The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
CO o GO O £= CU March 1982 Vol. 64, No. 3 CD CD a3 CO CD cr 3 Monetary Policy and Stock Returns: Are Stock Markets Efficient? CD “O CD 13 Monetary Policy and Short-Term Real Rates of Interest 20 Central Banks’ Demand for Foreign Reserves Under Fixed and Floating Exchange Rates The Review is published 10 times per year by the Research Department o f the Federal Reserve Bank o f St. Louis. Single-copy subscriptions are available to the public free o f charge. Mail requests fo r subscriptions, back issues, or address changes to: Research Department, Federal Reserve Bank o f St. Louis, P.O. Box 442, St. Louis, Missouri 63166. Articles herein may be reprinted provided the source is credited. Please provide the Bank’s Research Department with a copy o f reprinted material. Monetary Policy and Stock Returns: Are Stock Markets Efficient? LAWRENCE S. DAVIDSON and RICHARD T. FROYEN .^ ^ .N efficient m arket is one that quickly processes all relevant information. For exam ple, if m onetary policy affects stock returns, then an efficient stock m arket rapidly digests and incorporates all news about m onetary policy. C onsequently, past policy actions will have little value or explanatory pow er in understanding current stock returns. Previous tests of stock m arket efficiency have exam ined the rela tionship betw een the tim ing of the growth of money and stock returns. Although several early studies found th at stock returns lagged b eh in d m oney growth — evidence of stock m arket inefficiency — the results of recent studies have supported the efficient m arket hypothesis .1 The purpose of this article is to provide further evidence on the tim ing of the relationship betw een m onetary policy changes and stock returns by esti m ating m odels that express stock returns as func tions of anticipated and unanticipated m onetary policy m easures. T hese m odels extend previous work in several directions. First, past studies gen erally have divided m oney growth into anticipated and unanticipated com ponents in a m echanical or ad hoc fashion .2 We com pare these results with esti m ates of anticipated m oney growth m easured by the fitted values of previously estim ated m onetary policy reaction functions. This enables us to d eter m ine w h eth er the efficient m arket findings are robust across differing aggregates and decom po sitions of m onetary policy into anticipated and un anticipated com ponents. Second, previous studies focused on the rela tionship betw een m oney growth rates and stock returns. But, during m uch of the period covered by th ese studies, the F ed eral R eserv e’s short-run (month-to-month) operating target was the federal funds rate. Therefore, in addition to estim ating rela tionships betw een stock returns and m oney growth rates, we estim ate m odels relating stock returns and both anticipated and unanticipated m onetary policy actions using the federal funds rate. Again, antici pated and unanticipated policy actions will be de- Lawrence S. Davidson, an associate professor of business eco nomics and public policy at Indiana University, is a visiting scholar at the Federal Reserve Bank of St. Louis. Richard T. Froyen is an associate professor of economics at the University of North Carolina. 1Examples of studies that indicated a lag in the adjustment of stock returns to changes in m oney growth rates are: M ichael J. Hamburger and Levis A. Kochin, “Money and Stock Prices: The Channels of Influence,” Journal o f Finance (Decem ber 1971), pp. 1045-66; Michael W. Keran, “ Expectations, Money, and the Stock Market,” this Review (January 1971), pp. 16-31; and Beryl W. Sprinkel, Money and Stock Prices (Richard D. Irwin, Inc., 1964). Recent studies that support the market efficiency pos tulate include: Michael S. Rozeff, “ Money and Stock Prices: Market Efficiency and the Lag in Effect of Monetary Policy,” Journal of Financial Economics (September 1974), pp. 245-302; John Kraft and Arthur Kraft, “Determinants of Common Stock Prices: A Time Series Analysis,"Journal of Finance (May 1977), pp. 417-25, and “Common Stock Prices: Some Observations,” Southern