Abstract
This paper analyzes several issues regarding fiscal sustainability and fiscal adjustment in Brazil during 1990 and searches for econometric evidence of a monetary dominant regime during some subperiods. The following statistical data are also presented in detail: macroeconomic flows and balances, industrial production, consumer price index, relative public sector prices and tariffs, minimum wage statistics, financial system loans, monetary aggregates, exports by principal commodity groups, direction of trade, detailed balance of payments, total external debt, central government operations, and so on.
VI. Testing the Effectiveness of the Overnight Interest Rate as a Monetary Policy Instrument1
A. Introduction
1. This section analyzes the effectiveness of the overnight interest rate (the “SELIC” rate) as a monetary policy tool in Brazil. In particular, it addresses the question whether the SELIC rate that the Brazilian Central Bank (BCB) uses as a main policy tool is a good indicator of the monetary policy stance, and what effects it has on other market interest rates, output, and prices.2 Two main hypotheses are tested. First, with the annualized SELIC rate currently at 16.5 percent and average annual bank lending rates above 50 percent—implying both fairly high lending spreads and real interest rates—it may be argued that, apart from signaling, small changes in the SELIC rate, say by some 50 or 100 basis points, probably do not matter much. Second, as a corollary, one may also argue that the more significant effects of monetary policy, such as the recent pickup in bank lending, have been brought about by reductions in bank intermediation spreads (or simply lending spreads) and the ongoing structural reforms in the banking sector that have increased competition. If this were correct, one could ask if the SELIC rate has had a significant impact on these lending spreads, or whether it were other factors, like recent reductions in the reserve requirement on demand deposits, that have been shaping the effects of monetary policy.
2. These hypotheses are first discussed in some detail, largely using circumstantial evidence, and then tested using a simple vector-autoregression (VAR) model that traces the dynamic macroeconomic effects of monetary policy shocks that result from either changes in the SELIC rate or from changes in lending spreads (e.g., via reductions in the reserve requirement on demand deposits). VARs are a simple and convenient method for summarizing the dynamic relationship between variables. Also, once estimated, VARs can be used to simulate the response over time of any variable in the set to either an “own” disturbance or a disturbance to any other variable in the system.
3. This section is structured as follows. Following this introduction, we review the transmission mechanism of monetary policy; then we move on to discussing lending spreads and interest rates in Brazil; next, we present some statistical preliminaries for the VAR model, which is followed by a discussion of the results of the VAR model. The last part of this section presents some preliminary conclusions.
B. The Transmission Mechanism of Monetary Policy—A Review
4. The traditional view of the transmission mechanism of monetary policy has focused on the effects of real interest rates. If prices are sticky, an increase in short-term nominal interest rates will increase real interest rates and the real cost of capital. This, in turn, is likely to adversely affect spending on durable goods, such as fixed investment and inventories, and spending on housing and consumer durables. Also, in an open economy, monetary policy actions affect the trade balance (and therefore aggregate demand) through their effect on nominal and real exchange rates. In turn, changes in aggregate demand affect the level of production in the short run, and prices in the long run.
5. Beraanke and Gertler (1995), and Bernanke, Gertler, and Gilchrist (1998) noted that there is also a “credit channel” that complements the traditional channel by focusing on endogenous developments in credit markets that propagate and amplify shocks to the macroeconomy. The key mechanism involves the link between the “external finance premium” (the difference between the cost of funds raised in credit markets, and the opportunity cost of a firm’s internal funds) and the net worth of potential borrowers (i.e., borrowers’ liquid assets plus collateral value of illiquid assets less outstanding obligations).3
6. To illustrate the credit channel, suppose the central bank increases nominal interest rates. The traditional view predicts that firm profits drop because of a contraction in sales and increased costs of financing. Simultaneously, however, the firms’ asset value is likely to be reduced because future revenues are discounted at higher rates. Hence, firm balance sheets deteriorate, net worth is reduced, the amount of collateral is also reduced, and the probability of default on outstanding loans increases (“balance sheet effect”). At the same time, monetary policy affects lending spreads through the supply of credit relative to other forms of financing (“bank lending channel”). Specifically, banks will take into account the adverse impact of monetary tightening on their clients’ balance sheet, and accordingly tighten their supply of credit, thereby increasing lending spreads. As bank credit and other forms of financing are not perfect substitutes, firms cannot simply offset a reduction in bank credit, e.g., by issuing their own debt. As a result, some investments will not to be undertaken, which further depresses aggregate demand, output, and net profits (“bank lending effect”). Hence, the credit channel magnifies monetary policy actions by affecting firms’ balance sheets, net worth, collateral, risk of default, and, ultimately, the external finance premium.4
7. This general framework seems useful for studying monetary policy in Brazil, particularly given the prevailing large lending spreads. In particular, within this framework, it seems straightforward to consider the impact of monetary policy on economic activity, where “monetary policy” comprises both the SELIC rate and measures that affect lending spreads.5 In studying the macroeconomic effects of monetary policy, we abstract from the effects that monetary policy has on credit aggregates.6
C. Brazil: Bank Intermediation And Interest Rates
8. In this part we present largely circumstantial evidence on some main elements in the transmission mechanism for monetary policy in Brazil. In particular, we look at the role of the banking system in the monetary transmission mechanism and speculate on the role of the SELIC and of lending spreads in affecting the credit channel.
