Monetary Transmission Mechanisms in Spain: The Effect of Monetization, Financial Deregulation, and the EMS

Transcription

1 Monetary Transmission Mechanisms in Spain: The Effect of Monetization, Financial Deregulation, and the EMS Katarina Juselius University of Copenhagen and Juan Toro University of Oxford Abstract The paper presents a cointegrated VAR analysis of monetary transmission mechanisms and changes in them after Spain joined the EMS in Analyses of long-run price homogeneity within the I(2) model turned out to be crucial for understanding the joint behaviour of money, income, and prices. The empirical findings point to real monetization effects in the pre-ems period and the crucial role of shocks to nominal interest rates, in particular to the long-term bond rate. In addition, the paper provides further insights on the macroeeconomic effects of joining the EMS and financial deregulation. The increased economic integration within the EU is shown to have fundamentally changed the dynamics of the conventional IS- LM transmission mechanism and led to decreased effectiveness of monetary policy. Keywords: Cointegration, Long-Run Impact, Money Demand, IS LM, Monetary Policy, Capital Liberalization.

2 1 1. Introduction It was a starting hypothesis in Juselius (1998b) that the deviation from the PPP steady-state position and the degree of capital liberalization are crucial for how monetary mechanisms work for a country joining the ERM (Exchange Rate Mechanism) of the EMS (European Monetary System). For capital deregulated countries, which adopted the narrow 2.25% bands of the ERM and were above their PPP steady-state position, such as Germany and Denmark, the empirical results in Juselius (1996, 1998a, 1998b) demonstrated the impotency of monetary policy: Pure monetary shocks (changes in money stock) had minor effects (if any) on inflation. Changes in short interest rates had only short-run but hardly any long-run effects on inflation or on any of the intermediate instruments. For Italy which was capital regulated until 1989 and which adopted the broad currency bands of the ERM when its real (PPP) exchange rates were below its long-run steady-state value, Juselius (2001) demonstrated relatively higher effectiveness of monetary policy. Spain joined the EMS quite late, in It removed most restrictions on capital movements somewhat later and adopted the broad 6% bands of the ERM. As a result of speculative attacks it devalued its currency in September and November Juselius and Toro (1999) demonstrated that the real effective exchange rate was well below its steady-state value in the seventies and the eighties. Therefore, we expect monetary policy in Spain to be more effective before 1989 when it was capital regulated and below the PPP steady-state level. Moreover, as pointed out by Padoa-Schioppa (1995), free trade, free capital movements and (relatively) fixed exchange rates are inconsistent with independent fiscal and monetary policy (the so called inconsistent quartet). Therefore, we expect domestic interest rate control to be less effective after Spain joined the EMS in 1989 and, in particular, after it removed most capital regulations in This is more or less in line with the conclusions by Escriva and Malo de Molina (1991). They claim that the strict use of monetary objective has probably been ineffective in reducing monetary growth both in the pre- and post-erm period and point to a need for addressing the following important issues: (i) coordination of monetary and fiscal policy, (ii) the role of monetary targets when there are limits to the fluctuations of exchange rates, (iii) the managements of targets in term on money aggregates, (iv) the management of interest rate in the short term. Previous empirical work on monetary mechanisms in Spain has primarily focused on the estimation of aggregate money demand relations in a single equation

3 2 framework. See for example Dolado and Escriva (1990), Cabrero et. al. (1992), Vega (1995). All of them report various problems, of which implausible income coefficients and parameter instability seem particularly worrying from an econometric point of view. We attempt here to explain some of the previous empirical problems by adopting a multivariate approach to the analysis of the period before and after Spain joined the EMS and at the same time offer new empirical insight on the issues proposed by Escriva and Malo de Molina (1991). A system approach based on the cointegrated VAR model for money, income, prices, and interest rates allows us to estimate money demand and money supply as part of a monetary system characterized by long-run and short-run effects, interactions and feedback. By estimating the model separately for the pre- and post-ems regimes it is possible to shed some light on how capital deregulation and increased economic integration has affected the transactions channel of monetary policy. The organization is as follows: Section 2 discusses some methodological aspects of the econometric approach and Section 3 the cointegration implications of some hypothetical steady-state relations. Section 4 defines the statistical model and discusses its empirical properties. Section 5 studies co-movements between nominal money, prices and real aggregate income for the pre-ems and the EMS period in order to understand why long-run price homogeneity was rejected and how it affects a VAR analysis based on real money. Section 6 reports cointegration tests of the hypothetical long-run steady-state relations and selects for each period three steady-state relations to describe the long-run structure in the data. Section 7 reports an identified short-run adjustment model for each period, and discusses the basic findings based on a comparative analysis of the pre-ems and the EMS period. Section 8 asks whether inflation rate has been controllable by the monetary authorities and Section 9 summarizes and concludes. 2. Methodological considerations The last three decades have witnessed many reforms of the Spanish monetary procedure and several associated regime shifts (see the historical overview in Juselius and Toro, 1999). The monetary procedure was essentially based on monetary targeting of M3 (later the still broader measure, ALP 1 ) until 1995 when it was replaced by inflation targeting. At the first stage (from 1973 to 1987) the central 1 For a definition of the monetary aggregates see Juselius and Toro (1999).