Economic Journal (January 1977), pp. 1365-67; R. V. L. Cooper, “Efficient Capital Markets and the Quantity Theory of Money,” Journal o f Finance (June 1974), pp. 887-908; Richard J. Rogalski and Joseph D. Vinso, “Stock Returns, Money Supply and the Direction ofCausality, "Journal of Finance (September 1977), pp. 1017-30; James B. Kehr and David Leonard, “M one tary Aggregates, the Stock Market and the Direction of Causal ity. ” Journal o f the Midwest Finance Association (1980), pp. 47-57; and J. Ernest Tanner and John M. Trapani, “Can the Quantity Theory be Used to Predict Stock Prices — Or Is the Stock Market Efficient?” Southern Economic Journal (October 1977), pp. 261-70. 2Rozeff, “Money and Stock Prices,” for example, assumes that anticipated money growth in a given month depends on money growth in the past three months. 3 FEDERAL RESERVE BANK OF ST. LOUIS rived from an em pirical reaction function in which the federal funds rate is the d ependent variable. Third, we extend the tim e period in earlier studies through 1977. This allows us to exam ine the m one tary policy/stock return relationship in both a period o f low stable inflation (1954-65) and one of higher and more variable inflation and m oney growth (196677). Finally, for the period from 1974 through 1976, we estim ate m odels that relate weekly stock returns to the anticipated and unanticipated com ponents of weekly m oney growth. Most previous work on this topic used quarterly or m onthly data .3 Estim ates w ith w eekly data provide a finer test of the efficient m arket hypothesis. DO STOCK RETURNS LAG OR LEAD MONETARY POLICY? Several recent studies of the relationship betw een m oney growth rates and stock returns have found that fu tu re m oney growth rates affect current stock returns. Thus, stock returns appear to lead m oney growth rates .4 O ther studies, however, do not find such effects .5 T he finding that stock prices lead m oney growth has been interpreted in several different ways. O ne interpretation is that stock prices are a causal in fluence on m oney growth. However, as Rozeff points out, w ithin the general equilibrium setting of finan cial markets, it is arbitrary to single out stock returns as a causal variable .6 Rather, the evidence that future m oney growth rates affect current returns m ay be a reflection of the influence of other variables on both stock prices and m oney growth, w ith stock prices adjusting m ore quickly and, th erefo re, lead in g m oney growth rates. Another interesting interpretation of this finding is provided by the “reversed causation with accurate anticipations” m odel .7 In this m odel, causation runs 3One recent exception is Neil G. Berkman, “On the Significance of Weekly Changes in M l,” New England Economic Review (May/June 1978), pp. 5-22. 4See, for example, Rozeff, “ Money and Stock Prices;” Kraft and Kraft, “Determ inants of Common Stock Prices;” and Rogalski and Vinso, “Stock Returns, Money Supply and the Direction of Causality.” 5See, for example, Kehr and Leonard, “ Monetary Aggregates, the Stock Market and the Direction of Causality.” 6See Rozeff, “ Money and Stock Prices.” 7See Rozeff, “ Money and Stock Prices,” pp. 275-76. 4 MARCH 1982 from currently anticipated m oney growth to stock returns. The apparent effect of future m oney growth reflects the accurate anticipations of future m oney growth by the market. It is theseaccurate predictions of future m oney growth that affect current stock returns. SPECIFICATION OF THE MODELS This section describes two sim ple m odels of equity return determ ination. T obin’s theoretical m odel of the financial sector stressed the im portance of the return on capital as the link betw een the real and financial sectors .8 His m odel established a po tential causal connection betw een the exogenous variables of the com m odities and financial markets and the return on equities (ow nership claims on the capital stock). The first of the two m odels presented here is a sim ple version of T obin’s, originally esti m ated by Rozeff .9 This m odel stressed the linkage betw een m onetary aggregates and the equity return. It im posed the additional