Discussion Paper No. 32 UNSTABLE VELOCITY AND THE MONETARY APPROACH TO EXCHANGE RATE DETERMINATION: THE CANADIAN-U.S. DOLLAR EXCHANGE RATE By Stephen S. Poloz December, 1984 Economic Research Institute Economic Planning Agency Tokyo, Japan UNSTABLE VELOCITY AND THE MONETARY APPROACH TO EXCHANGE RATE DETERMINATION: THE CANADIAN-U.S. DOLLAR EXCHANGE RATE By Stephen S. Poloz* * Research Officer, Department of Monetary and Financial Analysis, Bank of Canada, Ottawa, Canada, KIA OG9. This paper was written while the author was a visiting researcher at the Economic Research Institute, Economic Planning Agency, Tokyo, Japan, November-December 1984. I would like to thank the members of the World Economic Model Group at the EPA for their hospitality, and to thank the Bank of Canada for granting leave to undertake the project. I am also grateful to David Longworth, Robert Lafrance and John Murray for their comments on the project proposal. The views expressed here are those of the author and no responsibility for these views should be attributed to the Bank of Canada, the Economic Planning Agency, or to the other individuals named above. Contents Page 1. Introduction ·············································································· 1 2. The Models and Data ·································································· 3 3. Adjusting Reported Money Stock Data for Shifts ···························· 11 4. Exchange Rate Model Estimation and Simulation Results ··············· 19 5. Further Testing of the DGA Model ·············································· 35 6. Conclusion ················································································ 41 Appendix ·························································································· 43 1. INTRODUCTION The purpose of this paper is to examine the empirical relevance of several alternative monetary models of exchange rate determination for the Canada-U.S. exchange rate. The principal motivation behind this study is the observation that most previous research has failed to take account of well-established results of the demand for money literature in implementing monetary exchange rate models, of which the demand for money function constitutes an integral part. In particular, it is argued below that instability in the demand for money has reduced the potential for success of monetary exchange rate models, and that preventing this distortion from biasing tests of these models requires that the problem of velocity shifts be dealt with explicitly prior to estimation. A monetary model of the exchange rate is one where the latter is defined as the relative price of two monies; early examples include Frenkel (1976) and Bilson (1978a). This study emphasizes this class of models to the exclusion of the more general asset-price approach of portfolio balance models and the more traditional trade-flaw approach. The latter theory, which emphasizes the numerous import and export demand functions which underlie the demand and supply curves for foreign exchange, is inconsistent with many of the observed empirical regularities of exchange rates (Mussa, 1979). The portfolio balance approach, which emphasizes imperfect substitutability of domestic and foreign assets, the supplies of which interact with a portfolio model of the demands to yield the market-clearing exchange rate (see Branson, Halttunen and Masson, 1977; Bisignano and Hoover, 1982), is judged to be inapplicable to the case at hand, since Canadian and U.S. dollar assets are perceived to be very close substitutes (Boothe et al., 1984). It is fair to say that the monetary approach, which may be viewed as a special case of the portfolio balance theory, has dominated the recent exchange rate literature, yet problems with its empirical implementation have prompted a steady succession of modifications. The result of this process is that more recent monetary models, while continuing to emphasize the role of money and the demand for money equation in the process of exchange rate determination, contain elements of both the portfolio balance and trade flow models; see Driskill (1981) and Hooper and Morton (1982), for example. The plan of this paper is as follows. In Section 2 we briefly summarize the chronology of monetary exchange rate models, and comment on the shortcomings which this study addresses. The data which are used are also described in Section 2, in general terms Section 3 focuses on the Canadian and U.S. demand for money equations, developing simple models which account for the level adjustments caused by financial innovation, and then constructing shift-adjusted money series therefrom. In Section 4 five exchange rate models are estimated and tested, using both published money stock data and the adjusted series constructed in Section 3 Section 5 -1- considers two special tests of the preferred model which emerges from the results of Section 4, while Section 6 offers some concluding remarks and suggestions for further research. Appendix A provides the detailed data definitions and sources, while Appendix B provides a bibliography. -2- 2. THE MODELS AND DATA 2.1 A Chronology of Monetary Exchange Rate Models The purpose of this section is to illustrate the sequence of modifications which have been made to the simple monetary model during the post decade. This purpose may be met by restricting attention to the various reduced forms and providing an intuitive discussion of the elements essential to each. For details of the derivations the reader is reterred to the original article. The analysis begins with the domestic and foreign demand for money equations which, for purposes of illustration only, are left in the following very simple form: (2.1) m = p + k0 + k1y - k2R (2.2) m* = p* + k0* + k1*y* - k2*R* where asterisks denote foreign variables, lower-case letters denote natural logarithms, the ki are positive parameters, m denotes money, y income and R the nominal rate of interest. The vast majority of work in this area assumes that ki = ki* for all i, an assumption which is convenient and will be retained for the purposes of this section. Then we may write the above two equations in relative form: (2.3) m-m* = p-p* + k1 (y-y*) - k2 (R-R*) To close the model we combine (2.3) with the assumption of continuous purchasing power parity (ppp), which may be represented as s = p-p*, where s is the natural logarithm of the spot domestic price of one unit of foreign currency. This gives the reduced form equation for the exchange rate: (2.4) s = m-m*- k1 (y-y*) + k2 (R-R*) This equation says that depreciations of the domestic currency (increases in s) are caused by relative expansionary monetary policy, high relative interest rates (representing, through the Fisher effect, a high relative inflation rate) or low relative income growth. It is recognized that the most critical assumption underlying (2.4) is ppp which, when tested with aggregate price indices, has been shown not to hold in the short run (Frenkel, 1981a; Officer, 1980). Despite some clever arguments that this failure is unimportant (see Bilson, 1978a), most subsequent models have attempted to allow explicitly for deviations from ppp. The path-breaking article in this spirit is Dornbusch (1976), who assumes that asset markets adjust instantaneously while prices are sticky. A slightly generalized version which -3- allows secular inflation is derived in Frankel (1979). The latter begins with the assumption of covered interest parity, which stutes that the forward premium (x) is equal to the interest differential (x = R-R*), and further assumes that x is a function of the gap between the current spot and the equilibrium exchange rate (s̅ ) and of the expected long-run inflation differential, (p-p*): (2.5) x = -k3 (s-s̅ ) + (p-p*) Combining (2.5) with covered interest parity and eliminating x gives an expression for the short-run spot rate: (2.6) s = s̅ -1/k3 (R-R*-p + p*) Notice that the term in parentheses is the real interest differential. The exchange rate is determined by ppp in the long run, which in turn is assumed to be determined by a long-run version of the relative demand for money equation (2.3) with (R-R*) replaced by (p-p*). However, it is assumed that the long run values of m, y and p may be taken as their current actual levels, which amounts to assuming that these variables follow a random walk, so the model essentially takes (2.4) with (R-R*) replaced by (p-p*) as a definition of s̅ , substitutes this into (2.6) and collects terms: (2.7) s = m-m*- k1 (y-y*) - 1/k3 (R-R*) + (1/k3 + k2)(p-p*) For our purposes (2.7) will be referred to as the Frankel model. The overshooting result of Dornbusch (1976) may be derived from (2.7) by noting that with p and y given in the shortrun and m exogenous, the endogenous variable in the relative demand for money equation (2.3) is (R-R*). Using (2.3) to eliminate (R-R*) from (2.7) yields a true reduced form which is free of the simultaneity between (R-R*) and s. (2.8) s = (1 + 1/k2 k3)(m-m*) - (k1 + k1/k2 k3)(y-y*) + (k2 + 1/k3)(p-p*) -1/k2 k3 (p-p*) Notice that the theoretical coefficient on (m-m*) now exceeds unity. For our purposes we will refer to equation (2.8) as the Dornbusch model. An alternative means of allowing for short-run deviations from ppp has been suggested in Bilson (1978a). Bilson defines (2.4) as the equilibrium exchange rate (s̅ ), which assumes that ppp holds and that the price level is determined by the interaction of money demand and exogenous money supply, and assumes that the exchange rate adjusts gradually to s̅ according to a partial adjustment hypothesis. -4- (2.9) s-s-1 = k4 (s̅ -s-1) , 0 < k4 < 1 Assuming that (2.4) defines s̅ and combining with (2.9) yields: (2.10) s = k4 (m-m*)- k4 k1 (y-y*) + k4 k2 (R-R*) + (1-k4) s-1 Function (2.10) we will refer to as the Bilson model. Notice that in this model, in contrast with the Dornbusch and Frankel models, the coefficient on the relative money supply is expected to be less than unity. At this point it is convenient to discuss a popular non-monetary model of the exchange rate, where (2.9) is assumed to hold but the demand for money equation is thought to be unreliable. Thus, rather than using (2.4) to define s̅ , the nonmonetary model uses ppp (s̅ = p-p*) so that the reduced form equation is given by: (2.11) s = k4 (p-p*) + (1-k4) s-1 In practice, any number of other exogenous variables which are perceived to influence the exchange rate in the short run may also be added to equation (2.11), such as real interest rate differentials or news about the current account, for example. Included in this class of model is the exchange rate equation which is in place currently in the Canadian Model of the EPA, as described in Cockerline (1984).1) To this point the models have assumed perfect capital mobility, so that any gradual adjustment takes place only in the goods market. This assumption is relaxed in the next class of model, known as stock-flow models, which contain elements of portfolio balance exchange rate theories. The relevant articles are Niehans (1977) and Driskill (1981). The approach drops the covered interest parity assumption of the Dornbusch (1976) model, replacing it with a net foreign asset demand function (F), a net trade flows equation (T) and a balance of payments identity, given by (2.12-2.14), respectively: (2.12) F = k5 (x-R + R*) (2.13) T = k6 (s-p + p*) - k7 (y-y*) (2.14) F = F-1 + T + A 1) The Cockerline equation begins with a form like (2.11) but constrains the adjustment to ppp somewhat by defining ppp as an eight-quarter moving average of (p-p*) One problem with this equation is that one of the additional explanatory variables is net official monetary movements, or intervention, which itself is largely determined by the exchange rate Thus, it is not surprising that intervention is correlated with the exchange rate; however, the problem of simultaneity is likely to be severe Indeed, the sign which is obtained indicates that the exchange rate causes intervention rather than the reverse, at least for quarterly data -5- where A includes all other autonomous flows, which are presumed constant. In addition, the model requires two equations from the real economy, which define the process of price level adjustment. (2.15) (p-p*)+1 - (p-p*) = k8 [(d-d*) - (y-y*)] + (p-p*) (2.16) (d-d*) = k9 (s-p + p*) + k10 (y-y*) - k11 (R-R* -p+p*) where d represents the logarithm of