Seasonal Cointegration Analysis of German Money Demand using Simple–Sum and Divisia Monetary Aggregates Klaus Eberl Diskussionsbeiträge der Katholischen Universität Eichstätt Wirtschaftswissenschaftliche Fakultät Ingolstadt Nr. 107 (ISSN 0938–2712) September 1998 Klaus Eberl Katholische Universität Eichstätt Wirtschaftswissenschaftliche Fakultät Ingolstadt Auf der Schanz 49 85049 Ingolstadt Germany Tel: ++49–841–937–1857 Fax: ++49–841–931–950 E–mail: [email protected] 1 Abstract In a multivariate seasonal cointegration framework four different money demand systems for Germany are investigated, using simple–sum and Divisia M3 as alternative measures of money. For each monetary aggregate both a nominal and a real money demand system is examined and from each system a real money demand relation is derived. In order to exclude diverging seasonal trends the Kunst and Franses (1998) approach is adopted. System stability is investigated by use of the Hansen and Johansen (1996) recursive estimation analysis. The Divisia aggregate systems yield consistent and stable real money demand relations, whereas the simple–sum M3 specifications show inconsistencies. Acknowledgements I gratefully acknowledge helpful comments and support from Josef Falkinger, Ulrich Küsters, Robert Kunst, Johannes Schneider, Franz Seitz, Winfried Vogt, and seminar participants in Ingolstadt, Regensburg, and Linz. 2 1 Introduction Since the introduction of Divisia monetary index numbers by Barnett (1980) there has been a debate if these indices could be serious alternatives to simple–sum monetary aggregates as measures for money in an economy.1 Some central banks, such as the German Bundesbank, still use a simple–sum monetary aggregate, namely M3, as an intermediate target in their conduct of monetary policy. Others, like the Federal Reserve or the Bank of England, have abandoned following a targeting strategy. The question whether a monetary aggregate is a useful target or not is clearly an object of empirical investigation. One major prerequisite is the existence and stability of a money demand relation where , , and denote the natural logarithm of a monetary aggregate, a transactions variable, and the price level, respectively, and denotes an opportunity cost measure. A sufficient condition for the existence of a money demand relation that has this special log–linear form is given if a quantity equation of money holds, where denotes the aggregate’s velocity of circulation which is assumed to be a stable nondecreasing function of opportunity costs only. Furthermore, the quantity equation concept imposes a nonpositive opportunity cost semi–elasticity , price homogeneity , and income homogeneity . In this paper we consider money demand relations using simple–sum and Divisia M3 as alternative monetary aggregates . We investigate four different money demand systems in a multivariate seasonal cointegration framework: for each aggregate one system in nominal terms with unrestricted, and one in real terms where price homogeneity is imposed. Most of the empirical literature on money demand considers real money, so the aim of this paper is to check if nominal vs. real system specifications give consistent results and can therefore be interpreted as evidence for the existence of a real money demand relation. The observation sample contains quarterly data from 1975 to 1997. Because of the evident seasonality in the data involved,2 all systems are investigated using seasonal cointegration analysis which was first introduced by Hylleberg et al. (1990) and was extended to a full information maximum likelihood method by Lee (1993). Here, the Kunst and Franses (1998) approach to seasonal cointegration is adopted which takes account of the fact that the data show non–diverging seasonal trends. Money demand relations are extracted from the non–seasonal long !" !$# %&# 1 2 For a more recent survey of the discussion, see Barnett et al. (1992). Graphical illustrations of the data are given in the appendix. 3 run part of each system, i.e. they are looked for in the cointegration space at the non–seasonal frequency. Furthermore, the observation period covers German reunification in 1990. Besides the fact that from July 1990 on the data concern the reunified Germany which manifests itself in an obvious jump in transactions and monetary data,3 this event could possibly have led to a structural break in the data generation process and, therefore, in the cointegrating relations. This issue is, for example, discussed in Hansen and Kim (1995), or in Wolters and Lütkepohl (1996). In both articles the authors find a stable real M3 demand relation in a single–equation non–seasonal cointegration framework. Here, a multivariate approach is chosen which gives a completer picture of the dynamics involved. In order to investigate system stability the recursive estimation procedure proposed by Hansen and Johansen (1996) is applied. This approach focuses on the evolution of long run relations while short run dynamics are held fixed. Here, we consider constancy of cointegration ranks and cointegration spaces at the non– seasonal frequency, and examine how estimated transactions and opportunity cost (semi–)elasticities evolve over time. The paper is organized as follows: In section 2 the monetary aggregate M3, the monetary index number Divisia M3 and two corresponding opportunity cost measures are introduced, section 3 reviews multivariate seasonal cointegration analysis, section 4 contains the empirical results, and section 5 concludes. 