Inflation Targeting in Latin America: Post-Crisis Changes (Part 2)

4 Results 4.1 Estimations for Full-Sample In this section we check the fit of our theory postulated Taylor-type reaction functions for each country since becoming Full Fledged ITers. Table 2 (shown in the Appendix) shows the estimations for three regressions of the model, testing the appropriateness of including forecasted inflation instead of observed inflation, and exchange rate movements.5 We compare each rule taking the higher adjusted R2 and lower coefficient p-values as indication of a better fit. Since these are relatively small emerging economies, we add a variable for nominal exchange rate movements in columns 3.C, 3.B, and 3.M. However, only Chile appears to be responsive to exchange rate fluctuations, while both Brazil and Mexico report insignificant coefficients. The results for Chile are not surprising given its idiosyncrasies. Chile’s economy is small and very open economy, and thus, fluctuations in the exchange rate can have important impacts on the domestic economy. Meanwhile, we would have expected Mexico, with its close integration with the US to be more sensitive to the exchange rate. Brazil’s results are also somewhat puzzling. The Central Bank of Brazil has been known for its large sterilized foreign exchange interventions and its ample discussion of exchange rate developments in its communications. Nevertheless, our results suggest that the central bank has managed to keep its interest rate policy fairly independent from the fate of the exchange rate, in line with Edwards’ (2007) proposal that exchange rate effects are fully captured by output and inflation variation.                                                                                                                 5 We also ran different combinations than the regressions shown, but the results did not change and the fit worsened.     21     The fundamental principle of a successful Taylor-type rule is that monetary policy should “lean against the wind”. In other words, if inflation increases by 1% then nominal interest rates should increase by more than 1% so as to increase real interest rates. Since monetary policy in these countries is quite gradual, we define the long-term inflation parameter 𝛼!"  as:6 ! 𝛼!" = !!! ! !!!   > 1. In fact, in the second and third set of columns, we see that the parameter for longterm inflation deviations is always greater than one. This is comforting as it suggests that these countries implement counter-cyclical policies to stabilize expected inflation. Considering the overall fit of our results in Table 2 we notice a significant consistency with previous studies. Table 3 shows a summary of the coefficients and their significance found in two other sources. Table 3 - Past Literature Authors Period 1991-2001 SchmidtHebbel and 1999-2001 Werner (2002) 1997-2001 1999-2006 Mello and 1999-2006 Moccero (2007) 1999-2006 Country Chile Brazil Mexico Chile Brazil Mexico R_lag 0.97*** 0.88*** 0.85*** 0.63*** 0.61*** 0.11*** y 0.03 0.2 0.07 0.11** 0.19** 0.15 pi 0.02** 0.29** -0.88 -0.02 0.01** 0.01 e LT Bonds 0.01 -0.04 0.1 0.3 0.03 0.88 -0.16 0.02 3.75*** Adj R^2 0.86 0.76 0.77 Note: Statistically significant * p < 0.1, ** p < 0.05, *** p < 0.01. Schmidt-Hebbel estimated using OLS and Mello and Moccero estimated using a DSGE model. Both estimations use monthly data and Mello and Moccero (2007) obtained the first differences of variables to make them stationary. From Table 3 we can see that our coefficients seem to be on a reasonable range; however, there are important differences that account for the variations in the results. On the one hand, these studies capture much of the transition period in every country. For                                                                                                                 6  By using two lags we are able to eliminate any evidence of serial correlation. We assume that this is a product of using monthly data and using two lags would then capture a quarter.   22     this reason, the significance of their coefficients is not consistent except in the case of the persistence parameter. On the other hand, Mello and Moccero (2007) have lower persistence coefficients due to the use of first differences of variables to achieve stationarity. As discussed earlier, we adopt a different strategy since we believe that first differences would distort the main economic intuition behind a response function. Instead, we addressed this problem by allowing for a non-constant neutral real interest rate, which – if anything – enhances the intuition of our model. It must be noted, however, that the high value of the lagged coefficient is not necessarily a product of nonstationarity, which we were able to avoid, but also of using monthly data. In fact, we find a high degree of persistence across months because since the more significant decisions are made on a less frequent basis. Still, since monetary policy meetings take place with an almost monthly frequency, our choice gives us more observations allowing a more detailed account of monetary policy. 