9. With the high and variable inflation that prevailed in Brazil in the 1980s and early 1990s, banks operated in a favorable environment: effective real interest rates were often negative; the real value of bank liabilities was eroded by inflation, and the inflation-induced high liquidity facilitated the repayment of loans. However, banks centered their operations not around lending, but around the “float,” which was inflation-related revenue based on various types of low cost liabilities (e.g., taxes collected, demand deposits, collateral against loans) that paid little or no interest for several days. Banks usually invested these almost free resources in short-term government securities that paid high nominal interest rates; in high inflation years, like 1991-92, about 40 percent of bank revenues were inflation-related.7
10. While the introduction of the Real plan in mid-1994 brought about macroeconomic stability, in and by itself it did little to reduce real interest rates or government (re)fmancing needs. As a result, banks, until most recently, felt no urgency to expand their domestic lending operations, and, as a result, the degree of competition for markets and customers remained low. More recently, with declining real interest rates, liberalization of banking sector activities, including opening to foreign competition, and an active program for restructuring and privatizing public banks, Brazilian banks have started to actively develop their lending portfolios.8 Against this background, it may be argued that, during much of the period since 1994, changes in the SELIC rate per se might not have had much impact on banking operations or economic activity in general, and that the functioning of the credit channel remained largely subdued.
11. Figure 6.1 plots the evolution of the annualized SELIC rate and annual average bank lending rates for individuals, enterprises, and a weighted average of the two.9 Accordingly, the three lending rates have evolved almost in parallel, suggesting that they all react to the same events, and the spread between bank lending rates and the SELIC rate has remained fairly constant. Only since late 1999, when the BCB started to address the determinants of high lending spreads and pursue options for reducing these, spreads have started to decline.
The Evolution of Annual Interest Rates: The SELIC and Average Bank Lending Rates
(Percent)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
The Evolution of Annual Interest Rates: The SELIC and Average Bank Lending Rates
(Percent)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
The Evolution of Annual Interest Rates: The SELIC and Average Bank Lending Rates
(Percent)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
12. A recent study by the BCB (1999) identified five main sources of the high lending spreads: the default rate for loans, the tax burden, administrative costs, profit margins in the banking industry, and minimum reserve requirements for various type of deposits. Most of these are not directly controlled by the BCB. The study concluded that loan defaults are the most important determinant of lending spreads, accounting for 35 percent of the total spread; followed by taxes (25 percent); administrative costs (22 percent); and bank profit margins (18 percent). The study defined lending spreads as the difference between effective bank lending rates and the banks’ effective cost of funds (taxa de captação) rather than the effective remuneration on demand deposits, and, hence, did not directly quantify the impact of changes in reserve requirements on demand deposits. Still, it acknowledged that lowering reserve requirements would induce a reduction in lending spreads.
13. The four factors whose impact on lending spreads was quantified—loan defaults, taxes, administrative costs, and profit margins—largely reflect macroeconomic, legal, or “cultural” factors. For example, loan defaults reflect macroeconomic conditions and the business cycle, but also moral hazard that may be created by imperfections in the legal system. Similarly, administrative costs reflect administrative efficiency, while profit margins reflect the degree of competition in the banking industry as well as demand factors (in this case, particularly government financing needs, which resulted in a typical situation of crowding out). Given the high spreads, the stock of credit to the private sector has remained consistently low in Brazil, and at end-1999 amounted to about 29 percent of GDP. In contrast, in Asian economies, like Thailand or South Korea, loans accounted for more than 130 percent of GDP following the 1998 crisis, but even in the United States they account for over 60 percent, and in the United Kingdom for about 100 percent of GDP.10
14. To help analyze these issues, Figure 6.2 plots the evolution of the reserve requirement on demand deposits and of lending spreads, as published by the BCB. The reserve requirement on demand deposits was kept at 75 percent between January 1997 and October 1999;11 since then, the BCB has gradually reduced it to its current level of 45 percent, thereby increasing loanable funds by over R$10 billion (about 1 percent of GDP) and allowing for a potential reduction of lending spreads. Figure 2 suggests that lowering reserve requirements on demand deposits may indeed have helped to reduce lending spreads. As a rough check on this relationship we ran a simple OLS regression of lending spreads (SPREAD) on the SELIC rate and the reserve requirement on demand deposits (RR), covering the period since January 1997. This resulted in the following estimate (with t-statistics in parentheses):12
The Evolution of Bank Lending Spreads (Percent), the SELIC rate (Percent), and Reserve Requirements
(Percent of Demand Deposits)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Source: Central Bank of Brazil.The Evolution of Bank Lending Spreads (Percent), the SELIC rate (Percent), and Reserve Requirements
(Percent of Demand Deposits)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Source: Central Bank of Brazil.The Evolution of Bank Lending Spreads (Percent), the SELIC rate (Percent), and Reserve Requirements
(Percent of Demand Deposits)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Source: Central Bank of Brazil.
15. The coefficients on both the SELIC rate and the reserve requirement are fairly large and significant at any level. This simplistic first estimate would suggest that a reduction in the reserve requirement on demand deposits by 10 percentage points reduces lending spreads by 5.9 percentage points, while a reduction in the SELIC by the same 10 percentage points reduces lending spreads by 5 percentage points.
16. More generally, in addition to reducing reserve requirements, there are several other measures that can be expected to reduce lending spreads. These would include tightening fiscal policies, strengthening property rights, fostering competition in the banking industry, and measures for reducing bank administrative costs. Fiscal adjustment, for example, reduces the government’s demand for credit, thereby allowing interest rates to fall; it also increases macroeconomic stability by reducing inflationary expectations and increases the economic growth outlook. In an environment of solid macroeconomic performance and protection of property rights, the loan default rate should decrease, adding further to a reduction of lending spreads. Also, increasing competition should foster efficiency in the banking sector; further reducing reserve requirements should reduce lending spreads by increasing the availability of loanable funds. Finally, reductions in bank administrative costs may also be brought about by central bank actions that facilitate the operations of the banking system.
17. These initial considerations suggest that lending spreads are determined by several factors, including also the SELIC rate, but they do not address the issue of the effectiveness of monetary policy via the SELIC rate. While Figure 6.1 showed that movements in the SELIC immediately imply affect all market interest rates, and while Figure 7.2 and the simple regression above showed that they also affect lending spreads, one may still question if a reduction in the SELIC rate by 100 or so basis points has a significant effect on the credit channel when average borrowing rates have remained above 50 percent while inflation has been running in the single digits. These issues will be explored further below.