4 3 bank primarily exerted direct control of M3 through changes in reserve requirements (quantity control). Thereafter (from 1987 onwards), as the increased capital liberalization made direct monetary control infeasible, indirect control through changes in the central bank interest rates (price control) became increasingly more important. After the 1990 reform which set the reserve requirement to a low 3%, and one year later to 2%, direct monetary control was essentially abandoned. Joining the EMS exchange rate system in 1989 and the associated capital liberalization seems by far the most important decision Spain has taken in the last decades. Our focus is here on how this affected the transactions channel of monetary policy, leaving the exchange rate channel for future research. In this sense we follow the principle of specific to general in the choice of variables, albeit general to specific in the choice of statistical model (Juselius and MacDonald, 2000, Juselius,1992,andGarratet.al., 2000) 2. To give the reader a possibility to assess the cointegration implications of this choice, Juselius and Toro (1999) extended the domestic IS LM model (Laidler, 1985) with BP effects and discussed how the hypothetical steady-state relations would have been modified correspondingly. Generally, a central bank is assumed to change its instruments to achieve a desired inflation rate, possibly also to influence real growth and purchasing power parity. The monetary transmission mechanism describes the causal links from a monetary intervention to the final target(s) via some intermediate targets and the behavioral relations linking the intermediate targets and the macroeconomy. Monetary control, direct or indirect, is based on the assumption that (i) there is a direct relationship between the level of money stock and the price level and (ii) the causation is from money to prices. Inflation control, acknowledging the practical difficulties of monetary control, specifies the monetary target directly as a desired inflationtargetaround anupperandlowerbound. Inthelattercaseno target exists for money stock, which is left to be determined by agents demand for money. More specifically we interpret the domestic monetary transmission mechanism to describe how changes in the intermediate targets, excess money stock (m m ) and the short-term interest rate (R s ), dynamically affect the domestic economy through the effect on the long-term interest rates (R l ), theoutputgap(y y ), and 2 The reason is that the VARapproach is very powerful for a detailed analysis of small systems, but almost unmanageable in large systems. A step-wise approach can also be justified because the cointegration property is invariant to increases in the information set. Any cointegration result found within the presently used set of variables would also be found in an extended analysis.

5 4 finally inflation rate ( p). The following simple diagram serves as an illustration: Central bank interventions : m ( ) m R (+) s (+) R l y ( ) y p ( ) (2.1) Empirical estimates of the transmission mechanism are usually obtained by fitting a VAR model to the relevant variables. Here we will utilize the integration and cointegration properties of the data to make inference on a number of important links determining the effectiveness of the monetary transmission mechanism, such as: 1. The strength of the association between a monetary instrument, an intermediate target and the ultimate target. 2. The dynamics of the short-run adjustment as a result of changes in the monetary instruments and in the relevant macroeconomic variables. 3. The long-run impact on inflation of shocks to aggregate liquidity and shortterm interest rate. A prior hypothesis is that the change in central bank independence combined with joining the Exchange Rate Mechanism (ERM) and the abolition of previous restrictions on capital movements has strongly affected the dynamics of the monetary transmission mechanism and the effectiveness of monetary policy. We expect both cointegration properties and short-run adjustment dynamics to differ in the two regimes possibly altering the direction of the arrows in (2.1). This gives the motivation for analyzing the two regimes separately. 3 Because of the sample split, inference on cointegration and integration properties is based on relatively short samples, so the interpretation of long-run has to be done with caution (Juselius, 1999). In the subsequent discussions we will use the concept of a long-run relation to mean a cointegrating relation between I(1) or I(2) variables and the notion of short-run adjustment when a stationary variable is significantly related to a cointegration relation or to another stationary variable. A necessary condition for a long-run relation to be empirically relevant in the model is short-run adjustment in at least one of the system equations. Basedonthisdefinition non-cointegrating relations incorrectly included in the model will eventually drop out: a stationary variable cannot significantly adjust to a nonstationary variable. 3 This was also supported by econometric testing (Juselius and Toro, 1999).

6 5 3. Steady-state relations In the first period with restrictions on capital movements and interest rate regulations it seems likely that Bank of Spain tried to lower inflationary pressure by reducing the level of aggregate liquidity in the economy. Its main tool was to change the reserve requirements of the private banks, a tool it used frequently in periods of substantial monetization of government debt. Thus, in this period we expect Bank of Spain to have attempted to keep the liquidity ratio, (m p y r ) t on a level consistent with an acceptable inflation level, π t. If (π t p t ) I(0), then an empirical relation can be formulated as: (m p y r ) t = b 01 + b 1 p t + u p,t (3.1) where m is M3, y r is real GDP, p is a consumer price index, and b 1 > 0. All through the paper lower cases denote logarithmic values and upper cases levels. If the central bank used (3.1) to control liquidity and inflation then we would expect money stock and inflation to be equilibrium correcting to this relation. In the second period restrictions on capital movements were lifted as were ceilings on the deposit rates of private banks. We expect two major effects from this: (i) observed money holdings are likely to be demand determined, reflecting agents transactions, precautionary, and speculative behavior, and (ii) the spread between the deposit rate (the yield on money holdings) and the alternative rate (the long-term bond rate) is likely to become smaller. The following steady-state relation describes agents demand for money: m t p t = y r t + b 2 (R bt R mt )+b 3 ( p t R mt )+b 02 + u mt (3.2) where R b is the yield on assets outside M3, essentially long-term bonds, and R m is an weighted average of the yields on the components in M3, and b 2 < 0,b 3 < 0. Indirect money control implies that the central bank attempts to control the level of aggregate liquidity by using short-term interest rate to influence the demand for money (3.2). The following conditions should be satisfied for this to work: 1. there is a direct link from central bank interest rates to the relevant market rates, i.e. there is cointegration between the central bank instrument and theintermediatetarget, 2. a change in the market interest rates changes the level of money demand (or excess aggregate demand) and

