Interventions in the Brazilian Foreign Exchange Market: An Empirical Investigation of the Determinants

Transcription

1 Interventions in the Brazilian Foreign Exchange Market: An Empirical Investigation of the Determinants Gabriela Fernandes* * Economist - Itaú Macro Research Team ABSTRACT Since the adoption of the floating exchange-rate regime in 1999, the Brazilian Central Bank (BCB) has intervened several times in the foreign exchange market, buying and selling dollars in the spot, futures and derivatives markets. What are the variables that led the central bank to intervene in the foreign exchange market? In our investigation of this question, we find that the behavior of some variables including, among others, the risk premium, the deviations of the real from its prior trend, comparison of the performance of the real with that of similar currencies, the volatility of markets and of the exchange rate itself strongly influence the likelihood of BCB intervention in FX. We also conclude that the monetary authority acts in blocks, and that the fact that it had intervened the day before increases the likelihood of a new intervention. We also note that the BCB interventions ( reaction function ) change over time, in accordance with different macroeconomic scenarios and administrations. The opinions expressed herein by the authors for Discussion do not necessarily reflect the position of Itaú Unibanco. Acknowledgments I would like to thank Aurélio Bicalho for his valuable assistance throughout this project and Ilan Goldfajn for his important during the preparation of this paper.

2 Table of Contents 1. Introduction Econometric Approach Ordered model Binary model Data Controls Results Robustness Test Differences Between Periods Conclusion References Appendix 1: Derivation of Models Ordered logit Binary logit Appendix 2: Database Considerations

3 1. Introduction Since the adoption of the floating exchange-rate regime in 1999, the Brazilian Central Bank (BCB) has intervened several times in the foreign exchange market, buying and selling dollars in the spot, futures and derivatives markets (Figure 1). Such actions have often been taken by central banks around the world in order to stabilize exchange rates, preventing excessive rate movements. The literature on the effectiveness of sterilized interventions i.e., interventions that do not generate an expansion or contraction of the monetary base, as they are counterbalanced by open-market operations is extensive. Some examples include: Dominguez (1993), Taylor (2004), Ostry et al. (2012) and Kohlscheen (2012). In this paper, we investigate the reasons that exchange market. What are the variables that make the BCB react with a purchase or sale of dollars? Does the risk premium or the recent trajectory of other currencies influence the decisions of the monetary authority? We seek, therefore, to estimate the the Central Bank, also evaluating whether this function is constant over time. 25 $ billion: dollar purchases (+) Brazil Central Bank's Interventions in the Foreign Exchange Market R$ R$ R$ 3.20 R$ sales (-) -35 Spot Forward 'Reverse' swap Swap Repo line Loans in USD Reais per dollar (rhs) R$ 2.00 R$ 1.40 Source: BCB, MCM Consultores, Itaú Figure 1: BCB interventions in the foreign exchange market We conclude that there exist a set of variables to which the Brazilian Central Bank seems to react when deciding whether or not to intervene in the exchange market: exchange rate deviations from recent trend, comparison to the movement of the currencies, variations in the price of the real effective US dollar exchange rate, volatility of markets and currencies, relative performance of the local and international stock exchanges, and the level of reserves. Furthermore, we also conclude that such variables may be asymmetric for purchase and sale of dollars: a strong dollar against its peers influences the purchase and not the sale, while the opposite occurs with the risk premium (represented by the CDS). We also conclude that the reaction function of the BCB changes over time: volatility, reserves and the stock exchange were very significant factors between 2003 and 2010, while dollar movements were very significant between 1999 and 2002 and from 2011 to The data also show that the reaction of the central bank is independent of the volume of the intervention; that is, that the decision to intervene in the market is little affected by the decision of how much to buy or sell. This paper is organized as follows. In Section 2, we discuss the empirical model. In Section 3, we present the data. In Section 4, we show the regression results, and in Section 5, we 3