4.2 Policy Rules and the Financial Crisis The recent global financial crisis and the ensuing recession were accompanied by significant changes in the way monetary policy was conducted around the world. In the aftermath of the crisis, a number of authors have advocated an updating of central bank policy rules.7 In this section we test whether there was a significant break in the policy rules around this time for Chile, Brazil and Mexico. In Table 4 we estimate reaction functions before and after September 2008 – the month when Lehman Brothers declared bankruptcy and financial markets crashed.                                                                                                                 7  See Brookings Institution: Rethinking Central Banking. Committee on International Economic Policy and Reform. September 2011.     23     The results in Table 4 show different responses across countries. First, we notice that the coefficients for Brazil did not change across periods other than expected inflation. This result is substantiated by the Chow test that fails to reject the null hypothesis of stability in the parameters. However, in Chile and Mexico the Chow test produces F-statistics rejecting stability at a 5% level. After September 2008, Chile became more responsive to output variation but saw no significant change with regards to its short-run response to expected inflation. Moreover, even though the policy rule showed a lower degree of smoothing, the longterm coefficients on expected inflation still decreased from 2.337 to 1.852 in the second period. Finally, the Central Bank of Chile doubled its response to nominal exchange rate movements, although in both cases the coefficient is insignificant. On the other hand, Mexico became more gradual in its implementation of monetary policy. Hence, although the short-run response to expected inflation did not change, the long-run coefficient increased. Moreover, the Central Bank of Mexico increased its response on output gap and nominal exchange rate movements. Therefore, we conclude that there is evidence of a significant change in the reaction function for Mexico, and some evidence of a change in the reaction function of Chile. 4.3 Stability Although we have determined that there was a structural break of the reaction function parameters for Chile and Mexico, it is worth looking at the stability of the parameters in other periods as well. In order to do so, we follow Woglom (2003) in using CUSUM and   24     CUSUM of squares test to identify other potential structural breaks in the sample.8 Figure 4 contains the pair of graphs for each country. Although the CUSUM graph for Chile does not step out of the confidence interval, the sudden drop in late 2008 and early 2009 seems to suggest a break. Indeed looking at the CUSUM of squares graph, we confirm a break around this time. This break is reflected in the Chow test that we ran in Table 4. On the other hand, Brazil only shows a slight increase in the values in late 2008, but a much more significant change in 2002/2003. Lastly, Mexico did see a significant change in late 2008/early 2009, which is consistent with the evidence from the Chow test. Unfortunately, since the power of CUSUM and CUSUM of squares tests is not as strong as the Chow test or other stability tests, we limit ourselves from making further interpretations about the structure and length of the break. This notwithstanding, the test adds evidence that Chile and Mexico have experienced changes in the reaction function parameters around the time of the Great Financial Crisis.                                                                                                                 8 The CUSUM and CUSUM of squares tests were developed by Brown, Durbin and Evans (1975) and they are based on the out-of-sample forecasting power of the regression as it is estimated over the period. The null hypothesis of stability will be rejected if we see the test values outside the confidence interval.   