7 6 3. a drop in money stock corrected for productivity trend (or a drop in aggregate demand) results in a corresponding drop in prices. The first condition, related to the interest rate channel of the central bank interest rates (see Johansen and Juselius, 2001), is not empirically investigated here, but assumed to hold. 4 The second condition is crucially related to integration and cointegration properties of the data. If (R bt R mt ) I(0), (astheexpectation shypothesiswould predict) then influencing the intermediate targets by indirect monetary control is not likely to be very effective, because even if R mt goes up after a monetary intervention, R b,t will adjust accordingly and the opportunity cost (R bt R mt ) will not change much. If on the other hand ( p t R mt ) I(0) (consistent with the Fisher parity), then increasing R m,t will result in a corresponding adjustment in inflation rate (i.e. a cost-push effect). Thus, the effectiveness of indirect monetary targeting is crucially related to the empirical performance of the term structure of interest rates and the Fisher parity. The third condition is related to the presence of long-run price homogeneity (a necessary condition) and requires that a permanent shock to money stock has a positive long-run impact on prices. A further condition for monetary policy rules to be effective is (i) that the effect of a central bank interest rate intervention is transmitted from the shortterm money market to the long-term bond market (i.e. shocks to the short-term interest rates are driving the long-term interest rates) and (ii) that an increase in the latter lowers the demand pressure in the economy through its effect on longterm investment and consumption. In this case we would expect real aggregate income to adjust to the IS steady-state relation: yt r = b 03 + b 4 trend + b 5 R bt + b 6 p t + u yt (3.3) where b 4 0, b 5 < 0, and b 5 = b 6. If, on the other hand, output gap is associated with demand-pull inflation in a short-run Phillips curve relationship, then reducing the output gap by increasing the interest rate level will also lead to lower inflation. Therefore, we would expect inflation to be positively adjusting to the output gap, i.e. p t = b 04 + b 7 (y r t b 4 trend+) + u pt (3.4) 4 Actually, the interest rate channel did not work as expected for US data (Johansen and Juselius, 2001)

8 7 where b 7 > 0. In the empirical analysis we find that p t I(1) and, hence, p t I(2). The quantity theory of money predicts that the price level is related to (m y r ) t, i.e. to monetary expansion in excess of real productive growth. If (m p y r ) t I(1),we expect the inflation rate to adjust to deviations from this steady state. According to the short-run Phillips curve, inflation increases with excess aggregate demand. Both mechanism can potentially be relevant possibly with different weights in different regimes. 4. The empirical model ThebaselineVARmodelisgivenby: 2 x t = Γ x t 1 + Πx t 1 + Φ 1 Ds t + Φ 2 Dp t + Φ 3 Dt t + Φ 4 S t + µ 0 + µ 1 t + ε t, ε t N p (0, Σ ), t=1,..., T (4.1) where x t is a p 1 vector of variables in the system, t is a deterministic trend, Ds t is a vector of shift dummy variables, Dp t is a vector of permanent blip dummies, Dt t is a vector of transitory blip dummies, S t is a vector of centered seasonal dummies and the parameters Θ = {Γ 1, Π, Φ 1, Φ 2, Φ 3, Φ 4,µ 0, Σ} are unrestricted. If the data are cointegrated then Π = αβ 0 and µ 1 = αγ 1, Φ 1 = αϕ 1 are restricted to be in the cointegration relations. Model (4.1) is estimated with two lags. The variables are defined by: x 0 t =[m, p, y r,r m,r b ] t,t= 1977:4-1996:3 (4.2) where m is the log of the monetary aggregate M3 5, p is the log of the CPI, y r is the log of real GDP, R m istherateofreturnonm3, R b is the yield on mediumterm bonds and the first two observations are used as initial values. Both interest rates are divided by 400 to make the estimated coefficients comparable with logarithmic quarterly changes. The model contains the following dummy variables to account for significant reforms and interventions: Dt791 t = 1 in 1979:1, -1 in 1979:2, 0 otherwise, accounts for the start of the EMS with 9 participating countries, Ds871 t = 1in 198 7:1-1988:4, 0 otherwise, accounts for the effect on interest rates when the reserve requirements were increased from a maximum of 7% to 19%, Dp872 t = 1 5 The empirical analysis is based on M3 instead of ALP because of the very close substitutability between assets inside and outside ALP discussed in Juselius and Toro (1999).

9 8 Table 4.1: Roots of the VAR process Period I Eigenvalues of the Π-matrix: Modulus of 5 largest roots Unrestricted model: r = r = Period II Eigenvalues of the Π-matrix: Modulus of 5 largest roots Unrestricted model: r = r = in 1987:2, 0 otherwise, accounts for the removal of the ceiling on interest rates that credit institutions could pay on sight deposits, Dp924 t = 1 in 1992:4, 0 otherwise, accounts for the devaluation of the peseta in September and November Thesampleisdividedintwo periods,1977:4-1988:4 and 1989:1-1996:3 corresponding to the pre-ems and the EMS exchange rate system. The split coincides also with strong econometric evidence of parameter non-constancy at around 1989 of β and α (see Juselius and Toro, 1999)). As a general check of the statistical adequacy of the model both multivariate and univariate misspecification tests of autocorrelation, normality and ARCH were calculated for each sub-sample. Except for some weak evidence of ARCH effects in the second period the VAR model seemed well-specified for both periods (Juselius and Toro, 1999). The cointegration rank can be seen as an indication of how well markets adjust and, therefore, of market deregulation. Based on the arguments in Juselius (1998a, 1999) we expect that at least two common stochastic trends are driving this system. Thus, our preferred hypothesis is {r =3,p r =2}, butinthefirst more regulated period r might be lower. The inflation rate was found to be I(1) for both periods 6 (cf. Figure 5.1) which gives the rational for analyzing long-run price homogeneity in the I(2) model. In 6 Note that this does not imply that inflation is structurally I(1), but that there is no statistically significant mean reversion over the chosen sample period.