restriction that only un anticipated changes in the growth rate of m oney (gu) cause unanticipated m ovem ents in the equity return (Ru). R ozeff s “predictive m onetary portfolio” m odel relates the unanticipated current return on equities (R“) to past u nan ticipated changes in m onetary growth rates, that is, (1) Rtu = f(gt1-l,...,gt1n-) + €„ w here R }1 is the unanticipated m ovem ent in the equity return, defined as the actual return (Rt) m inus the expected return conditioned on all available past inform ation (E [R t/B t-i]). U n an ticip ated m oney growth in period t-i, gt-i, is m easured as the change in the m oney growth rate betw een t-i and t-i-1. The error term , et, is assum ed to be a norm ally distrib uted random variable w ith a m ean o f zero and a constant, finite variance. Rozeff assum ed that the expected value of the nom inal eq uity retu rn is constant (E[Rt/Bt_i] = Co) and the m onthly em pirical counterpart of the predictive m odel is: 16 (2) Rt = C 0 + 2 a; gi'j + e2t, i= l w here Co and ai are param eters to be estim ated. 8James Tobin, “A General Equilibrium Approach to Monetary Theory,” Journal o f Money, Credit, and Banking (Februarv 1969),'pp. 15-29. 9See Rozeff, “Money and Stock Prices,” pp. 255-66. FEDERAL RESERVE BANK OF ST. LOUIS To evaluate the relative im portance of the m ost recent m onetary information, Rozeff also estim ated the nonpredictive m onetary portfolio m odel. In this m odel, the contem poraneous m oney surprise is added; the lag on the m onetary surprises starts at zero instead of one: 16 (3) R, = C0 + 2 a; g"i + e3t. i=o A final variant of this model assum es that m arket participants form expectations of future changes in m onetary growth. If these expectations are at least unbiased, then future m onetary growth rates would cause changes in current equity returns. Rozeff’s em pirical nonpredictive m onetary portfolio m odel with anticipations adds eight leads (negative lags) to equation 3 :10 16 (4) Rt = Co + 1 a; g“j + e4t. i= - 8 To test w hether past inform ation about unex pected m onetary growth influences current stock returns, we exam ine the statistical significance of the lagged unanticipated m oney growth terms in the predictive m odel (equation 2). If the stock m arket is efficient, the coefficients on the lagged term s should be equal to zero ( a j = 0 , i = l,...,n). An F-test is used to test this hypothesis; an F-value significantly greater than 1.0 w ould suggest that the stock m arket was in efficient, since past inform ation w ould affect current stock returns. On the other hand, a significant F-value for a sim ilar test of the coefficients in the nonpredictive m odels (equations 3 or 4) does not indicate m arket inefficiency. The finding that only current m onetary growth affects returns sim ply establishes the im por tance of m onetary variables in equity return deter m ination. If future, but not past, m oney growth affects cu rren t returns, this suggests a forwardlooking propensity of the m arket w hich also is not inconsistent w ith an efficient market. The second m odel of equity returns considered here is referred to as the Fam a approach .11 In this 10Future values of unanticipated money growth should not cause current stock market returns to change. However, the exact interpretation of gtu+i is not unambiguous. It could be rein terpreted as the perfectly correct anticipated future change in money growth. In that case, it would be an indicator of the forward-looking propensity of the market. “ This approach is set out in Eugene F. Fama, “Short-Term In terest Rates as Predictors of Inflation,” American Economic Review (June 1975), pp. 269-82. MARCH 1982 m odel, the nom inal return on stocks (Rt) is assum ed to be com posed of the real return (rt) and a prem ium for expected inflation (7rt) — a Fisher effect for stock returns: (5) Rt = rt + TTt. From equation 5, w e can write the expected value of the nominal return conditioned on inform ation available from period t-1 (Bt-i), as (6) E(Rt/B,.1) = E(rt/Bt_i) + E(fft/Bt,1). If we assum e a constant real m ean of stock returns (c0), we can rew rite equation 6 as (7) E(Rt/Bt.,) = c0 + E ^ /B