4 sample range. In the last section, we conclude. 2. Econometric Approach There are at least two ways to econometrically estimate the drivers of BCB interventions. The first would be an ordinary least squares estimation, where the dependent variable is the amount purchased (plus sign 1 ) or sold (minus sign) that day or zero if the BCB does not act. This (47.5% of the sample), which could be solved with a Tobit model. The second way of estimating BCB interventions in the FX market would be to abstract from the volume placed and focus on the three possible action options: purchase, sale or neutrality. In this approach, which is one we have adopted here, we use discrete response models with a comprises the three options available to the BCB (ordered logit). Then we test binary logits, which allow only purchase or non-purchase (and sale or non-sale), seeking to verify asymmetries. 2 That is, we admit the possibility of a variable being a determin and vice versa. Other types of estimation are discussed in Section 5, in which we test the robustness of our results Ordered model In the ordered logit model, the dependent variable may only assume discrete values. In our case, these values will be equivalent to each of the three possible actions of the Central Bank: buy, sell or inaction. The value of each variable is irrelevant, provided that there is an ordering between them. In this paper, we define the dependent variable as: Then, the model, derived in Appendix 1, calculates probabilities for each scenario, given the known variables, so that in which is the set of independent variables. Therefore, the action predicted by the model is the one to which the largest probability is assigned: (1) The coefficients of the model have no comparable magnitude, but a positive coefficient shows that the respective variable increases the probability of (purchase) and a negative sign acts in the higher probability of (sale). Therefore: 2.2. Binary model The binary logit may be considered a subcase of the multinomial logit in which there are only two options to the Central Bank. In this article, one of the states of nature will be the purchase or sale, while the other will be its complement ). The interpretation is straightforward: the BCB is never "split" between the three actions. Without losing generality, when the central 1 In Section 3, the me 2 Kohlscheen (2012) uses a similar binary approach in a test of robustness of his model of volatility and flows. 4

5 bank makes a sale, the only other possible response to the same set of variables would be to remain off the market. 3. Data Our database is daily and starts in January It includes five types of intervention already used by the Brazilian Central Bank: spot transactions, forward transactions, foreign exchange swaps, loans in foreign currency and repurchase lines. 3 Unlike the other four types 4 of interventions, the foreign exchange swaps do not directly alter the position of international reserves, as the latter are settled in dollars. However, because our intention is to evaluate BCB measures that influence the level of the exchange rate, 5 foreign exchange swaps are included, and indeed are rather important in our sample. Furthermore, for the same reason, each transaction was accounted for on the auction day and not on the settlement day. Data pertaining to interventions in the spot, forward and foreign exchange swap markets 6 The database pertaining to liquidity lines was cing these auctions and their results. The classification as sale, purchase or inaction was made according to the following criteria: each day, the amounts placed on auctions of all these instruments are added up. In calculating this sum, spot, forward and reverse foreign exchange swap purchases i.e., actions that promote the devaluation of the exchange rate are assigned a positive value (plus sign). Meanwhile, spot, forward, traditional foreign exchange swaps and liquidity lines i.e., measures that act to strengthen the real against the dollar are assigned a negative value (minus sign). Whether the sum of all these instruments is positive or negative defines the classification: if positive, it was a purchase; if zero, no action was taken; and if minus, it was a sale Controls As controls for the regression, we use: the level of the exchange rate (measured as reais per dollar), the ceiling and floor of the daily rate, its deviation from trend (measured as a Hodrick- Prescott filter) 7 and its volatility 8 over the last ten, 50 and 100 days. Other market variables used are: Ibovespa, S&P 500, the volatility index (VIX), the dollar index (DXY 9 ), the spread of highyield corporate bonds against ten-year U.S. Treasury bonds, the Brazil risk premium (measured by credit default swaps, the CDS rate 10 ) and the daily position of reserves. 11 Comparisons of the daily performance of the real with a basket of other currency pairs, also called commodity currencies, were also used. This set is composed of the following currencies: 3 Henceforth, foreign currency loans and times of crisis ( and ), during shortages of dollar liquidity in the Brazilian banking system. Further details on the use of these operations can be found in Stone et al. (2009). 4 As a result of liquidity operations, we divided Brazilian international reserves into two categories: cash and liquidity. In the cash category, these transactions are accounted for; in the liquidity category, they are not. 5 It may be argued that the liquidity lines were not intended to influence the trajectory of the exchange rate, but only to provide liquidity to the banking system. For this reason, in Annex 2, we re- estimate the model without such actions. It may be anticipated that the difference is irrelevant. 6 The BCB has made daily data on spot purchases and sales available in its time series system, but only for transactions since September Therefore, we chose to use the more complete database. Additionally, the BCB database includes only the date of settlement of these transactions, which generally falls two business days after the auction. 7 A caveat: the trend measured by the HP filter is clearly endogenous in the medium and long term, because BCB operations generally influence the exchange rate and, as a result, the filter also moves. However, we are using data with daily frequency, which in general do not suffer from such influence. A comparison of the rate with the trend serves only to identify quick shifts of the currency from its most recent trajectory. 8 Volatility is measured as the moving standard deviation of logarithmic differential (between the closing of each day close to close and within the daily window high to low). 9 The DXY is the proxy of the nominal effective exchange rate of the United States and is comprised of: Euro (57.6%), yen (13.6%), British pound (11.9%), Canadian dollar (9.1%), Swedish krona (4.2%) and Swiss franc (3.6%). 10 As the CDS series starts only in 2001, variations of Emerging Markets Bond Index + of JP Morgan were used to simulate the data for the rest of the sample. 11 Daily foreign exchange flows, as disclosed by the central bank, and several exchange coupon maturities and their spread over LIBOR, yield of the 10- year US Treasury bond and effective Selic rates and Fed Funds also represented variables, tested without success. 5

