Survey forecasts in Brazil: A prismatic assessment of epidemiology, performance, and determinants

Journal of International Money and Finance 31 (2012) 1371–1391

Contents lists available at SciVerse ScienceDirect

Journal of International Money and Finance journal homepage: www.elsevier.com/locate/jimf

Survey forecasts in Brazil: A prismatic assessment of epidemiology, performance, and determinants Fabia A. Carvalho*, André Minella Central Bank of Brazil, Research Department SBS Qd 3 Bloco 3 Ed Sede do Bacen, Brasília, DF 70074-900, Brazil

a b s t r a c t JEL classification: E47 E49 E58 Keywords: Survey forecasts Inflation expectations Inflation targeting Credibility

This paper assesses the behavior of survey forecasts in Brazil during the inflation targeting regime, when managing expectations is one of the cornerstones of the conduct of monetary policy. The distinctive database of the survey conducted by the Central Bank of Brazil (BCB) among professional forecasters allows for a thorough investigation of the epidemiology, determinants, and performance of forecasts. The main results are: i) top performing forecasters are influential to other forecasters; ii) survey forecasts perform better than vector autoregressive model-based forecasts; iii) common forecast errors prevail over idiosyncratic components across respondents; iv) inflation targets play an important role in inflation expectations; and v) agents perceive the BCB as following a Taylor rule consistent with inflation targeting. The last two suggest high credibility of the monetary authority. Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction Modern macroeconomic theory and central bank practice have made the management of expectations a key ingredient in a successful monetary policy.1 In particular, taming expectations is at the heart of the conduct of monetary policy in countries where the construction of credibility is warranted, which is usually the case in emerging market economies. That calls for a thorough understanding of the inner workings of expectations formation in the economy. The purpose of this paper is to understand the nature of these expectations in Brazil, focusing on aspects that are of particular relevance to managing expectations by the monetary authority. To this * Corresponding author. Tel.: þ55 6134141050; fax: þ55 6134143913. E-mail address: [email protected] (F.A. Carvalho). 1 See, for instance, Blinder et al. (2008) and Woodford (2001). 0261-5606/$ – see front matter Ó 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.jimonfin.2012.02.006

1372

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

end, we use the forecasts from the survey conducted by the Central Bank of Brazil (BCB) since the implementation of inflation targeting in 1999. The dataset is distinctive as it combines two key features for our analysis. First, it samples from an accredited survey conducted among professional forecasters. The survey is broadly used in the economy and its excellence is internationally recognized.2 Second, it refers to a period when the monetary authority is building up credibility. Data on expectations are mostly obtained from surveys d usually conducted among professional forecasters, academics or households d or from financial instruments, in which case the extraction of embedded expectations calls upon a number of assumptions regarding the statistical properties of the risk premium and of shocks to the term structure. However, in general, financial instruments in emerging market economies are not duly informative about inflation expectations as the risk premium is usually volatile and long-term bond markets are not well developed. On the other hand, households do not offer specialized forecasts, whilst academics do not have as many incentives as professional forecasters to provide accurate forecasts. Therefore, surveys from professional forecasters seem to be the most appropriate data source on macroeconomic expectations in emerging markets. High inflation in Brazil lasted for about two decades and ended in 1994, just a few years before the adoption of inflation targeting. As the country had earned a reputation of weak monetary and fiscal institutions, building up credibility emerged as a key issue in the conduct of monetary policy in Brazil under the new regime.3 Against this background, managing inflation expectations became even more important for the country. The old management adages “what you do not measure you cannot manage” and “to measure is to know” apply to the management of expectations, in our opinion. With that in mind, this paper attempts to understand the features of expectations in Brazil. Initially, we refer to Carroll (2003) to assess whether published forecasts of the top-five performing forecasters contaminate the remainder of professional forecasts in a systematic way. Next we compare survey forecasts’ predictive power with autoregressive model-based forecasts (ARMA, VAR and BVAR models). Then, we test the joint performance of survey participants’ forecasts for a multivariate set of macroeconomic variables using the methodology developed in Bauer et al. (2003, 2006) and Eisenbeis et al. (2002). Finally, we empirically investigate the factors driving the levels and dispersion of survey forecasts for selected macroeconomic variables. The main results found in the paper can be summarized as follows: i) the top-five performing forecasters are influential in the survey of inflation, interest rate and exchange rate forecasts; ii) survey inflation forecasts had similar or better forecast performance than autoregressive model-based forecasts with standard information sets; iii) in the decomposition of forecast errors for inflation, interest rate and exchange rate, the common forecast error component prevails over the idiosyncratic component across survey respondents; iv) the joint performance of survey forecasts in Brazil has improved over the past years, albeit not monotonically; v) inflation uncertainty is positively related to increasing inflation and to country-risk premium; vi) credibility in inflation targets is noteworthy, as targets have played an increasingly important role in explaining inflation expectations after the country overcame the 2002–2003 confidence crisis; and vii) private agents perceive the BCB as following a Taylor-type rule that is consistent with the inflation targeting framework, with policy interest-rate expectations depending mostly on deviations of inflation expectations from the target. The investigation of inflation and interest-rate forecast determinants provides evidence that the BCB has built up credibility during the inflation targeting regime. The paper is organized as follows. Section 2 provides some details on the database used and gives an overview of the behavior of survey forecasts for selected macroeconomic variables. The section that follows investigates the dissemination of published top-five forecasts to the sample of professional forecasts. Section 4 evaluates the predictive power of survey forecasts and assesses the joint performance of survey forecasts. Section 5 estimates the determinants of forecasts, and provides

2 The survey, for instance, won the second prize of the II Regional Award for Innovation in Statistics 2009/2010 in Latin America and Caribbean, offered by the World Bank. 3 See Fraga et al. (2003) and Minella et al. (2003).

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1373

some evidence on the determinants of inflation forecast dispersion. The last section concludes the paper. 2. Inflation targeting and survey forecasts: overview and database For the purposes of this paper, we can divide the last decades of economic policy regimes in Brazil into three periods. Up to 1994, the country experienced increasing and high inflation rates and a sequence of failing stabilization programs. The monthly inflation rate in June 1994 reached 47.4% per month, corresponding to annualized 10,444.6%. The Real Plan, launched in 1994, put an end to the high inflation period. However, the Balance of Payments crisis in late 1998 and early 1999 disrupted the prevailing regime, which had been based on a crawling peg system. The exchange rate soared. Measured as domestic currency units per USD units, the exchange rate went from 1.21 to a peak of 2.16 in less than two months. Analysts predicted a double-digit inflation rate for 1999, with some forecasts reaching as high as 80%. In a very short period of time, a new regime was implemented4 to stand on three pillars: inflation targeting, floating exchange rate, and fiscal soundness (with an important change in the fiscal regime). In this context, anchoring inflation expectations became crucial. The inflation index chosen as target was the Broad National Consumer Price Index (IPCA). The target value is announced in June for the calendar year that starts one and a half years later. The implementation of inflation targeting involved several institutional arrangements, such as: i) timely release of the minutes of COPOM (Committee for Monetary Policy); ii) publication of a quarterly Inflation Report; iii) investment in research and modeling at the BCB; and iv) implementation of a reliable survey-based forecasting system. The BCB’s survey comprises around 100 professional forecasters, which are expected to provide forecasts for a number of macroeconomic variables over different horizons. In the sample used in this paper, there are approximately 70 professionals providing regular forecasts for inflation. The respondents are financial institutions, consulting firms, asset managers and other legal entities with an assigned economist responsible for the information passed on to the BCB. Since 2001, analysts have been providing their forecasts through an on-line system. The forecast database is updated daily, and aggregate measures, such as median, mean, standard deviation, are published weekly in the “Focus – Market Readout”. Historical data on a daily basis is also available on the BCB’s website. The forecast horizon varies across variables. For the IPCA inflation forecasts, respondents may supply forecasts for monthly inflation over the current and next 13 months5 and for the 12-month ahead inflation (measured as year-on-year price increase), in addition to forecasts for the current and next four calendar years. Similar horizons apply to the policy interest rate and the exchange rate, with the distinction that, in these cases, the forecasts refer to values at the end of the month or of the calendar year, depending on the frequency of the forecasted variable, and to the average value over the calendar year. As new participants are often added to the survey, and others drop out, the panel of survey forecasts is unbalanced. The question asked is regarding the analyst’s forecast for the corresponding variable.6 We should stress that the private sector in Brazil is reasonably well developed in terms of resources allocated to economic analysis. An incentive mechanism was designed to stimulate participants to provide thoughtful figures. The names of the top-five forecasters are published regularly, according to the forecast performance in short-, medium- and long-term horizons. The mean and median forecasts of the top-five forecasters are published in the same weekly release that contains the aggregate forecasts. The BCB survey has gained prominence in the economic debate throughout the country, in many instances serving as a benchmark for forecasts used in budget planning and price adjustments. The weekly release of its numbers is part of the domestic news.