10 9 this model the (p r) stochastic trends need to be classified into the number of I(1) trends, s 1, and I(2) trends, s 2 = p r s 1. We expect the former to be associated with permanent real shocks and the latter with permanent nominal shocks. Because of the sample split the trace test is based on rather few degrees of freedom and the asymptotic distributions are usually poor approximations in such cases. But even if one corrects for the small sample size the power is often miserably low. However, the roots of the characteristic polynomial of the VARmodel provide additional information on the number of unit roots in the data we also report the former in Table 4.1. When there are I(2) or near I(2) components in the data the number of unit roots is s 1 +2s 2 implying that there will be one additional unit root in the model for each I(2) trend. Only if there are no I(2) components in the data is the number of unit roots (or near unit roots) in the characteristic polynomial equal to the number of singular values, i.e. zero eigenvalues, of the Π matrix. The intuition is that when x t I(2) and x t I(1), Γ in addition to Π contain information on unit roots in the data. InTable4.1thecharacteristicrootsarereportedfortheunrestrictedVAR and the VAR restricted to r =2and 3. In the first period there are three quite large roots in the unrestricted model. When restricting r to be 2 or 3 a large root remains. Because, restricting the rank of Π cannot remove a unit root in Γ, this is strong evidence of one stochastic I(2) trend, i.e. { p t, m t } I(1) in the first period. In the second period there are two pairs of large complex roots. Restricting r to3leavesonepairoflargecomplexrootsinthemodel 7. The order of integration and cointegration was formally tested based on the likelihood procedure (Juselius and Toro, 1999). For both periods our prior hypothesis {r =3,s 1 =1,s 2 =1} was statistically supported. However, for the second period the case {r =2,s 1 =2,s 2 =1} obtained equally strong support. As discussed above the tests results should be interpreted with caution because of the small sample properties. Nevertheless, {r =3,s 1 =1,s 2 =1} seems empirically defensible and we continue the empirical analysis with this choice. 7 These are probably related to the large cyclical swings in the growth rates in this period associated with the severe recession in the beginning of the nineties (cf. Figure 5.1)

11 10 5. Long-run price homogeneity and the nominal to real transformation Most empirical models of money demand use real money based on the implicit assumption of long-run price homogeneity. In this case the nominal to real transformation removes an I(2) trend in nominal money and prices, thus avoiding the more complicated I(2) model. Because, long-run price homogeneity is a crucial assumption underlying the monetary explanation of price inflation it is often assumed to hold without formal testing. When it does not hold but is imposed anyway, the econometric analysis can be seriously affected. Section 5.1 tests this assumption and finds that it is rejected. Section 5.2 discusses possible explanations and demonstrates some of the cointegration implications of analyzing real variables in this case. Section 5.3 discusses the nominal to real transformation Testing long-run price homogeneity In the I(2) model the hypothesis of long-run price homogeneity can be formulated as: β 0 i = [a i, a i,,, ], i=1, 2, 3, (5.1) β 0 1 = [b, b,,, ], (5.2) β 0 2 = [c, c, 0, 0, 0]. (5.3) where (β 1,β 2 ) are the orthogonal complements of β, β 0 x t and β 0 1x t define the CI(2, 1) relations, and β 2 define the variables which are affected by the I(2) trends (Johansen, 1995, Juselius, 1999). 8 Note that conditions (5.1) and (5.2) are implied by (5.3) (Kongsted, 2002). Table 5.1 reports the estimates of β,β 1,β 2 for both periods. The condition (5.1) was tested with a likelihood ratio test, distributed as χ 2 (3), and was rejected based on a test statistic of 9.54(0.02) in the first period and 9.31(0.03) in the second period. Although rejecting one of the hypotheses is sufficient for rejecting price homogeneity, we note that the estimated coefficients of β 1 further strengthen the rejection by suggesting that m and p 8 For simplicity we have disregarded that the dimension of β in the first period is 3 7 (instead of 3 5) because of the step dummy Ds871 andthetrendterm.

12 11 Table 5.1: Estimates of the β directions in the nominal model Period I ˆβ 1 ˆβ2 ˆβ3 ˆβ3.r ˆβ 1 ˆβ 2 m p y r R m R b D t Period II m p y R m R b t have moved in opposite directions in the first period. The graphs in Figure 5.1 (b)showthatthishasfrequentlybeenthecase. 9 The estimates of β 2 provide additional information about the magnitude of the I(2) trend and how it has affected the variables. Consistent with the rejection of price homogeneity the first two coefficients of β 2 indicate that the I(2) trend has affected nominal M3 with a larger weight than prices. Furthermore, the coefficient to real income is quite large in magnitude in both periods, so it might be close to I(2) Absence of price homogeneity: cointegration implications The validity of the nominal to real transformation is crucial for the analysis of real variables. Rejecting long-run price homogeneity is, therefore, likely to have important implications both for the econometric analysis and the interpretation of the results in such models. To shed some light on this issue Table 5.1 reports the following representation of β : 9 A possible explanation is that Bank of Spain reduced money stock by changing the reserve requirements to curb inflationary pressure as a result of increased government spending.