6 the Australian dollar, the Canadian dollar, the South African rand, the Chilean peso, the Mexican peso, the Norwegian krone and the New Zealand dollar Results We estimate the model with and without inertia i.e., with or without lags in the performance of the monetary authority for almost all specifications used. The reason for the inclusion of lags is empirical. In the database, it is noticeable that the BCB usually operates through purchase or sale blocks, 13 causing the probability of action on a given day to depend considerably on the actions taken in the immediately preceding days. The results of the ordered models (shown in Table 1) demonstrate that, in fact, the coefficient signs are consistent with economic intuition. First, we can observe the power of lags: four days. Table 1: Ordered logit. Dependent variable: central bank action. With inertia Without inertia Action (-1) 1.81 (0.00)*** Action (-2) 1.17 (0.00)*** Action (-3) 0.77 (0.00)*** Action (-4) 0.63 (0.00)*** BRL/Trend, log (0.01)** (0.00)*** BRL/Basket, log DXY, log (0.00)** (0.00)*** VIX/average(VIX), log (0.00)** (0.00)*** S&P500/Ibovespa, log (0.00)** (0.00)*** Reserves, log (0.00)** (0.00)*** 50 day volatility, log (0.00)** (0.00)*** Pseudo-R Observations In parenthesis, p-values of estimations. *, ** and *** denote statistical significance at 10%, 5% and 1%, respectively. Our conclusions for the other explanatory variables may be summarized as follows: 12 The tickers of the series used can be obtained from the author days. 6

7 Table 2: Results of the ordered logit model Act toward sale (purchase): Real more depreciated (appreciated) than trend Real more depreciated (appreciated) than peers Dollar appreciated (depreciated) relative to peers High (Low) market volatility Brazilian stock exchange with worse (better) performance than North American stock exchanges Volatility (stability) of the recent exchange rate trajectory In Table 3, it is possible to evaluate the predictive power of the model. It is noticeable that although the percentage of right answers of the model as a whole is quite high, 88.1%, this success is not evenly distributed among the three possible answers. While the model hits 94.8% of purchases, it only correctly predicts 72.7% of sales. Furthermore, the accuracy falls considerably when the component inertia is removed. Table 3: Evaluation of predictions With inertia Action Obs. Correct Incorrect Right (%) Wrong (%) Sale Inaction Purchase Total Without inertia Action Obs. Correct Incorrect Right (%) Wrong (%) Sale Inaction Purchase Total First column shows the number of observations for each type of action. The next two columns show the number of times that the two last columns show the proportion of right and wrong (i.e., second and third columns over the first column). Table 4 shows the results of the purchase binary logit model. Again, time lags of up to three days are quite significant. One variable that contributes significantly to the purchase is the deviation of the real in relation to its trend, towards appreciation. A low risk premium in the market, represented by lower spreads on high yield bonds, also helps. The third significant variable in all of the models is the differential between the Brazilian and the North American 500, the BCB also tends to buy, according to the model. Possible explanations for this are an indication of flow or optimism regarding Brazil. Finally, there are those variables that are only significant in one of the types of models (with or without inertia). While movements of the U.S. dollar relative to its peers, represented by DXY, are significant only in the model without inertia, the movement of the real sign in the model with inertia. 7