4

For the transition between the two regimes, see Ötker-Robe et al. (2007). Since 2010, monthly forecasts can be input for up to 17 months ahead. 6 Additional information on the survey can be found in Carvalho and Bugarin (2006), Marques et al. (2003), and at http:// www.bcb.gov.br/?INVESTOR, which also provides the series of the forecast aggregate measures. 5

1374

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

14

12

10

%

8

6

4

2 00

01

02

03 04 Surveyed Date

05

06

07

08

12-Month Ahead Inflation Expectations 12-Month Ahead Inflation Target Fig. 1. Inflation forecasts and target (12 months ahead) – 2000:1–2008:7.

Fig. 1 shows survey inflation forecasts and inflation targets in Brazil.7 The dynamics of inflation expectations have not been stable, and that calls for fragmenting the sample into three periods. The first period covered the implementation and first years of the inflation targeting regime. Inflation targeting in Brazil was put in place rapidly and at a moment of great uncertainty. Yet, inflation expectations became relatively well anchored, closely tracking the declining targets of that period. The second period, starting in mid-2001, was marked initially by a deviation of expectations from the target, triggered by the rationing of electric energy and exchange rate depreciation. However, the subsequent, and most important, event was a confidence crisis at the end of 2002 and beginning of 2003, characterized by an increased perception of a disruption in the policy regime. That, together with a high exchange rate depreciation, made inflation expectations soar to reach 12% p.a., whereas the target was 4%. However, with a strong reaction of the BCB and increasing confidence that the policy regime would be maintained, inflation forecasts receded to around 7% in the middle of 2003 and to less than 6.0% at the end of the year. The third period was characterized by the end of the confidence crisis and its effects, the return of expectations to values approaching the target, and the consolidation of the inflation targeting regime. Inflation forecast errors also portray those movements (Fig. 2).8 The unexpected inflationary shocks of 2001 and mainly of 2002 led to forecast errors as high as 12 percentage points. On the other hand, the forecasts made at the very end of 2002 and first half of 2003 overestimated future inflation, a likely result of the perception of a relevant probability of a change of the policy regime, which did not materialize. In the following years, forecast errors were relatively lower, reflecting the construction of a more stable economic environment in the country. With the inflation rise in 2008, however, forecasts that had been made one year earlier ended up underestimating future inflation.

7 In this paper, inflation forecasts always refer to the IPCA. For every period depicted, the forecast value shown refers to the median of the 12-month ahead inflation forecasts (cumulative inflation from t through t þ 11), surveyed on the eve of the release of the IPCA-15 index, a date which we refer to as “critical date”. The inflation target series takes into account the fact the BCB pursued an adjusted target for 2003 and 2004, and an off-center objective within the original target range for 2005, announced in September 2004, which constituted the main reference for agents. We refer to this series as “adjusted target series”. 8 Forecast errors are defined as actual minus forecasted inflation. Note that, since forecasts refer to a 12-month horizon, forecast errors tend to have high serial correlation.

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1375

16

12

p.p.

8

4

0

-4 2000

2001

2002

2003 2004 2005 Surveyed Date

2006

2007

Fig. 2. Inflation forecast errors (12-month actual inflation minus forecasted inflation) – 2000:1–2007:8.

Figs. 3 and 4 show forecasted values for the Selic interest rate and for the change in the exchange rate.9 The Selic rate corresponds to the policy interest rate in Brazil d its target is set in the COPOM meetings. It is the rate on overnight repurchase operations between banks, with government securities as collateral. The Selic rate forecasts tend to closely track inflation forecasts, a relationship we elaborate further on in Section 5. The exchange rate, in turn, is defined as units of domestic currency per dollar; so an increase in the exchange rate means a depreciation of the domestic currency. Expectations of depreciation prevailed in most of the sample, except for the confidence crisis period and most of 2007. 3. Epidemiology of the survey forecasts Economic literature abounds in theories of how expectations are formed (Evans and Honkapohja, 2001; Colucci and Valori, 2005; among many others). In empirical studies on the formation of expectations, it is common to test whether they conform to the rational expectations hypothesis. In the case of Brazil, studies point to the absence of systematic forecast biases, although the evidence suggests that available information could have been used more efficiently.10 In this section, however, we are concerned about a different question. We test whether a selected group of top forecasters in BCB’s survey is influential to the remaining survey participants. Should top forecasters be a focal point of professional forecasters, monetary authority should weigh the option of using this group to help transmit its policy intentions and economic assessment to a wider audience. On this issue, Blinder et al. (2008) stress that “central bank communication is (.) a two-way street: it must have both a transmitter and a receiver, and either could be the source of uncertainty or confusion.”

9 Both are surveyed on the same dates used for inflation forecasts. The values for the Selic, starting in 2002:1, are calculated as the arithmetic average, from month t þ 1 through t þ 12, of the forecasts for the Selic interest-rate target at the end of each month. The values for the exchange rate change, in turn, are calculated as the average exchange rate forecasted from t to t þ 11 divided by its average value in t  1. 10 See Carvalho and Bugarin (2006), which used a set of different methods, including panel data, and, for more recent data, Carvalho and Minella (2009), which focused on aggregate measures. Guillén (2008) tested a set of expectations theories and found that the median forecast is more likely to conform to the sticky-information theory. The evidence on other countries is mixed (Keane and Runkle, 1990). For Chile and Mexico, Carvalho and Bugarin (2006) found that expectations usually passed the tests for unbiasedness and efficiency, but also with some exceptions for efficiency, mainly for Mexico in the case of the interest rate.

1376

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391 24 22 20 18 % 16 14 12 10 2002

2003

2004

2005

2006

2007

2008

Surveyed Date

Fig. 3. Selic interest-rate forecasts (forecast of the average Selic target over a 12-month horizon) – 2002:1–2008:7.