2 Monetary Aggregates and Opportunity Cost Measures The German Bundesbank uses the monetary aggregate M3 to monitor the demand for liquidity and transactions services within the German economy. M3 is defined as ')( *,+.0/ 13- 254 0 6 * 26 4 6 ** == currency, deposits, 487 6 * = demand deposits with up to 4 years time to maturity, 489 6 * = time savings deposits at three months notice, 4 held by domestic non-banks during period : . where The simple–sum aggregation procedure implies perfect substitutability between the components of M3, which is implausible since some of its components – namely time and savings deposits – bear interest and therefore are suited to con3 Again, see the appendix. 4 tribute to the accumulation of wealth. The motivation to hold these assets goes beyond the pure transactions purpose. In order to “concentrate out“ the liquidity–relevant part of money components, Barnett (1980) proposes the use of Divisia monetary index numbers4 instead. Divisia M3 is defined as Q Q Q Q ; <=5> ?A@ B5C)D E FG; <H= > ?I@ B CJD E K L FMOPQ N L%TS U V W E K LX V W E Y > ; <=Z W E G; <H=%Z W E K5L F R with the liquidity cost shares Q Q Q E M\[ L W E Z E W E E V W ] ^N R [ ^ W Z ^ W Q E where [ W equals the user cost of money component _ . In an intertemporal representative consumerQ model with perfect foresight, this Q user cost is5 E Ma` E X Gb E W E [ W Q ` S b where is the own rate of interest on component _ and ` is the interest rate on a default–free reference asset which yields no transactions services and therefore reflects the opportunity costs of holding money. In a nutshell, Barnett’s argument is as follows: In an intertemporal model where the liquidity services of money components contribute to the utility of the representative agent in terms of a linear homogeneous subutility function, the (continous) Divisia index tracks this subutility exactly. This implies that the Divisia index mirrors liquidity preferences for a wider range of models than the very special and unlikely case of perfect substitutes. Of course, the formation of Divisia M3 as a measure of transactions utility presupposes that the components of M3 can indeed be aggregated in a subutility function and are therefore weakly separable from other monetary components or from consumption goods. This assumption does not seem to cause any problems, since nonparametric tests for the validity of GARP do not reject weak separability.6 However, even if one does not ”believe” in the microfoundation framework the Divisia index nevertheless seems to be a useful measure for money, simply because of the way it is constructed. The growth rate of Divisia M3 is a weighted average of the growth rates of M3 components with time varying weights that are determined by opportunity costs. If the own rate of a component is close to the reference rate, it is likely to be a close substitute to longer term investments and 4 The expression “Divisia index“ is used throughout this paper to refer to what more precisely should be called the “Törnqvist–Theil discrete approximation of the continous time Divisia index“. 5 See Barnett (1978). 6 The nonparametric testing strategy is developed in Varian (1982) and Varian (1983). For an application of nonparametric tests to M3 components, see Scharnagl (1996). 5 therefore does contribute only a small amount to liquidity demand. If the own rate is low compared to the reference rate, the component is likely to contribute a large amount to liquidity, since it is likely to be held only contemporaneously. Simply, one could say that Divisia M3 measures the liquidity content of M3. For empirical considerations, I choose the following interest rates as own rates and reference rate, respectively: = 0%, = 0.5%, = rate on three months time deposits, from 100,000 DM to 1,000,000 DM, = rate on savings deposits at three months notice, = yield on bonds with at least 4 years time to maturity (Umlaufsrendite), during period .7 In times of an inverted term structure, which for instance was the case in some parts of 1992 and 1993, the calculation of Divisia user costs has to be modified in order to yield empirically useful results. If own rate exceeds the reference rate , user cost and user cost share would become negative. In order to