n ). Since E(Rt/Bt-i) is equal to the actual nom inal return on stocks (Rt) minus its unanticipated com ponent (Rt1), we can transform equation 7 into an expression for the actual nominal stock return: (8) Rt = c„ + Ri' + E(7rt/B,.1). Equation 8 then can be converted into a rela tionship betw een money growth and nom inal stock returns if we express (as in equation 1) the unan ticipated com ponent of stock returns as a function of unanticipated changes in m oney growth and if, fur ther, we express the expected inflation rate as a function of expected m oney growth. W ith these as sum ptions, our expression for nom inal stock returns becom es (9) Rt = c0 + f(g“, g“i,...,g“.ni) + h(gt*, gt*i,...,gt*.m2) + vt, w here gt is the expected rate of growth of the m oney stock, and h is the function relating expected m oney growth to expected inflation. T he em pirical counter part to equation 9 used in our estim ation is ni n2 (10) Rt = c0 + 2 b; i=0 + 2 d, g*,_j + vt, j =0 w here various lag lengths and several different m easures of anticipated and unanticipated m oney growth are em ployed. Additionally, one test uses the federal funds rate rather than a m onetary aggregate as the m onetary policy variable. The effects of this substitution on the theoretical interpretation of ourm odels of equity return are discussed below. Using the Fam a (or Fisher) m odel of stock returns, we can also test for m arket efficiency. M arket effi ciency im plies that lagged unanticipated changes in 5 FEDERAL RESERVE BANK OF ST. LOUIS money growth rates would not affect current stock returns (bj = 0 for i > 0 in equation 10). In the Fam a approach, however, lagged anticipated changes in m oney growth rates m ight affect current stock re turns through an effect on expected future inflation. This result w ould not violate m arket efficiency; it w ould sim ply be an elem ent of E(Rt/Bt-i) and would not provide a basis for any profitable trading rules .12 This effect of anticipated m onetary policy on stock returns is another channel by w hich m onetary policy may affect stock prices — even in an efficient m arket — an effect w e test for in the following section. ESTIMATES OF THE MODELS MARCH 1982 policy variables to last-day-of-the-m onth activity. Changes in the average m onthly value w ould appear to be the proper m easure of the shift in m onetary policy from month to month. We relate this to the cum ulative change in stock prices for the month. This does m ean, how ever, that w hile the dependent and independent variables pertain to the same tim e period, they w eight daily observations w ithin the tim e period differently. O ur tests w ith weekly data therefore provide more intra-m onth precision. Unanticipated Money Growth and Stock Returns: Alternative Specifications o f the Basic Models Five sets of m odel estim ates are presented. In all five, the m easure of the nom inal equity return is the percentage change (m easured from the last business day in each m onth or week) in the overall index of all stock prices on the New York Stock Exchange .13 T hese tests em ploy a variety of m onetary policy m easures .14 These include: 1) percentage changes in actual, anticipated and unanticipated M l and the m onetary base, and 2 ) anticipated and unanticipated values of the federal funds rate. The policy m easures in all the tests, except those with w eekly data, are changes in average m onthly values. Returns are changes betw een the last busi ness day of each m onth. This specification relates the cum ulative stock price change from the end of one m onth to the next to the average month-to-month change in the m onetary policy variable. As a result, the stock return variable is m ore sensitive than the The m odels in equations 2-4 specify that unan ticipated m oney growth affects the unanticipated stock return. Rozeff’s tests m ake the following two explicit assum ptions: 12See Rozeff, “Money and Stock Prices,” p. 260. 13An alternative measure includes dividends, but because its variance is so dom inated by stock price changes, it performs almost identically to the index which contains only prices. This alternative measure is not used in our tests. 