25     Figure 4 – CUSUM and CUSUM of squares Chile 40 1.2 30 1.0 20 0.8 10 0.6 0 0.4 -10 0.2 -20 0.0 -30 -40 -0.2 2002 2003 2004 2005 2006 CUSUM 2007 2008 2009 2010 2011 2002 2003 2004 5% Si gnificance 2005 2006 2007 CUS UM of S quare s 2008 2009 2010 2011 5% Sign ifi can ce Brazil 40 1.2 30 1.0 20 0.8 10 0.6 0 0.4 -10 0.2 -20 0.0 -30 -40 -0.2 00 01 02 03 04 05 CUSUM 06 07 08 09 10 11 00 01 02 03 5% Sig nifica nce 04 05 06 CUS UM of Sq uares 07 08 09 10 11 5 % Signi ficance Mexico 30 1.2 1.0 20 0.8 10 0.6 0 0.4 -10 0.2 -20 -30 2004 0.0 2005 2006 2007 CUSUM   2008 2009 2010 2011 -0.2 2004 2005 2006 2007 2008 CUS UM of S quares 5% Si gnificance 26   2009 2010 2011 5% Sign ifican ce   5 A More Structural Approach Perhaps one of the shortcomings of estimating reduced-form reaction functions is the difficulty in interpreting whether the changes were due to a different structure of shocks, in the transmission mechanism of the economy or, ultimately, in the preferences of policy makers. Indeed, this has led many researches – including Caputo and Liendo (2005), McDermott and McMenamic (2008), Mello and Mocerro (2011), among others – to advocate the use of structural models, developed from micro-foundations as general equilibrium models (specifically DSGE models). Such models have the additional advantage of being able to address the Lucas Critique (Svensson (2003) provides a detailed discussion). However, these models require a comprehensive specification of the economy and preferences which, in turn, could be the source of misspecification and provide misleading interpretations of the dynamics at play. A less structured approach consists of specifying Structural Vector Autoregressive models (SVAR) where, with the imposition of some identification restrictions, researchers can draw some structural conclusions.9 In light of this, we opted to specify a SVAR that follows a traditional neokeynesian structure to further shed some light on the dynamic forces that drove the changes observed in the previous section. Our SVAR specification attempts to identify the nature of the changes that took place after the global financial crisis. Our emphasis will be placed on whether there are observable changes that have taken place in the volatility of shocks affecting the                                                                                                                 9  See Sims and Zha (2004) and Bernanke and Binder (1992).     27     economy, if there are signs of a differential dynamic response to shocks, and whether there are clear signs of noisier policymaking. 5.1 SVAR Specification The specification we use is an augmented version of the structural model described by Equations (1)-(3), (5) where we add an expectations formation rule. As is typical of VARs, we take an agnostic approach to the matrix of coefficients allowing for significant dynamic cross-correlations across variables. The model we use is, in matrix form: 𝑌! = 𝐴 𝐿  𝑌! + 𝐵𝜀! Where the vector of endogenous variables is 𝑌! = [𝑦! , 𝜋! , 𝑅! , 𝜋!! , 𝑒! ], (L) is a lag operator and εt are i.i.d. shocks to each variable. Our identification restrictions consist of assigning a specific correlation structure to the innovations Bεt. The simplest of such structures of identification restrictions results in a structural choice of the ordering of equations such that the B matrix becomes lowertriangular. The resulting recursive ordering for identification assumes: (1) Output gap responds to lagged shocks and variables but is only contemporaneously affected by shocks to output. The rationale is that output is unlikely to respond instantaneously to any of the other economic drivers under consideration. (2) Inflation is affected by all lagged shocks but is also contemporaneously influenced by output and inflation innovations. Once again, the intuition we pursue is that inflation will tend to respond to current shocks to output but that it will take at least   28     a month to respond to changes in policy, expectations or to reflect pass-through from the exchange rate. (3) The policy rate reacts to all lags but also to contemporaneous shocks to policy, output and inflation. The intuition is that policymakers do observe the economic environment and manage to rapidly respond to unfolding events. However, because of observational lags, they only respond with a one month lag to shocks to inflation expectations or exchange rates. (4) Expected inflation is affected by all lags and contemporaneous policy, output, inflation and expected inflation innovations. The assumption is that inflation expectations, as measured in central bank surveys are quick to respond to relatively slow moving economic and policy variables, yet because of the way the data is collected, they do not reflect contemporaneous movements in exchange rates. (5) Finally, the exchange rate is assumed to be the faster market variable to respond to all shocks to the economy. While the identification restrictions are based on economic reasoning, the fact that we estimate the model in a monthly frequency reduces some of the potential discomfort that might emerge from our assumed correlation structure. To model the dynamic structure for A(L), we use three-lags for Brazil, six-lags for Chile, and two-lags for Mexico as suggested by the Akaike Information Criterion.   