avoid this undesirable feature it is assumed that in this case the money component does not contribute to transactions services. In order to achieve this the user costs are redefined to if if c d ef c g ef c h ef c i ef j f k j f m l ef c l ef n l ef dx m l e foqprs t{ ru v s w s j fzyc l e f j fz| c l e f The intuition behind this modification is that a money component cannot be “less illiquid“ than the reference asset.8 In the subsequent empirical analysis the modified costs will be used. For money demand considerations appropriate opportunity cost measures are needed for both monetary aggregates. As an opportunity cost measure for simple– sum M3, the difference between reference rate and a weighted arithmetic average of own rates i l ef j c f} o f3~ J c l ef l d f is used. 7 There seems to be consensus about the choice of these interest rates for the analysis of German (Divisia) money demand, e.g. see Wolters and Lütkepohl (1997), Gaab (1996). 8 A more consistent formulation of user costs could possibly be obtained by extending the intertemporal consumer model. On the one hand, it is not clear what has to be taken as reference rate. One could think of combining stock indices and longer term bonds in a reference group of assets to give a better measure of opportunity costs of holding money. On the other hand, an inverted term structure implies declining interest rates in the nearer future, with short rates declining faster than long rates. Taking account of this inherent uncertainty could lead to a more appropriate definition of user costs. Barnett, Liu, Xu, and Jensen (1996) derive user costs under uncertainty in a CAPM setting, which is a first attempt of research in this direction, but it seems that a lot more has to be done. 6 From the Divisia quantity index an opportunity cost measure for Divisia M3 can be derived by forming the dual price index H I J H I J 5 . % ¡ Hz¡ 5 ¢ ¡ a£ ) causes Unfortunately, the possibility of an inverted term structure (with consistency problems, so the following user cost aggregate will be used instead: ¤ ¤ ¥ ¦ . ¡ . HA§ § z¨ ¤ where the summands with § are set to zero.9 3 Seasonal Cointegration © Consider a –dimensional vector of quarterly time series sonal first differences « ª with stationary sea- ¬ ª q ª ª ª 3 J®¯ , that follows a VAR –process, ²± ¢ ¢ ¢ ²±z³ ³%´ ª ° ª ª (1) ´ ° where is Gaussian white noise and contains all deterministic components including seasonal trends. Reparametrization gives the seasonal error correction model10 « with µ° ³¶A ·3 « ,¶I¸ · ¸ ,¶I¹ · ¹ 5¸¶ · ¹ 8º » 3 ¼ » » ´ ½ ª ª · ¾ ¬% ¬ ¸¸ ¸ 5¹ ·3¸ ¾ ¬%¸ ¬ ª ª ª , ª 5¸ ª ¹ ª ·3¹ ¾ ¬ ª ª ª 5¸ ª ª ª ª (2) · ¹ In order to be able to estimate the system by canonical correlation analysis one has to impose the restriction that the so–called asynchronous cycles Ô. 9 10 Note that ¿À3ÁÂ Ã Ä À implies ÅHÃ Ä À ÆJÇ and È Ã Ä À ÆJÇ . In this case É Ê Ë!Ì Í Î Ï ÐzÑ5È Ã Ä À É Ò Ó ÅHÃ Ä À ÆJÇ See Hylleberg, Engle, Granger, Yoo (1990). 7 is taken as summand ÛÜ ÕÖ ÕÖ × ØIÙÕÖ ×5Ú !Ý ÜÞàß á Ø Þ Ý!æIÞ& ç ß èæzÞ$é êë5ì í Ýîæ ï á,æ ßð$èæzñ&ò Ý!æ do not have any influence on , i.e. .11 Under this assumption the time series are said to be seasonally cointegrated at frequency , , or , if the corresponding matrix . Let . In the case of seasonal cointegration at frequency we have , and can be decomposed as ß áâIÞ&ã á Ú Þ&ãäå ò8öè æ óæ ôæ Ý æ Þó æ ô3æ õ èæ óæ ôæ with –matrices and , both of rank .12 is called loading matrix, matrix of cointegrating relations. Note that the main object of our interest is the vector of moving annual averages, , because it is the only –variable which both has a simple economic meaning and whose dynamics is reduced to (at most) a single unit root process. Kunst and Franses13 have pointed out that without any restrictions on the deterministic seasonality contained in , model (1) implies diverging seasonal trends in general. In order to avoid this unlikely feature of seasonal time series they propose to restrict seasonal trend variables to values that cannot lead to common stochastic trends with different drift terms. Under this restriction the seasonal error correction model has to be written as ÷ Ø øÖ ÷ ù Û Ü Õ Ö ÞµúAûÚ ó Ø Ú ô Øõ Ú÷ ø Ö Ø × ø Ö ×5Ø ûóÖâ í ô âõ ÷3â ø ÖÖ ×5Ø ûüýÖþ ï û ûIó í ô× ØÜõ ÷ â Õ ûÖ ÿ × ý ûÖ ý ï û ÛÜ û where ú (3) ýÖ þÖ Þ í ã ï ý Þ í â ï Ù ý Ö Þ í â í ï ï and denotes remaining deterministic components.14 Efficient estimators for all parameters can be obtained from canonical correlation analysis between and plus relevant ,15 given ( ), lagged fourth differences of and all other deterministic terms. From the associated eigenvalue problems estimates of ranks and seasonal cointegration matrices are derived by use of trace and maximum eigenvalue test statistics. In a second– step OLS–regression estimates of , and are calculated. Û ÜÕÕÖÖ ÷ æ ø Ö × Ø ýÖ èæ ú ô5æ 11 ÷ ø Ö ×5Ø Þ ç óæ This restriction is necessary because otherwise #! $ "% cannot be decomposed in a cointegration and a loading matrix. For a detailed discussion of the complications in the general case , see Johansen and Schaumburg (1997). 