The tests in this section are inspired on Carroll’s (2003) epidemiology model, with a few distinctions. In Carroll’s baseline model, people can either maintain their inflation expectations unchanged from one period to another or update them based on professional forecasts published in newspaper articles. This results in a testable equation that relates current inflation expectations to current professional forecasts and past inflation expectations. Carroll tests this model by using the mean forecast in the Michigan Survey of Households as a proxy for people’s expectations and the mean forecast of the Survey of Professional Forecasters as a proxy for professional forecasts published in newspapers. The first modification we introduce regards the proxies we use to conduct the epidemiology tests. We apply the epidemiology tests to each participant in the Brazilian professional forecast survey, not only to the mean forecast. By using disaggregated data, we are able to analyze the micro behavior of forecasts more straightforwardly.11 We let the median of the forecasts of the top-five professional forecasters be the analog of Carroll’s newspaper articles. The BCB has been announcing top-five forecasters since July 2001. Professional forecasters are ranked according to their performance under three different forecast horizon periods. Short-, mediumand long-term top-five forecasters are those with the best performance considering one-, one- to sixand 12-month horizons, respectively,12 and their forecasts are published on a weekly basis. Since we focus on 12-month ahead forecasts, we use the best long-term forecasters in our tests. A second modification we introduce here refers to the model of expectations formation itself. We allow for the possibility that, when confronted with their forecast errors, people react to them. This is to account for the evidence presented in Carvalho and Bugarin (2006) of forecast error correction in the formation rule of median professional forecasts in two major Latin American countries for the period from 1999 to 2005. We also introduce a few controls in the testable equations, to allow for the possibility that people also employ their own statistical methods to help fine tune their forecasts. We also extend our analysis beyond inflation forecasts. We test for epidemiology of top-five forecasts in three variables: inflation, average (Selic) policy interest rate, and average exchange rate. For each forecasted variable xt, the first specification we test is the following:

  e;j e;top5 e;j e;j xt;12m ¼ a þ b1 xt;12m þ b2 xt1;12m þ b3 xt13;12m  xt12;t1 ; where the terms

11 12

e;j xt;12m ,

e;j xt1;12m ,

and

e;j xt13;12m

(1)

represent forecasts over a 12-month horizon made at t,

Guillén (2008) uses aggregate measures of the Brazilian survey to test for a distinct type of epidemiology, based on information costs. The horizon used has changed slightly over time.

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1377

12

8

4 % 0

-4

-8 2002

2003

2004

2005 2006 Surveyed Date

2007

2008

Fig. 4. Exchange rate change forecasts (forecast of the average exchange rate change over a 12-month horizon) – 2002:1–2008:7.

e;top5

t-1 and t-13, respectively; xt;12m are forecasts made by the top-five forecasters, collected a week before the analyst reports its current prediction for the variables; and xt12;t1 are the actual values from t-12 to t-1. In particular, we are interested in testing whether respondents’ forecasts react to the top-five forecasts (b1 different from zero), controlling for their own past forecasts and forecast errors. To control for other forecasting sources, we test a second specification in which the set of regressors is extended to include current forecasts from Bayesian vector autoregressive (BVAR) models of standard variables. The BVARs serve as a proxy for the statistical model that supports professional forecasters. In the case of inflation forecasts, we also include inflation targets in the set of controls, capturing relevant information that is not included in the information set of the vector autoregressive models. Hence, the estimated equation for the interest rate and exchange rate forecasts is the following:

  e;j e;top5 e;j e;j e;VAR xt;12m ¼ a þ b1 xt1;12m þ b2 xt;12m þ b3 xt1;12m þ b4 xt13;12m  xt12;t1

(2)

and for inflation forecasts:





e;j e;j pt;12m b pe;VAR b pe;j b pt13;12m ¼ a þ b1 pe;top5  pt12;t1 þ b5 ptarget t1;12m þ 2 t;12m þ 3 t1;12m þ 4 t;12m ;

(3)

target where VAR refers to the mean BVAR forecast, and pt;12m refers to the inflation target.

For each specification, we restrict the sample to disaggregated series exceeding 20 observations. This leaves us with 40 institutions for inflation, 44 for the interest rate, and 36 for exchange rate forecasts. In every specification, some regressions exhibit serially correlated residuals, and are thus excluded from our analysis. Table 1 shows the results for the subsamples that do not present serial correlation. It records the share of statistically significant regressors in this sample. The results show that adaptive components are important in the formation rules of Brazilian expectations. A great share of respondents has formation rules with statistically significant reactions to lagged levels of their forecasts (more than two thirds of the individuals) and their past forecast errors (ranging from nearly one quarter to more than a half).13

13 The significant presence of adaptive components might indicate some sort of learning behavior in the formation of survey expectations. It is beyond the scope of this paper to identify which learning rule best fits the survey data.

1378

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

Table 1 Epidemiology Tests

Inflation forecasts Equation # 1 3 Interest rate (Selic) forecasts Equation # 1 2 Exchange rate forecasts Equation # 1 2

# of regressions with non-auto correlated residuals

% share of significant coefficients associated with the regressors: Top-five forecasts of the previous week

Analyst's lagged forecast

Analyst's lagged forecast error

35 39

74.3 71.8

91.4 92.3

40.0 23.1

8.6 7.7

42.9 53.8

28 33

96.4 97.0

78.6 69.7

46.4 54.5

25.0 30.3

50.0 42.4

34 35

97.1 82.9

88.2 88.6

47.1 37.1

8.8 11.4

47.1 48.6

Share of participants who react to top-five forecasts but not to their own lagged forecasts

Share of participants who react to top-five forecasts but not to their own lagged forecast errors

Note: Each cell from the third column on refers to the percent share of regressions that show statistically significant coefficients at the 10% confidence level associated with each regressor indicated in the headline column. The two columns to the right refer to the share of regressions that have significant coefficients associated with the top-five forecasts but not with their own lagged forecast term (sixth column) or forecast error (seventh column).

The results also provide evidence of an important influence of top-five forecasts on survey responses for the three variables. The coefficient of the top-five forecasts is statistically significant in a large part of the regressions. Regardless of the specification, top-five forecasts have a significant coefficient in more than 70% of the regressions. In addition, more than 40% of survey participants do not react to their forecast errors but react to the top five,14 and, in the case of the interest rate, about a quarter of the survey participants disregard their own past forecasts but rely on the top five.15 To draw conclusions on whether these results should be of concern to the central bank, we refer to Cornand and Heinemann (2008a,b).16 In their theory, “information with low precision should be partially withheld from the public; information of high precision should always be released with full publicity”.17 Even at the risk that the market puts too much emphasis on public information, Cornand and Heinemman argue that “the higher the precision of public signals, the lower is the probability that an exaggerated weight reduces welfare”. In the next section, we show that BCB’s survey forecasts perform as well as or even better than standard forecasting procedures. As such, publicity of survey forecasts would be welfare enhancing. BCB’s publicity of top forecasters can thus be interpreted as even more welfare improving, given that these forecasts are even more accurate and thus help reduce the costs that bear upon the market and firms attempting to form higher order expectations. 4. Performance of survey forecasts In this section, we first focus on the performance of the inflation forecasts and then assess the joint performance of the forecasts for inflation, the exchange rate and the interest rate. In the case of

14

This can be seen in the last column of the table. This can be seen in the sixth column of the table. 16 Economic literature has recently been debating on the optimal degree of publicity and precision of public information (e.g. Cornand and Heinemann, 2008b; Morris and Shin, 2002, 2007). 17 Cornand and Heinemann (2008a). 15

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1379

Table 2 Root mean squared errors of inflation forecasts. Dates when forecasts were produced 2002:1–2007:6

2004:1–2007:6

Mean market forecast

4.43

1.22

Dependent variables in log-levels Univariate forecasting ARMA(6,6)