D. Modeling the Transmission Mechanism: Statistical Considerations
18. To explore further the role of the SELIC rate in the transmission mechanism, a simple VAR model was used to trace the dynamic effects of monetary policy. While the main focus here is on the effects of shocks to the SELIC rate, we also evaluate the macroeconomic effects of reducing the spread between the SELIC rate and other market interest rates (by, for instance, reducing the reserve requirement on demand deposits, reducing taxes on financial transactions, or allowing for more liquidity in the system). VARs are generally a convenient method for summarizing the dynamic relationship between variables. Once estimated, they can be used to simulate the response over time of any variable in the set to either an “own” disturbance or a disturbance to any other variable in the system. Before proceeding to estimate a VAR, several standard econometric issues need to be addressed, particularly stationarity of the variables and causality.
19. The variables included in the VAR presented here are the SELIC rate (SELIC), the average bank lending spread for individuals and enterprises (SPREAD), real output as represented by a series of monthly GDP proxies (OUTPUT),13 the consumer price index (as measured by the IPCA that is also used for the purpose of inflation targeting) (IPCA), and money (M1). The reason for having a monetary aggregate in the VAR is to include a money market equilibrium (LM-type) equation in the model. The VAR is applied to monthly data from January 1995 to August 2000, yielding 68 observations. While the Real Plan was introduced in July 1994, the January 1995 starting date was chosen to assure that variables had stabilized following the period of hyperinflation that had prevailed earlier.14
20. For each series included in the VAR, unit root tests were performed to check for stationarity.15 The null hypothesis of existence of a unit root could not be rejected at the five-percent level of significance for OUTPUT, SELIC, SPREAD, and M1, using either the Augmented Dickey-Fuller (ADF) or Phillips-Perron (P-P) tests. For the IPCA, the presence of a unit root was rejected at the five-percent level of significance under the P-P test; this result probably owes to the low power of the P-P test, and we assume a unit root to be present, as suggested by the ADF test. The hypothesis of a second unit root was rejected for all series except for the IPCA at the five-percent level of significance. Hence, for all series except for the IPCA, taking the first difference would induce stationarity. For the IPCA, the presence of a unit root in the first difference was rejected at the five-percent level of significance, using both the ADF test and the P-P test, but not at the 1 percent level. Overall, the data suggest that the first differences of OUTPUT, SELIC, SPREAD, and M1 are stationary, and that the second difference of the price level (i.e., the acceleration rate of prices) is also stationary.
21. Next, pairwise Granger causality tests for levels and first differences were run to check the dynamic behavior of the variables in the system. Table 6.1 shows the Granger causality tests for the model in levels and first differences. The variables in the original model, which we will refer to as the “level” specification are OUTPUT, SELIC, SPREAD, the inflation rate as indicated by the IPCA, and M1. The causality tests were carried out for the level specification and the first difference of the level specification. We chose to run the tests for lags of one, two, and three months.
Granger Causality Tests 1/
The variables included in the model in levels are those in left-hand column; for the first-order differences specification, the results refer to first order differences of the variables in the left-hand column; e.g., the first-order difference of output being the change in output, and the first-order difference of inflation being the change in inflation, i.e., the acceleration rate of prices. The table shows the results of Granger causality tests at one to three month lags, with the direction of causality shown by the arrow. “No” means the Null hypothesis of no Granger causality could not be rejected; “Yes” means rejection of the same hypothesis at the 5 percent significance level; “Yes*” means rejection of the same hypothesis at the 10 percent level.
Granger Causality Tests 1/
| Levels (Lag Length in Number of Months) | First Order Differences (Lag Length in Number of Months) | |||||
|---|---|---|---|---|---|---|
| One | Two | Three | One | Two | Three | |
| Impact of output | ||||||
| Output ⇒ SELIC rate | No | No | No | No | Yes | Yes |
| Output ⇒ spread | No | No | No | No | No | No |
| Output ⇒ inflation | Yes | Yes | Yes | No | No | No |
| Output ⇒ money | Yes | Yes* | Yes* | No | No | No |
| Impact of the SELIC rate | ||||||
| SELIC rate ⇒ output | Yes | Yes | Yes | Yes | Yes | Yes |
| SELIC rate ⇒ spread | Yes | Yes | Yes | No | Yes* | No |
| SELIC rate ⇒ inflation | Yes | Yes | Yes* | No | No | No |
| SELIC rate ⇒ money | No | No | No | No | No | No |
| Impact of spread | ||||||
| Spread ⇒ output | Yes | Yes | Yes | Yes | Yes* | Yes |
| Spread ⇒ SELIC rate | No | Yes | No | No | No | No |
| Spread ⇒ inflation | Yes | Yes | Yes | Yes | Yes | Yes* |
| Spread ⇒ money | Yes* | No | No | No | No | No |
| Impact of inflation | ||||||
| Inflation ⇒ output | Yes* | No | No | No | No | No |
| Inflation ⇒ SELIC rate | No | No | No | No | No | No |
| Inflation ⇒ spread | Yes* | No | No | No | No | No |
| Inflation ⇒ money | No | No | Yes | No | No | No |
| Impact of money | ||||||
| Money ⇒ output | Yes | No | No | No | No | No |
| Money ⇒ SELIC rate | No | No | No | No | No | No |
| Money ⇒ spread | No | Yes* | Yes* | No | Yes | Yes |
| Money ⇒ inflation | Yes | Yes | Yes | No | No | No |
The variables included in the model in levels are those in left-hand column; for the first-order differences specification, the results refer to first order differences of the variables in the left-hand column; e.g., the first-order difference of output being the change in output, and the first-order difference of inflation being the change in inflation, i.e., the acceleration rate of prices. The table shows the results of Granger causality tests at one to three month lags, with the direction of causality shown by the arrow. “No” means the Null hypothesis of no Granger causality could not be rejected; “Yes” means rejection of the same hypothesis at the 5 percent significance level; “Yes*” means rejection of the same hypothesis at the 10 percent level.