13 12 β0 1 β 0 2 = β 0 3 1, 1,,, 0, 1, 1,, 0,,,,, 0, 0, It appears that β 1 and β 2 satisfy long-run price homogeneity, but not β 3 for which the coefficients to m and p are unrestricted. Two just-identifying restrictions have been imposed on each β i, so no testing is involved. By additionally imposing long-run homogeneity on the third vector while keeping ˆβ 1 and ˆβ 2 fixed at their previous values, the cointegration implications of incorrectly transforming from nominal to real are described by the difference between ˆβ 3.r and ˆβ 3 in Table 5.1. In both periods price homogeneity was rejected because nominal money (corrected for income and interest rates) has grown faster than prices as illustrated by the graphs in Figure 5.1 (a). Except for the first few years after the oil crisis, money stock (M3) has grown faster than prices over the whole sample period. The adjustment coefficient to ˆβ 0 3x t (not reported in the table) showed that in the first period this relation had a significant effect on money stock (0.66, t-ratio 3.3) andonthetwointerestrates(-0.02,t-ratio-2.9and-0,11,t-ratio-3.3),butno significant effects on prices and real income, whereas in the second period it had an error-correcting effect on real income (0.17, t-ratio 3.4) and the two interest rates (-0.01, t-ratio 3.1 and -0.02, t-ratio 5.5), but no significant effect on money and prices. That prices did not adjust significantly to excess aggregate liquidity is an interesting result that needs an empirical interpretation. This period coincides with a strong expansion of the government sector as a result of building up the Spanish welfare state. The obligation of Bank of Spain to monetize government debt until the end of the eighties can explain the strong common component in the growth of money stock and real income in the first period. Thus, the conventional explanation that the long-run trends in prices have been caused by excess money expansion does not seem appropriate to explain the development of nominal money and prices in this period. In the pre-ems period thefastexpansionofthegovernmentsectorfacilitatedbytheobligationofthe central bank to monetize government debt seemed to have played an important role for the rapid growth of GDP without a proportional increase in prices..

14 13 5 Nominal money and prices m Changes in nominal money and prices prices D4m D4p Real money and income m-p 9.2 y-p Changes in real income and nominal money.15 D4m D4y Figure 5.1: The graphs of nominal M3 and implicit price deflator (upper lhs), yearly changes in nominal M3 and prices (upper rhs), real GDP and real M3 (lower lhs), and yearly changes in real GDP and M3 (lower rhs) The nominal to real transformation In the ideal case of overall long-run price homogeneity with real money and real income being I(1), the model could have been transformed into the I(1) model without loss of information by using either of the following data vectors: or x 0 t =[m p, y r, p, R m,r b ] t (5.4) x 0 t =[m p, y r, m, R m,r b ] t (5.5) However, overall long-run homogeneity was rejected in both periods and real

15 14 income was found to be close to I(2). A model based on (5.4) may approximately correspond to an I(1) model, but some information (described by the difference between ˆβ 3 and ˆβ 3.r ) relevant for explaining the non-homogeneous movements in nominal money and prices will be lost. Furthermore, such a model is likely to contain a near unit root, because both real income and real money were found to be close to I(2). See Kongsted (2002) for further details. The choice of Juselius and Toro (1999) was, therefore, to analyze the liquidity ratio m p y r together with inflation rate and real income growth, i.e. to base the VAR model on x 0 t =[m p y r, p, y r,r m,r b ] t.nevertheless, the LR test of the liquidity ratio transformation was strongly rejected and imposing it implied loss of information. In particular the implicit unit income elasticity is problematic given the strong evidence in Table 5.1 of a non-unit real income coefficient. Also, previous studies of the demand for money in Spain usually find an income elasticity between 1.5 and 2.0. The choice here is to continue the VAR analysis based on (5.4), but leave out the first few years after the oil crises. The results in Juselius and Toro (1999) indicated that the Spanish economy was far away from steady state in these years explaining some of the near I(2) problems. Nevertheless, for r =3a large root (0.94 in the first and 0.90 in the second period) remains in the model. Because the residuals are well behaved (approximately normal, no significant autocorrelation) the estimated coefficients should be consistent and unbiased, but the distribution is likely to be somewhere between Student s t and the Dickey-Fuller t (Hendry and Juselius, 2000). Thus, when interpreting the statistical significance we need to allow for somewhat higher t ratios for a given significance level. 6. Cointegration tests Here we use our empirical model as a (simplified) description of the domestic transmission mechanism to assess the hypothetical steady-state relations of Section 3 using cointegration analysis. The latter is also used to evaluate the strength of association between the target variable, i.e. the inflation rate, and the monetary instruments as well as between the monetary instruments and the intermediate targets. In doing so we note that cointegration, in particular between an intermediate and a final target variable, is crucial for the long-run impact of a policy action (Johansen and Juselius, 2001). Based on the obtained results we are able to formulate a fully identified cointegration structure. The single cointegration tests reported in Table 6.1 are used to discover where