aggregate demand and we have followed Lafrance and Racette (1984) in adding a secular inflation term to (2.15). The latter two equations originate with Dornbusch (1976), and have been written in relative form, implying that identical specifications exist for both the domestic and the foreign economy. Equations (2.15) and (2.16) may be solved for (p-p*): (2.17) (p-p*) = k12 (p-p*)-1 + k13 (y-y*)-1 + k14 (m-m*)-1 + k15 (p-p*)-1 + k16 s-1 The detailed expressions underlying the coefficients are not needed for our purposes and therefore have been ignored. The stock-flow model solves (2.12-2.14) for x, combines the result with (2.5) and solves for s̅ . The latter is combined with the long-run form of (2.4) (where R-R* is replaced with (p-p*) to eliminate s̅ and the result is combined with (2.17). Finally, the remaining (R-R*) term is solved out using (2.3) and terms are collected to yield an expression of the following form: (2.18) s = k17 + k18 s-1 + k19 (m-m*) + k20 (m-m*)-1 + k21 (p-p*)-1 + k22 (y-y*) + k23 (y-y*)-1 + k24 (p-p*) + k25 (p-p*)-1 Driskill (1981) points out that the true reduced form exchange rate equation of Dornbusch (1976), which may be found by combining (2.17) with (2.8) to eliminate (p-p*) from the latter, is identical in form with (2.18) but has different inplications for the coefficients. In particular, although k18 + k19 + k20 + k21 = 1 for both interpretations, the Dornbusch view implies that k18 < 0, k19 > 1, k20 < 0 and k21 < 0, whereas the Driskill view implies that k18 < 1, k19 > 0, k20 0 and k21 > 0. In other words, the Driskill stock-flow model may be regarded instead as a Dornbusch sticky-price model with perfect capital mobility but with a freer dynamic structure. As noted above the principal motivation behind the sticky-price and stock-flow models is the dropping of the continuous ppp assumption. However, no explicit allowance is, made in the various models for permanent or even long-lived deviations of the real exchange rate from control. Frankel (1982) and Hooper and Morton (1982) have attempted to modify this -6- aspect of the monetary model by reintroducing the current account into the exchange rate equation. Frankel does so by assuming that wealth is an argument of the demand for money equation, and that wealth will be influenced by accumulation of net foreign assets which, by definition, is the cumulated current account. Hooper and Morton introduce the cumulated current account in two ways, arguing that it can effect expectations of the long-run real exchange rate, as well as the risk premium, a wedge which they introduce into the covered interest parity relationship. Adding the cumulated current account to any of the models described above will enable a test of these hypotheses, but discriminating between them is very difficult. To conclude this subsection, then, in the work to follow we propose to examine the empirical properties for the Canada U.S. exchange rate of five models: the non-monetary model based on ppp (2.11), the Frankel model (2.7), the Dornbusch model (2.8), the Bilson model (2.10), and the Driskill model (2.18) whih itself may be viewed as a generalization of the Dornbusch model (2.8). In each case the cumulated current account variable will also be tested The innovations which this study brings to the exchange rate literature are discussed in detail in the next subsection. 2.2 Outstanding Issues A detailed review of the empirical evidence relating to these models would serve little purpose here, a selection of articles which implement these models may be found in the bibliography, and for an excellent survey the reader is referred to Khan and Willett (1984). It is fair to conclude from this evidence that these models have empirical content, and that each of the modifications which have been described above have resulted in improved specifications. However, the overall success of these models can at best be described as mixed, with the strongest evidence against the approach provided by the work of Meese and Rogoff (1983a, b). These authors have compared the out-of-sample forecasting accuracy of (2.4), (2.7) and (2.7) plus the cumulated current account with that of the simple random walk model. The random walk model consistently provides a superior forecasting performance. Most of this literature may be criticised on the grounds that the reduced form exchange rate equations which have been employed imply assumptions regarding the underlying demand for money equations which are not supported by the data, as demonstrated in the extensive demand for money literature. The point has been made by, for example, Hakkio (1982) and by Smith and Wickens (1984). The problems are essentially four in number. First, a convenient assumption is that the domestic and foreign variables enter the exchange rate equation in differential or relative form. This assumes that the demand for money equation parameters are the same for both countries; for the sticky-price and stock- -7- flow models this assumption is extended to other structural equations as well. The importance of this assumption was first discussed in Haynes and Stone (1981), and they found that the results of Frankel (1979) were altered in important ways once the restrictions were dropped. In later work the assumption has been dropped by, for example, Frankel (1982), and Meese and Rogoff (1983b). Lafrance and Racette (1984) were unable to reject this hypothesis for the Canada-U.S. exchange rate over the 1970’s. The implications of rejecting the equal-coefficients hypothesis are far-reaching. Consider in particular the implications of rational expectations for simple monetary models such as (2.4), as discussed in Mussa (1976), Barro (1978), Driskill and Sheffrin (1981) and Hoffman and Schlagenhautf (1983). Assuming uncovered interest perity allows one to replace the interest differential with the expected change in the exchange rate, and implies that the current exchange rate depends upon the current expected future value of itself. Assuming modelconsistent or rational expectations, therefore, leads to repeated forward substitutions, until the result is obtained that the current spot exchange rate depends on the expected future path to infinity of the exogenous variables of the equation. Clearly, this argument rests on acceptance of the equal coefficients hypothesis; it also rests on the assumption of uncovered interest parity, which has been placed in doubt by the recent exchange market efficiency literature (see Longworth, 1981; Boothe, 1983, and Longworth, Boothe and Clinton, 1983). In any case, operationalizing rational expectations requires assumptions about the future paths of the exogenous variables. A common assumption, used by Frankel (1979), Driskill and Sheffrin (1981) and (after first testing the assumption) Lafrance and Racette (1984) is that (y-y*) follows a random walk while (m-m*) follows a random walk around (p-p*) which also follows a random walk. Then, for purposes of estimation the actual current values are used. Testing the equal-coefficients hypothesis, then, requires that one have available expected series for the individual variables, rather than for the differentials. The most obvious means of doing so would be to construct time series models of the series in question, as in Hoffman and Schlagenhauf (1983), but a model-consistent series may be generated from the reduced form of a structural macro model, if one is available. A second problem with the standard treatment of the relative money demand equation is related to the first, and that is that the two money demand equations are assumed to contain the same explanatory variables. Even in the simplest of money demand models, as in Goldfeld (1973) and Clinton (1973) for the U.S. and Canada, respectively, there have been differences in specification documented in the literature. If a simple common form is desired for ease of implementation of an exchange rate model, then some assessment of the empirical cost of doing so should be offered. A third problem with equation (2.3) is that the two demand for money equations are -8- assumed to be static, which runs counter to volumes of evidence. The demand for money literature has allowed explicitly for dynamics in estimation, usually by including a lagged dependent variable. This implies that lagged real balances should appear in the reduced-form equation for the exchange rate as well. The empirical importance of this point is easily tested. A fourth problem with the assumptions underlying equations (2.1) and (2.2) is that the demand for money conventionally defined has not been stable over time, particularly in the cases of the U.S. and Canada. For detailed discussion the reader is referred to Goldfeld (1976), Judd and Scadding (1982) and Simpson (1984) for the U.S.; Landy (1980), Bank of Canada (1982, 25-29) and Freedman (1983) for Canada. The main driving force behind these instabilities seems to be financial innovation, the main implication of which is that traditional measures of liquidity must be redefined. Some effort to do so has been made in both countries, leaving definitions of money whose demand equations are still subject to level adjustments, both in the past and currently. This problem may have very important implications for the monetary approach to exchange rate modelling. Suppose, for example, that (2.1) represents the Canadian demand for money, and (2.2) that for the U.S., and suppose that the latter shifts downward while the former does not. Assuming that m* is allowed to fall, it will be lower with p*, y* and R* unchanged, and the various monetary models would predict an increase in s which never materializes because money supply has changed one-for-one with demand. Notice that under a regime of money stock targeting m* would not decline but in the long run the foreign price level and hence s would change by the amount of the downward shift. However, the relationship between s and the measured money supply still will have been displaced by the amount of the shift. The implication is that money stock measures which enter exchange rate equations should be corrected for shifts to enable undistorted tests of their performance. The problem of money demand instability in this context has been alluded to by Frankel (1981, 1982), Messe and Rogoff (1983b) and Bilson (1978a, 1979b). However, explicit treatment of the problem is restricted to the latter two papers. In the first Bilson has added a time trend, and in the second a squared time trend, to the relative demand for money equation, and therefore to his final exchange rate equation as well, to account for the perceived downward drift in relative money demand. Although these variables turn out to be significant determinants of the Dollar/DM exchange rate it seems likely that more precise estimates could be obtained if one were to analyze the demand for money equations separately so as to determine the exact timing and nature of the shifts and then altering the model appropriately. To sum up this discussion, then, there seems to be some scope for improving upon the performance of monetary models of the exchange rate through more conscientious prior modelling of the demand for money. In the following sections this approach will be imple- -9- mented to the maximum extent possible in developing a monetary model of the Canada-U.S excharge rate. 2.3 The Data Details of data sources are provided in Appendix A. All mnemorics to be used in the presentation below may also be found there. For real income and the price levels GNP and the GNP deflators are used, and inflation rates are four-quarter differences of the logs of the GNP deflators. The cumulated current account for Canada cumulates the data from 1926 to avoid any level problems, the Canada/U.S. exchange rate is presumed to be independent of the overall U.S. trade balance. GNP, the GNP deflator and the current account are quarterly and seasonally adjusted at source. The exchange rate is the closing spot rate. The money supply variables are discussed in detail in Section 3. Interest rates for both countries are 90-day commercial paper rates. Money supply, interest rate and exchange rate data are monthly at source, and only the money supply is seasonally adjusted; the monthly series are converted to quarterly by selecting the third monthly observation from each quarter and collapsing. -10- 3. ADJUSTING REPORTED MONEY STOCK DATA FOR SHIFTS In this section