4.44

2.69

Multivariate forecasting Optimal lag choice (AIC) Standard VAR Bayesian VAR – Minnesota prior Bayesian VAR – harmonic decay Bayesian VAR – geometric decay

6.86 5.51 5.54 5.76

3.59 3.59 3.76 3.97

Fixed 6 lags Standard VAR Bayesian VAR – Minnesota prior Bayesian VAR – harmonic decay Bayesian VAR – geometric decay

9.89 5.59 5.32 5.66

2.83 2.75 2.88 3.72

4.17 4.09 4.13 4.13

2.47 2.47 2.47 2.47

5.10 4.31 4.12 4.05

3.22 2.81 2.73 2.54

Dependent variables in first log-differences Multivariate forecasting Optimal lag choice (AIC) Standard VAR Bayesian VAR – Minnesota prior Bayesian VAR – harmonic decay Bayesian VAR – geometric decay Fixed 6 lags Standard VAR Bayesian VAR – Minnesota prior Bayesian VAR – harmonic decay Bayesian VAR – geometric decay

inflation, we compare the predictive power of the mean survey forecast over a 12-month ahead horizon with that of inflation forecasts obtained from alternative forecasting methods. The purpose is not to find the best forecasting method, but to assess the relative performance of survey forecasts, which use a large information set, compared to autoregressive models with standard datasets, which are usually considered good forecasting tools. Table 2 shows the root mean squared errors of survey forecasts and of forecasts arising from univariate autoregressive moving average processes (ARMA) and different vector autoregression (VAR) and BVAR specifications, all of them estimated recursively.18 The VARs and BVARs were estimated using the standard set of Brazilian monthly data on the inflation rate, the Selic interest rate, the exchange rate, and industrial production,19 either in logs or in first log-differences. The sample begins in August 1999, and the first forecast exercise starts in December 2001. We tested the out-of-sample predictive power of specifications with six fixed lags and with optimal lag choices using both AIC and SBC criteria. Since estimations with SBC optimal lags underperformed those with AIC, we report only the latter. Due to the magnified effect of the 2002–2003 confidence crisis on the path of macroeconomic variables, we split the data sample into two periods to assess the forecast performance. In the full

18 Recursive estimations better approximate the possible inference from the information set that was available to forecasters at the moment they produced their forecasts. It is advisable to use real-time data in these comparisons, but those were not available for this study. 19 We also tested alternative VAR and BVAR specifications that included fiscal and other real sector variables in the set of endogenous variables or that took the sovereign spread (EMBIþ Brazil) as an exogenous variable, but the out-of-sample predictive power of these alternatives was also outperformed by the specifications presented in Table 2.

1380

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

16

12

40

p.p.

30

8

20

4

10 0 0 -4 -10 -20 2002

2003

2004 2005 Surveyed Date

2006

2007

Inflation Forecast Error (right axis) Exchange Rate Change Forecast Error (left axis) Fig. 5. Forecast errors for inflation and exchange rate change over a 12-month horizon period.

surveyed sample (January 2002 to June 2007), professional forecasters produced root mean squared forecast errors very close to those from ARMA, VARs and the BVARS that used first log-differenced data. In the subsample that excludes most of the shocks that stemmed from the confidence crisis, i.e., January 2004 to June 2007, the root mean squared error of the mean survey forecast was substantially lower than the best econometric forecast obtained using only autoregressive models. This result might imply that professional forecasters were able to use information other than that obtained from autoregressive models. In fact, economic agents might have combined forecasts from autoregressive models with those from other models and from judgment, relying on an information set significantly larger than that used in the autoregressive models. Furthermore, during the confidence crisis, agents assigned a relevant probability of a change in the policy regime, which did not materialize, but resulted in an overshooting of inflation forecasts. Inflation forecast errors seem to be partially related to exchange rate forecast errors. Fig. 5 shows the forecast errors (defined as actual minus forecasted values) for 12-month ahead inflation and exchange rate changes (as defined in Fig. 4). The correlation coefficient is 0.90. For the first half of the sample, this relationship is very strong, i.e., errors in the forecast of the exchange rate change can explain a large part of inflation forecast errors. However, over the last three years of the sample, inflation forecast errors have decreased significantly, and exchange rate forecast errors do not seem to have played an important role.20 Initially, due to the confidence crisis, exchange rate forecasts largely underpredicted the actual exchange rate change. The peak in forecast errors was in 2002:5, when agents forecasted a 9.0% average increase in the exchange rate from t to t þ 11, but the actual value was 40.7% (the inflation forecast error peaked at 12.7 p.p. in the following month). From 2002:10 through the end of the sample (2007:7), however, forecast errors were negative. Either agents underpredicted the magnitude of the exchange rate appreciation (forecasts made at the end of 2002, beginning of 2003 and in 2007) or expected some depreciation while the exchange rate appreciated (from the second half of 2003 to 2006). So far we are following the common procedure used in empirical literature, which focuses on analyzing univariate forecast series. However, Bauer et al. (2003, 2006) and Eisenbeis et al. (2002) developed a method to assess the joint accuracy of multivariate survey forecasts. The method is based on statistical theory and weighs forecast errors for distinct macro variables according to their

20

Estimation exercises (not shown) point to the same conclusion.

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1381

predictability. Forecast errors for easier-to-forecast variables are more penalized in the calculations; in addition, the correlations amongst the variables being forecast are properly taken into account. By normalizing forecast errors, we obtain a univariate measure of multivariate forecasting capacity. This measure allows for both intertemporal and cross-sectional analysis of forecast accuracy. The method assumes that forecast errors made by institution j for a vector Yt of economic variables are normally distributed with a zero mean and covariance Ut:

Yt  Yte ðjÞwað0; Ut Þ;

(4) Yte ðjÞ

where Yt and Yt ðjÞ are m  1 vectors, m is the number of economic variables being forecast, and is the vector of economic forecasts made by institution j for vector Yt. The Gaussian assumption for forecast errors implies that the normalized variable ðYt  Yte ðjÞÞ0 U1 t ðYt  e Yt ðjÞÞ has a c2 ðmÞ distribution. The p-value that is associated with this measure, calculated as 1  c2cdf ðmÞ, can thus be interpreted as the probability of observing a worse forecast error. The p-value, referred to as “accuracy score” in Bauer et al. (2006), is used as an indicator of joint forecast accuracy. These scores allow for the ordering of forecasts according to their overall performance within the cross section of forecasts and also for the assessment of the time evolution of forecast performance. The dynamic covariance matrix Ut is obtained from a decomposition of forecast errors into common and idiosyncratic components:





0
Yt  Yte ðjÞ ¼ E Yt  Y t þ Y t  Yte ðjÞ Yt  Y t þ Y t  Yte ðjÞ  
0
c i ¼ E ðYt  Y t Þ0 ðYt  Y t Þ þ E Y t  Yte ðjÞ Y t  Yte ðjÞ hUt þ Ut ;