16 15 and how strongly cointegration is present. The hypotheses are of the form β = {Hφ 1,ψ 1,ψ 2 }, i.e. we test whether a single restricted relation is in sp(β) leaving the other two relations unrestricted (Johansen and Juselius, 1992). If the hypothetical relations exists empirically, then this procedure will find it. Each hypothesis is tested in both periods and changes in cointegration properties are related to possible changes in monetary transmission mechanism as a result of joining the EMS. Under the headline Monetary relations we report tests of hypotheses related to (3.1) and (3.2), under Demand pressure relations hypotheses related to (3.3) and (3.4), and under Interest rate relations, hypotheses related to term structure of interest rates and the Fisher hypothesis. A necessary condition for a steadystate relation to have empirical support is that it corresponds to a stationary cointegration relation (or a combination of stationary relations). Since cointegration by itself does not say anything about the direction of causality crucial for the monetary transmission mechanism, the test results of Table 6.1 need to be evaluated jointly with an inspection of the α coefficients. At this stage we will only refer to the latter information informally in the text and leave a more formal treatment to Section 7. The inflation - excess liquidity relation (3.1) is tested in H 1 and H 13. This hypothesis obtains strong support in the first period where also ˆα gives evidence of equilibrium correction in the inflation equation and, thus, of monetary inflation effects in the first period. In the second period H 13, though only borderline accepted, estimates a negative relation between inflation and excess liquidity, i.e. money demand effects. H 2 and H 14 is a test of (3.2) with b 2 = 0,b 3 < 0. The hypothesis seems acceptable in both periods with similar estimates of b 3. The estimates of ˆα (not reported here) show similar adjustment coefficients for both periods; money stock has been equilibrium error correcting to this relation. Altogether, agents seem to have been less willing to hold their wealth in money in periods of high inflation than in periods of low inflation. H 3 tests (3.2) with b 2 < 0, b 3 =0 10. It is borderline acceptable but b 2 has the wrong sign for a money demand relation to be. The estimated ˆα suggests that R b R m has a negative effect on the change of money stock, but not on the level of the liquidity ratio. Because the expansion of government sector in this period was financed partly by issuing government bonds and partly by money printing, 10 Because the interest spread was found to be stationary in the second period this hypothesis is only relevant for the first period.

17 Table 6.1: Cointegration properties m-p y r p R m R b D87 trend χ 2 (υ) p.val. Period I Monetary relations H (1) 0.87 H (1) 0.55 H (1) 0.08 H (0) - Demand pressure relations H (1) 0.33 H (1) 0.72 H (1) 0.57 Interest rate relations H (2) 0.02 H (2) 0.02 H (2) 0.34 H (2) 0.03 H (2) 0.02 Period II Monetary relations H (1) 0.04 H (1) 0.29 H (0) - Demand pressure relations H (1) 0.09 H (1) 0.90 H (0) - Interest rate relations H (2) 0.01 H (2) 0.01 H (2) 0.25 H (2)

18 17 the results suggest two interrelated effects: a negative effect on money holdings from issuing bonds and a positive effect on the liquidity ratio from monetization of government debt. Thus, condition 2 in Section 3, page 5 does not seem to be empirically valid. Previous money-demand analyses have reported income coefficients greater than 1 and in H 4 and H 15 we relax the unit coefficient restriction on y r. Two justidentifying restrictions are imposed. In period I the estimated income coefficient is 1.9 (close to previously reported estimates) but the coefficient to the opportunity cost has the wrong sign. In period II the income coefficient has dropped to 0.45 (cf. the estimate of β 3.r in Table 5.1) and there is a strong positive effect from own interest rate on money. The effect from the bond rate on the liquidity ratio has altogether been insignificant in this period. Thus, in the pre-ems period direct monetary targeting seemed to have been potentially effective as a means to control inflation, but not indirect monetary targeting. This was because we could not find a stable money demand relation with plausible opportunity costs of holding money. In the second period the spread was stationary, but only R m seemed to have an effect on the liquidity ratio, possibly because of the close substitutability between money and alternative assetsinthisperiod. Inflationary pressure can also be related to demand-pull effects in periods of excess aggregate demand (possibly as a result of monetary expansion). This hypothesis is tested by the short-run Phillips-curve relationship (3.4) in H 7 and H 17. In both periods cointegration is accepted but with the wrong sign in period I. Demand pull inflation can be brought down by increasing the long-term interest rate, thereby lowering agents investment and consumption demand. The IS curve relationship (3.3) is tested in H 5 and H 16, both of which are accepted. The estimated ˆα showed that inflation, but not real income, was adjusting in the first period, whereas in the second period real income was equilibrium error correcting to this relation. Thus, inflation seemed to react to excess demand pressure in the pre-ems period, but not in the EMS period. Monetary expansion is generally not assumed to have real effects, except for possibly in the short run. The hypothesis that real income is positively related to real money and negatively to the government bond rate is tested in H 6 and H 18 and is strongly accepted in the first period. A similar effect was found in Italian data for the pre-ems period, suggesting some (possibly quite large) effects on real GDP from the monetization of government debt. In the EMS period of central bank independence and free capital movements, H 18 is just-identified but