12 See Hylleberg, Engle, Granger, and Yoo (1990), Lee (1992). 13 See Franses and Kunst (1996), Kunst and Franses (1998). 14 See Franses and Kunst (1996). 15 None if , if , and if . & #(' ) #* & #%+ ) #%, - . & #%/ 8 4 Empirical Results 4.1 Data and Model Specifications The sample period considered is 1975 through 1997. The starting year 1975 is often justified by two historical facts: the end of Bretton Woods in early 1973, and the first announcement of a monetary target by the German Bundesbank at the end of 1974. Both events could possibly have led to a structural break in German money demand. For this reason most of the literature on German money demand choose 1975 as sample start.16 M3 components and interest rates are taken from the monthly report of the German Bundesbank, which contains end–of–month series for money and monthly averages for interest rates. The Divisia M3 index and the opportunity cost measures were constructed from these series on a monthly basis, Divisia M3 with starting value 0 0 13254 687 95:%;3< =?>AB @ C H < 31 2JI C H < C DEGF in period 0, which is January 1975.17 Monetary and opportunity cost measures were then transformed into quarterly time series by calculating arithmetic averages over the corresponding quarter. German GDP in prices of 1991 and the GDP–deflator, which are used as transactions variable and price index, respectively, are available on a quarterly basis and are taken from the database of the Federal Statistical Office (Statistisches Bundesamt). From July 1990 on the series include data concerning the reunified Germany, which leads to a level shift in money, transactions and price variables. Levels and first differences of the data are shown in the appendix, Figures A1 and A2. and are considered in absolute values, monOpportunity cost measures etary aggregates, GDP and the deflator are considered in logarithms, with the abbreviations , , , and GDP–deflator . Four different money demand systems are investigated in what follows: 0 K ML 3K 0 M N L 0 > 3 1 5 2 4 % : ; = > 1 G 2 4 R 6 7 G 9 S : ; = > M PQO M M T M 13254 U86RV M = O M = M 0 12G4 Y M >Z4 O M T M W M K ML X Model 2 (nominal Divisia M3): Y M >\4 PQO M T M X Model 3 (real M3): Y M >Z4 O M ] W M T M K ML = [ X Model 4 (real Divisia M3): Y M >Z4 PQO M ] W M T M X Model 1 (nominal M3): 16 =[ WM > W M 3K M N L = [ K MN L = [ Scharnagl (1996) gives an overview of recent publications on German money demand. This choice, which is nothing else but a weighted geometric average of M3 components, enables us to interpret the value of directly as the (log of the) transactions–relevant “part“ of M3. This implies that represents a monetary aggregate rather than merely denoting an index number. 17 ^ _ `3a b c d eRf g h ^ _ `3a bc d eRf g h 9 For each system a seasonal error correction model (3) with restricted seasonal dummies is estimated, including a constant and an impulse dummy as deterministic terms. The impulse dummy has a value of 1 in 1990–3 and 0 otherwise in order to account for the level shift due to German reunification. TABLE 1 A KAIKE I NFORMATION C RITERION i j k l m noj j k l m sAj j u l m noj j u l m poj 4 M ODEL 1 M ODEL 2 M ODEL 3 M ODEL 4 k l m poj k l m toj u l m sAj u l m toj 5 k l m kAjk q m nojk q m r k l m uAjk q m nojk q m s u l m roju l m uoju l m q u l m uAju l m lvju q m n 6 7 8 Table 1 gives the values of an AIC for each system.18 From these results, for each model a VAR–degree of was chosen.19 Model misspecification test statistics are given in Table 2. The Lagrange multiplier (LM) test for residual autocorrelation, the normality test and the vector heteroscedasticity test using squared residuals are taken from Doornik and Hendry (1997). The test statistics were calculated using Ox 1.20a developed by Jurgen Doornik.20 wxzy TABLE 2 M ODEL M ISSPECIFICATION T ESTS21 M ODEL 1 autocorrelation LM(1) LM(4) LM(8) LM(1–4) normality heteroscedasticity M ODEL 2 {"| l p } l n k ~5 l mt s 0.84 2.08* l mt s 1.13 0.92 {J| p r3} l q ~G 1.21 1.26 G | n ~ 29.7** 29.9** {"| k u q } u p p ~5 0.55 0.53 M ODEL 3 M ODEL 4 l mn t 1.41 1.42 0.62 0.37 0.72 {J| } l p q ~G {"| k p } l p ~5 1.18 G | p ~5 20.3** 17.8** {J| l r r3} u r s ~G 0.90 0.82 0.74 There are only minor problems with autocorrelated or heteroscedastic residuals. Vector normality is clearly rejected in each case, but this fact seems to be of less importance since the full information maximum likelihood method applied here for seasonal cointegration analysis is robust against deviations from normality. 18 See Lütkepohl (1993). Note that a seasonal error correction model requires a VAR–degree of at least 4. 20 See Doornik (1996). 