Ut hE Yt  Yte ðjÞ


0

(5)

where Uc is the common component, Ui is the idiosyncratic component, and the vector Y t is assumed to be the mean of both Yt and Yte .21 The theoretical vector Y t is a latent variable, which is proxied by BVAR projections or by the mean survey forecast. This decomposition segregates the variance of the errors c that are attributed to unpredicted events ðU Þ from the variance of forecast errors that are due to the use of an underperforming forecast model ðUi Þ by individual forecasters. The estimation of the common component covariance can be performed in a number of ways, and the results might be sensitive to the choice selected. Bauer et al. (2006) report the performance of different choices of common covariance matrices. They consider a model-based covariance matrix, estimated from a Bayesian VAR, and a survey-based covariance matrix, estimated as the forecast error covariance of the mean forecaster in the survey. In the case of US data, they show that the mean forecast approximates the true model of the economy better than a VAR forecast does. The results of the test using a survey-based common covariance are also more in line with expected statistical properties of the score distribution. They also report that a time-varying common covariance matrix does not outperform a constant covariance estimated with the full sample, implying that common forecast errors in the US are covariance-stationary. We also estimate the common component covariance matrix using both methods, i.e., based on VAR and BVAR estimations and on the survey. We use time-varying common covariance matrices to calculate accuracy scores. To better approximate the information set available to forecasters at each moment, in the case of the model-based common covariance, we estimate VARs and BVARs recursively. This is done in order to build the covariance matrices from the forecast errors obtained with this recursive estimation for the horizon in question. i The idiosyncratic covariance component ðUt Þ, in turn, is estimated as the sample covariance matrix of forecast errors across individual forecasters in BCB’s survey, as proposed in Eisenbeis et al. (2002). The average vector Y t used in the calculations is the same vector employed to calculate the common covariance matrix Uc . We chose forecasts of three key macroeconomic variables to conduct the multivariate assessment: inflation, the exchange rate, and Selic interest rate. Fig. 6 shows the evolution of the mean accuracy

21 Implicit in the derivation of equation (5) is the assumption that forecast errors inherent in the economy ðYt  Y t Þ are uncorrelated with those made by forecasters ðY t  Yte ðjÞÞ.

1382

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

May-07

Jan-07

Sep-06

Jan-06

May-06

Sep-05

Jan-05

May-05

Sep-04

Jan-04

May-04

Sep-03

May-03

Jan-03

Sep-02

Jan-02

May-02

P-value

Mean Accuracy Score 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

Surveyed date +/- 1SD

Mean P-value

Fig. 6. Mean accuracy score using BCB’s survey mean forecast as a proxy for the best forecasting model.

score over time using the survey to estimate the common covariance matrix U , proxied by the covariance of the mean forecast errors. The overall performance of forecasts was highly volatile throughout the sample. At the beginning of the series the mean accuracy score was very low. 2002 coincides with the period of the confidence crisis in the monetary policy regime. Forecasts made in the first half of 2002 did not anticipated the surge in inflation, exchange rate and interest rate that took place in the second half of the year. Later on, by envisaging the possibility of a structural break in the conduct of monetary policy, and thus in inflation dynamics, forecasters might have been attributing a higher probability to a change in the underlying model of the economy. Over most of 2003, with a new government in office and a tight monetary policy stance, the accuracy of survey forecasts improved slightly. However, the improvement was accompanied by a great dispersion of scores within the cross section of survey participants. Over 2004, the market was persistently bearish on its exchange rate forecasts with high bets on a looser monetary policy. As this scenario did not materialize over the forecast period (year 2005), and also because of initially unexpected and then longer lasting commodity and oil price shocks that fed mostly into inflation alone, forecast accuracy significantly worsened over this period. Later, after monetary policy initiated a tightening cycle in the end-2004, and economic growth slowed toward a pace that was more consistent with non-inflationary supply conditions, forecast accuracy steadily improved. This took place in spite of important uncertainties regarding oil prices and the contagion of external shocks to the exchange rate throughout most of 2005. In September 2005, the central bank resumed interest-rate cuts. After mid-2006, forecast accuracy worsened as demand gained momentum amid uncertainties regarding the time lag of the monetary transmission channel in Brazil. Important uncertainties prevailed during this period regarding oil prices, the contagion of asset price volatility to the Brazilian exchange rate, and the monetary transmission. The idiosyncratic component played a smaller role in total forecast errors, as shown in the forecast error decomposition of each macroeconomic series, using the survey mean prediction error as a proxy for the true unpredictable value (Fig. 7). In other words, a large part of each agent’s forecast errors was common to all respondents. In the case of inflation, the idiosyncratic component prevailed only in times of large forecast dispersion across respondents, such as at the end of 2002, or when the average forecast error was low, such as in the second half of 2004. During most of the period, however, the common forecast error component prevailed, indicating the important role of aggregate shocks to the economy. This result may also be rationalized as an attempt by respondents to align their forecasts to those of their peers so as to avoid making large desynchronized forecast errors. c

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1383

18

12

16 14

10

12

8

10 6

% p.y.

Standard Error

Inflation 14

8

4

6

Jan-07

May-07

Sep-06

Jan-06

May-06

Sep-05

Jan-05

May-05

Sep-04

Jan-04

May-04

Sep-03

Jan-03

May-03

Sep-02

2

Jan-02

4

0

May-02

2

Survey Date Idiosyncratic Error

Common Error

Actual 12 months ahead (right axis)

4.0 3.5 2.5 2.0 1.5

BRL per USD

3.0

1.0

May-07

Jan-07

Sep-06

May-06

Jan-06

Sep-05

Jan-05

May-05

Sep-04

May-04

Jan-04

Sep-03

May-03

Jan-03

Sep-02

Jan-02

0.5 May-02

Standard Error

Exchange Rate 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

0.0

Survey Date Idiosyncratic Error

Common Error

Actual 12 months ahead (right axis)

10

6 4

May-07

Jan-07

Sep-06

May-06

Jan-06

Sep-05

May-05

Jan-05

Sep-04

May-04

Jan-04

Sep-03

May-03

Jan-03

Sep-02

0

Jan-02

2 May-02

Standard Error

8

30 28 26 24 22 20 18 16 14 12 10

% p.y.

Policy (Selic) Interest Rate 12

Survey Date Idiosyncratic Error

Common Error

Actual 12 months ahead (right axis)

Fig. 7. Decomposition into common and idiosyncratic forecast errors using the forecast error of BCB’s survey mean forecast as the common error.

1384

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

The trough in accuracy performance over 2004 was not so intense in model-based calculations of accuracy scores (Fig. 8). Model-based calculations confirm the low accuracy performance at the beginning of the series, but do not support the depth of the second most important trough observed in Fig. 6. 5. Determinants of survey forecasts and their dispersion This section estimates the determinants of inflation and interest-rate forecasts and of the dispersion of inflation forecasts. In the case of forecasts for the inflation level, we are particularly interested in the role played by the inflation target. Regarding interest-rate expectations, our objective is to check whether agents perceive the BCB as following some rule. The section also outlines some insights about the behavior of inflation expectations dispersion across survey respondents, a measure that can be considered as a proxy for inflation uncertainty. 5.1. Determinants of inflation forecasts We estimate the determinants of inflation forecasts following the approach in Bevilaqua et al. (2008). The choice of variables to include in the regression is grounded on the assumption that the market believes inflation dynamics to be portrayed by a basic Phillips curve. In other words, variables that are usually founddempirically or theoreticallydto be determinants of inflation are tested to explain the behavior of inflation expectations. In particular, we include the output gap, past inflation, exchange rate change and commodity price change in the set of regressors. The inflation target also enters as a regressor because it is expected to work as an anchor for inflation expectations. The estimated equation is the following:


Et pt;tþ11 ¼ b1 þ b2 pt;tþ11 þ b3 pt12;t1 þ b4 ðet1  et7 Þ þ b5 pct1  pct7 þ b6 yt2 þ 3 t ;

(6)

where Et is the expectations operator, pt;tþ11 is cumulative inflation from t to t þ 11, pt;tþ11 is the twelve-month inflation target in t þ 11 (calculated using interpolation), pt12;t1 is cumulative inflation from t  12 to t  1, et is the nominal exchange rate, pct is commodity prices (measured by the All Commodities Index of the IMF), yt is the output gap, and 3 t is the residual. The dependent variable is 12-month ahead survey inflation forecasts. For the inflation target, we use the 12-month adjusted target series (defined in Section 2). The output gap is estimated recursively

Surveyed Date +/- 1SD

Mean P-value

Fig. 8. Model-based accuracy scores using a time-varying common covariance.