19 18 the coefficients are no longer plausible and ˆα is essentially zero in the real income equation. Thus, real monetization effects were no longer present in the EMS period. Condition 2 at p. 5 requires that the transmission takes place from central bank interventions via the short-term money market to the long-term bond market. In H 8 H 12 and H 19 H 22 we test various versions of the Fisher parity (stationary real interest rates) and the expectations hypothesis (stationary interest rate spread). There is essentially no support for stationary real interest rates in either periods, but the interest rate spread (H 21 ) is found to be stationary in the second period. Nevertheless, adding inflation to the interest rate spread (H 22 ) increases the p- value from 0.25 to The sign of the inflation coefficient is consistent with a monetary policy rule. In the first period the two interest rates were found to cointegrate with a very small bond rate coefficient. Similar results were found in Juselius (1998a) for Danish data where the yield on M3 was similarly measured as a weighted average over all components, some of which gave zero, others very high interest. Also private bank deposites were highly regulated in this period preventing a market adjustment to take place. Based on the single hypotheses tests of Table 6.1 we now test a fully identified cointegration structure consisting of three long-run relations. The hypothesis is of the form β r = {H 1 φ 1,H 2 φ 2,H 3 φ 3 }, where the design matrices H i imposes at least two just-identifying restrictions on each cointegration vector. The test of overidentifying restrictions is based on the LR test procedure described in Johansen and Juselius (1994) and is asymptotically distributed as χ 2 (υ), where υ is the number of overidentifying restrictions. For the firstperiodwehavechosen a representation that consists of H 4, H 6, and H 10 andforthesecondperiodof H 14, H 17, and H 22. The joint hypothesis of the first period imposes five overidentifying restrictions and of the second period three. Both are accepted with high p-values (0.66 and 0.94). The estimated results are reported in Table 6.2. Thus, the first period is characterized by a strong IS relationship (with positive real money balance effects) determining real activity, by inflation responding with a small (and only borderline significant) coefficient to trend-adjusted excess liquidity, and by regulations of the banking sector. The second period is characterized by money demand behavior (although insignificant bond rate effects), by inflation responding to excess aggregate demand (the output gap), and by market determined interest rates. We interpret the results to imply that inflationary pressure was more strongly

20 19 Table 6.2: An identified long-run structure for the EMS period m p y r p R m R b D87 trend Period I β (0.10) β (0.03) (0.47) 0.05 (0.03) (0.0008) (0.0001) 0 β (0.02) (0.0003) The LR test for overidentifying restrictions: χ 2 (5) = 3.25 (p.val. =0.66) β (0.03) β (0.006) Period II (1.1) (0.0002) (0.0002) β (0.03) The LR test for overidentifying restrictions: χ 2 (3) = 0.41 (p.val. =0.94) Standard error are given in brackets related to excess monetary expansion (monetization of government debt) in the first period, but to excess aggregate demand pressure in the second period. In the first period interest rates did not move together in a stationary spread because of ceilings on the deposit rates, but in the second period of capital liberalization this was the case. 7. The dynamics of the monetary transmission mechanism 11 Table 7.1 reports the estimated dynamic adjustment structure for each period conditional on the estimated long-run relations of Table 6.2. (Johansen and Juselius, 1994). For each period insignificant lagged variables from the full system have been removed based on a F -test. Lagged changes in real income exhibited no significant effects in either period and were altogether omitted from the system. Lagged changes of the short-term interest rate could be omitted from period I but not period II. The residual covariance matrix showed a fairly high cross correlation coefficient between the short-term interest rate and the long-term bond rate in both periods, 11 The estimates are computed with PcFiml (Doornik and Hendry, 1998).

21 and between real money stock and real income in the first period. All attempts to model them as simultaneous effects in the full system failed. We believe this is because the model does not contain sufficient information to economically identify a causal relationship based on current quarterly changes. The long-term bond rate was weakly exogenous for the long-run parameters β in both periods and, in addition, it was not Granger caused by the remaining variables. To simplify the short-run structure the bond rate was, therefore, treated as exogenous in both periods. Real income was similarly weakly exogenous and not Granger caused in the EMS period and was also treated as exogenous in model II. All dummy variables except Di79.1 became insignificant when conditioning on the exogenous variables and were left out. The overidentified short-run structures in Table 7.1 have been determined by removing 14 insignificant coefficients in period I and 11 in period II. These were tested with a Likelihood Ratio test and accepted based on χ 2 (14) = 11.9(0.61) and χ 2 (11) = 8.88(0.63) respectively. It appears that most of the explanatory power of the two models comes from the error-correction mechanisms. The growth of real aggregate M3 is negatively related to lagged changes in inflation rate in period I, but positively in period II. The first effect might be evidence of central bank policy reactions to increasing inflation (alternatively a negative demand for money effect). In both periods real money growth is positively related to changes in the long-term bond rate. In the first period it seems likely that the positive coefficient describes an increase in money stock resulting from financing government debt partly by monetization and partly by issuing bonds, whereas in the second period it is more likely to describe a speculative money demand effect due to short-run expectations of further (similar) changes in the bond rate. Also, the close substitutability of assets inside and outside M3 is likely to blur the relationship between the growth in money stock, its own return and the return on long-term alternative assets. The growth of real GDP in period I reacted positively to current and lagged changes in the bond rate, lagged changes in real money stock and was equilibrium correcting to ecm1 (the IS steady-state relation with quite large real money stock effects). The positive short-run relationship between real GDP growth and changes in the bond rate reflects that a significant proportion of the former was linked to government expenditure financed by issuing bonds. In period II these effects were no longer significant and real GDP growth had essentially become exogenous to the domestic monetary transmission mechanism. This points to the importance of the expansion of the government sector in the first period, partly 20

