inflation variability are taken into account, the announcement and implementation of inflation targets has little additional effect on the standard deviation of inflation forecasts. The middle panel of Table 3 presents these results. Most of the coefficients on target-related dummy variables do not differ significantly from zero. One oddity is that in first year of inflation targets in the AR(1) specification, there is even some weak evidence of a significant increase in next-year inflation uncertainty, not the result desired by monetary authorities at the announcement of targets. By the fourth and fifth years of targets the signs on ‘‘year-based’’ target dummies are negative but not generally significant. When the coefficients on ‘‘country-based’’ target dummies are examined, in New Zealand targets significantly increase the dispersion of forecasts in the AR(1) specification but not in the specification without the AR(1) correction. The New Zealand result is obtained with quite limited data, prior to targets New Zealand has only three forecasts in 1984 and 1988, four forecasts in 1985, five forecasts in 1986 and six forecasts in both 1987 and 1989. New Zealand and the Netherlands have the fewest number of individual forecasts. However, Table 3 suggests that targeting, beyond an important indirect effect through successful reduction in the level of past inflation, has no additional effect on the standard deviation of expected inflation. The last investigation in this study compares the behavior of average absolute forecast errors in targeting and non-targeting countries. These results are presented in the right-hand side panel of Table 3. To this point in the history of inflation targets, there is very little evidence that targeting regimes experience lower average absolute value next-year forecast errors relative to the non-targeting countries. Figs. 1 and 3 leave little doubt that the reduction in inflation in targeting countries was unexpected. This is the standard result of previous studies. Italy, Japan, France and the United States, non-targeting countries, also experienced unexpected disinflation. The dependent variable in (1) is the average absolute next-year forecast error across forecasters. The country-specific control variables are the average level and standard deviation of inflation over the past 5 years. Higher lagged inflation is related to higher average forecast errors in some countries. But the effects are weak and unsystematic. Thus, the reduction in the level of inflation associated with targets 15
These models were also estimated with similar results when lagged inflation or its standard deviation is calculated over the previous 10 years of observed inflation.
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
1537
indirectly reduces average forecast errors in some countries by lowering past inflation levels. Any additional effects of the target announcement are quite weak. The only observable direct effect of targets on absolute forecast errors is that the coefficients on target-year dummies do become negative later in the targeting process. By years 5–8 after the announcement of targets, there may be some small all effect of targets in reducing average forecast errors. The summary coefficient on Ti;t is negative but not significant. Although targeting does not directly reduce forecast errors in targeting countries relative to those in non-targeting countries, recall the strong evidence that targets reduced the level of expected inflation in targeting countries after the announcements. Fig. 1 shows that actual inflation fell more rapidly in targeting countries than in non-targeting countries. Targeting countries, through the effect of targets on the level of expected inflation, avoided even larger forecast errors than would have occurred in the absence of targets. For a larger disinflation, targeting countries had similar forecast errors. This is a success of targets.
5. Conclusions One purpose in introducing inflation targets was to alter the behavior of expected inflation. The evidence is very strong that the period after the announcement of inflation targets is associated with a large reduction in the level of expected inflation. That reduction is of both economic and statistical significance, at least one to two percentage points in most years after targets. That reduction took place in all 5 countries with inflation targets. This is an important success of inflation targets. Targets were most successful in New Zealand where other reforms may have played a role in reducing expected inflation. The actual rate of inflation falls in both targeting and non-targeting countries. With lower actual inflation, the dispersion of inflation forecasts falls in all countries. But there is little or no additional reduction in the dispersion of inflation forecasts associated with the period after the announcement of inflation targets. Targets are not yet a success in this dimension. Finally, in the disinflations of the 1990s both targeting and non-targeting countries experience large forecast errors. Inflation targets did not reduce absolute average forecast errors in targeting countries relative to those in non-targeting countries. However in targeting countries expected inflation falls more rapidly with inflation targets. These events are reconciled if inflation targets prevented forecast errors in targeting countries from being larger than they would have been without targets. Inflation targets allowed a larger disinflation with smaller forecast errors to take place in targeting countries. This is also an important success of inflation targets.