May-08

Jan-08

Sep-07

May-07

Jan-07

Sep-06

Jan-06

May-06

Sep-05

Jan-05

May-05

Sep-04

Jan-04

May-04

Sep-03

Jan-03

May-03

P-value

Accuracy scores - BVAR with Geometric Decay

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1385

Table 3 Determinants of inflation forecasts. Dependent variable: 12-month ahead inflation forecasts (2000:1–2008:6). Regressors

Specifications I

II

Constant Inflation target (12-month ahead) 12-Month inflation (1) Six-month nominal exchange rate change (1) Six-month commodity price change (1) Output gap (2) Dummy for 2002:11–2003:1 R-squared Adjusted R-squared

0.42 (0.89) 0.66** (0.26) 0.23*** (0.07) 0.07*** (0.02) 0.02* (0.01) 0.06 (0.08)

0.53 (0.64) 0.71*** (0.19) 0.17*** (0.06) 0.04*** (0.01) 0.01* (0.01) 0.01 (0.05) 4.86*** (0.47) 0.8877 0.8806

0.7439 0.7306

Notes: Estimation using two-stage least squares. Instrumental variables: inflation target (1 and 2), 12-month inflation (1 and 2), six-month nominal exchange rate change (1 and 2), six-month commodity price change (1), output gap (2), Selic interest rate (1 and 2), and the dummy (in the case of specification II). Standard errorsdshown in parenthesesdwere corrected by Newey–West heteroskedasticity and autocorrelation consistent covariance matrix estimator since estimation residuals present autocorrelation and heteroskedasticity. *, ** and *** indicate the coefficient is significant at the 10%, 5%, and 1% levels, respectively.

using an HP filter applied to the industrial production series,22 and enters the equation with a twomonth lag because of the presence of lags in the release date. The past inflation regressor refers to the 12-month price change since we want to capture less noisy movements. The same applies to the exchange rate and commodity prices, for which we use six-month changes. Table 3 shows the estimation results, which are in general in line with those found in Bevilaqua et al. (2008). We also present a second specification, which includes a dummy variable for the peak of the confidence crisis (2002:11–2003:1). The objective is to verify the results when we control for an abnormal period, marked by expectations of a change in the policy regime. We would not expect inflation targets to play an important role when agents believe that there is a relevant probability that the policy regime will changedeither by replacing inflation targeting by another framework or by following a less committed policy regime. Both specifications illustrate the important role played by inflation targets in the formation of inflation expectations in Brazil. The coefficient is statistically significant. Adding the dummy variable slightly increases the importance of inflation targets in the estimation as it generates higher point estimates in a tighter confidence interval, reduces the point estimates of the other regressors and tightens their confidence intervals. In fact, the dummy variable refers to a period where expectations deviated strongly from the targets. In both specifications, using a Wald test, we cannot reject the null that the coefficient on the target is equal to one. Past inflation, exchange rate change and commodity price change also enter significantly, whereas the output gap is not significant.23 A novelty in the results, compared to previous papers, was to find a statistically significant coefficient in commodity prices (at the 10% level), although with a low magnitude. In the inflation targeting regime, inflation expectations should converge to the target in a mediumrun horizon. Note, however, that the dependent variable refers to the cumulative inflation from t through t þ 11 and not to the inflation prevailing twelve months (or further) ahead. Therefore, because of the presence of lags in the transmission mechanism of monetary policy, we would expect that other variables in addition to the inflation target would affect inflation over the next twelve months. Furthermore, the presence of a tolerance interval around the target allows the central bank to accommodate some shocks or neutralize only part of their effects. Full neutralization of all shocks

22 In other words, output gap at time t is estimated with the series up to time t. However, it is not equivalent to real-time data because we are using revised data. 23 Similar results were obtained using an output gap filtered from the rate of industrial capacity utilization, estimated by Getulio Vargas Foundation (FGV).

1386

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

would generate undue volatility. Actually, in the case of temporary shocks, a full reaction could produce stop-and-go movements because of the lags in the transmission mechanism. Recursive estimates of both specifications (not shown) yield stable estimated coefficients when the sample exceeds 42 observations. However, the confidence shock may have enacted changes in the coefficients that can be best captured contrasting the estimations using more recent samples with those from the early periods of inflation targeting. In order to capture those possible changes over time we estimate 48-month rolling window regressions. We use the specification without the dummy because the inclusion of the dummy weakens the effect of the confidence crisis on the set of coefficients. The results are presented in Fig. 9, and confirm the increasing importance of inflation targets and the effect of the 2002–2003 confidence crisis. For the subsample that includes the confidence crisis, the inflation target is less significant. In fact, at the end of 2002 and in January 2003, expectations were far from the target, which pushed down coefficient estimates to around zero. With the adoption of the adjusted targets and the increasing confidence that, in fact, the policy regime would not change, targets recovered their role. For the sample starting in 2004, when doubts about the maintenance of the policy regime had largely faded away, estimates of the coefficient on the inflation target are around one or even greater than that. The coefficient on past inflation falls abruptly in 2003 because of the introduction of the adjusted target that year. It increases over the rest of the sample, but remains lower than in the initial subsamples. Considering full sample estimation and the rolling window exercise, the point estimates range from 0.15 to 0.30. Taking into account the arguments listed above with respect to lags in the transmission mechanism of monetary policy and room for shock accommodation, those estimates seem consistent with an inflation targeting framework. The same reasoning applies to the exchange rate. Their estimates, however, clearly indicate a reduction in the pass-through from the exchange rate to inflation expectations, with a tighter interval. This result, however, should be analyzed with caution because the greater magnitude of the estimates is affected by the confidence crisis period.

Panel A

Panel C Coefficient on the Exchange Rate Change

Coefficient on the Inflation Target 0.14 0.12 0.10 0.08 0.06 0.04 0.02 0.00 -0.02 -0.04 -0.06

Jan-00 Mar-00 May-00 Jul-00 Sep-00 Nov-00 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 Jan-02 Mar-02 May-02 Jul-02 Sep-02 Nov-02 Jan-03 Mar-03 May-03 Jul-03 Sep-03 Nov-03 Jan-04 Mar-04 May-04 Jul-04

Jan-00 Mar-00 May-00 Jul-00 Sep-00 Nov-00 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 Jan-02 Mar-02 May-02 Jul-02 Sep-02 Nov-02 Jan-03 Mar-03 May-03 Jul-03 Sep-03 Nov-03 Jan-04 Mar-04 May-04 Jul-04

3.0 2.5 2.0 1.5 1.0 0.5 0.0 -0.5 -1.0

Beginning of the sample

Beginning of the sample Coefficient

Lower Bound - CI

Upper Bound - CI

Coefficient

Panel B Coefficient on Past Inflation

1 0.8

0.10

Lower Bound - CI

Upper Bound - CI

Panel D Coefficient on Commodity Price Change

0.05

0.6 0.00

0.4 0.2

-0.05

0

-0.10

-0.2 Jan-00 Mar-00 May-00 Jul-00 Sep-00 Nov-00 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 Jan-02 Mar-02 May-02 Jul-02 Sep-02 Nov-02 Jan-03 Mar-03 May-03 Jul-03 Sep-03 Nov-03 Jan-04 Mar-04 May-04 Jul-04

Jan-00 Mar-00 May-00 Jul-00 Sep-00 Nov-00 Jan-01 Mar-01 May-01 Jul-01 Sep-01 Nov-01 Jan-02 Mar-02 May-02 Jul-02 Sep-02 Nov-02 Jan-03 Mar-03 May-03 Jul-03 Sep-03 Nov-03 Jan-04 Mar-04 May-04 Jul-04

-0.15

-0.4

Beginning of the sample

Beginning of the sample Coefficient

Lower Bound - CI

Upper Bound - CI

Coefficient

Lower Bound - CI

Upper Bound - CI

Fig. 9. Coefficients in the regression for the 12-month ahead inflation forecasts in a 48-month rolling window estimation.