Granger Causality Tests 1/
| Levels (Lag Length in Number of Months) | First Order Differences (Lag Length in Number of Months) | |||||
|---|---|---|---|---|---|---|
| One | Two | Three | One | Two | Three | |
| Impact of output | ||||||
| Output ⇒ SELIC rate | No | No | No | No | Yes | Yes |
| Output ⇒ spread | No | No | No | No | No | No |
| Output ⇒ inflation | Yes | Yes | Yes | No | No | No |
| Output ⇒ money | Yes | Yes* | Yes* | No | No | No |
| Impact of the SELIC rate | ||||||
| SELIC rate ⇒ output | Yes | Yes | Yes | Yes | Yes | Yes |
| SELIC rate ⇒ spread | Yes | Yes | Yes | No | Yes* | No |
| SELIC rate ⇒ inflation | Yes | Yes | Yes* | No | No | No |
| SELIC rate ⇒ money | No | No | No | No | No | No |
| Impact of spread | ||||||
| Spread ⇒ output | Yes | Yes | Yes | Yes | Yes* | Yes |
| Spread ⇒ SELIC rate | No | Yes | No | No | No | No |
| Spread ⇒ inflation | Yes | Yes | Yes | Yes | Yes | Yes* |
| Spread ⇒ money | Yes* | No | No | No | No | No |
| Impact of inflation | ||||||
| Inflation ⇒ output | Yes* | No | No | No | No | No |
| Inflation ⇒ SELIC rate | No | No | No | No | No | No |
| Inflation ⇒ spread | Yes* | No | No | No | No | No |
| Inflation ⇒ money | No | No | Yes | No | No | No |
| Impact of money | ||||||
| Money ⇒ output | Yes | No | No | No | No | No |
| Money ⇒ SELIC rate | No | No | No | No | No | No |
| Money ⇒ spread | No | Yes* | Yes* | No | Yes | Yes |
| Money ⇒ inflation | Yes | Yes | Yes | No | No | No |
The variables included in the model in levels are those in left-hand column; for the first-order differences specification, the results refer to first order differences of the variables in the left-hand column; e.g., the first-order difference of output being the change in output, and the first-order difference of inflation being the change in inflation, i.e., the acceleration rate of prices. The table shows the results of Granger causality tests at one to three month lags, with the direction of causality shown by the arrow. “No” means the Null hypothesis of no Granger causality could not be rejected; “Yes” means rejection of the same hypothesis at the 5 percent significance level; “Yes*” means rejection of the same hypothesis at the 10 percent level.
22. Three results stand out from the Granger causality tests on the model in levels, independent of the number of lags. First, the SELIC rate Granger-causes lending spreads (SPREAD) but not vice versa, although, curiously, there is some indication of Granger-causality running from SPREAD to SELIC at a two-period lag, which may be spurious. Note that Granger causality tests do not pick up contemporaneous causality. Figure 6.1 showed a strong contemporaneous correlation between the SELIC rate and bank lending rates, and it seems likely that changes in the SELIC rate cause changes in lending rates (and therefore lending spreads) contemporaneously, and not with a lag.
23. Second, both the SELIC rate and lending spreads Granger-cause inflation and output. As a result, there seems to be a transmission channel of monetary policy that goes from adjustments in the SELIC rate and lending spreads to economic activity. This relationship holds at all lags, suggesting that monetary policy actions via the SELIC rate or via lending spreads have a significant effect on economic activity. At the same time, there is no strong indication of Granger causality running from the inflation rate to either the SELIC rate or to lending spreads. This is probably due to the fact that the SELIC rate reacts to changes in inflation expectations rather than inflation outcomes; since the SELIC impacts also on lending spreads, the same would also apply to these spreads. Also, during the fixed exchange rate regime that prevailed until January 1999, monetary policy was aimed at maintaining the exchange-rate parity. Hence, prior to January 1999, we would only expect inflation to have impacted on the SELIC rate to the extent that inflationary pressures were threatening the sustainability of the exchange rate regime, which was generally not the case.
24. Third, there is strong evidence that money Granger-causes inflation. The relationship between output and inflation is less clear, with some Granger-causality running in both directions. There is some evidence for the standard money demand relationship, with Granger causality running from both output and spreads to M1, although the latter is fairly weak and with equally weak Granger causality running the other way around. Finally, Granger causality between output and inflation is less clear with causality running in both directions. The fact that there is some indication of Granger causality running from money to output with a one period lag, suggests that money may have at least a short term effect on real variables. However, this result should be interpreted with care, considering the significant remonetization of the Brazilian economy in the context of the Real plan.
25. Granger causality tests on first differences of the model show that most causalities vanish. For instance, there is no longer any Granger causality between changes in output and changes in inflation. However, changes in either the SELIC rate or lending spreads continue to Granger-cause changes in output; output changes in turn Granger-cause changes in the SELIC rate with a lag. Also, changes in lending spreads are shown to Granger-cause inflation, and changes in money supply Granger-cause lending spreads with a lag.