we document the methods which were used to shift-adjust the Canadian and U.S. money stock series. In both cases the estimated adjustments are based on a particular model of the demand for money. Hence, readers who have reservations about the underlying demand for money equations naturally will object to the particular shift adjustments which have been employed. To some extent this is inevitable, but we have tried to minimize this problem by using simple models of the demand for money which, although imperfect, are widely regarded as adequate. (a) CANADA From late-1975 until the autumn of 1982 the Bank of Canada was following announced targets for M1, which comprises currency outside banks and non-government demand deposits net of private sector float. Some outright economization of transactions balances by large firms during 1976-1977 led to a small downward shift in the M1 demand equation, but otherwise M1 appeared to be a stable function of nominal spending and interest rates. However, after 1979 the demand for M1 began to shift downward quite rapidly. This was due to institutional developments on both the personal and the non-personal side (see Bank of Canada, 1982, 25-29; Freedman, 1983): (i) On the non-personal side, the cash management techniques which had been developed for large firms in the mid-1970s began to spread to much smaller firms. These developments included consolidation of geographically dispersed corporate accounts into one central account, automatic transfer of excess balances into interest-bearing notice accounts near the end of each business day, and automatic paydowns of credit lines with excess balances. The adoption of these arrangements became widespread in mid1981, when market rates of interest reached historical peaks. (ii) On the personal side, the introduction of daily interest non-chequable savings accounts in late-1979 led to some economization of transactions balances. Before this time personal notice deposits had paid interest only on the minimum monthly balance. Subsequently, banks began to offer daily interest chequable savings accounts, which offer free chequing and a market rate of interest when a minimum balance requirement is met. This has led to a further shift out it M1. These developments led to the abandonment of the M1 target late in 1982. Although no new targets have been announced to date, the Bank of Canada has begun to publish a new aggregate, M1A, which adds to M1 the non-personal notice deposits often used for cash management -11- Figure 3.1 M1 and M1A for Canada purposes, and daily interest chequable savings deposits. The two series M1 and M1A are compared in Figure 3.1. It is clear that the two series are very similar until mid-1981, Based on the work of Cockerline (1984), the EPA World Economic Model currently uses M1A rather than M1. Unfortunately, M1A has recently become very difficult to model, because daily interest chequable savings accounts have been growing very rapidly, largely at the expense of conventional savings accounts. Hence, the concept of M1A is beginning to over-compensate for the downward shifts in M1 discussed above. In addition, it is unclear where or when this process will conclude. Without some survey data it will be very difficult to estimate the proportion of daily interest chequable savings accounts which relate to transactions motives, and therefore to model M1A properly. For these reasons it has been decided to use shiftajusted M1, rather than M1A, for the purposes of the present study. Either approach should provide essentially the same amount of information, but shift-adjusting M1 will be accomplished more readily. The basic Model which is used for M1 is essentially the same as that used in Clinton (1973) for M1, and in Cockerline (1984) from M1A. The simple real partial-adjustment model of money demand, popularized by, for example, Goldfeld (1973), is assumed; this implies that the natural logarithm of real money balances is a linear function of a constant, logarithms of -12- real income and a short-term money market interest rate, and a lagged dependent variable. In preliminary estimation the data revealed a slight preference for this double-logarithmic form, as opposed to the semi-logarithmic form used by Cockerline (1984) for M1A. In addition to these variables we need to model three downward shifts 1976:01 (economization by large firms), 1979:04 (introduction of daily interest non-chequable savings accounts) and 1981:02 (widespread adoption of cash management techniques by firms and the strong growth in daily interest chequable savings accounts). The first shift is modelled as a 0-1 dummy variable in the Cockerline (1984) M1A equation, we have retained this method and have followed the same approach for the other two shifts, adding 0-1 dummy variables which become unity in 1979:04 and 1981:02, respectively. This approach may be criticised on the grounds that it is too simple and can only approximately capture the various shifts. However, using the forecast errors of the unadjusted equation to formulate better-fitting shift variables would open us to the criticism of data mining. Ideally, one would allow all parameters to shift by including slope-dummy interaction variables. However, in practice it is rarely possible to distinguish empirically between intercept and slope shifts before accumulating many observations of post-shift data. Before considering the results it should be noted that the equation does not include dummy variables to account for the effects of postal strikes. This is because the use of end-ofquarter data virtually eliminates this problem from the data. Formal definitions of all data used below are given in Appendix A. The M1 equation with shift dummy variables included was estimated using all available data following the 1967 Bank Act revision, 1968:02 - 1984:02. The following result was obtained: (3.1) log(M1/PGNP) = -1.44220 + 0.25377* log (GNP)-0.05182*log (RCP) (3.46) (4.25) (5.11) + 0.70640 * log (M1.-1/PGNP.-1)-0.02940 * DSHIF76 (9.85) (2.94) -0.01710 * DSHIF79-0.03789 * DSHIF81 (1.49) (3.07) R2C = 0.96185 SE = 0.01691 DW= 2.48 Dynamic RMSE = 1.89% Long-run elasticities: GNP = 0.8643, RCP = -0.1765 All variables BLOCK = CA 68:02-84:02 All parameters have the expected signs and all except that on the 1979 shift dummy are significant at the 0.95 level. However, the t-statistic on DSHIF79, at 1.49, is sufficiently high -13- to be indicative. The long-run elasticities are within the ranges predicted by theory. The intrasample dynamic RMS simulation error of 1.89% compares favourably with a value for M1A of 2.29% as reported in Cockerline (1984) over the 78:01-82:04 period. The Durbin-Watson statistic of 2.48 is worrisome, especially since its value is biased towards two in the presence of a lagged dependent variable. However, it is believed that this indicates not a systematic correlation in the estimated residuals but rather the inevitable element of misspecification which arises by using the simple 0-1 dummy variables to model a phenomenon which obviously is more complex. To verify that this interpretation is valid, equation (3.1) was reestimated over the 1968:02-1979:03 period, with DSHIF79 and DSHIF81 excluded, and the DurbinWatson statistic was a more respectable 2.15. The full result is given in (3.2): (3.2) log(M1/PGNP) = -1.31010 + 0.22328 * log (GNP)-0.05584 * log(RCP) (3.21) (3.65) (6.32) + 0.75503 * log (M1.-1/PGNP.-1)-0.02582 * DSHIF76 (10.22) (2.78) R2C = 0.97047 SE = 0.01481 All variables BLOCK = CA DW= 2.15. 68:02-84:02 More importantly, all of the parameters in equation (3.2) are within one standard error of the corresponding parameters in equation (3.1). This indicates that for practical purposes the shift dummies adequately capture the intended effects. Equation (3.1) was simulated dynamically over the entire sample period, 1968:02- 1984:02, twice - once as shown and once with the three shift parameters constrained to zero. The difference between the two simulated M1 series is our estimate of the total shift in the demand for M1; this total shift adjustment was then added to actual M1 to generate shift-adjusted M1. Actual M1 and adjusted M1 are shown in Figure 3.2. Notice that the two series are the same prior to 1976:01. After this time the two series diverge by an increasing amount, with shift-adjusted M1 given by the dotted line. The two series exhibit very similar movements, since the smooth geometric estimates of the shifts affect the level gradually and preserve the short-term variations of the original series. It seems likely from the magnitude of the gap between actual and shift-adjusted M1 that the two series will generate non-trivially different results when used as explanatory variables in exchange rate equations below. (b) UNITED STATES The U.S. monetary aggregate M1 now comprises currency, demand deposits and other chequable deposits, the latter category consisting mainly of negotiable order of withdrawal -14- Figure 3.2 Actual and Shift-Adjusted M1 for Canada (NOW) accounts which became available nationwide in January 1981. Thus, U.S. M1 is now similar in concept to Canadian M1A. The strong growth in NOW accounts since 1981 has been at the expense of both demand deposits and savings deposits. Hence, just as was the case for Canadian M1A, adding NOW accounts to old M1 over-compensates for the reduction in demand deposits due to switching to NOWs. Unlike Canadian M1A, however, which has never been adopted as a policy target, in the United States “new” M1 has become official. The details of the various innovations in the U.S. monetary sector during the late 1970’s and early 1980’s are well-summarised in Brayton (1983), so need not occupy us further here. In addition to the effect of NOW accounts, of course, one must consider the welldocumented downward shift in the demand for M1 which occurred in 1974-1976 (see Goldfeld, 1976, see also Judd and Scadding (1982) for a review of the literature which Goldfeld’s ‘missing money’ prompted). This shift may be handled conveniently using a 0-1 binary variable which becomes unity in 1974:01, as was done for the 1976 shift in Canadian M1 above. This gives the estimated shift a smooth and gradual interpretation. The net upward shift in M1 demand due to the nationwide introduction of NOW accounts, on the other hand, is more difficult to model. For this purpose we have chosen to draw on the work of Brayton (1983), who has constructed individual models for the three major components of M1 -15- cur- rency, demand deposits, and other chequable deposits. The simultaneous downward and upward shifts in demand deposits and other chequable deposits, respectively, have been captured using a variable which represents the fraction of the population holding NOW accounts Brayton constructs this variable, which he denotes ‘ALPHA’, as the product of three elements, the income-weighted fraction of the U.S. for which NOW accounts were available; the demand deposit weighted fraction of the population eligible to hold NOW accounts; and the proportion of those eligible to hold NOW accounts who actually open NOW accounts. The variable therefore depends essentially upon the number of states which offered NOW accounts, beginning with New Hampshire and Massachusetts in 1974, and ending with the entire nation in 1981. The other two elements of ALPHA are estimated from survey evidence. A plot of the series ALPHA is given in Figure 3.3; for additional details the reader is referred to Appendix C of Brayton (1983). Figure 3.3 Now Availability Variable (ALPHA) for U.S. The M1 equation which is estimated for the purposes of this study is essentially that of Goldfeld (1973), modified to account for the two shifts. Thus, the logorithm of real M1 balances is regressed on a constant, the logarithms of real GNP, the 90-day commercial paper rate and the rate on savings deposits; the 1974, 0-1 dymmy variable, ALPHA, and the lagged -16- dependent variable. In preliminary estimation the usefulness of a time trend and the rate on NOW accounts was tested without success. This means that the equation as estimated is without an own-rate of interest, which, since the introduction of NOW accounts, has been non-zero. However, it is unlikely that the own-interest elasticity could be distinguished empirically from the effect of ALPHA itself. The somewhat arbitrary starting point of 1968: 02 was chosen, simply because this is the same as that used for the Canadian equation. Ultimately, the exchange rate equations will be estimated from a later start date in any case. The final equation, estimated to 1984:02, is as follows. (3.3) log(M1/PGNP) = 0.32068 + 0.06874* log (GNP)-0.01654* log (RCP) (1.37) (2.76) (4.28) -0.03302* log (RSD) + 0.86846 * log (M1.