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1387

The estimates of the coefficient in commodity price changes are statistically significant only in the last subsamples. This indicates the role played by those prices in domestic inflation in the more recent period. Concerning the results for the other two coefficients (not shown), the output gap is significant only in some subsamples. In addition, the constant coefficient is high when we include the confidence crisis period. 5.2. Determinants of interest-rate forecasts We also estimate the determinants of policy interest-rate forecasts. In particular, we are interested in verifying whether expectations of the Selic interest rate are consistent with the inflation targeting framework. Estimations for behavior of the actual Selic rate provide evidence that the BCB responds to deviations of expected inflation from the target and allows for some interest-rate smoothing.24 The issue here is whether the market has also incorporated this evidence into their forecasts. In some sense, it is a way of checking the credibility of the BCB as we are assessing whether agents believe that the BCB fulfills its commitment to the inflation targeting regime. According to a basic Taylor rule that incorporates only inflation and lagged interest rate, the nominal policy rate evolves according to:

it ¼ ait1 þ ð1  aÞ½gðEt ptþh  ptþh Þ þ ptþh þ r;

(7)

where it is the policy interest rate, Et is the expectations operator, ptþh is inflation over a horizon h, ptþh is the inflation target, and r is the equilibrium real interest rate. The coefficient g measures the response of the interest rate to deviations of inflation expectations from the target. According to the “Taylor principle”, it should be greater than one: positive deviations of inflation from the target call for higher real interest rates. The coefficient a measures the degree of interest-rate smoothing. In the steady state, where Et ptþh  ptþh ¼ 0, the nominal interest rate should be equal to the target plus the equilibrium real rate. In our estimations, the dependent variable is the 12-month ahead survey forecast of the average Selic rate target (which was calculated using the forecasts for the end of each month). Therefore, according to equation (7), those expectations will depend on the deviation of inflation expectations from the target. The forecast horizon will depend on the value of h, but because of data availability and for the sake of simplicity, we use 12-month ahead survey inflation forecasts. We estimate, using twostage least squares (2SLS), the following equation:

 
Et itþ1;tþ12 ¼ ait1 þ ð1  aÞ g Et pt;tþ11  pt;tþ11 þ pt;tþ11 þ c ;

(8)

where Et itþ1;tþ12 is the 12-month ahead forecasts of the average Selic rate target from t to t þ 12, it1 is past Selic target, Et pt;tþ11 is the 12-month ahead inflation forecasts (including month t), pt;tþ11 is the 12-month ahead target (calculated using interpolation),25 and c is a constant. The results are shown in Table 4, which includes two specifications: with and without the interestrate smoothing component. The estimate of g (the coefficient of the term corresponding to deviations of inflation expectations from the target) is larger than one. Using a Wald test, we can reject the null that this coefficient is equal to one in both specifications. The interest-rate smoothing component significantly improves the fit of the regression. The coefficient value of 0.55 tends to be lower than that in estimates of the Taylor rule because the dependent variable here refers to expectations of the Selic rate over a 12-month horizon. The constant would indicate an equilibrium real interest rate of 7.9% per year. Fig. 10 shows the forecast errors for inflation and interest rate, confirming the role played by inflation expectations. The correlation coefficient of the two errors is 0.95, which implies that interestrate forecast errors are in large part related to inflation forecast errors.

24

For example, Fraga et al. (2003), and Minella and Sousa-Sobrinho (2009). For this estimation, we are using the CMN targets without any adjustment. Using the adjusted target series, the results are similar, although we do not reject the null of g ¼ 1 (the p-value is 0.14 in specification II). The point estimates are 1.22 and 1.31 for specifications I and II. 25

1388

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

Table 4 Determinants of Selic interest-rate forecasts. Dependent variable: 12-month ahead average Selic interest-rate forecasts (2002:1– 2008:7). Regressors

Specifications

Constant 12-Month ahead inflation forecast minus target Selic interest-rate target (1) R-squared Adjusted R-squared Wald test: coefficient on inflation forecast ¼ 1 (p-value shown)

I

II

9.57*** (0.44) 1.48*** (0.25)

7.89*** (0.54) 1.28*** (0.14) 0.55*** (0.07) 0.9301 0.9282 0.05

0.6817 0.6774 0.06

Notes: Standard errorsdshown in parenthesesdwere corrected by Newey–West heteroskedasticity and autocorrelation consistent covariance matrix estimator since estimation residuals present autocorrelation and heteroskedasticity. *, ** and *** indicate the coefficient is significant at the 10%, 5%, and 1% levels, respectively. Instrument variables: constant, inflation target, selic interest-rate target (1), 12-month ahead inflation forecasts (1 and 2), 12-month actual inflation (1), 6-month exchange rate change (1).

Therefore, there is strong evidence that private agents perceive the BCB as following a behavior that is consistent with the inflation targeting regime. When agents expect a higher inflation rate in the future, they also expect that the nominal Selic rate will rise and in greater magnitude than that of the inflation forecast, implying, therefore, an increase in the real interest rate. This result confirms the credibility built up by the BCB over the investigated period. In fact, one way of measuring credibility is verifying whether agents believe that the monetary authority will behave in line with its commitment, which seems to find support in this estimation.

5.3. Expectations dispersion The BCB’s survey also provides information on forecast dispersion across respondents. Dispersion of forecasts for inflation, the exchange rate and the Selic rate are highly correlated, as Fig. 11 shows. The correlation is around 0.6–0.8, reflecting a strong relationship among those variables. 16

12

p.p.

8

4

0

-4

-8 2002

2003

2004

2005

2006

Surveyed Date

Inflation Forecast Error Selic Rate Forecast Error Fig. 10. Forecast errors for inflation and Selic rate 12 months ahead.

2007

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1389

35 30 25

%

20 15 10 5 0 2002

2003

2004 2005 Surveyed Date

2006

2007

2008

Inflation expectations Exchange rate expectations Selic rate expectations Fig. 11. Dispersion of forecasts (coefficient of variation).

Table 5 Inflation forecast dispersion. Dependent variablea: coefficient of variation of the 12-month ahead inflation forecasts (2002:1–2008:7). Regressors

Coefficients

Constant 12-Month inflation (1) minus 12-month inflation (7) Embi(1) Dummy for 2002:10 R-squared Adjusted R-squared

6.44*** (0.57) 0.18** (0.08) 0.56*** (0.10) 15.25*** (1.47) 0.7784 0.7696

Notes: Standard errorsdshown in parenthesesdwere corrected by Newey–West heteroskedasticity and autocorrelation consistent covariance matrix estimator since estimation residuals present autocorrelation and heteroskedasticity. *, ** and *** indicate the coefficient is significant at the 10%, 5%, and 1% levels, respectively. a Defined as (standard deviation/average) * 100.