22 21 Table 7.1: An estimated short-run adjustment structure for each period Period I Period II m r t yt r 2 p t R m,t m r t 2 p t R m,t yt r 0.47 (8.0) R b,t 1.79 (1.6) m r t (0.10) 2 p t ( 3.4) 0.15 (5.3) 0.40 ( 2.9) 0.31 (2.3) 0.84 (0.16) 0.47 (8.0) R m,t ( 3.7) R b,t (3.2) ecm11 t (2.8) 2.16 (1.7) 0.67 ( 6.2) 2.42 ( 2.9) ecm12 t ( 8.4) ecm13 t (5.9) 3.41 ( 1.7) (1.6) 5.01 ( 4.3) 5.07 (3.2) 0.08 ( 1.6) ecm21 t ( 7.4) ecm22 t ( 4.7) 0.31 (8.9) 9.50 (3.5) 3.50 ( 3.7) 0.36 (4.3) 1.22 ( 7.9) 0.07 (6.9) ecm23 t ( 11.0) Di ( 6.2) ecm11 t 1 = y r 0.47(m p)+3.04r b 0.005t ecm12 t 1 = p 0.05(m p y r ) t ecm13 t 1 = R m 0.35R b D87 ecm21 t 1 = (m p) 0.44y r 19.6R m t ecm22 t 1 = p 0.04y r t ecm23 t 1 = R m R b p Approximate t-values are given in brackets

23 22 financed by issuing bonds, partly by monetization of government debt. Inflation adjusts negatively to the lagged change in real money stock in period I, whereas in period II there is no adjustment. In both periods inflation adjusts negatively to the lagged change in long-term interest rate. This can be interpreted as decreasing demand-pull inflationary pressure as a result of increases in the longterm bond rate and consequently declining investment demand. In the first period inflation equilibrium error-corrects to excess money (ecm1 ) and tends to decrease when the short rate is higher than (a fraction of) the bond rate (ecm3 ). In the second period inflation is error-correcting to the demand pressure relation (ecm2) and positively related to excess demand for money (ecm1 ). Thus, in both periods there is evidence of some, but not very large, monetary policy effects on inflation. In period I the short-term interest is positively related to excess aggregate demand and excess money but not in an error-correcting manner, whereas in period II it is equilibrium correcting to the long-short spread. The results indicate that the use of short-term interest as an instrument in the first period was more effective than in the second. A plausible explanation is that capital deregulations and the membership of the ERM tended to neutralize the effect of the interventions made by the central bank. 8. Has inflation been controllable in this period? Johansen and Juselius (2001) showed that a necessary condition for inflation to be controllable is that a shock to a monetary instrument has a significant longrun impact on inflation. However, the interpretation of the VAR residual as a monetary intervention is not straightforward, particularly not in a model based on temporally aggregated data 12. With this reservation in mind we have used the residual ˆε it from (4.1) as an estimate of the unanticipated shock to variable x i. The estimated long-run impact of these shocks reported in Table 8.1 have been calculated from the estimates of the VAR model in the following way: C = e β α 0 where e β = β (α 0 Γβ ) 1,β and α,β are the (p p r) orthogonal complements of α and β (Johansen, 1996). Because the estimates of the long-run impact 12 For example, over a quarter money stock and short-term interest are likely to have been both exogenously determined by the central bank interventions, and endogenously by changes in the key variables of the system.

24 23 Table 8.1: The estimated long-run impact of unanticipated shocks to the system First period Second period The C matrix ( t values are given in brackets) ˆε m p ˆε y r ˆε p ˆε Rm ˆε Rb ˆε m p ˆε y r ˆε p ˆε Rm ˆε Rb m p 1.1 (2.3) 0.5 (1.9) 0.1 (0.2) 48.3 (3.2) 0.4 (1.1) 0.6 (3.0) 0.2 (0.3) 3.2 (0.3) 8.7 (2.1) y r 0.7 (1.3) p 0.0 (0.9) R m 0.0 ( 0.1) 0.3 (1.0) 0.0 (0.7) 0.0 ( 0.1) 0.1 (0.1) 0.0 (0.1) 0.0 ( 0.0) 30.2 (1.8) 1.0 (1.2) 0.2 ( 0.2) 6.1 ( 1.4) 11.0 ( 2.2) 0.3 (1.3) 0.8 (1.9) 0.5 (0.9) 0.0 (1.1) 0.0 (0.3) 0.8 (2.6) 0.0 (3.2) 0.0 (0.8) 0.2 (0.2) 0.0 (0.3) 0.0 (0.1) 4.2 (0.3) 0.1 (0.4) 0.1 (0.1) 15.8 (2.5) 0.4 (2.2) 0.9 (2.9) R b ( 0.2) ( 0.2) ( 0.0) ( 0.3) (2.0) (0.3) (0.9) (0.1) (0.1) (2.9) ˆσ ε are based on non-standardized residuals we report the residual standard errors separately. Standard errors of the estimated long-run impact are calculated using Paruolo (1997). Coefficients with a p-value of 0.10 or less are indicated with bold face. The C-matrix can be read column- or row-wise: the estimates of column ˆε i show the long-run impact of a shock to variable i on all the variables of the system whereas the row of x i show which of the shocks ˆε i,i=1,..., p have had a long-run impact on this particular variable. When assessing the effectiveness of monetary policy the long-run impact of shocks to money stock and short-term interest rate on inflation is of particular interest. In both periods shocks to real money stock have had a very small and positive, but insignificant, effect on inflation. Similarly, the long-run impact of shocks to the short-term interest on inflation is positive but insignificant. Thus, based on the present model neither money nor interest rate seem to have been effective instruments for inflation control in this period. Increasing the short rate might even have had a positive rather than a negative long-run impact on inflation (the so called price puzzle). See also Juselius (1998b), Juselius (2001), Johansen and Juselius (2001) for similar results. It is interesting to notice that shocks to real money had only a moderate long-run impact in period I and essentially no impact in period II. Similar lack of pushing behavior is also found for inflationrateinbothperiods.