8 Table 4: Binary logit. Dependent variable: BC purchases. With inertia Without inertia (1) (2) (3) (4) Action (-1) Action (-2) Action (-3) BRL/Trend, log High Yield Spreads, log BRL/Basket, log S&P500/Ibovespa, log DXY, log (0.97) (0.00)*** Observations In parenthesis, p-values of estimations. *, ** e *** denote statistical significance at 10%, 5% and 1%, respectively. le average Looking at the two models that have statistical significance and the coefficients of which are consistent with intuition, (1) and (4), Table 5 shows the correctness of the predictions: again, it is clear that inertia is an extremely important component in the equations, with over a 10% loss in accuracy when comparing models with and without inertia. Table 5: Assessment of predictions With inertia model (1) True value Purchase=0 Purchase=1 Total Pr(purchase) >0, Total % Right Without inertia model (4) True value Purchase=0 Purchase=1 Total Pr(purchase) >0, Total % Right Columns show the number of observations of each type of action. The first line shows observations for which the model indicated that there would not be a sale that is, for which the calculated probability of a sale is less than 50%. The second line shows the opposite: times when the model indicated a purchase. Table 6 shows the results of binary models applied to sales. As expected, the real being more depreciated than its prior trend, greater currency volatility and greater country risk all increase the likelihood of sales. Again, the comparison of stock exchanges proved relevant. The comparison with the basket of currency pairs, by contrast, had its significance rejected by both models, with and without inertia. 8

9 Table 6: Binary logit. Dependent variable: BC sales. With inertia Without inertia (5) (6) (7) (8) Action (-1) Action (-2) Action (-3) BRL/Trend, log (0.12) (0.00)*** CDS, log (0,00)*** (0.00)*** 10 day volatility, log BRL/Basket, log 0.65 (0.18) 0.08 (0.69) S&P500/Ibovespa, log Observations In parenthesis, p-values of estimations. *, ** and *** denote statistical significance at 10%, 5% and 1%, respectively. As for the predictive performance of the model presented in Table 7, it is clearly inferior to the other models especially in the case of no inertia, where its predictive power over sales is below 50%. Table 7: Evaluation of predictions With inertia model (6) True value Sale=0 Sale=1 Total Pr(sale) >0, Total % Right Without inertia model (8) True value Sale=0 Sale=1 Total Pr(sale) >0, Total % Right Columns show the number of observations of each type of action. The first line shows observations for which the model indicated that there would not be a sale, that is, the calculated probability of a sale is less than 50%. The second line shows the opposite: times when the model indicated purchase. 5. Robustness Test As discussed in Section 2, we opted to use discrete choice models in our investigation, but this was not the only possible choice. If we wish to take into account the difference in the size of each auction, defined as the volume of dollars traded, two alternative options may be used: i) an ordinary least squares model, which allows treatment of both purchases and sales; or ii) a censored response model (Tobit), which implies a restriction to one type of operation only. 9