21 Throughout this paper, , *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 19 10 TABLE 3 S EASONAL C OINTEGRATION T EST S TATISTICS J ? rank M ODEL 1 M ODEL 2 M ODEL 3 M ODEL 4 4.2 frequency 0 trace 60.2** max ev 30.6* 23.6* 5.8 0.3 42.9** 8 8S 8 6.0 8 0.3 8 70.3** 8S 27.4 8 6.0 5.7 8 0.3 0.3 8 26.9 22.6* 8S 4.3 3.7 8 0.6 0.6 8 35.0* 29.1** 8S 6.0 5.8 8 0.2 0.2 trace 144.6** 96.1** 48.2** 19.5** 140.9** 81.6** 39.4** 14.4** 110.9** 62.2** 23.0** 103.3** 55.5** 16.9** max ev 48.5** 47.9** 28.6** 19.5** 59.2** 42.2** 25.0** 14.4** 48.7** 39.2** 23.0** 47.8** 38.5** 16.9** trace 127.8** 82.8** 46.9** 21.8** 126.8** 70.7** 36.9** 14.1* 105.4** 54.8** 22.5** 96.7** 39.6** 13.7* max ev 44.9** 35.9** 25.0** 21.8** 56.1** 33.8** 22.8* 14.1* 50.6** 32.3** 22.5** 57.1** 26.0** 13.7* Multivariate Seasonal Cointegration Analysis Table 3 contains trace and maximum eigenvalue test statistics for each model at each frequency. Critical values are taken from Franses and Kunst (1996) who tabulate critical values for a sample with size .22 For each model a maximum number of cointegrating relations at the seasonal frequencies and is accepted at a 5% significance level or higher. This means that the matrices and have full rank and there is no stochastic seasonality within neither system. The absence of multivariate stochastic seasonality implies that none of the univariate time series has a seasonal unit root at or . For the non–seasonal frequency 0 in both nominal money demand models a is accepted, whereas in the real money demand cointegration rank of models one cointegrating relation is found. Standardized cointegration vectors are given in Table 4, with coefficients of the monetary aggregates normed to . Since the cointegration space is spanned by these vectors a money demand relation has to be looked for in the set of all linear combinations of cointegrating relations. In order to select an appropriate relation for models 1 and 2 which is consistent with monetary theory, the real money demand relation that lies in the respective cointegration space is depicted as money demand relation. Note that for two cointegrating relations this real money demand relation is uniquely defined. The (implied) real money demand relations are given in Table 4. z\ 3 3 ¡R ¢R£ ¤5¥¦ ¡R 22 §¨© © ª« Although these values are only approximations – the observation period 1975–1 through 1997–4 covers a sample of size only (note that ) – it can be expected that the error is negligible because there is only little difference between the critical values Franses and Kunst simulated for sample sizes of and . See Franses and Kunst (1996), Table 1. §S% §¨ 11 TABLE 4 S TANDARDIZED C OINTEGRATING R ELATIONS AT FREQUENCY 0 M ODEL 1 0.20 2.06 1.02 1.31 implied real money demand 1.35 1 M ODEL 2 1.61 0.51 1.25 0.91 implied real money demand 1.17 1 M ODEL 3 1.10 M ODEL 4 1.18 ´ ¬® µ °¯² ´¶ ±· ¸ ³ ¹ ´µ ´¶ · ¶ º ´µ 0.22 » ´¬® ° ¼ ¯ ±½³ µ 0.16 ´µ ´¶ · µ ¾ ´µ ´¶ · ¿ ¿ ´ ¬ ´¯À µ ´¶ ±· Á ³ ¿ » ¬ ´ ´¯À µ ´¶ ± ·½  ³ ¾ The real money demand relation obtained from the nominal M3 model 1 shows an income elasticity which is slightly higher than one. This finding is quite in accordance with theory since M3 holdings may partly contribute to the accumulation of wealth. A serious problem arises from the positive semi–elasticity of the opportunity cost measure which stands in stark contrast to any theoretical conception of a quantity equation of money. The implied real money demand in model 2 also yields an income elasticity exceeding unity, but to a smaller amount than the value in model 1, which confirms that Divisia M3 expresses more closely the transactions relevant part of M3 money. In addition, the opportunity cost semi–elasticity shows the right sign. The real money demand obtained from model 3 is seemingly contradictory to the findings in model 1. The parameter values fit better to economic theory than in the nominal model. Obviously, the nominal and real M3 systems yield inconsistent results, which shades doubt on the usefulness of M3 with regard to the formulation of money demand relations. Results are much more homogeneous for Divisia M3. The real money demand relation estimated in model 4 shows only minor differences to the implied real money demand obtained from model 2. The inconsistency concerning M3 does not appear here. In order to be able to interpret the estimation results as money demand relations we yet have to examine the role of the cointegration matrix within the full system. For this issue the two–step estimation approach described in section 3 is applied to each model. First all cointegration matrices are estimated by canonical correlation analysis. Because the objects of interest are the real money demand relations already considered, for models 1 and 2 the first column of the cointegration matrix is replaced by the respective implied real money demand relation. Note that this operation does not change the cointegration space. In a second step Ã5Ä ÃQÅ 12 TABLE 5 L OADING V ECTORS FOR R EAL M ONEY D EMAND R ELATIONS (t–values in parentheses) M ODEL 1 0.02154 0.01351 ( ) (2.55) (3.21) ( ) M ODEL 2 0.01422 0.80115 (0.46) ( ) ( ) (3.56) M ODEL 3 0.00391 0.08839 (1.17) ( ) (4.05) M ODEL 4 0.02505 0.15888 (2.67) ( ) (2.66) ÆJÇ É ÆJÇ ÊË Î Ï Ð Ï JÆ Ï Ï Ç Ñ È Ò ÎÏ Ð Æ?