26. The unit root tests suggested that, to have stationary variables, the model should be estimated in first differences, i.e., where all variables are expressed in first differences of their levels, and the price level is expressed in second differences. However, the fact that many Granger causalities vanish when taking first differences, may indicate overdifferentiation: while the series become stationary, possible cointegrating relationships may no longer hold.16 While having stationary series is important in other econometric contexts, it is not clear that it is strictly necessary in a VAR system, which to some extent, is driven by cointegrating relationships between variables, irrespective of their stationarity.17
E. Analyzing the Monetary Policy Transmission Mechanism in Brazil
27. The VAR model was estimated in a nonstructural, recursive way, for both the level specification and the first differences specification.18 As just argued, working with nonstationary variables is frequently harmless when cointegration relationships are present. Also, working with nonstationary variables in a VAR context frequently also convenient: as stationary variables are often first or second order differences of levels, they are somewhat harder to interpret. In our case, for example, it may be easier to discuss the impact of interest rate shocks on output, than to discuss the impact of shocks to the change in interest rates on changes in output. Since the differences between the two specifications were, for the most part, not significant, only the results from the level specification are discussed here.19
28. In contrast, the ordering of variables in a VAR is always important, since it has strong implications for the identification of macroeconomic disturbances. In particular, it is crucial that variables enter in a logical sequence.20 There would seem to be two main logical orderings of variables for our purposes. The first ordering would be “output, inflation, SELIC rate, lending spreads, money,” with the rationale being that exogenous supply and demand shocks affect prices and output, and these (plus intrinsic shocks) in turn affect the SELIC rate and lending spreads; money then accommodates, given the value of all other variables. The second ordering would be “SELIC rate, lending spreads, output, inflation, money,” implying that overnight interest rates and lending spreads are determined in anticipation of supply and demand shocks; money would adjust, given the behavior of all other variables in the system.
29. With “reduced form” VAR equations it is not possible to distinguish pure supply and demand shocks unless some long-run restrictions are imposed, as, for example, in Blanchard and Quah (1989). An alternative would be to estimate a structural VAR, as, for example, Bernanke (1986) did. Still, shocks to the SELIC rate and to lending spreads, which are a main focus of this section, are perfectly identified assuming that shocks to the SELIC rate also affect lending spreads contemporaneously, but not vice versa. In the discussion that follows, we interpret shocks to the SELIC as “monetary policy shocks” and shocks to lending spreads “other financial market shocks.” The model was estimated using three lags of every endogenous variable. The choice of the number of lags is consistent with the Granger causality tests presented above, and seems sufficient to induce white noise residuals.
30. Figure 6.3 shows the impulse responses for the VAR with the first ordering suggested above, and Figure 6.4 shows the same for the second ordering.21 The impact of shocks to the SELIC rate and to lending spreads are the crucial point of this exercise
Brazil: Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—First Ordering of Variables
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil: Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—First Ordering of Variables
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil: Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—First Ordering of Variables
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil. Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—Second Ordering of Variables
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil. Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—Second Ordering of Variables
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil. Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—Second Ordering of Variables
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
31. In Figure 6.3, increases in both the SELIC rate and in lending spreads cause significant negative deviations of output from its long run value. Both magnitude and persistence of the output response are larger for shocks to the SELIC rate than for shocks to lending spreads; however, the output response peaks after five to seven months for both. Figure 6.3 also confirms a strongly positive and persistent reaction of lending spreads to the SELIC rate, which peaks after four months. The fact that the effects taper off only slowly may be related to the fact that, in the past, upward adjustments in the SELIC rate (and consequently lending spreads) were implemented rapidly and usually in crises situations, whereas downward adjustments were made over a period of many months. Hence, the initial upward adjustment in the SELIC rate had a persistent effect on subsequent levels of lending spreads. There is also a significant, but relatively small feedback in the opposite direction, where the SELIC rate responds positively to movements in lending spreads; this somewhat counterintuitive response also peaks after four months. A possible explanation for this may be that it reflects the strong contemporaneous correlation between the two variables, as suggested in Figure 6.1. Also, the apparent feedback from lending spreads to the SELIC rate is fairly small, and strictly speaking, only significant between lags three and five.
32. According to these results, increases in the SELIC rate would have both a direct adverse impact on output, and a significant but smaller adverse indirect impact on output via lending spreads. This transmission mechanism generally peaks after four to six months, attesting to a rapid response of output to interest rate shocks.
Brazil: Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—Using Lending Rates (TXAP)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil: Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—Using Lending Rates (TXAP)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
Brazil: Impulse Response to One Standard Deviation Innovations (± 2 Standard Errors)—Using Lending Rates (TXAP)
Citation: IMF Staff Country Reports 2001, 010; 10.5089/9781451805901.002.A006
33. Although not strongly significant, Figure 6.3 also reveals a somewhat puzzling response of inflation to changes in either the SELIC rate or lending spreads, where increases in either of these two variables seems to result in higher inflation. This effect is not uncommon in the VAR literature, and is referred to as the “price puzzle effect,” which describes a situation where inflation and interest rates seem to move in the same direction as a result of interest rates reacting to inflation expectations. For example, if future inflation is expected to increase, say due to the expected impact of a bad harvest on food prices, the central bank may tighten monetary policy in anticipation, which may mitigate the increase in inflation, but interest rates and inflation would probably still move in the same direction.22
34. The results in Figure 6.3 also suggest that shocks to output have a significant effect on output itself, but that the persistence is rather low. This may reflect the low, below potential, output growth rates in recent years, and may not be stable once output starts to grow closer to its potential rate. All other effects of output shocks are insignificant. Similarly, shocks to the inflation rate cause a significant but low-persistence increase in subsequent inflation rates. Again, all other effects of shocks to the inflation rate are not significant.
35. The response of money to other variables in the system is insignificant except for a negative response of money demand to increases in the SELIC rate, as suggested by economic theory. A contraction in money demand, in turn, appears to have a small adverse effect on output, although this is not strictly significant. The other variables in the model do not seem to be affected significantly by changes in money, although inflation seems to drop slightly following a contraction in money, with the reaction peaking after three months. As already mentioned, the small reaction of inflation to money may simply reflect the remonetization of the Brazilian economy under the Real plan, where the money supply initially increased rapidly; it may also reflect that money aggregates were a “passive” element under the fixed exchange rate system that prevailed until January 1999.