10 As we know, the OLS model is consistent, despite the large quantity of zeros. Nevertheless, in this case, the assumption that is not reasonable. On the other hand, if we use the Tobit model, we have to separate purchases from sales, similarly to the binary logits. 14 The Tobit model can be described as: (2) In this case, and the interpretation of the coefficient sign is similar to the OLS model. The results are shown in Table 9. The Tobits estimation shows the strength of the independent variables used. Over both purchase and sale, almost all tested variables are significant, at 1%, and have the correct sign, except for two lags in the sale equation and except for deviations from trend in the purchase case. For the OLS model, as we expected, the results are mixed. Table 9: Estimations in Tobit and OLS Tobit OLS Purchase Sale General Purchase Sale Constant (0.93) (0.61) (0.62) Action (-1) (0.00)*** Action (-2) (0.00)*** (0.01)** (0.00)*** Action (-3) (0.00)*** (0.49) (0.00)*** Action (-4) (0.00)*** (0.96) (0.01)** BRL/Trend, log (0.45) (0.00)*** (0.17) (0.68) (0.01)** BRL/Basket, log (0.00)*** (0.08)* (0.00)*** DXY, log (0.00)*** (0,00)*** (0.49) (0.90) (0.42) VIX/average(VIX), log (0.00)*** S&P500/Ibovespa, log (0.00)*** (0.01)** (0.71) (0.43) (0.73) Reserves, log (0.32) (0.20) (0.95) 50 day volatility, log (0.04)** (0.49) (0.03)** Observations In parenthesis, p-values of estimations. *, ** e *** denote statistical significance at 10%, 5% and 1%, respectively. 6. Differences Between Periods over time, we have split the studied period and analyzed the differences. The selected criteria 14 Sales were estimated with a plus sign i.e., the model is left truncated. 10

11 were the three different administrations of the BCB that took place between 1999 and We broke down our sample between those three central bank administrations, two of which are incomplete, as follows: 03/04/1999 to 12/31/2002; 01/01/2003 to 12/31/2010; and 01/01/2011 to 06/22/2012. The table below shows the number of days and actions taken by each administration. Table 10: Sub-samples studied and interventions in FX: 3/4/1999 to 12/31/2002 1/1/2003 to 12/31/2010 1/1/2011 to 6/22/2012 Purchase Neutral Sale Total The estimation results are shown in Table 11. Clearly, there is a large difference in the number of interventions and, more importantly, in the composition of these interventions. Whereas the sample in the first period has only two purchase actions, there are only 12 sale actions in the current administration sample. The intermediate period, in turn, is more balanced between purchase and sale actions. Nevertheless, it is possible to draw some conclusions. The lags are less important in the last period, and the dollar-index was not significant in the period between 2003 and 2010, in contrast to other periods. Probably because it corresponds to a time when there was a great accumulation of international reserves, the volume of international reserves is only statistically significant in the intermediate period. Table 11: Comparison between periods Total 3/4/1999 to 12/31/2002 1/1/2003 to 12/31/2010 1/1/2011 to 6/22/2012 Action (-1) (0.00)*** (0,00)*** (0.00)*** (0.01)** Action (-2) (0.00)*** (0.00)** Action (-3) (0.00)*** (0.03)** (0.00)*** (0.15) Action (-4) (0.00)*** (0.01)** BRL/Trend, log (0.01)** (0.16) (0.09)* (0.06)* BRL/Basket, log (0.03)** (0.87) DXY, log (0.12) (0.00)*** VIX/average(VIX), log (0.00)*** (0.09)* (0.00)*** (0.23) S&P500/Ibovespa, log (0.00)*** (0.33) (0.00)*** (0.41) Reserves, log (0.00)*** (0.23) (0.07)* (0.28) 50 day volatility, log (0.00)*** (0.23) (0.02)** (0.02)** Pseudo-R , Observations In parenthesis, p-values of estimations. *, ** e *** denote statistical significance at 10%, 5% and 1%, respectively. 11

12 It is worth noting the important macroeconomic transformations that the country went e beginning of the period analyzed, for instance, international reserves were at 36 billion dollars, or 15% of gross external debt. At the end of 2011, reserves had increased almost tenfold, reaching 352 billion dollars, or 118% of debt. The risk premium, for its part, has been much higher in other periods: the 5-year CDS reached almost 4,000 points in Country-risk: 5-year CDS basis points 35% 30% 25% 20% 15% 10% 5% 0% Net External Debt % GDP billion dollars Net External Debt International Reserves (rhs) % Source: Bloomberg, Itaú -10% Source: BCB, Itaú 7. Conclusion In this article, we observed that certain variables are part of of the Brazilian Central Bank that influences its intervention in the market by means of dollar purchases or sales. Through an ordered logit, we have found that variables such as the volatility of the market and the volatility of the currency, the Brazilian stock exchange relative performance to U.S. stock exchange, the real being more depreciated than its prior trend and its peers, and a foreign currency. We also observed the strong influence that immediately preceding actions exert over the concluded that there may be differences between the variables that influence the purchase or sale of dollars. The dollar being stronger relative to its peers influences purchases and not sales, while the opposite pattern is observed for the risk premium, represented by the CDS. We concluded that for different BCB administrations correspond different reaction functions, and that some variables may exert different influence at distinct times. Between 2003 and 2010, that same period, however, there was no significant sensitivity to the strength of the dollar against its peers, even though this is a variable that has a high and significant coefficient during the administrations that came before and after the said period. Finally, we saw that the estimates are robust to the inclusion of the amount of each intervention (Tobit). Therefore, the decision to intervene precedes the decision of how much to buy or sell in each intervention. 12