-1/PGNP.-1) (1.87) (16.94) -0.01192* DSHIF74 + 0.09828 * ALPHA (2.78) (3.21) R2C = 0.93084 SE = 0.00801 DW= 2.37 Dynamic RMSE = 0.83% Long-run elasticities. GNP = 0.5227, RCP = -0.1257, RSD = -0.2511 All variables BLOCK = US 68:02-84:02 All coefficients have the expected signs and all except that on RSD are significant at the 0.95 level. The coefficient on RSD, however, with a t-statistic of 1.87, is indicative. The long-run elasticities of 0.52 for income and -0.13 and -0.25 for the two interest rates are within the ranges given by theory. The intrasample dynamic RMS simulation error of 0.83% compares favourably with that presented in Brayton (1983), which was 0.94% for the 1970:01-1981:04 period, and was based on the sum of the three component equation simulations. As was the case for the Canadian M1 equation, the Durbin-Watson statistic is a little high considering that its value is biased towards 2 in the presence of a lagged dependent variable. However, as was the case for Canada, we hesitate to correct for this by assuming an AR1 error structure because we suspect that the problem is concentrated in the period of shifting, which can only be approximately modelled. To verify that this was indeed the case, the equation was reestimated over the 1968:02-1979:04 period over which ALPHA is virtually inactive. The results were as follows: (3.4) log(M1/PGNP) = 0.14712 + 0.06695 * log (GNP)-0.01679 * log (RCP) (0.53) (2.43) (3.82) -0.03614 * log (RSD) + 0.90346 * log (M1.-1/PGNP.-1) (1.57) (13.09) -17- -0.01078 * DSHIF74 + 0.12986 * ALPHA (2.34) (0.79) R2C = 0.93241 All variables SE = 0.00616 DW= 2.19 BLOCK = US Here the Durbin-Watson statistic is a more respectable 2.19. More importantly, the estimated parameters of (3.4) are all within one standard error of those presented in (3.3). Hence, it would seem that ALPHA does an adequate if imperfect job of explaining the NOW shift. Equation (3.3) was simulated dynamically over the entire sample period twice - once as written and again with the coefficients on the 1974 shift and on ALPHA constrained to zero. The difference between the simulated values of M1 in the two cases is our estimate of the total shift in M1 demand. This estimate of the shift is then added to actual M1 to form shiftadjusted M1. These two series are shown in Figure 3.4. Notice that in this case the upward shift due to NOWs has been sufficient to more than offset both the 1974 downward shift in demand deposits and that due to switching into NOWs. Hence, the adjusted series, given by the dotted line in Figure 3.4, crosses the actual series (given by the solid line) at the beginning of 1982. The difference between the two series is sufficient to generate quite different predictions of the exchange rate within the context of a monetary model. The empirical relevance of this distinction will be examined in the next section. : Figure 3.4 Actual and Shift-Adjusted M1 for U.S. -18- 4. EXCHANGE RATE MODEL ESTIMATION AND SIMULATION RESULTS In the following five subsections the five models discussed at the end of Section 2.1 above are estimated, respectively. In each case we consider both the model as written in theory and a generalized form which imposes only those restrictions which are acceptable to the data. The intra- and post-sample simulation results of each model are summarized in Table 4.1 at the end of the Section. In Subsection 4.6 the collection of results is discussed with the objective of choosing a model which is preferred overall. All data definitions are given in Appendix A. 4.1 Non-Monetary Model In this model the exchange rate follows a partial-adjustment process towards relative ppp and reacts in the short run to the relative nominal rate of interest and the cumulated current account. The latter is a negative number over the sample period, so rather than taking its logarithm we have normalized it on nominal GNP. Estimation of this and all other models uses the sample period 1970:03-1982:04. The estimation results for this simple model are as follows:2) (4.1) log (s) = -1.65720 + 0.33943 * log (CA-PGNP/US-PGNP) + 0.66057* log(s.-1) (4.94) (4.92) (9.58) -0.06039* log(CA-RCP/US-RCP)-0.64761* (CA-CCA/CA-NGNP) (3.81) (3.30) R2C = 0.30976 (transformed) SE = 0.01743 DW= 1.84 70:03-82.04 According to equation (4.1) the speed of adjustment to ppp is relatively fast, on the order of 34% per quarter. The coefficient on the cumulated current account is statistically significant and of the expected sign a current account surplus, which increases the variable CA-CCA, induces an appreciation of the Canadian dollar relative to the U.S. dollar. The coefficient on the relative nominal interest rate is statistically significant but bears a sign opposite to that suggested by the simple monetary theory. In the latter a higher nominal interest rate relative to the foreign rate is associated with a higher relative inflation rate and therefore a rising ex2) In order to impose the theoretical constraint that the coefficients on log (CA-PGNP/US-PGNP) and log (s.-1) be a1 and (I-a1), respectively, equation (4.1) was estimated in the form (log (s) -log (s.-1)) = a0 + a1 * log ((CA-PGNP/US-PGNP)/s.-1) + . . . Because of this transformation the R2C statistic is not comparable with those of other equations to be presented below. However, because s is still treated as the endogenous variable for simulation, the RMS simulation errors are comparable. -19- change rate. Perhaps the opposite effect is found because the Canadian and U.S. inflation rates are sufficiently highly correlated so that the relative interest rate term is dominated by movements in relative real interest rates. When added to equation (4.1) the Canada-U.S. relative inflation rate was not statistically significant, thus providing some support for this interpretation. We should also note that the parameter in question may be subject to simultaneous equations bias, with endogeneity coming by way of the monetary policy reaction function. The extent of this problem will be examined in Section 5 below. Simulation results are provided in the first row of Table (4.1). The intrasample dynamic simulation, with a RMS of 2.04%, shows a degree of precision which is similar to that of the exchange rate equation of Cockerline (1984). The equation predicts the four observations of 1983 quite well, but does not capture the sudden depreciation which occurred during the first half of 1984. These results will serve as a point of reference for comparison of the monetary models below. 4.2 Frankel Model In this model the nominal exchange rate is regressed on the relative money supply, relative real income, relative nominal interest rates and the relative inflation rate. Since the equation includes the money supply as an exogenous variable, there are two versions of the model; Version 1 uses actual published money supply data, while Version 2 uses the shift-adjusted money supplies constructed in Section 3 above. In order to capture possible movements in the underlying real exchange rate we have also included the cumulated current account as an explanatory variable. The estimation results are as follows; both versions have been estimated using Cochrane-Orcutt AR-1 estimation. (4.2) Version 1 - Unadjusted Money log (s) = 2.05430 + 0.08014 * log (CA-MI/US-MI) - 0.44182* log(CA-GNP/ (1.45) (0.59) (1.48) US-GNP) - 0.04038 * log (CA-RCP/US-RCP) - 0.01751 * log (CA-PDOT/ (2.08) (1.01) US-PDOT) - 0.93125 * (CA-CCA/CA-NGNP) (1.13) R2C = 0.13142 SE = 0.01955 70:03-82:04 -20- DW = 2.38 RHO = 0.98862 (4.3) Version 2 - Adjusted Money log (s) = 0.33092 + 0.52229 * log (CA-MIADJ/US-MIADJ) - 0.55940* log (0.33) (8.75) (2.30) (CA-GNP/US-GNP) -0.04788 * log(CA-RCP/US-RCP) - 0.02578 * log (CA-PDOT/US-PDOT) (2.23) (1.48) -1.25700 * (CA-CCA/CA-NGNP) (2.29) R2C = 0.79295 SE = 0.01905 DW = 2.03 RHO = 0.56475 70:03-82:04 All of the parameters have the expected signs except that on the relative inflation rate, which should be positive according to Frankel’s model. One possible explanation for this result is that during a major part of our estimation period increases in inflation often gave rise to expectations of tighter monetary policy, and hence higher real interest rates. It is immediately apparent that Version 2 is statistically superior to Version 1. The degree of statistical significance is much higher for Version 2, particularly with respect to the relative money supply variable itself. This result lends support to the arguments in favour of shift-adjustment of money stock data presented in Section 2 above. A similar comparison is true of the simulation results of equations (4.2) and (4.3), which are presented in Table 4.1. Although the postsample simulation results for the two equations are very similar, the adjusted version clearly fits the estimation period much better. Implicit in equations (4.2) and (4.3) is the assumption that domestic and foreign variables bear coefficients which are of opposite sign but equal in absolute value. A joint test of these restrictions yields F-statistics of 11.18 and 9.78 for versions 1 and 2, respectively, which, when compared with the critical F value at the 0.95 level of 2.61, implies rejection of these restrictions by the data. The unrestricted versions of (4.2) and (4.3) were then taken as our new maintained hypotheses; we then tested the joint significance of three additional variables not found in these equations but implied by the estimated demand for money equations of Section 3, namely, the lagged real domestic and foreign money supplies, and the U.S. savings deposit interest rate. For both versions these three variables were jointly insignificant at the 0.95 level. It should also be noted that in versions of (4.2) and (4.3) in which the domestic-foreign equality constraints have been relaxed no correction for autocorrelation was necessary. Having discarded the lagged real money stocks and the U.S. savings rate as explanatory variables, then, we proceeded to impose statistically acceptable restrictions on the generalized equations so as to improve efficiency. The final estimated equations are as follows. -21- (4.4) Version 1 - Unadjusted Money log (s) = 4.77460 + 0.22170*log (CA-MI) + 0.29661 * log (US-MI) (4.27) (2.77) (4.70) -1.19870* log (CA-GNP) (7.05) + 0.74577* log (US-GNP) + 0.02891* log (US-RCP) - 0.05949* log (6.38) (3.45) (8.12) (CA-PDOT/US-PDOT) R2C = 0.97402 SE = 0.01395 DW = 2.15 70:03-82:04 (4.5) Version 2 - Adjusted Money log (s) = 5.24240 + 0.21816* log (CA-MIADJ) + 0.20969* log (US-MIADJ) (7.79) (3.02) (1.64) -1.17190 * log (CA-GNP) (13.78) + 0.70520* log (US-GNP) + 0.03869* log (US-RCP) - 0.05425* log (6.49) (5.35) (6.61) (CA-PDOT)/US-PDOT) R2C = 0.97472 SE= 0.01377 DW=2.05 70:03-82:04 Both equations retain the same set of restrictions and therefore have the same form. The two versions are now very similar in terms of statistical significance and explanatory power, but the summary statistics remain slightly better for Version 2. The most disturbing aspect of the results is that the coefficient on the U.S. money stock bears a sign opposite to that predicted by the theory. The Canadian short-term interest rate is not individually significant; possibly this is because most of the information contained in that variable may be found either in US-RCP or in CA-PDOT. Attempts to retain CA-RCP by putting it in the form of a relative interest rate (CA-RCP/US-RCP), were rejected by the data. The cumulated current account variable also dropped out of the specification. Tests of the final sets of restrictions implicit in (4.4) and (4.5) relative to the generalized forms of (4.2) and (4.3) yield F-statistics of 0.73 and 0.13, respectively, which should be compared with the critical value at the 0.95 level of 2.84 (F(3, 40)). The simulation results of these two versions of the Frankel model are given in Table 4.1. Comparing these with the previous results we note that the intrasample simulation errors have been reduced by generalizing the equations and then imposing only statistically accept- -22- able restrictions. The two versions perform similarly in this respect. In post-sample simulation, both models tend to overpredict the exchange rate; that is, to predict more depreciation relative to what actually occurred. The errors mode by the version which uses adjusted money, however, are much smaller than those of Version 1. On the other hand, both equations (4.4) and (4.5) perform much worse in post-sample simulation than do their more restrictive counterparts, (4.2) and (4.3). This may be because the post-sample simulation period is too short for the bias implicit in (4.2) and (4.3) to make itself evident; instead, the efficiency which is gained in exchange for this bias dominates the summary statistics. The same reasoning is consistent with the opposite ranking which is found for the intrasample period, which is comparatively long. Thus, the ranking of the various models explored so far and to be investigated below may depend upon how long the typical forecast period is likely to be once the model is implemented. 