29     5.2.1 Results 5.2.1 Change in Residuals Once we estimate the SVAR for the whole sample, we first consider the estimated structural residuals as shown in the Appendix. Our first test looks at whether there is a statistically significant difference in the variance of the structural shocks of the key equations. A finding of heteroskedasticity could imply that there is either a change in the fundamental equation or that there is a change in the distribution of shocks. We find that in Chile, all structural shocks exhibit statistically larger variance after the crisis. In particular, the variance of output and inflation shocks increase, suggesting that shocks became more acute. Similarly, the variance of the policy rate residual increases which can be taken as a-priori evidence of greater policy noise, or a change in policymakers’ preferences. To further understand the source of the increase in variance, we checked a plot of residuals. Interestingly, the output residual of the SVAR shows that for output the Earthquake that rocked Chile in March 2010 explains the jump in the variance. Except for this outlier, the volatility of output appears to be similar to that evidenced by the prefinancial crisis period. On the other hand, volatility in inflation appears to have increased meaningfully, suggesting that either there are specification difficulties or a shift in the distribution of inflation shocks (the period was characterized by large commodity price shocks).   30     The Chilean policy reaction function residuals illustrates that with the exception of a sharp negative shock in Q1 2009, the distribution of shocks appears similar to those pre-crisis. Indeed, an analysis of the policy minutes of the central bank illustrates that in January 2009, the central bank conducted an unusual reduction of 100bp, followed by two extraordinary cuts in February and March of 250bp, only to bring the pace of reductions to a more typical 50bp in April. Such an extraordinarily rushed response was conducted against a global backdrop of significant financial difficulties which had not, at the time, fully reflected in domestic conditions. The Figure also suggests that once this crisis-response period was over, the distribution of policy shocks returned to that observed pre-September 2008 and, therefore, portrays a central bank with no change in preferences. In the case of Brazil, the statistical tests suggest that output shocks were more volatile after the crisis, yet inflation and policy shocks were less so. In line with our previous findings on the change in the Taylor-rule type reaction function, the volatility pre-2005 appears to have been much greater in Brazil and this probably hides other smaller, less noticeable changes that may have taken place after the 2008 crisis. A quick evaluation of the residuals shows that output volatility appears driven by a sharp negative shock right after the global financial crisis. On the other hand, inflation and policy noise appear much more volatile in the 2002-04 period as Brazil experienced its home-grown crisis. In any event, the graph of residuals and statistical test suggest there was significant continuity in the policy reaction function post-crisis. Finally, the tests for Mexico show that there were no significant changes in the distribution of output shocks, while there were statistically significant increases in the   31     variance of inflation and policy shocks. A look at the residuals illustrates a remarkable circumstance in Mexico. Despite two very negative shocks, linked to the global financial crisis and the A1H1 influenza outbreak in the fourth quarter of 2008 and third quarter of 2009 respectively, the volatility is somewhat similar to that observed earlier. Indeed, positive shocks of relatively similar magnitudes appeared to have taken place in Q1 2006 and Q4 2004. Meanwhile, as in Chile, inflation seems to have been somewhat more volatile in the post-crisis period, perhaps a reflection of the commodity price shocks that affected the region at the time. On the other hand, the residuals to the policy reaction function point to substantial volatility in the months following the crisis. Once again, rather than a permanent change towards a noisier reaction function, it appears that Banco de Mexico responded with unusual strength to the crisis: Cutting 50bp in January 2009, 25bp in February, a sequence of three 75bp in March-May, and 50bp in June to close the cycle with one final 25bp in July, having held policy rates constant since. The above discussion suggests that policy reaction functions may have changed after the crisis, but they hardly seem to have become noisier, once the extraordinary immediate policy right after the crisis is accounted for. Furthermore, the results suggest that inflation has become more volatile and that this may be both an explanation for and a result of those potential changes in reaction functions. 