13 References Dominguez, Kathryn M. and Jeffrey A. Frankel. -Exchange The American Economic Review, 83(5), pp Kohlscheen, E. Sterilized Interv Ostry, Jonathan D., Atish R. Ghosh and Marcos Chamon. Staff Discussion Note. Stone, M., W.C. Walker and Y. Yasui Foreign Exchange Liquidity Easing in Brazil in Response to the Global Shock of IMF Working Paper 09/259. Taylor, Mark P Economica, 71(281), pp Taylor, Mark P. and L. Sarno. Journal of Economic Literature, XXXIX, pp Wooldridge, Jeffrey M conometric A 2nd edition. The MIT Press, MA. 13

14 Appendix 1: Derivation of Models Ordered logit In order to derive the ordered logit model, assume that there exists a latent variable and determined by the following simple linear estimation: That is, this subjacent decision on depends on the decision effectively taken on and the vector of exogenous variables. Let us denote as the entire informational set known on by the BCB. represents the logistic distribution function with average 0 and variance 1. Taking abuse of notation into account, henceforth we also denote the accumulated distribution function of such uniform logistic distribution as, that is: In addition, is a vector (K x 1) and x does not contain a constant. We also define, such that and: That is, the actual decision taken on will depend on the distance of from the subjacent coefficients e. Therefore: Therefore 14

15 Binary logit The binary logit may be deemed a subcase of the multinomial logit where only two options are available to the BCB. Econometrically,, and, therefore,. In this case, is the intercept of the model. 15 Appendix 2: Database Considerations It may be argued that the liquidity lines implemented during the crisis were not intended to influence the exchange rate, but only to provide dollars to the banking system during a credit crunch. Others may argue that forward purchases are virtually irrelevant and that the only such transaction representing a significant amount, 7 billion dollars in February 2012, served only to counterbalance a large foreign issue in the same amount from a state company that would absorb this amount all at once. Therefore, we removed the above instruments from our sample and proceeded to the same main estimation shown in Table 1. As we can see in the second column of Table 8, the removal of liquidity lines and forward operations had virtually no impact on the output of the model. The model output was also practically unchanged by the removal of swaps made in the opposite direction to that taken by the expiration of another swap. For example, on June 1, 2012, 650 million dollars in reverse swaps were due to expire. On May 18, therefore, as a means to avoid the depreciation pressure on the exchange rate from the expiration of these reverse swaps, the BCB announced a traditional swap of equal value maturing on the same date, zeroing the flow. Aggregating all the information about swap volumes and maturities, we at the opposite end of maturities of other swaps, with a net position inferior to 500 million dollars on the referred day. Thus, if a traditional swap valued at 1 billion dollars is maturing on a given date, and the BCB carries out a reverse swap auction on the same day in the amount of 500 million dollars to 1.5 billion dollars, this second swap is removed from the sample. As shown in the third column of Table 12, the removal of such operations did not cause any ults. 15 This is the reason why wasn t allowed to have intercept. 15

16 Table 12: Robustness test Original Without lines and forward Without zeroing swaps Action (-1) (0.00)*** Action (-2) (0.00)*** Action (-3) (0.00)*** Action (-4) (0.00)*** BRL/Trend, log (0.01)** BRL/BasketPairs, log (0.00)*** DXY, log (0.00)** VIX/average(VIX), log (0.00)** S&P500/Ibovespa, log (0.00)** Reserves, log (0.00)** 50 day volatility, log (0.00)** Pseudo-R Observations In parenthesis, p-values of estimations. *, ** e *** denote statistical significance at 10%, 5% and 1%, respectively. 16