36. Figure 6.4 shows the impulse response when the second variable ordering is used. While the SELIC rate now enters the model first, the results do not change much. Increases in either the SELIC rate or lending spreads have a strong and persistent adverse effect on output, with the impact of the SELIC rate again being much stronger. The price puzzle is still present, although it remains not fully significant. The SELIC rate continues to have an adverse impact on money. Shocks to output, inflation, and money have the same effects as in the previous ordering, which suggests that the model is fairly robust. As in Figure 6.3, Figure 6.4 continues to show a two-way feedback between the SELIC rate and lending spreads, which is surprising, also given the results from the Granger causality tests.
37. Overall, the VAR results presented here convey a consistent picture concerning the effectiveness of monetary policy. First, they show that positive shocks to either the SELIC rate or to lending spreads adversely affect output, and do so with a delay of at least a quarter. Second, they also show that shocks to the SELIC rate have a relatively stronger and more persistent impact on output than shocks to lending spreads. Third, they show that shocks to the SELIC rate also have a significant and persistent indirect effect on output, through their effect on lending spreads.
F. Concluding Remarks
38. This section has explored the effectiveness of the overnight interest rate in the monetary transmission mechanism in Brazil. In particular, it analyzed the effects of movements in the overnight interest rate (the SELIC rate) and in bank intermediation spreads (or lending spreads) on overall economic activity. In September 2000, annualized lending spreads averaged 37 percent, compared to an annualized SELIC rate of 16.5 percent, implying that the SELIC rate accounted for not even one third of average bank lending rates.
39. In general, given the sheer magnitude of lending spreads in Brazil, it could be argued that monetary policy via the “direct” channel, i.e., changes in the SELIC rate, may be less effective than monetary policy via various “indirect” channels, i.e., the determinants of lending spreads, which include, for example, administrative costs, minimum reserve requirements, banking system liquidity, loan default risks, the degree of competition in the banking sector, and taxes.
40. Two main hypotheses were tested in this regard. First, given still relatively high real interest rates and the fact that the largest part of bank lending rates consists of lending spreads, it may be argued that, apart from signaling, small changes in the SELIC rate, say by some 50 or 100 basis points, probably do not matter much. Second, one may also argue that the more significant effects of monetary policy, such as the recent pickup in bank lending activities, have been brought about by reductions in lending spreads and ongoing structural reforms in the banking sector that have increased competition.
41. As it turns out, however, both hypotheses can safely be refuted. The evidence presented in this section, which ranges from casual empiricism and circumstantial evidence to a more formal analysis using a VAR model, suggested that movements in the SELIC rate do matter, and that the SELIC rate itself is also an important determinant of lending spreads. In particular, movements in the SELIC rate have a significant impact on all market interest rates, and therefore on lending spreads.
42. To analyze the broader issue of the effectiveness of monetary policy via the SELIC rate, a simple VAR model was used. In particular, the model was meant to quantify the importance in the monetary transmission mechanism of movements in the SELIC rate and the bank lending rate, and comprise both the “traditional channel” and the “credit channel” of monetary policy. The estimation results provide compelling evidence that changes in the SELIC rate have a powerful, i.e., significant and persistent, effect on output and lending spreads, and are probably a more effective tool of monetary policy than other policies that affect lending spreads. The model results appeared robust, and, in particular, were not significantly affected by changes in the definition of variables or the ordering of variables in the VAR model.
43. At least two small puzzles remained. First, it seemed difficult to explain why, as suggested by the VARs, bank lending rates would feed back into the SELIC; this effect was largely absent from the Granger causality tests. A possible explanation may be that the result could owe to the fact that the BCB has been implementing simultaneously measures to reduce both lending spreads and the SELIC rate. Second, there was also a small and not very significant “price puzzle” effect, where increases in either the SELIC rate or lending spreads seemed to increase inflation. The interpretation offered here was that interest rate increases may have coincided with exogenous shocks that increased inflationary expectations, such as, exchange rate developments, oil price changes, administered price changes, or minimum wage changes. For example, the SELIC rate was increased when the Real was floated in January 1999, which was followed by a period of somewhat higher inflation than had previously been the case. More generally, with changes in the SELIC rate determined by changes in inflation expectations (where the latter is usually correlated with inflation outcomes), it may well appear to be the case within a simple time series framework that interest rate increases go hand in hand with higher inflation. This paradox may already explain part of the price puzzle that showed up in the VAR. More detailed research in these two issues would clearly be desirable.
APPENDIX I Results of the Unit Root Tests
44. For the unit root tests, the monthy series for real output (OUTPUT), consumer prices (IPCA), wholesale prices (IPADI), and M1 were transformed by taking the natural logarithm. The real output series was also deseasonalized using the XII-additive procedure. The sample period used was January 1995 to August 2000. Two standard tests were used, the Augmented Dickey-Fuller (ADF) test, and the Phillips-Perron (P-P) test.
Unit Root Test Results
Unit Root Test Results
| Additional | Additional | Additional | Additional | |||||
|---|---|---|---|---|---|---|---|---|
| ADF | Regressors | P-P | Regressors | ADF | Regressors | P-P | Regressors | |
| SELIC | -1.51 | 2 | -1.61 | -6.44* | 1 | -6.22* | ||
| SPREAD | -1.79 | c,1 | -1.81 | c,2 | -7.30* | c | -7.32* | c,1 |
| OUTPUT | -2.16 | c,1 | -3.10 | c, t | -12.6* | c | -13.1* | c |
| IPCA | -3.28 | c, t,1 | -4.01** | c, t | -3.47** | c, t,1 | -3.49** | c, t |
| M1 | -2.27 | c, t | -2.53 | c, t | -8.89* | c | -8.82* | c |
Unit Root Test Results
| Additional | Additional | Additional | Additional | |||||
|---|---|---|---|---|---|---|---|---|
| ADF | Regressors | P-P | Regressors | ADF | Regressors | P-P | Regressors | |
| SELIC | -1.51 | 2 | -1.61 | -6.44* | 1 | -6.22* | ||
| SPREAD | -1.79 | c,1 | -1.81 | c,2 | -7.30* | c | -7.32* | c,1 |
| OUTPUT | -2.16 | c,1 | -3.10 | c, t | -12.6* | c | -13.1* | c |
| IPCA | -3.28 | c, t,1 | -4.01** | c, t | -3.47** | c, t,1 | -3.49** | c, t |
| M1 | -2.27 | c, t | -2.53 | c, t | -8.89* | c | -8.82* | c |
APPENDIX II Model Specification and Data Sources
45. Different model specifications were tried out initially. Based on the considerations discussed in this section, the model was defined to include the SELIC rate; the overall lending spread for enterprises and individuals; the natural logarithm of seasonally adjusted real output; inflation (IPCA), as defined by the first difference of the natural logarithm of the price level; and the natural logarithm of money (M1).