4.3 Dornbusch Model The Dornbusch model addresses the problem of endogeneity in the relative interest rate term of the Frankel model by solving out the variable using the relative demand for money function. Although the interpretation of the various parameters changes as a result, the outcome in terms of explanatory variables is to replace the relative interest rate variable with the ratio of the domestic to the foreign price level. The estimation results for the two versions (using unadjusted and adjusted money) of this model are given below. As was the case for the Frankel model, it was necessary to use Cochrane-Orcutt AR-1 estimation for both versions. (4.6) Version 1 - Unadjusted Money log (s) = 3.33590 + 0.12347*log (CA-MI/US-MI) - 0.63693* log (CA-GNP/ (1.16) (0.88) (2.19) US-GNP) -0.07259*log (CA-PGNP/US-PGNP) - 0.02126*1og (CA-PDOT/ (0.15) (0.96) US-PDOT) R2C = 0.05508 SE = 0.02039 DW = 2.37 RHO = 0.99202 70:03-82:04 (4.7) Version 2 - Adjusted Money log(s) = 0.55810 + 0.35895*log (CA-MIADJ/US-MIADJ) - 0.99589*log (0.51) (2.81) (4.46) (CA-GNP/US-GNP) + 0.53006*log (CA-PGNP/US-PGNP) - 0.05854*log (CA-PDOT/ (2.04) (3.32) US-PDOT) -23- R2C = 0.73315 SE = 0.01952 DW = 2.15 RHO = 0.62140 70:03-82:04 The cumulated current account variable was tested but was not statistically significant in either equation. The coefficients on the money stock and real income variables bear the expected signs; however, as with the Frankel model, the inflation rate term obtains a negative sign, which is contrary to the theory. In addition, in the Dornbusch model the relative price level is expected to bear a negative sign, a situation which is true of Version 1 but not of Version 2. More importantly, only the relative income term attains statistical significance in Version 1, while the overall degree of significance in Version 2 is much higher. There is no evidence in either equation of the Dornbusch overshooting hypothesis. The simulation results from equations (4.6) and (4.7) are given in Table 4.1. As noted previously for the Frankel model, the intrasample performance of Version 2 is much stronger than that for Version 1. However, in this case we see that Version 1 is much more precise in post-sample simulation than is Version 2. Indeed, the post-sample RMS simulation error of Version 1, at 1.36%, is the lowest observed from all of the models considered thus far. Both equation (4.6) and equation (4.7) impose linear restrictions on the parameters which are statistically unacceptable to the data. In particular, testing the implied set of restrictions that the domestic and foreign variables enter the equations with coefficients equal in absolute value but opposite in sign results in F-statistics for Versions 1 and 2 of 10.93 and 8.01, respectively, which should be compared with the critical value at the 0.95 level of 2.61. Accordingly, these generalized forms of equations (4.6) and (4.7) were adopted as our new maintained hypotheses. Subsequently, the joint significance of three additional variables suggested by our money demand equations - domestic and foreign lagged real money balances, and the U.S. interest rate on savings deposits - was tested and rejected. We then imposed sequentially a series of statistically acceptable restrictions on the two generalized forms so as to increase efficiency. The final equations are as follows: (4.8) Version 1 - Unadjusted Money log(s) = 4.12720 + 0.11634*log (CA-MI) + 0.38137*log (US-MI) (3.64) (1.59) (6.56) -1.19760*log (CA-GNP) + 0.91049*log (US-GNP) (6.41) (6.76) -0.04988*log (CA-PDOT) + 0.07723*log (US-PDOT) (5.72) (7.62) R2C= 0.97124 SE = 0.01468 DW = 2.23 -24- 70:03-82:04 (4.9) Version 2 - Adjusted Money log(s) = 6.12730 + 0.33976*log (CA-MIADJ) (7.71) (12.82) -1.54210*log(CA-GNP) + 1.18230* log(US-GNP) (13.63) (10.75) -0.03879*log (CA-PDOT) + 0.09215* log (US-PDOT) (4.50) (8.80) R2C = 0.97000 SE = 0.01500 DW = 2.25 70:03-82:04 Notice that relaxing the restrictions implicit in (4.6) and (4.7) enables the use of ordinary least squares in estimation for both versions. In neither version has the relative price level term been retained, and the U.S. money supply variable has been dropped from Version 2. As was the case with the Frankel model, the U.S. money supply bears a coefficient of sign opposite to that predicted by the theory in Version 1. The restrictions implicit in (4.8) and (4.9) relative to the generalized forms of (4.6) and (4.7) were tested; the F-ratios were 0.56 and 0.72, which should be compared with F (3, 40) = 2.84 and F (4, 40) = 2.61, respectively However, notice that when compared with the final estimates of the Frankel model, equations (4.4) and (4.5), it is evident that equations (4.8) and (4.9) omit a statistically significant explanatory variable, US-RCP. Thus, according to the Frankel model, equations (4.8) and (4.9) are misspeeified. The simulation results for the two final Dornbusch equations are given in Table 4.1. The intrasample performance of these equations is much better than that of their more restrictive counterparts. The post-sample performance of Version 1, however, is worse than for its restricted form, whereas the post-sample performance of equation (4.9) is much better than that for (4.7), and must be regarded as very good, in general. This latter result is quite surprising given the point made above with respect to misspecification relative to the Frankel model. 4.4 Bilson Model The Bilson model is similar in spirit to the non-monetary model considered above, where the exchange rate follows a partial-adjustment process around ppp; however, the Bilson model solves the relative price level out of the reduced form equation using the relative demand for money equation. As a result, the exchange rate is regressed on the relative money supply, relative real income and nominal interest rates, and the lagged exchange rate. To allow for changes in the equilibrium real exchange rate we have also included the cumulated current account. Because of the presence of the lagged dependent variable, correction for autocorrelation was unnecessary. The estimation results for two versions, the first using published money supply data and the second using shift-adjusted money, are as follows: -25- (4.10) Version 1 - Unadjusted Money log(s) = -1.13110 + 0.08739* log(CA-MI/US-MI) + 0.16536*log(CA-GNP/ (1.30) (1.45) (0.72) US-GNP) -0.04339*log (CA-RCP/US-RCP) - 0.51424* (CA-CCA/CA-NGNP) (2.38) (1.67) + 0.87421*log (s.-1) (16.56) R2C = 0.94885 SE = 0.01958 DW = 1.92 70:03-82:04 (4.11) Version 2 - Adjusted Money log(s) = -0.97586 + 0.22963* log (CA-MIADJ/US-MIADJ) - 0.00228* log (1.68) (4.75) (0.02) (CA-GNP/US-GNP) -0.06872*log (CA-RCP/US-RCP) - 0.86943* (CA-CCA/CA-NGNP) (4.13) (3.58) + 0.55619* log (s.-1) R2C = 0.96458 SE = 0.01629 DW = 2.00 70:03-82:04 As was the case for previous models, we note that the equation which uses shift-adjusted money fits the data better than does Version 1. Once again the relative interest rate takes on a negative sign, which is contary to the theory. Relative real output is not statistically significant in either equation, while the relative money stock is only significant in Version 2. We also note that the coefficient on the relative money stock lends support to the interpretation of the Bilson model, in which it is the speed of adjustment to ppp. as opposed to the Dornbusch model, where by the overshooting hypothesis this parameter is expected to exceed unity. The simulation results for this pair of equations are given in lines (4.10) and (4.11) of Table 4.1. As we have noted with previous models, a stronger intrasample performance is rendered by Version 2. Indeed, equation (4.11) has the lowest intrasample RMS percentage simulation error than any of the other basic models examined so far, 1.65%. Moreover, unlike previous models, this dominance by Version 2 over Version 1 extends to the post-sample simulation as well. The post-sample simulation of equation (4.11) is very good except for 1984:02, where it fails to predict the magnitude of the sudden depreciation which occurred at that time. Indeed, the equation predicts a 1% appreciation in 1984:02, rather than the 3% depreciation which occurred. Tests of the restrictions implicit in (4.10) and (4.11) that domestic and foreign variables -26- enter with coefficients equal in absolute value but opposite in sign yielded F-statistics of 9.09 and 2.48 for Versions 1 and 2, respectively, which should be compared with a critical value at the 0.95 level of 2.84. Thus, in the case of Version 2 we cannot reject these restrictions. However, we subsequently tested the relative inflation rate as an additional explanatory variable in (4.40) and (4.11), thus allowing the equations to distinguish between movements in real and nominal rates of interest, and found empirical support for such inclusion, especially in equation (4.11). Then, retesting the equal-but-opposite-signs hypothesis for these two expanded versions of equations (4.10) and (4.11) resulted in F-statistics of 11.19 and 4.34, respectively, which, when compared with the critical value at the 0.95 level of 2.61, imply rejection of the hypothesis for both versions. Thus, for both Versions 1 and 2 our maintained hypotheses became (4.10) and (4.11) with the domestic and foreign variables entered separately, and with the domestic and foreign inflation rates added as explanatory variables. At this point we tested the joint significance of lagged real domestic and foreign money and the U.S. savings deposit rate, as suggested by the work of Section 3, and found these variables to be jointly insignificant for both versions. We then imposed statistically acceptable vestrictions on our maintained hypotheses in sequence, beginning with the most acceptable and stopping when no additional linear restriction was acceptable to the data. The final equations which result are as follows: (4.12) Version 1 - Unadjusted Money log(s) = 2.44650 + 0.12793* log (CA-MI) + 0.27475* log (US-MI) (2.34) (1.71) (3.29) -0.74073* log (CA-GNP) + 0.47748* log (US-GNP) - 0.02990* log (3.83) (2.88) (2.02) (CA-RCP/US-RCP) -0.04321* log (CA-PDOT/US-PDOT) + 0.22950* log (s.-1) (4.06) (1.94) R2C = 0.97094 SE = 0.01476 DW = 2.17 70:03-82:04 (4.13) Version 2 - Adjusted Money log(s) = 2.77860 + 0.13655* log (CA-MIADJ) + 0.19684* log (US-MIADJ) (3.17) (1.81) (1.42) -0.66282*log (CA-GNP) + 0.35214* log (US-GNP) - 0.04558* log (4.12) (2.21) (3.12) (CA-RCP/US-RCP) -0.03025* log (CA-PDOT/US-PDOT) + 0.31540* log (s.-1) (3.23) (3.22) R2C = 0.97168 SE = 0.01457 DW = 2.21 -27- 70:03-82:04 The only variable which eventually was dropped from the equations was the cumulated current account, and it is clear that the equality restrictions were being rejected because of the money stock and real income coefficients. As was the case in previous models, the U.S. money stock enters with the wrong sign. The other parameters either take on the expected sign or may be interpreted as discussed for the Frankel model above. Notice that relaxing the unsupported restrictions has reduced the reliance of these equations on the lagged dependent variable for their explanatory power. The final set of restrictions implicit in (4.12) and (4.13), relative to generalized versions of (4.10) and (4.11) which include both domestic and foreign inflation as explanatory variables, were tested and found to be acceptable to the data; the relevant F-statistics were 2.38 and 2.22 for Versions 1 and 2, respectively, which should be compared with critical values at the 0.95 level of F (3, 39) = 2.85. The simulation results for equations (4.12) and (4.13) may be found in Table 4.1. There we note that the intrasample performance is about the same for both versions, and is comparable in precision with that of the Frankel model. Both versions have improved their intrasample performance relative to their restricted counterparts, (4.10) and (4.11). Post-sample performance is improved for Version 1 but worsened for Version 2, and both (4.12) and (4.13) simulate post-sample worse than their Frankel and Dornbusch counterparts, and also worse than the non-monetary model. Both equations (4.12) and (4.13) overpredict the exchange rate throughout the post-sample simulation period. 