Inflation forecast dispersion can be regarded as a proxy for inflation uncertainty.26 Table 5 shows the results of an OLS estimation where the dependent variable is the coefficient of variation of inflation forecasts.27 Inflation forecast dispersion depends positively on the sovereign spread (dubbed EMBI Brazil) and on the change in inflation. Periods when the country-risk premium is high or the inflation rate is increasing tend to be characterized by higher forecast dispersion. This result seems consistent

26 Using the Survey of Professional Forecasters in the US, Giordani and Soderlind (2003) compute the aggregate inflation uncertainty as the combination of individual uncertainty (average standard deviation of individual histograms) and disagreement on the point forecast. In the case of the Brazilian survey, however, the respondents provide only point forecasts instead of probabilities for different intervals; thus the available measure of uncertainty is the disagreement among participants. In spite of this limitation, disagreement seems to capture to a large extent the degree of inflation uncertainty, as it is known to move together with individual uncertainty in the US survey (correlation of 0.6). 27 The coefficient of variation is defined as (standard deviation/average of inflation expectations) * 100. To be coherent with previous estimations, we use the values on the critical dates.

1390

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

with the historical evidence. Episodes of higher country-risk premium in Brazil were characterized by greater general uncertainty. Similarly, at moments when inflation was increasing, such as in 2001– 2003, there was significant uncertainty about future inflation values. 6. Conclusion Expectations management is one of the cornerstones of inflation targeting. This paper takes a broad perspective, evaluating the expectations obtained from a rich dataset for Brazil in a multitude of aspects: epidemiology, forecast performance of individual macroeconomic variables and of multivariate forecast sets, in addition to determinants of levels and dispersion of macroeconomic forecasts. The epidemiology estimations indicate that top performing forecasters in Brazil have played an important role as focal points for survey participants. In addition, survey respondents also exhibit significant adaptive responses to their own forecast errors and past forecast levels. The assessment of the joint forecast performance of inflation, interest rate, and exchange rate shows that forecasts are usually made with significant alignment amongst respondents since the common error component across survey participants prevails over the idiosyncratic component in the decomposition of forecast errors. The joint forecast performance in Brazil has improved over the sample period, although not monotonically. In the case of inflation forecasts, they had similar or better forecast performance than those from autoregressive models, suggesting that forecasters are using more information than that used in autoregressive models with standard information sets. The dispersion of inflation expectations seems to be related to the country-risk premium, which gives some indication as to the level of confidence in the economy, and to inflation rate changes. Moments when inflation rises are usually associated with more uncertainty about its future values. Brazil has a history of weak fiscal and monetary institutions. Notwithstanding, the estimation concerning the determinants of inflation expectations provide evidence of the prominent role played by inflation targets. Expectations of the policy interest rate, in turn, are in line with what is expected in an inflation targeting framework, depending basically on deviations of inflation expectations from the target. In other words, agents believe that the central bank reacts to inflation expectations and those depend on the inflation targets. This result confirms the credibility built up by the BCB over the period investigated. Acknowledgments The views expressed here are those of the authors and not necessarily those of the Central Bank of Brazil (BCB). We are thankful to Carlos Hamilton Araújo, Emanuel Kohlscheen, Mario Mesquita, and Klaus Schmidt-Hebbel for insightful comments, to the Investor Relations Group of the BCB for kindly providing the data used in this paper, and to the participants of the 2008 meeting of the Latin America and Caribbean Economic Association (Lacea) and of the BCB XI Inflation Targeting Seminar. We are also grateful to two anonymous referees for their comments and suggestions. References Bauer, A., Eisenbeis, R., Waggoner, D., Zha, T., 2003. Forecast evaluation with cross-sectional data: the Blue Chip surveys. Federal Reserve Bank of Atlanta Economic Review Q2, 17–31. Bauer, A., Eisenbeis, R., Waggoner, D., Zha, T., 2006. Transparency, expectations, and forecasts. European Central Bank, Working Paper Series no. 637, June. Bevilaqua, A.S., Mesquita, M., Minella, A., 2008. Brazil: taming inflation expectations. In: Bank for International Settlements (Ed.), Transmission Mechanisms for Monetary Policy in Emerging Market Economies, pp. 139–158. BIS Papers no. 35, Jan. Blinder, A., Ehrmann, M., Fratzscher, M., De Haan, J., Jansen, D., 2008. Central Bank communication and monetary policy: a survey of theory and evidence. Journal of Economic Literature 46 (4), 910–945. Carroll, C., 2003. Macroeconomic expectations of households and professional forecasters. Quarterly Journal of Economics Feb. Carvalho, F.A., Bugarin, M.S., 2006. Inflation expectations in Latin America. Economía (Washington), 101–145. Carvalho, F.A., Minella, A., 2009. Market forecasts in Brazil: performance and determinants. BCB Working Paper Series no. 185, Apr. Colucci, D., Valori, V., 2005. Error learning behaviour and stability revisited. Journal of Economic Dynamics & Control 29 (3), 371–388. Cornand, C., Heinemann, F., 2008a. Can Central Banks Talk Too Much?, 27 May 2008, at: www.voxeu.org. Cornand, C., Heinemann, F., 2008b. Optimal degree of public information dissemination. Economic Journal 118 (April), 718–742.

F.A. Carvalho, A. Minella / Journal of International Money and Finance 31 (2012) 1371–1391

1391

Eisenbeis, R., Waggoner, D., Zha, T., 2002. Evaluating Wall Street Journal survey forecasters: a multivariate approach. Business Economics 37 (3), 11–21. Evans, G.W., Honkapohja, S., 2001. Learning and Expectations in Macroeconomics. Princeton University Press. Fraga, A., Goldfajn, I., Minella, A., 2003. Inflation targeting in emerging market economies. In: Gertler, M., Rogoff, K. (Eds.), NBER Macroeconomics Annual 2003, vol. 18. MIT Press, pp. 365–400. Giordani, P., Soderlind, P., 2003. Inflation forecast uncertainty. European Economic Review 47, 1037–1059. Guillén, D., 2008. Ensaios sobre Expectativas de Inflação no Brasil. Master’s thesis, Puc-Rio. Keane, M., Runkle, D., 1990. Testing the rationality of price forecasts: new evidence from panel data. American Economic Review 80, 714–735. Marques, A.B.C., Fachada, P., Cavalcanti, D.C., 2003. Sistema Banco Central de Expectativas de Mercado. Notas Técnicas do Banco Central do Brasil, no. 36, Junho. Minella, A., Freitas, P.S., Goldfajn, I., Muinhos, M.K., 2003. Inflation targeting in Brazil: constructing credibility under exchange rate volatility. Journal of International Money and Finance 22 (7), 1015–1040. Minella, A., Souza-Sobrinho, N.F., 2009. Monetary channels in Brazil through the lens of a semi-structural model. Central Bank of Brazil, Working Paper Series no. 181, Apr. Morris, S., Shin, H., 2007. Optimal communication. Journal of the European Economic Association 5, 594–602. conference proceedings. Morris, S., Shin, H., 2002. Social value of public information. American Economic Review 52 (5), 1522–1534. Ötker-Robe, I., Vavra, D., a team of economists, 2007. Moving To Greater Exchange Rate Flexibility: Operational Aspects Based on Lessons from Detailed Country Experiences. IMF. Occasional Paper no. 256. Woodford, M., 2001. Monetary policy in the information economy. In: Economic Policy for the Information Economy. Federal Reserve Bank of Kansas City, Kansas City, pp. 297–370.