Appendix A. Data All forecasts of inflation except for the United States were drawn from economic forecasts: a monthly worldwide survey, its successors Consensus Forecasts and
1538
D.R. Johnson / Journal of Monetary Economics 49 (2002) 1521–1538
Asia-Pacific consensus forecasts between 1984 and 1998. Forecasts for the United States are the ASA-NBER forecasts available at the Reserve Bank of Philadelphia website: www.phil.frb.org. Actual inflation rates and actual unemployment rates for all countries are drawn from the OECD database and updated from various national websites to February 2000. Unemployment rates are seasonally adjusted. These are not the OECD standardized unemployment rates. They are the unemployment rates reported by the national statistical agencies to the OECD. In New Zealand unemployment rates from a labour force survey are available only after December 1985. Before 1985, a proxy for the unemployment rate is constructed by dividing the number of registered unemployed by employment and unemployment.
References Ball, L., 1994. What determines the sacrifice ratio? In: Mankiw, N.G. (Ed.), Monetary Policy. University of Chicago Press, Chicago, pp. 155–182. Berg, C., Lundkvist, P., 1997. Has the inflation process changed? Sveriges Riksbank Quarterly Review 5, 5–25. Bernanke, B.S., Laubach, T., Mishkin, F.S., Posen, A., 1999. Inflation Targeting: Lessons from the International Experience. Princeton University Press, Princeton, NJ. Bowen, A., 1995. British experience with inflation targetry. In: Leiderman, L., Svensson, L.E.O. (Eds.), Inflation Targets. Centre for Economic Policy Research, London, pp. 53–68. European Central Bank, 1999a. Editorial, European Central Bank Monthly Bulletin, January 1999, pp. 9–10. European Central Bank, 1999b. The stability-oriented monetary policy strategy of the Eurosystem. European Central Bank Monthly Bulletin, January 1999, pp. 39–50. Feldstein, M., 2000. The European Central Bank and the Euro: the First Year, National Bureau of Economic Research Working Paper, 7517. Fischer, A., 1995. New Zealand’s experience with inflation targets. In: Leiderman, L., Svensson, L.E.O. (Eds.), Inflation Targets. Centre for Economic Policy Research, London, pp. 32–52. Greene, W.H., 2000. Econometric Analysis, 3rd Edition. MacMillan, New York. Groenveld, J.M., 1998. Inflation patterns and monetary policy: lessons for the European Central Bank. Edward Elgar, Cheltenham. Johnson, D.R., 1997. Expected inflation in Canada 1988–1995: an evaluation of Bank of Canada credibility and the effect of inflation targets. Canadian Public Policy 23, 223–258. Johnson, D.R., 1998. The credibility of monetary policy: international evidence based on surveys of expected inflation. In: Macklem, T. (Ed.), Price Stability, Inflation Targets and Monetary Policy. Proceedings of a Conference held by the Bank of Canada, May 1997. Bank of Canada, Ottawa, pp. 361–395. Johnson, D.R., 2000. The effect of targeting inflation on the level of expected inflation in five countries. Wilfrid Laurier University Working Paper 2000–01. Laidler, D.E.W., Robson, W.B.P., 1993. The Great Canadian Disinflation: The Economics and Politics of Monetary Policy in Canada 1988–1993. C.D. Howe Institute, Toronto. Leiderman, L., Svensson, L.E.O., 1995. Inflation Targets. Centre for Economic Policy Research, London. Svensson, L.E.O., 1995. The Swedish experience of an inflation target. In: Leiderman, L., Svensson, L.E.O. (Eds.), Inflation Targets. Centre for Economic Policy Research, London, pp. 69–89. Svensson, L.E.O., 2000. The first year of the Eurosystem: inflation targeting or not? National Bureau of Economic Research Working Paper No. 7598.