4.5 Driskill Model As demonstrated in Section 2 above the Driskill model is essentially a generalization of the Dornbusch model. The logarithm of the nominal exchange rate is regressed on logarithms of the relative money supply, the lagged relative money supply, relative and lagged relative real income, lagged relative price level, current and lagged relative inflation, and the lagged exchange rate Estimates of the two versions of this model over the 1970:03-1982:04 period are as follows: (4.14) Version 1 - Unadjusted Money log(s) = -0.36544 + 0.18188* log (CA-MI/US-MI) - 0.15480* log (0.56) (1.27) (1.07) (CA-MI.-1/US-MI.-1) -0.51423* log (CA-GNP/US-GNP) + 0.14054* log (CA-GNP.-1/ (1.76) (0.43) US-GNP.-1) -0.01090* log (CA-PDOT/US-PDOT) - 0.01246* log (CA-PDOT.-1/ (0.62) (0.59) US-PDOT.-1) -28- + 0.41085* log (CA-PGNP.-1/US. PGNP.-1) + 0.70551* log (s.-1) (2.72) (6.98) R2C = 0.95129 SE = 0.01911 DW = 2.02 70:03-82:04 (4.15) Version 2 - Adjusted Money log(s) = 0.48463 + 0.23064* log (CA-MIADJ/US-MIADJ) - 0.01464* log (0.83) (1.53) (0.09) (CA-MIADJ.-1/US-MIADJ.-1) + 0.45864* log (CA-GNP/US-GNP) - 0.12372* log (CA-GNP.-1/ (1.70) (0.38) US-GNP.-1) -0.02924* log (CA-PDOT/US-PDOT) - 0.01162* log (CA-PDOT.-1/ (1.66) (0.59) US-PDOT.-1) + 0.27119* log (CA-PGNP.-1/US-PGNP.-1) + 0.50299* log (s.-1) (1.80) (4.09) R2C = 0.95706 SE = 0.01794 DW = 1.98 70:03-82:04 The cumulated current account variable was tested in preliminary estimation and was not statistically significant in either version. With the exception of the lagged relative real income variable the two sign patterns are the same, and these results, although in many cases not statistically significant, lend support to the Driskill interpretation of this reduced form, as opposed to that of Dornbusch. Interestingly, the theoretical implication that the coefficients on relative money, lagged relative money, lagged relative prices and the lagged exchange rate sum to unity, although not imposed in estimation, would be easily accepted by the data, for the point estimates sum to 1.14 and 0.99 for versions 1 and 2, respectively. There is no evidence of autocorrelation. The simulation results for equations (4.14) and (4.15) are given in Table 4.1. As we have noted for the Frankel and Dornbusch models, Version 2 simulates better than Version 1 over the estimation period, but the reverse is true in the post-sample period. However, we would not expect (4.14) and (4.15) to simulate well post-sample, given the inclusion of several statistically insignificant variables. The restrictions implicit in (4.14) and (4.15) that the domestic and foreign variables enter the equations with coefficients equal in absolute value but opposite in sign were tested and rejected by the data; the calculated F-ratios were 4.81 and 3.57 for Versions 1 and 2, respectively, which should be compared with a critical value at the 0.95 level of 2.30. Since the equations already include the lagged money stocks and lagged price levels we tested only -29- the U.S. savings deposit rate in addition, and it was found to add no significant explanatory power to either equation. Thus, our maintained hypotheses became (4.14) and (4.15) with the domestic and foreign variables entered separately. On these equations we imposed in sequence restrictions which individually and jointly were acceptable to the data at the 0.95 level. The final equations which emerged from this process are as follows: (4.16) Version 1 - Unadjusted Money log(s) = 4.52690 + 0.15257* log (CA-MI/US-MI) + 0.51504* log (US-MI.-1) (4.08) (2.16) (11.97) -1.26300* log (CA-GNP) + 0.92880* log (US-GNP) (6.94) (7.25) -0.03218* log (CA-PDOT) + 0.05681* log (US-PDOT) (2.41) (4.18) -0.02184* log (CA-PDOT.-1/US-PDOT.-1) (1.67) R2C = 0.97405 SE = 0.01395 DW = 2.15 70:03-82:04 (4.17) Version 2 - Adjusted Money log(s) = 6.09790 + 0.21631* log (CA-MIADJ/US-MIADJ) + 0.42716* log (8.10) (2.58) (6.11) (US-MIADJ.-1) -1.45990* log (CA-GNP) + 1.05450* log (US-GNP) (13.22) (8.61) + 0.04801 * log (US-PDOT) - 0.03676 * log (CA-PDOT.-1/ (4.19) (4.32) US-PDOT.-1) R2C = 0.97352 SE = 0.01409 DW = 2.09 70:03-82:04 The only difference between the two final forms is that CA-PDOT has been dropped from Version 2 but not from Version 1. Lagged real income, the price levels and the lagged exchange rate have all been dropped from the equations. Other parameters generally bear the expected signs, except those on the inflation rate variables. One possible explanation for this result, as noted above for another equation, is that over a substantial portion of our sample period an increase in U.S. inflation prompted expectations of tighter U.S. monetary policy, higher real interest rates and an appreciating U.S. dollar. The final sets of restrictions implicit in (4.16) and (4.17) relative to generalized versions of (4.14) and (4.15) were tested, and resulted in F-ratios of 0.36 and 0.46 for Versions 1 and 2, respectively, which should be compared -30- with critical values at the 0.95 level of F (8, 34) = 2.23 and F (9, 34) = 2.19, respectively. There is no evidence of first-order autocorrelation in either equation. The simulation results of equations (4.16) and (4.17) are given in Table 4.1. As with previous models the intrasample results have been improved somewhat by relaxing the restrictions which are rejected by the data and eliminating variables which are not statistically significant. In post-sample simulation equation (4.16) performs worse than its more restricted counterpart, whereas equation (4.17) performs much better. Indeed, equation (4.17) records the best post-sample simulation results of any of the models tested. 4.6 Discussion Several new insights have emerged from the analysis of this section. First, it is clear that monetary models which use shift-adjusted money supply data fit the sample period better and have better statistical properties overall than do models which use unadjusted money supply data. This was true of all four models which use the money supply as an explanatory variable. Second, we have seen that the residual autocorrelation found in most exchange rate models may be eliminated by relaxing the constraint that domestic and foreign variables enter the reduced form with coefficients which are equal in absolute value and opposite in sign, a constraint which was rejected by the data for every model considered. The interpretation of these rejections depends on which theoretical model is the maintained hypothesis; however, which element of the underlying structure is the cause of this rejection is immaterial from an econometric standpoint. In each case linear restrictions were rejected in favour of a more general form whose data matrix contained the same information and whose structure was equally consistent with the underlying theory. Third, we have noted that for many of the equations examined relaxing unacceptable restrictions and imposing acceptable ones improves the intrasample fit but worsens the postsample predictive performance. This may be an indication of one or both of the following factors: (a) the post-sample period may be too short for the bias caused by inappropriate restrictions to outweigh the reduction in variance which results, or (b) the equations for which this problem arises are misspecified in any case, in the sense that they exclude statistically significant variables, so that although the restrictions implicit in these equations are acceptable to the data, relative to a misspecified alternative, the end result is that the final equations are simply more precisely biased than their original counterparts, so their predictive ability suffers. The estimation and simulation results of the Driskill model, which does not seen to be affected by this problem and which includes regressors omitted by the other models, lend support to hypothesis (b). -31- The superior performance of the Driskill model should not be surprising, since this model begins with the most general maintained hypothesis of those considered. The final equation says that the exchange rate is determined by the relative shift-adjusted money supply, the lagged U.S. shift-adjusted money supply, domestic and foreign real income, foreign inflation, and lagged relative inflation. Thus, equation (4.17) is essentially a Dornbusch exchange rate model with extra terms which allow a freer determination of the equations dynamics. The signs of the coefficients on the inflation rate variables are bothersome but may be rationalized. Equation (4.17) is based on a well-considered theoretical structure, has strong statistical properties, achieves the (apparent) minimum intrasample RMS error of 1.3%, and records the strongest post-sample simulation performance. For these reasons, the Driskill-generalizedadjusted (hence forth, DGA) model (4.17) is our preferred working model. One of the questions which has been posed in the literature is whether any exchange rate model can outperform the simple random walk model. The latter is given as follows: (4.18) log(s) = log (s.-1) Table 4.1 gives the results of both dynamic and static simulations of model (4.18). The most relevant comparison is between (4.17) and the static simulation of (4.18), since the DGA Figure 4.1 Intrasample and Post-sample Simulation of DGA Model -32- model does not include a lagged dependent variable. There we see that the DGA model outperforms the static random walk model both within and post-sample. However, it is interesting to note that in post-sample simulation the static random walk model outperforms 15 out of 17 models in terms of percentage RMS error. A plot of the intrasample and post-sample simulation of the DGA model (4.17) is provided in Figure 4.1. Actual data are represented by the solid line, the intrasample simulation by the dotted line, and the post-sample simulation by the dashed line. There is a gap between the intrasample and post-sample simulations at 1982:04. -33- Table 4.1 Dynamic Simulation Errors* (Percent) Intra-sample Equation MEAN RMS Post-sample 83:01 83:02 83:03 83:04 84:01 84:02 MEAN RMS (4.1) Non-monetary 0.07 2.04 -0.30 0.59 0.21 -1.15 -4.08 -7.53 -2.04 3.54 (4.2) Frankel-basic-unadjusted 2.56 3.65 -0.08 0.59 0.34 0.73 -1.51 -4.38 -0.72 1.93 (4.3) Frankel-basic-adjusted 0.03 2.17 0.05 0.78 1.15 1.74 -0.63 -4.34 -0.21 2.01 (4.4) Frankel-general-unadjusted 0.01 1.29 3.16 4.68 4.91 8.15 5.78 3.74 5.07 5.32 (4.5) Frankel-general-adjusted 0.01 1.27 1.78 2.87 2.75 5.40 2.93 0.82 2.76 3.09 (4.6) Dornbusch-basic-unadjusted 3.82 5.45 -0.31 (4.7) Dornbusch-basic-adjusted (4.8) (4.9) 0.93 1.02 1.79 0.01 -2.45 0.17 1.36 -0.02 2.36 0.61 2.45 3.96 6.57 4.35 1.99 3.32 3.83 Dornbusch-general-unadjusted 0.01 1.36 3.03 4.33 4.63 7.52 5.61 3.88 4.83 5.04 Dornbusch-general-adjusted 0.01 1.41 1.01 1.52 1.47 3.23 1.72 -0.32 1.44 1.78 (4.10) Bilson-basic-unadjusted 0.73 2.76 -1.94 -2.54 -4.24 -6.47 -10.34 -15.00 -6.75 8.18 (4.11) Bilson-basic-adjusted 0.02 1.65 0.21 1.37 1.25 0.35 -2.30 -6.12 -0.87 2.78 (4.12) Bilson-general-unadjusted (4.13) Bilson-general-adjusted 0.00 1.31 3.58 6.11 6.71 9.08 7.19 4.77 6.24 6.48 -0.01 1.29 2.61 4.80 4.93 6.01 3.90 1.11 3.90 4.22 0.01 2.21 0.34 2.32 2.89 3.53 3.19 1.14 2.24 2.51 -0.02 1.81 1.57 3.84 4.99 7.14 6.68 4.55 4.79 5.14 (4.16) Driskill-general-unadjusted 0.01 1.27 1.63 4.03 4.62 6.60 6.47 3.98 4.56 4.86 (4.17) Driskill-general-adjusted 0.01 1.30 -1.08 1.03 0.58 0.26 1.82 -1 22 0.23 1.11 (4.18) Random Walk-dynamic -3.51 8.84 -0.67 0.12 -0.28 -1.25 -3.71 -6.70 -2.08 3.18 (4.18) Random Walk-static -0.32 2.09 -0.67 0.80 -0.40 -0.98 -2.48 -3.11 -1.14 1.73 (4.14) Driskill-basic-unadjusted (4.15) Driskill-basic-adjusted * Intrasample refers to 1970:03-1982:04 sample period; all errors are (estimated-actual) as a percentage of actual. Notice that only for the equations which include a lagged dependent variable will a dynamic simulation differ from a static one; this is the case for (4.1),(4.10)-(4.13) and (4.14)-(4.15). -34- 5. FURTHER TESTING OF THE DGA MODEL 5.1 Functional Form As noted in Section 3, when estimating the demand for money equations there was a slight preference for a double-logarithmic (DL) functional form over a semi-logarithmic (SL) one on the basis of fit. As a result, interest rates were also used in logarithmic form when estimating the exchange rate equations; by extension, and since one would like to decompose the effect of interest rates into real and nominal parts for some models, inflation rates have been treated similarly. Most theoretical work, on the other hand, has assumed an SL form, with both interest rates and inflation rates entered in level form. Moreover, although this has not been a problem during the period under investigation, it is possible that in future the inflation rates will become negative, in which case taking a logarithm will require adding a positive constant to the original series. The purpose of this subsection is to examine the importance of this issue for the special case of the DGA model (4.17). The DGA model was reestimated after changing the inflation rates from logarithms to levels, with the following result: (5.1) log(s) = 6.51640 + 0.20172* log (CA-MIADJ/US-MIADJ) + 0.45423* log (8.80) (2.26) (6.22) (US-MIADJ.