Ñ Ó Ç Ô Ì Ô Í Õ ÎÏ Ð Ï Õ ÎÓ3Ð Ï Ò Æ?Ç Ö È Î Ï Ð Ï Ï ÆJÒ ÇÙ É ÑAÎ Ï Ð Ï Ù ÆJÚ ÇÙ Ê× Ú Æ?Ç Ì Ø Í Î Ï Ð Ñ Ï Î Ò ÐÔ Ó Î ÆJÇ Û È Ê3Ü Î Ï Ð Ï Ï JÆ Ô ÇÒ É Ù Æ?Ç Ì Í ÎÑ Ð Ú Ï Î Æ?Ç Û Ö È Ê3Ü Î Ï Ð Ï Ú ÆJÏ ÇÝ É Þ Æ?Ç Ì Ø Í Î?Ú Ð Ó Ô ßà estimators for the loading matrices and all other parameters are obtained from an OLS–regression. Table 5 gives the estimate of the first column of (in row form) for each model. The estimated coefficients reflect the short run adjustment of each variable to a deviation from the “long run equilibrium“ defined by the cointegrating real money demand relation. The model 1 adjustment coefficients hardly fit to a money demand interpretation of the corresponding cointegrating relation. If a positive (negative) monetary shock occurs, income and prices significantly adjust downwards (upwards) and opportunity costs adjust upwards (downwards), which is in contradiction to economic theory.23 The coefficient on the monetary aggregate is found to be insignificant. Model 2, however, is in accordance with theoretical considerations. Both coefficients on money and income growth are insignificant. All adjustment to a positive (negative) monetary schock takes place through a rise (fall) in prices and a fall (rise) in opportunity costs.24 In models 3 and 4 all coefficients show the expected sign. In model 3 adjustment to a deviation from long run equilibrium takes place through income and interest rates, whereas in model 4 only the coefficients of money and opportunity costs are significantly different from zero (at a 5% level). This means that in the M3 model a monetary shock implies real effects, whereas Divisia money is neutral. Furthermore, if we compare the coefficients of each M3 model with its Divisia counterpart we find that absolute values of (significant) coefficients in the Divisia models are about twice to five times as large. Therefore, the Divisia systems adjust to deviations from equilibrium at a faster rate than the M3 models. ßá ß ß 23 Note that money enters the cointegration relations with negative sign. The opportunity cost measures are constructed basically from interest rate spreads and are therefore real variables which should not be affected by inflation. 24 13 4.3 Recursive Estimation of Money Demand Relations One of the main prerequisites for a money demand relation is its stability which means that in an empirical analysis the relevant parameter values should remain constant over time. Hansen and Johansen (1996) give a guideline along which stability analysis for the long run part of a multivariate cointegrated system can be conducted. The idea is as follows: Take a subsample as a starting point, successively extend the sample back to its full size and observe the evolution of long run characteristics of the system – eigenvalues, statistics and cointegrating vectors. As Hansen and Johansen suggest, this can be achieved either by successive reestimation of the full system, or by keeping the short run parameters at their full sample values and only calculate eigenvalues and cointegrating vectors recursively.25 Since for short sample periods the loss of degrees of freedom might become a severe problem the latter method will be applied in this study. Furthermore, the Hansen and Johansen approach is adopted in a modified way which accounts for seasonality. Not only short run dynamics but also long run dynamics at the seasonal frequencies are considered constant, and only the system’s long run properties at the non–seasonal frequency are investigated by recursive canonical correlations analysis. The sample is extended from 1975–1 through 1984–4 to 1975–1 through 1997– 4 by successively adding four quarters. This choice takes account of German reunification in 1990 which could possibly have led to a structural break in the considered money demand relations. Two sets of questions will be addressed: First, we examine if the cointegration rank is constant over time. Second, the constancy of the cointegration space is considered. We are especially interested in how the estimated real money demand relations evolve. Cointegration rank constancy is examined by consideration of recursive trace statistics.26 In Figure 1 rescaled trace statistics are plotted, where a value of zero means significance at the 10% level.27 First, from all models a phase of system instability during the late eighties can be detected when no cointegrating relation is empirically found, whereas evidence towards cointegration becomes stronger in the period after 1991. Second, the nominal money demand models 1 and 2 show similar features with regard to recursive cointegration rank diagnostics, whereas the real models are more distinctive. At the 10% level no cointegrating relation is detected for any subsample specification in the real M3 model, but there is strong evidence towards cointegration in the real Divisia M3 model, especially when the subsample ends after German reunification. Third and perhaps most importantly, we observe that trace statistics according to 25 For details, see Hansen and Johansen (1996). Maximum eigenvalue statistics do not yield further insights and are therefore omitted. 