5.3.1 Dynamic Response to Shocks One of the advantages of estimating SVARs is that they provide a series of impulse response functions (IRF), which can be structurally interpreted as the dynamic response of a dependent variable to a shock. The shape, pace of convergence and sign of   32     such IRFs are of interest to us. Since our objective is to evaluate whether there are changes in the policy rule, we will skip the presentation of the IRFs for the whole period (they have a similar structure to the IRFs for the sub-periods) and focus on the key subperiods under analysis. Furthermore, we will concentrate the presentation on the policy rule dynamic response, sparing the reader of a longer discussion of the IRFs for the other structural equations of the model. The results are presented in the different panels in the Appendix. Chile The impulse response functions suggest a more responsive central bank of Chile, rather than a clear cut case of change in the weights that the central bank places on inflation, output or even the exchange rate. Comparing the two periods suggests that there is now a sharper and more immediate response to output shocks (statistically significant as it appears to fall outside the original confidence interval). Importantly, however, the response of policy rates to a positive (one standard deviation) shock to expected inflation is also much faster and more aggressive than in the earlier part of the sample (once again statistically significant change). On the other hand, in the earlier part of the sample, the central bank did seem to respond gradually to higher headline inflation, a pattern of response absent after the crisis. At the same time, in neither period, shocks to exchange rates do not seem to have much influence on policy decisions. All in all it seems that, if anything, the central bank of Chile has evolved towards a purer expected-inflation targeting regime and one where the central bank is faster and more aggressive in its response. Had the central bank shifted its preferences, we would   33     have expected to observe either: (a) an increase in the policy rule noisiness; or (b) a tilt in the way that the policy rule responded to output/inflation shocks. Neither one seems to be the case. Rather, the response appears to be simply more forceful on both fronts. This pattern of response is, in fact, aligned with the theoretical discussions on inflation uncertainty. As Srour (1999) shows, under uncertainty over the response of inflation to policies (in contrast to uncertainty over output shocks), central banks ought to become more forceful and speedier in their response. As we showed in the previous section, inflation volatility did seem to increase in Chile in the post-crisis period, while output volatility did not (except for the earthquake outlier). We therefore interpret the results in Chile as suggesting a rational evolution of IT, in response to shocks and changes in the economic backdrop, rather than a change in policy-makers’ preferences. Brazil Remarkably, the central bank of Brazil seems to have shifted its preferences evidencing a somewhat higher (lower) sensitivity to expected inflation (output) shocks. The Brazilian panel of IRFs shows that in the pre-crisis period, the central bank tended to respond in a hump-shaped manner to output shocks more meaningfully than in the postcrisis. In both cases the response seems to peak at around 10 months, though it seems mightier in the first period. The response to expected inflation is, however, somewhat slower in the second period, though eventually more powerful. It appears that the central bank, when faced with increases in inflation expectations, tends to “fall behind the curve” in the post-crisis period and is forced to over-compensate with stronger policy actions.   34     Brazil’s central bank seems to show some subtle changes in preferences. Faced with lower inflation volatility and similar in output, it is difficult to argue that Brazil’s modest changes represent a rational response to a changing environment. Rather, they seem to show that, if anything, the central bank is now more sensitive to what markets expect, but that it tends to be a bit slower to respond, forcing the authorities to compensate later with bigger policy adjustments and perhaps, larger output costs. Mexico The Central Bank’s responses show a central bank that has become less sensitive to economic shocks. In line with our Taylor-rule results, Mexico’s central bank appears insensitive to movements in expected inflation in both samples (in fact, in the post-crisis period the IRF is negative, though statistically we cannot reject that as nil). The