-1) -1.47870* log (CA-GNP) + 1.01820* log (US-GNP) (12.98) (7.63) + 0.00812* US-PDOT - 0.00458* (CA-PDOT.-1/US-PDOT.-1) (4.43) (3.90) R2C = 0.97092 SE = 0.01476 DW = 1.97 70:03-82:04 The standard error of equation (5.1) is only slightly (about 5%) higher than that for equation (4.17), and there is no substantive difference between the two equations. The parameter estimates other than those on the inflation rates are all within one standard error of their counterparts in equation (4.17). The similarity between the two sets of results may be underlined by considering the simulation results of equation (5.1), which are as follows: Simulation Results for Equation (5.1), percent (a) Intrasample (b) Post-sample mean 0.01 RMS 1.36 83:01 -0.47 83:02 1.83 -35- 83:03 1.62 83:04 1.03 84:01 1.31 84:02 -1.20 mean 0.69 RMS 1.32 Not surprisingly, given that its standard error is slightly higher, the SL equation has a slightly higher intrasample RMS simulation error. The same comparison is valid for post-sample simulation, but the precision of these predictions remains high. Thus, although this is a very limited test of the DL/SL issue, it seems unlikely that the choice of functional form has seriously affected the inferences made in previous sections. 5.2 Endogenity and Expectations We noted in Section 2 that in some of the models considered there exists the potential for simultaneous equations bias. Indeed, given that the exchange rate is determined within a two-country structural macro model, strictly speaking all of the right-hand variables should be treated as endogenous; if the small-country assumption applies to one of the economies, then the foreign variables may be treated as exogenous. The degree of endogeneity depends also on the varying lag lengths in the various relationships. For example, if the exchange rate is given a non-zero weight in the reaction function of the monetary authorities, we would expect the simultaneity between the exchange rate and the short-term rate of interest to be higher than, say, that between the exchange rate and real income or the rate of inflation. In addition, to the extent that the nominal rate of interest is manipulated in order to affect the exchange rate, the money supply also becomes increasingly endogenous. From an econometric viewpoint the problem of endogeneity means that shocks which impact on the economy and affect the exchange rate also affect to varying degrees the variables on the right-hand side of the exchange rate equation. This means that these explanatory variables may be correlated with the error term in the equation, which gives rise to simultaneous equations bias. The correct means of addressing this problem in estimation is to estimate a full structural macromodel using a full information method. Such an approach is beyond the scope of the present study. A more practical means in a single-equation context is to use a two-step procedure, whereby the explanatory variables are regressed on a set of instruments and the fitted values used in the estimation of the exchange rate equation. This is the approach which is used below. There is a second reason for considering such an approach. Most of the models considered above, including the DGA model, have used some assumptions about expectations in their -36- derivation. In particular, most of the models used expected versions of the relative demand for money equation to derive the expected price level and therefore the expected exchange rate as given by expected ppp, toward which the actual exchange rate is assumed to adjust. This means that the reduced form exchange rate equations in fact should contain expected values of the money supply, output and the inflation rate; actual values have been used by assuming that there are equal coefficients across countries and that the relative variables follow random walks. However, in the DGA model we were forced by the data to relax the equal coefficients restriction, and it is much less likely that the individual variables follow random walks. The correct means of estimation in this situation, then, is to estimate the entire structure with the expectations derived by solving the model; that is, rational or model-consistent expectations. Since this requires estimating a multi-equation model, this, too, is beyond the scope of this study. However, once in a single-equation context, we note that from an econometric viewpoint the endogeneity problem discussed above and the expectations problem are essentially the same. To account for the expectations problem we would like to replace the actual values with unbiased expected values in estimation; that is, the expected values would differ from the actual values by a random error. The econometric problem with using actual values is that the included random error may, in fact, be correlated with the residuals of the exchange rate equation, which is a classical errors-in-variables situation. This problem is therefore not distinguishable from the endogeneity problem in a single-equation context. To account for both problems, therefore, we have conducted the two-step procedure described above. A convenient means of performing the first step is to construct a time series model for each variable. Indeed, one can approach the efficiency of model-consistent expectations by estimating the exchange rate equation and the time series equations jointly, subject to the cross-equation constraints implied by rational expectations, such as in Hoffman and Schlagenhauf (1983). However, this procedure requires computing capabilities beyond those which were available when this study was completed. Using the two-step procedure means that the variance-covariance matrix which is estimated at the second stage are incorrect, but in any case our present concern is with the problem of bias. In order to reestimate the DGA model, we require estimated series for shift-adjusted Canadian and U.S. money, Canadian and U.S. GNP, and the U.S. inflation rate. The Canadian inflation rate enters the equation in lagged form and therefore is treated as exogenous. Univariate time series models for each variable were fitted over the 1970:03-1984:02 sample period for orders up to eight, and the Akaike final prediction error criterion was used to choose the final order for each series. The results are presented in Table 5.1. It is evident that the time series equations fit the data very well. Only the U.S. shift-adjusted money stock may -37- Table 5.1 Time Series Models of The Explanatory Variables* (1970:05-1982:04) CA-MIADJ CA-GNP US-MIADJ US-GNP US-PDOT Constant 0.15490 (2.84) 0.31578 (2.43) 0.07657 (3.09) 0.07500 (0.78) 0.70365 (2.27) LAG 1 0.73131 (5.64) 1.30080 (10.22) 0.98750 (232.30) 1.33320 (10.26) 1.42380 (10.29) LAG 2 0.25636 (1.99) -0.32734 (2.65) - - -0.34297 (2.65) -0.36943 (1.68) LAG 3 - - - - - - - - 0.19584 (1.04) LAG 4 - - - - - - - - -0.85260 (4.61) LAG 5 - - - - - - - - 0.75068 (3.47) LAG 6 - - - - - - - - -0.25838 (1.86) LAG 7 - - - - - - - - - - LAG 8 - - - - - - - - - - R2C 0.99836 0.99333 0.99898 0.98998 0.94141 SE 0.01759 0.01032 0.00841 0.01123 0.49592 DW 2.09 1.98 1.88 2.08 2.02 RMS(%) 1.71 1.00 0.82 1.10 0.46 * All except US-PDOT were estimated in logarithms. -38- be approximated by a random walk. The fitted values from these equations were used to replace their counterparts in the DGA equation, which was reestimated and resimulated with the following results; (5.2) log(s) = 6.02370 + 0.17462* log (CA-MIADJE/US-MIADJE) + 0.50327 * log (6.25) (1.68) (6.21) (US-MIADJ.-1) -1.32540*log (CA-GNPE) + 0.81442*log (US-GNPE) (10.29) (6.29) + 0.02702 + log (US-PDOTE) - 0.03386*log (CA-PDOT.-1/ (2.11) (3.62) US-PDOT.-1) R2C = 0.96095 SE = 0.01711 DW = 1.86 70:03-82:04 Simulation Results, Percent (a) Intrasample - mean 0.01 RMS 1.59 83:01 1.98 83:02 2.21 83:03 2.91 83:04 1.84 84:01 1.68 84:02 0.04 mean 1.78 RMS 1.98 (b) Post-sample As anticipated, there has been some loss of precision in the parameter estimates, but only that on the relative money supply has become statistically insignificant at the 0.95 level (it remains significant at the 0.90 level). The standard error of the equation has been increased by about 20%. Nevertheless, five of seven coefficient estimates lie within one standard error of their counterparts in equation (4.17) and the remaining two are within two standard errors. Thus, the bias contained in equation (4.17) due to errors-in-variables or endogeneity must be regarded as statistically small. The precision of the simulation has been reduced is well, but is still very good in an absolute sense. It is likely that a more precise and more appealing test of these issues could be performed once such an equation has been placed within the context of a complete structural model. Nevertheless, the above results are suggestive that the DGA model will not be severely affected -39- by these considerations. It bears repeating, however, that this conclusion would come less easily for other exchange rate models which have short-term interest rates as explanatory variables. For such models to succeed it will be necessary to estimate jointly at least the exchange rate equation and the monetary policy reaction function. -40- 6. Conclusion The purpose of this study was to examine the applicability of the class of monetary models of exchange rate determination to the Canada-U.S. exchange rate. As there had been previous studies purporting to do the same, the principal motivation of this study was the observation that previous researchers had failed to take account of well-established results of the demand for money literature when estimating monetary exchange rate models. The most important of these was the failure to recognize the problem of instability in the demand for money equation. The first step in the analysis, then, was to shift-adjust the money supply data; this proved to be straightforward, as conventional partial-adjustment models of the demand for money were found to explain the 1968-1984 data adequately once known structural shifts had been accounted for. Both measures of money were then used in a full comparison of four different monetary models. The conclusions regarding exchange rate models which follow from this study may be stated very simply. First, the hypothesis that monetary models of the exchange rate would have more attractive statistical properties if shift-adjusted money measures were used in place of published data was confirmed in every case. Second, the hypothesis of equal coefficients across countries was rejected in every case; relaxing this restriction was found to eliminate the problem of first-order serial correlation when it arose in some of the restricted models. Third, the cumulated current account made statistically significant contributions to the equations only when the unacceptable equal-coefficients restrictions were imposed; the variable does not appear in any monetary equation which is acceptable to the data. Fourth, using these results it was possible to develop