27 Again, critical values are taken from the small sample simulations by Franses and Kunst (1996). There is a systematic error from the fact that these critical values were simulated for a sample size of 100 observations, but this error does not seem to be severe as was argued before. 26 14 âãäåçæzè âãäåçæ\é the null hypotheses and are higher in average in the Divisia models than in their simple–sum counterparts. This means that the cointegration relationship between variables is more intense in the Divisia specifications. Figure 2 shows recursive –statistics of a likelihood ratio test for restricted cointegration where the full sample cointegration matrix is taken as restrictions matrix.28 This test checks whether the full sample cointegration space equals the considered subsample cointegration space and therefore serves as an indicator for system stability. For recursive analysis, in the subsample specifications for models 1 and 2 a constant cointegration rank of 2 and for models 3 and 4 one cointegrating relation was assumed. The statistics were rescaled by their respective 5% critical values, so a value less than zero means insignificance. For each model the null hypotheses of a constant cointegration space is accepted, i.e. the cointegration space does not change significantly. In order to investigate the stability of cointegrating relations at the parameter level we finally consider the recursive shape of (implied) real money demand relations. Figure 3 shows income elasticities and opportunity cost semi–elasticities obtained from the recursive estimation procedure. Apparently, the estimates derived from the Divisia M3 models are much less volatile than the simple–sum M3 estimates (except for an outlier at sample end dates 1988–4 and 1989–4 in model 2). Furthermore, nominal and real Divisia money demand systems yield approximately equal elasticities whereas the simple–sum M3 demand systems show large discrepancies. Estimated opportunity cost semi–elasticities for M3 range from to . Especially, the finding of opportunity cost semi–elasticities about with opposite signs in models 1 and 3 for subsample ends since 1990–4 confirms the inconsistency result we already found for the full sample. Hence, the formulation of a nominal M3 system does not seem to be an appropriate approach to modelling money demand. êë ìRé í8èQî ï 28 For a description of the likelihood ratio test in a multivariate cointegration framework, e.g. see Johansen (1995). 15 16 1995 -1 -1 (c) M ODEL 3 -.75 -.75 1995 -.5 -.5 1990 -.25 -.25 1985 0 0 .5 .25 r<=0 r<=1 r<=2 .25 .5 (a) M ODEL 1 -1 -.75 -.75 1990 -.5 -.5 1985 -.25 -.25 -1 0 0 .5 .25 r<=0 r<=1 r<=2 r<=3 .25 .5 1985 r<=0 r<=1 r<=2 1985 r<=0 r<=1 r<=2 r<=3 F IGURE 1 R ECURSIVE T RACE S TATISTICS AT F REQUENCY 0 (rescaled by 10% significance levels) (d) M ODEL 4 1990 (b) M ODEL 2 1990 1995 1995 17 -.8 -1 -.8 -1 (c) M ODEL 3 -.6 -.6 1995 -.4 -.4 1990 -.2 -.2 1985 0 0 (a) M ODEL 1 -.8 -.8 1995 -.6 -.6 1990 -.4 -.4 -1 -.2 -.2 -1 0 0 1985 1985 1985 = COINTEGRATION MATRIX OF (rescaled by 5% significance levels) (d) M ODEL 4 1990 (b) M ODEL 2 1990 FULL SAMPLE F IGURE 2 STATISTICS OF RESTRICTED COINTEGRATION TEST RESTRICTIONS MATRIX R ECURSIVE ñð 1995 1995 18 1985 1990 1995 (c) M ODEL 3 -1 1995 -1 1990 0 0 1985 1 2 1 2 r^m 1995 2 1 1990 1995 1 1985 (a) M ODEL 1 1990 1.5 y 1985 -1 0 1 1.5 2 -1 0 1 1 1 r^m 1.2 1.2 r^dm y r^dm y 1985 1985 1985 1985 IN REAL MONEY DEMAND RELATIONS 1.4 y F IGURE 3 (d) M ODEL 4 1990 1990 (b) M ODEL 2 1990 1990 INCOME AND OPPORTUNITY COST ( SEMI –) ELASTICITIES 1.4 R ECURSIVE 1995 1995 1995 1995 5 Conclusion In a multivariate seasonal cointegration framework we investigated four different money demand systems for Germany, covering the period from 1975 to 1997. For both a simple–sum and a Divisia M3 aggregate a nominal and a real money demand system was specified and from each system a real money demand relation was derived. Stability analysis was performed by recursive canonical correlation analysis at the non–seasonal frequency. There is a discrepancy between the nominal and real simple–sum M3 systems. Especially the nominal M3 system seems to be inappropriate for modelling money demand because the derived real money demand relation implies a positive opportunity cost semi–elasticity. Recursive analysis confirms