central bank seems to target, instead, actual inflation. However, the response appears to be somewhat more muted after the crisis, though a bit faster peaking after 5 months instead of the previous 8 months. The response to output shocks is also faster, peaking at 5-6 months while before the crisis it would peak at 10-12 months with a similar response in both cases peaking at (0.2). Another response that is somewhat different is the IRF to foreign exchange shocks. After the crisis, the response initially tends to be fairly negative (higher policy rates when faced with a depreciating exchange rate), and turns positive after 6 months leading to an easing of policy as the exchange rate tightens monetary conditions. The Central Bank of Mexico’s policy, in the post crisis period, holding interest rates unchanged since the last move in July 2009, is clearly illustrating a less sensitive   35     central bank. This is evident in the inflation shock IRF. All in all, our empirical estimates offer some mild evidence of a change in preferences. The reaction function appears to show a somewhat more dovish, more responsive, central bank when faced with output shocks rather than inflation considerations while, at the same time, they pay increased attention to exchange rates. 6 Conclusion In this paper we set out to evaluate the implementation of IT in three important Latin American countries. In particular, we wanted to check whether the reaction function of policy makers showed any evidence of having shifted after the global financial crisis. Our results show that there is not a uniform behavior across countries, yet there are some signs of changed behavior. One of the important innovations of our research is to allow for a time-varying neutral real interest rate. Our exercises are suggestive that changes in neutral real rates are explained by improvements in central bank credibility and a reduction in inflation risk premia, and a decline in international neutral real rates. Once we adjust for the trend decline in neutral real rates, the policy rates pursued by the countries becomes stationary and regression based policy rules have an improved fit. We test reduced-form Taylor type policy rules and find mixed evidence of changes in regime. In Chile, we find that after the crisis, the policy response became speedier. In Mexico, we find some evidence of increased responsiveness to the output gap. Finally, in Brazil we find limited evidence of changes and, if anything, the CUSUM tests suggest that a more meaningful change in regime took place in 2003. Our regression   36     analysis also finds that Brazil and Chile respond to inflation expectations while Mexico appears more focused on actual inflation. Only Chile seems somewhat sensitive to exchange rates. When we analyze the evidence through structural vector autoregressive models we obtain further understanding of what changed after the 2008 global financial crisis. We find very limited evidence of changes in the policymakers’ preferences in the case of Chile. Rather, our results suggest that there was a rational response to increased uncertainty over inflation. We also find a modest change in behavior in Brazil, where the central bank has tilted towards targeting expected inflation. Simultaneously, the institution emerges as slower to respond to increases in expected inflation to, eventually, be forced to overcompensate its policy adjustment. In turn, Mexico’s central bank offers mild evidence of changes in preferences. In this case, the results suggest the central bank has become somewhat more dovish. Our research suggests opportunities for future research. If one pursues the avenue of structural VARs, perhaps we should include a richer set of variables, allowing for the inclusion of other variables such as fiscal policy or terms of trade. It is also desirable to extend the specification to allow for more elaborate forms of interaction across variables. For instance, one could further explore the presence of non-linear responses of the central bank to exchange rates (i.e. only reacting when exchange rates cross certain thresholds). The relatively short sample period after the crisis suggests that, as more data becomes available and further inflation cycles develop, the analysis should be further refined to obtain more conclusive evidence. We attempted to complement the   37     econometric estimates with an analysis of the actual policy communication, which tended to validate our conclusions. A greater emphasis on out-of-model information could help sharpen the interpretation further. All in all, our research finds that inflation targeting remains strong in the region, despite some modest changes in policy-makers preferences. The benefits (and costs) associated with the regime, should continue to accrue.   38