46. Monthly (end-of-period) data for the annualized overnight interest rate (SELIC), the various bank lending spreads, base money, and M1 (end-of-period) were obtained from the various BCB press releases; the general methodology for collecting the lending rate data is described in BCB (1999). The series for M2 was defined as M1 plus time and savings deposits; notably it excludes public debt instruments, that are included in the national definition of M2. Data on the nominal exchange rate of the real against the U.S. dollar were also obtained from the BCB. The model was run with various alternative specifications of these variables. For example, using average bank lending rates instead of bank lending spreads, did not alter the general conclusions presented in this section; similarly, the outcome was largely indifferent to whether M1 or M2 was used.
47. Output was proxied by a series of monthly GDP proxies, prepared by the BCB, and, alternatively by industrial production, as obtained through IBGE (Brazilian Statistical Institute). Both series were deflated using the general price index (IGP-DI). Working with the series of monthly GDP proxies seemed an improvement over the industrial output series as it also takes into account other productive sectors. However, qualitatively, the two variables yielded fairly similar results.
48. Consumer price inflation (IPCA) was taken directly from the Statistical Institute (IBGE). The index provides information on consumer prices in 11 metropolitan regions for families earning between 1 and 40 minimum wages. The IPCA is the main index for the BCB’s inflation targeting framework. Alternative price indices that were tried out, particularly wholesale prices (IPA-DI and IPA-OG), were obtained from the Getulio Vargas Foundation. Since the IPCA is used for inflation targeting, it was decided that it would be the best index to use for the purpose of this section.
References
Averbug, André, and Fabio Giambiagi, 2000, “The Brazilian Crisis of 1998/1999: Origins and Consequences,” (May) draft.
Baer, Werner, and William Miles, 2000, “Introduction” Quarterly Review Of Economics And Finance Vol.40, No.1, pp. 1-2.
Baer, Werner, and William Miles, and Nader Nazmi, 2000, “Privatization and Restructuring of Banks in Brazil” Quarterly Review of Economics and Finance, Vol.40, No.1, pp.3-24.
Banco Central do Brasil, 1999, “Juros e Spread Bancário no Brasil”, Departamento de Estudios e Pesquisas.
Bernanke, Ben, 1986, “Alternative Explanations of the Money-Income Correlation”, Carnegie-Rochester Conference Series on Public Policy, Vol.25, pp.49-100.
Bernanke, Ben, and Mark Gertler, 1995, “Inside the Black Box: the Credit Channel of Monetary Policy Transmission” Journal of Economic Perspectives, Vol.9, No.4, pp.27-48.
Bernanke, Ben, and Mark Gertler, and Simon Gilchrist, 1998, “The Financial Accelerator in a Quantitative Business Cycle Framework,” Working Paper, No. 6455, National Bureau of Economic Research (Cambridge, MA).
Blanchard, Olivier, and Danny Quah, 1989, “The Dynamic Effects of Aggregate Demand and Supply Disturbances,” American Economic Review, Vol. 79, No. 4., pp. 655-673.
Doan, Thomas, 1996, RATS User’s Manual, Version 4, Estima (Evanston, IL).
Grilli, Vittorio, and Nouriel Roubini, 1996, “Liquidity Models in the Open Economy: Theory and Empirical Evidence” European Economic Review, Vol.40, No.3-5, pp. 847-859.
Makler, Harry M., 2000, “Bank Transformation and Privatization in Brazil: Financial Federalism and Some Lessons about Bank Privatization” Quarterly Review Of Economics And Finance, Vol. 40, No.1, pp.45-69.
Nazmi, Nader, 1999, “Financial Reform, Banking Crisis, and the Exchange Rate Anchor: An Assessment of Latin American Experiences,” draft (available at www. lfc. edu/~nazmi/).
Nelson, Charles R., and Charles I. Plosser, 1982, “Trends and Random Walks in Macroeconomic Time Series” Journal of Monetary Economics, Vol.10, pp.139-162.
Ness Jr., Walter L., 2000, “Reducing Government Bank Presence in the Brazilian Financial System—Why and How” Quarterly Review Of Economics And Finance, Vol. 40, No.1, pp.71-84
Sims, Chris A., James Stock, and Mark Watson, 1990, “Inference in Linear Time Series Models with Some Unit Roots”, Econometrica, Vol.58, pp. 113-144.
Studart, Rogério, 2000, “Financial opening and deregulation in Brazil in the 1990s—Moving towards a new pattern of development financing?” Quarterly Review Of Economics And Finance, Vol.40, No.1, pp.25-44.
Zandamela, Rogerio, 1998, “The Transmission of Monetary Impulses in Brazil: Some Theoretical and Empirical Issues”, Brazil—Recent Economic Developments, International Monetary Fund (SM/98/27).
Prepared by Pau Rabanal and Gerd Schwartz.
The SELIC rate is the relevant interest rate in the overnight interbank market when public sector debt instruments are used as collateral, and it is the most important reference rate in the Brazilian economy; also see Zandamela (1998).