this feature which is present for varying sample sizes. Furthermore, adjustment to deviations from long run equilibrium takes place in a way that is contradictory to theoretical considerations. On the other side the Divisia M3 model specifications give results according to each other. Both models yield an income elasticity of about unity and a moderate and negative opportunity cost semi–elasticity. In these models adjustment to deviations from long run equilibrium takes place through prices and interest rates only, as would be expected from monetary theory. Parameter values are approximately constant over time which means that both Divisia money demand systems picture a stable relationship. Despite several theoretical objections against Divisia monetary indices, concerning e.g. the choice of money components or of an adequate reference rate, empirical evidence seems to confirm that Divisia M3 is a serious alternative to the simple–sum aggregate for modelling money demand within the German economy. 19 References William A. Barnett (1978). The User Cost of Money. Economics Letters 1, pp. 145–149. William A. Barnett (1980). Economic Monetary Aggregates. Journal of Econometrics 14, pp. 11–48. William A. Barnett, Douglas Fisher, and Apostolos Serletis (1992). Consumer Theory and the Demand for Money. Journal of Economic Literature 30, no. 4 (December), pp. 2086–2119. William A. Barnett, Yi Liu, Haiyang Xu, and Mark Jensen (1996). The CAPM Risk Adjustment Needed for Exact Aggregation over Financial Assets. Working Paper, Washington State University St. Louis. Jurgen A. Doornik (1996). Object–oriented Matrix Programming using Ox. International Thomson Business Press, London. Jurgen A. Doornik and David F. Hendry (1997). Modelling Dynamic Systems Using PcFiml 9.0 for Windows. International Thomson Business Press, London. Philip H. Franses and Robert M. Kunst (1996). On the Role of Seasonal Intercepts in Seasonal Cointegration. Discussion Paper TI 96–175/7, Tinbergen Institute, Rotterdam. Werner Gaab. On the demand for Divisia and simple–sum M3 in Germany, 1960– 93. In: Andrew Mullineux (ed.), Financial Innovation, Banking and Monetary Aggregates, pp. 160–186. Elgar Publishing, Cheltenham, 1996. Gerd Hansen and Jeong-Ryeol Kim (1995). The Stability of German Money Demand: Tests of the Cointegration Relation. Weltwirtschaftliches Archiv 131, pp. 286–301. Henrik Hansen and Søren Johansen (1996). Recursive Analysis of Eigenvalues in Cointegrated VAR–Models. Working Paper, Institute of Mathematical Statistics, University of Kopenhagen. Helmut Herwartz and Hans-Eggert Reimers (1996). Seasonal Cointegration Analysis for German M3 Money Demand. Diskussionspapier, Sonderforschungsbereich 373, Humboldt Universität Berlin. Sven Hylleberg, Robert F. Engle, Clive W. J. Granger, and B. S. Yoo (1990). Seasonal Integration and Cointegration. Journal of Econometrics 44, pp. 215– 238. Ottmar Issing, Karl–Heinz Tödter, Heinz Herrmann, and Hans–Eggert Reimers (1993). Zinsgewichtete Geldmengenaggregate und M3 – ein Vergleich. Kredit und Kapital 26, no. 1, pp. 1–21. Søren Johansen (1995). Likelihood–Based Inference in Cointegrated Vector Autoregressive Models. Oxford University Press, Oxford. 20 Søren Johansen and Ernst Schaumburg (1997). Likelihood Analysis of Seasonal Cointegration. Working Paper ECO No. 97/16, European University Institute. Robert M. Kunst and Philip H. Franses (1998). The Impact of Seasonal Constants on Forecasting Seasonally Cointegrated Time Series. Journal of Forecasting 17, pp. 109–124. Hahn S. Lee (1992). Maximum Likelihood Inference on Cointegration and Seasonal Cointegration. Journal of Econometrics 54, pp. 1–47. òó Helmut Lütkepohl (1993). Introduction to Multiple Time Series Analysis. 2 edition. Springer, Berlin. Michael Scharnagl (1996). Geldmengenaggregate unter Berücksichtigung struktureller Veränderungen an den Finanzmärkten. Diskussionspapier 2/96, Volkswirtschaftliche Forschungsgruppe der Deutschen Bundesbank, Frankfurt am Main. Hal R. Varian (1982). The Nonparametric Approach to Demand Analysis. Econometrica 50, pp. 945–973. Hal R. Varian (1983). Nonparametric Tests of Consumer Behaviour. Review of Economic Studies 50, pp. 99–110. Jürgen Wolters and Helmut Lütkepohl (1997). Die Geldnachfrage für M3: Neue Ergebnisse für das vereinigte Deutschland. IFO Studien 43, pp. 35–54. 21 22 1975 6.25 6.5 1975 6.75 5 5.5 6 1975 6.5 7 7.5 y dm m 1980 1980 1980 1985 1985 1985 1990 1990 1990 1995 1995 1995 1975 4.25 4.5 4.75 1975 1 1.5 1975 2 1 1.5 F IGURE A1 DATA , LEVELS p r^dm r^m 1980 1980 1980 1985 1985 1985 1990 1990 1990 1995 1995 1995 Appendix 23 1975 0 .1 1975 0 .05 .1 1975 0 .05 .1 Dy Ddm Dm 1980 1980 1980 1985 1985 1985 1990 1990 1990 1995 1995 1995 1975 -.025 0 .025 1975 .05 -.2 0 .2 1975 -.2 0 .2 Dp 1980 1980 Dr^dm 1980 Dr^m F IGURE A2 DATA , FIRST DIFFERENCES 1985 1985 1985 1990 1990 1990 1995 1995 1995
© Copyright 2025 Paperzz