an exchange rate equation based on the Driskill (1981) stock-flow hypothesis with secular inflation which explains the sample period very well and which predicts the six post-sample observations very accurately. Furthermore, this performance was found to dominate that of the random walk model, to be independent of the double-log/ semilog functional form issue, and to be only slightly perturbed by accounting for endogeneity in the explanatory variables. Taken together, these results indicate that the monetary approach to exchange rate determination has empirical content and may be usefully applied to the Cnada-U.S. exchange rate. There are, of course, a number of potentially fruitful avenues for further research. The concept of shift-adjusted money might usefully be applied elsewhere, in monetary policy reaction functions, for example. In addition, a number of interesting implications of the various theoretical models were not tested here and, more importantly perhaps, the implications of rational expectations have been tested in only the most rudimentary fashion. Both of these limitations were imposed by software constraints. In any case, the most satisfactory means of -41- invoking the rational expectations hypothesis is within the context of a complete structural macro model. Indeed, it may be counter-productive to invoke rational expectations on a singleequation basis for one equation which will then form part of a structural model, for the remainder of which expectations continue to be based on perfect foresight or adaptive behavioural assumptions. It is hoped that this study has provided a theoretically attractive working model of the Canada-U.S. exchange rate which will be useful both before and after the introduction of rational expectations into the EPA World Economic Model. -42- Appendix A - Data Definitions and Sources In this appendix CANSIM refers to Statistics Canada main database, and BCR refers to the Bank of Canada Review. S : Closing spot price of U.S. dollar in Canadian dollars, end of month data converted to quarterly by choosing every third end of quarter observation. (CANSIM B3414, BCR Table 65) CA-M1 : Canadian MI, average of Wednesday monthly seasonally adjusted data converted to quarterly by choosing every third end of quarter observation. (CANSIM B1627, BCR Table 9) CA-MIADJ : Shift-adjusted quarterly Canadian MI CA-GNP : Canadian quarterly real GNE, seasonally adjusted at annual rates (CANSIM D40593, BCR Table 53) CA-PGNP : Canadian quarterly GNE deflator, seasonally adjusted (CANSIM D40625, BCR Table 54) CA-NGNP : CA-GNP * CA-PGNP/100 CA-RCP : Canadian 90-day prime corporate paper rate, last Wednesday of each month, converted to quarterly data by choosing every third end of quarter observation. (CANSIM B14017, BCR Table 20) CA-PDOT : 100*[log (CA-PGNP)-log (CA-PGNP.-4)] CA-CCA : Cumulated Canadian current account balance; cumulated from 1926 using annual data, then quarterly seasonally adjusted data from 1950. (quarterly current account series CANSIM D60555, BCR Table 69) CA-DSHIF76 : = 0 until 1975 : 04, = 1 afterward CA-DSHIF79 : = 0 until 1979 : 03, = 1 afterward CA-DSHIF81 : = 0 until 1981 : 01, = 1 afterward US-MI : U.S. MI, seasonally adjusted monthly average of weekly data, converted to quarterly by choosing every third end of quarter observation. (taken from EPA database) US-MIADJ : Shift-adjusted quarterly U.S. MI. US-GNP : U.S. quarterly seasonally adjusted GNP (taken from EPA database) US-PGNP : U.S. quarterly seasonally adjusted GNP deflator (taken from EPA database) -43- US-RCP : U.S. 90-day commercial paper rate, adjusted to annual yield, last Wednesday of the month, converted to quarterly by taking every third end of quarter observation. (CANSIM B54412, BCR Table 20) US-PDOT : 100* [log (US-PGNP)-log (US-PGNP.-4)] US-RSD : Quarterly U.S. Maximum rate paid on savings deposits, taken from DRI database (mnemonic RMSD) US-DSHIF74 : = 0 until 1973 : 04, = 1 afterward US-ALPHA : Availability variable for Now accounts, based on Brayton (1983) (taken from EPA database). -44- Appendix B - Bibliography Ahtiala, P. (1984), “A Synthesis of the Macroeconomic Approaches to Exchange Rate Determination”, European Economic Review 24,117-36. Backus, D. (1982), “Empirical Models of the Exchange Rate: Separating the Wheat from the Chaff”, Discussion Paper #463, Institute for Economic Research, Queen’s University. Bank of Canada (1982), Annual Report of the Governor to the Minister of Finance, Ottawa: Bank of Canada. Barro, R.J. (1978), “A Stochastic Equilibrium Model of an Open Economy Under Flexible Exchange Rates”, Quarterly Journal of Economics 92,149-64. Bilson, J.F.O. (1978a), “The Monetary Approach to the Exchange Rate - Some Empirical Evidence”, IMF Staff Papers 25, 48-75. Bilson, J.F.O. (1978b), “The Current Experience with Floating Exchange Rates: An Appraisal of the Monetary Approach”, American Economic Review, 68,392-7. Bilson, J.F.O. (1979a), “Recent Developments in Monetary Models of Exchange Rate Determination”, IMF Staff Papers 26, 201-23. Bilson, J.F.O. (1979b), “The Deutsche Mark/Dollar Rate: A Monetary Analysis”, Carnegie-Rochester Con- ference Series on Public Policy 11, 59-101. Bisignano, J. and K. Hoover (1982), “Some Suggested Improvements to a Simple Portfolio Balance Model of Exchange Rate Determination with Special Reference to the U.S. Dollar/Canadian Dollar Rate”, Weltwirtschaftliches Archiv 118,19-38. Boothe, P. (1983), “Speculative Profit Opportunities in the Canadian Foreign Exchange Market, 1974-78”, Canadian Journal of Economics 16, 603-11. Boothe, P., Clinton, K., Coté , A. and D. Longworth (1984), “International Asset Subsutitutability”, memo, Bank of Canada. Branson, W.H., Halttunen, H. and P. Masson (1977), “Exchange Rates in the Short Run: The DollarDeutschemark Rate”, European Economic Review 10, 303-24. Brayton, F. (1983), “A Model of the U.S. Monetary Sector”, Discussion Paper No. 26, Economic Research Institute, Economic Planning Agency, Tokyo, Japan. Caves, D.W. and E.L. Feige (1980), “Efficient Foreign Exchange Markets and the Monetary Approach to Exchange Rate Determination”, American Economic Review 70, 120-34. Clinton, K. (1973), “The Demand for money in Canada, 1955-70: Some Single Equation Estimates and Stability Tests”, Canadian Journal of Economics 6, 53-61. Cockerline, J. (1984), “A Model of the Canadian Financial and Capital Account Sectors”, Discussion Paper No. 29, Economic Research Institute, Economic Planning Agency, Tokyo, Japan. Dornbusch, R. (1976), “Expectations and Exchange Rate Dynamics”, Journal of Political Economy 84, 1161-76. -45- Dornbusch, R. (1980), “Exchange Rate Economics: Where Do We Stand?”, Brookings Papers on Economic Activity 1, 143-205. Dornbusch, R. and S. Fischer (1980), “Exchange Rates and the Current Account”, American Economic Review 71, 1068-74. Driskill, R.A. (1981), “Exchange Rate Dynamics: An Empirical Investigation”, Journal of Political Economy 89, 357-71. Driskill, R.A. and S.M. Sheffrin (1981), “On the Mark: Comment”, American Economic Review 71, 1068-74. Frankel, J.A. (1979), “On the Mark: A Theory of Floating Exchange Rates Based on Real Interest Differentials”, American Economic Review 69, 610-22. Frankel, J.A. (1981), “On the Mark: Reply”, American Economic Review 71, 1075-82. Frankel, J.A. (1982), “The Mystery of the Multiplying Marks: A Modification of the Monetary Model”, Review of Economics and Statistics 64, 515-19. Freedman, C. (1983), “Financial Innovation in Canada: Causes and Consequences”, American Economic Review 73, 101-6. Frenkel, J. (1976), “A Monetary Approach to the Exchange Rate: Doctrinal Aspects and Empirical Evidence”, Scandinavian Journal of Economics 78, 255-76. Frenkel, J. (1981a), “The Collapse of Purchasing Power Parities During the 1970s”, European Economic Review 16, 145-65. Frenkel, J. (1981b), “Flexible Exchange Rates, Prices and the Role of “News”: Lessons form the 1970s”, Journal of Political Economy 89, 665-705. Fukao, M. (1983), “The Risk Premium in the Foreign Exchange Market”, National konomie Journal of Economics Suppl. 3, 99-125. Girton, L. and D. Roper (1977), “A Monetary Model of Exchange Market Pressure Applied to the Postwar Canadian Experience”, American Economic Review 67, 537-48. Girton, L. and D. Roper (1981), “Theory and Implications of Currency Substitution”, Journal of Money, Credit and Banking 13,12-30. Goldfeld, S.M. (1973), “The Demand for Money Revisited”, Brookings Papers on Economic Activity 3, 577-638. Goldfeld, S.M. (1976), “The Case of the Missing Money”, Brookings Papers on Economic Activity 3, 683-730. Haas, R.D. and W.E. Alexander (1979), “A Model of Exchange Rates and Capital Flows”, Journal of Money, Credit and Banking 11,467-82. Hakkio, C.S. (1982), “Exchange Rate Determination and the Demand for Money”, Review of Economics and Statistics 64, 681-6. Haynes, S.E. and J.A. Stone (1981), “On the Mark: Comment”, American Economic Review 71, 1060-67. Hoffman, D.L. and D.E. Schlagenhauf (1983), “Rational Expectations and Monetary Models of Exchange -46- Rate Determination: An Empirical Examination”, Journal of Monetary Economics 11, 247-60. Hooper, P. and J. Morton (1982), “Fluctuations in the Dollar: A Model of Nominal and Real Exchange Rate Determination”, Journal of International Money and Finance 1, 39-56. Judd, J.P. and J.L. Scadding (1982), “The Search for a Stable Money Demand Function: A Survey of the Post1973 Literature”, Journal of Economic Literature 20, 993-1023. Khan,W. and TD.Willett (1984), “The Monetary Approach to Exchange Rates: A Review of Recent Empirical Studies”, Kredit und Kapital 17,199-222. Lafrance, R. and D. Racette (1984), “The Canadian-U.S. Dollar Exchange Rate: A Test of Alternative Models for the Seventies”, Journal of International Money and Finance (forthcoming). Landy, L. (1980), “Financial Innovation in Canada”, Federal Reserve Bank of New York Review 5, Autumn, 1-11. Longworth, D. (1981), “Testing the Efficiency of the Canadian-U.S. Exchange Market under the assumption of No Risk Premium”, Journal of Finance 36,43-49. Longworth, D., Boothe, P. and K. Clinton (1983), “A Study of the Efficiency of Foreign Exchange Markets”, Bank of Canada. Makin, J. (1984), “Exchange Rates and Taxes”, Working Paper #1350, National Bureau of Economic Researeh. Makin, J. and R.D. Sauer (1984), “Exchange Rate Determination With Systematic and Unsystematic Regime Change: Evidence from the YEN/DOLLAR Rate, working paper #1406, National Bureau of Economic Research. Meese, R.S. and K. Rogoff (1983a), “Empirical Exchange Rate Models of the Seventies-Do they Fit out of Sample?”, Journal of International Economics 14, 3-24. Meese, R.A. and K. Rogoff (1983b), “The Out-of-Sample Failure of Empirical Exchange Rate Models: Sampling Error or Misspecification?”, in J.A. Frenkel (Ed), Exchange Rates and International Macroecono- mics, Chapter 3, National Bureau of Economic Research, Chicago: University of Chicago Press. Mussa, M. (1976), “The Exchange Rate, The Balance of Payments and Monetary and Fiscal Policy Under a Regime of Controlled Floating”, Scandinavian Journal of Economics 78, 229-48. Mussa, M. (1979), “Empirical Regularities in the Behaviour of Exchange Rates and Theories of the Foreign Exchange Market”, Carnegie-Rochester Conference Series on Public Policy 11, 9-58. Mussa, M. (1982), “A Model of Exchange Rate Dynamics”, Journal of Political Economy 90, 74-104. Niehans, J. (1977), “Exchange Rate Dynamics with Stock/Flow Interaction”, Journal of Political Economy 85,1245-58. Officer, L.H. (1976), “The Purchasing Power Parity Theory of Exchange Rates: A Review Article”, IMF Staff Papers 23, 1-60. Officer, L.H. (1980), “Effective Exchange Rates and Price Ratios over the Long-Run: A Test of the Purchasing Power Parity Theory”, Canadian Journal of Economics 13, 206-30. -47- Rodriguez, C.A. (1980), “The Role of Trade Flows in Exchange Rate Determination: A Rational Expectations Approach”, Journal of Political Economy 88,1148-58. Schadler, S. (1977), “Sources of Exchange Rate Variability: Theory and Empirical Evidence”, IMF Staff Papers 24, 253-96. Simpson, T.D. (1984), “Changes in the Financial System: Implications for Monetary Policy”, Brookings Papers on Economic Activity 1, 249-72. Smith, P.N. and M.R. Wickens (1984), “An Empirical Investigation into the causes of the Failure of the Monetary Model of the Exchange Rate”, Discussion Paper No. 7, Centre for Economic Policy Research, London. -48-