The external finance premium may broadly be viewed as bank lending spreads; changes in interest rates tend to change the external finance premium in the same direction. For a more detailed review of the credit channel in Brazil, see Zandamela (1998).
A similar argument can be applied to consumption. The traditional effect predicts that higher interest rates will provide incentives for savings and disincentives for consumption. However, individuals engaged in the purchase of durable goods, such as housing, will have to repay their debt at a higher interest rate. This higher debt burden will cause a decrease in the consumption of other goods.
In equilibrium, the marginal productivity of capital equals the risk free real interest rate. The external finance premium may be approximated by the spread between the risk free interest rate and the effective lending interest rates charged by banks. The risk-free interest rate can be approximated, alternatively, either by the overnight interest rate or by the cost of funds for banks, both of which are fairly similar.
While some of the VAR literature has tried to disentangle “balance sheet effects” and “bank lending effects” of the credit channel, it is difficult to separate the two effects from aggregate data, as pointed out by Bernanke and Gertler (1995). It is simply assumed here that there is a “credit channel” that goes from the SELIC rate and from lending spreads to overall economic activity, and an attempt is made to quantify the overall importance of this effect.
A recent issue of the Quarterly Review of Economics and Finance contained several good papers on these and other banking issues in Brazil, including Baer and Miles (2000), Baer and Nazmi (2000), Makler (2000), Ness (2000) and Studart (2000). Also see Nazmi (1999).
This view is also expressed in various trade publications. See, for example, BBA Icatu’s Monthly Report on Banks, of September 2000, which, under the header “Are Banks Heading to New Frontiers?” discusses bank efforts to develop their client base.
See BCB (1999) for a description of the methodology to compute these interest rates. Data reflect a survey of 17 major banks operating in Brazil.
Also see Averbug and Giambiagi (2000), and Baer and Nazmi (2000).
Previously, since mid-1994, the reserve requirement for demand deposits had been reduced gradually from initially 100 percent.
A constant term was not significant, and was therefore excluded from the regression.
Qualitatively, the various results presented here were largely invariant to using either the monthly series of GDP proxies or industrial production; see Appendix 2 for data sources. Also, note that the model estimated here uses actual output rather than a measure of the output gap. The Brazilian economy has performed well below its potential for several years, and the available econometric techniques are unlikely to yield a realistic picture of the output gap. Also, some of these techniques are problematic conceptually. For example, it is common to compute an output gap as the residual between actual output and potential output, where the latter is estimated with a Hodrick-Prescott filter. However, this filter is double-sided, and, in taking moving averages of past, current, and future levels of output to compute potential output, it violates one of the main assumptions of OLS (and VARs estimated using OLS), namely that future values of output are uncorrelated with current shocks. As a result, estimates employing the technique would be biased and inconsistent. But even other alternatives, e.g., estimating the output gap as the residual of a regression of output on a constant, a linear trend, and a quadratic trend, would not yield a realistic picture of the output gap in an economy that has already for some time been performing below its potential.
The results were somewhat sensitive to the starting date. Earlier starting dates (e.g., August or September 1994) changed many of the results; later starting dates did not. This suggests that some variables took time to stabilize following the introduction of the Real Plan.
The relevant statistics for the unit root tests are presented in Appendix I.
Using Johansen’s cointegration test, three cointegrating relationships between the five variables in the system were identified. While this test is useful in identifying the number of cointegrating relationships, it does not offer guidance on which variables are actually cointegrated. Still, given the evidence on the presence of cointegration, we assume estimating the VAR in levels to be a valid strategy.
See the Doan (1996), Nelson and Plosser (1982), and Sims et al. (1990) for a discussion on estimating VARs when series are nonstationary and possibly cointegrated. In general, in this paper, estimating first differences of the original model yielded larger standard errors and rendered most of the effects statistically not significant.
Except for the prices, which enter as the inflation rate in the “level specification” and as changes in inflation in the “first differences specification” of the model.
Qualitatively, the various results presented here were not particularly sensitive to the following changes: using lending rates instead of lending spreads; using industrial production instead of the output proxy; using base money (monthly average) or M2 instead of M1; using quadratically detrended output (“output gap”) instead of output levels; and using lending spreads for firms instead of overall lending spreads. When using wholesale prices (IPA-DI) instead of consumer prices (IPCA), the impact on inflation of interest rate shocks and lending spreads became completely insignificant. Finally, two minor changes in the estimation procedure did not affect the main results: the first involved including a dummy variable indicating changes in the reserve requirements for demand deposits; the second involved including the reserve requirement as an exogenous variable.
In a sense, a VAR amounts to regressing separately every endogenous variable against lags of all other endogenous variables using Ordinary Least Squares (OLS). Then, using the Cholesky decomposition of the variance-covariance matrix of the errors, the underlying shocks are recovered from the reduced form residuals. However, the decomposition of the shocks heavily relies on the ordering of the variables in the system. Economic theory and some priors about the contemporaneous relationship between variables need to be used in deciding the ordering of the variables.
Figure 6.5 shows an alternative specification where (using the same ordering of variables as in Figure 6.3), lending rates are used instead of lending spreads. In general, as mentioned in footnote 19 above, the results were not significantly sensitive to either using lending spreads or lending rates.
See also Grilli and Roubini (1996), who emphasize that the price puzzle might arise from inflationary expectations. To separate monetary policy shocks from endogenous responses to inflationary pressures, they suggest to use the spread between the risk-free short-term and long-term interest rates. This might be a good measure of real interest rate changes since it captures movements of short term rates not due to inflation. Unfortunately, Brazilian data series for long-term government bonds are fairly short. Using the spread between the SELIC rate and the implicit interest rate for one year swaps (“Pre x DI” for 12 month contracts) as a proxy for the long term risk-free interest rate did not resolve the price puzzle.