14These measures of monetary policy each have limitations lor the testing of the efficient market hypothesis. Tests of this hypoth esis must distinguish betw een information which is currently known and used by market participants and that which is not. In fact, we do not know what information was available to and used by these agents. In this research, we have limited the monetary policy measures to those listed above. We have not tried narrow erorbroaderm easures of money like nonborrowed reserves or M2, nor have we used seasonally unadjusted versions of M l or the monetary base. Our tests have selectively em ployed both revised and initially announced seasonally adjusted versions ol M l. Since seasonally adjusted data are revised several times, it would seem preferable to use the initially announced numbers since those were the ones available to market participants. Furtherm ore, Courtenay C. Stone and Jeffrey B. C. Olson, “Are the Preliminary Week-to-Week Fluctuations in M l Biased?” this Review (Decem ber 1978), pp. 13-20, have shown with weekly data that the revised seasonally adjusted series is largely independent of the unrevised series and therefore is a poor proxy for that data. Our weekly aggregate tests, therefore, em ploy the unrevised growth rates of seasonally adjusted M l. This use of initially announced data is not without drawbacks. For example, since initial announcements have been shown to be unreliable indicators of how money is performing, market participants may either ignore seasonally adjusted data or they may modify it. One useful modification would discount the announcem ent with what agents think is the true seasonal ad justment. If they do this correctly, then they are using what turns out to be the actual revisions. If they use seasonal adjustment factors that are different from the true ones, they are using an unobservable series. Our monthly aggregate tests use the re vised, seasonally adjusted growth rates of M l. The monetary reaction function tests do not rely totally upon either revised or unrevised data. For example, the consumer price index and the unemployment rate, which are used to predict the monetary base, are not regularly revised. However, the monetary base itself, like M l, is revised frequently. Finally, the tests with the federal funds rate have no data revision prob lems since this series is not revised. 6 i) Rtu = Rt - Cn, and ii) gi1 = gt - gt-i- The unanticipated return is a deviation from a m ean (Rt - Co), w hile the unanticipated m oney growth rate is a first difference (gt - gt-i). This section compares the results based on these assum ptions w ith two alternative specifications. T he first of these we call the differenced model: iii) R“ = Rt - R n, iv) gi1 = gf- gt_i. The second is called the m ean deviation model: FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Table 1 Sum m ary Statistics for Lead-Lag M oney Growth (g) and Equity Return (R) M odels1 _______________1 9 5 4 — 1 965______________ _____ Lag (lead) M od e l s p e c ific a tio n M ixed 2 3 4 16 to 1 16 to 0 16 to (9) D iffe re n c e d 2 3 4 M ean D e via tio n 2 3 4 __________ 1 9 6 6 — 1977______________ N um ber of s ig n ific a n t c o e ffic ie n ts F R2 DW Lags Leads 1.382 1.296 1.488 .149 .150 .250 1.78 1.79 1.92 2 1 1 — 16 to 1 16 to 0 16 to (9) 1.849* 1.888* 1.285 .178 .192 .214 2.83 2.83 2.85 0 0 0 — 16 to 1 16 to 0 16 to (9) 1.748* 1.649 1.720* .170 .172 .267 1.80 1.82 1.95 4 4 1 — 0 2 0 0 0 1 N u m b e r of s ig n ific a n t c o e ffic ie n ts F R2 DW Lags .889 .859 2.790* .100 .103 .381 1.80 1.83 1.87 0 0 3 1.480 1.720 2.990** .147 .177 .386 2.90 2.88 2.84 0 0 0 1.150 1.070 2.940** .119 .119 .384 1.85 1.86 1.91 0 0 2 Leads N ote: In all cases th e d e p e n d e n t va ria b le s are som e tra n s fo rm o f th e e q u ity re tu rn , R. R2 is th e a d ju ste d c o e ffic ie n t o f d e te rm in a tio n . F is th e F-value, and DW is th e D u rb in -W a tso n s ta tistic. A * (**) im p lie s re je c tio n o f th e n u ll h y p o th e sis at th e 95% (99%) level. T h e n u ll h y p o th e sis states th a t th e e stim a te d c o e ffic e n ts o f th e in d e p e n d e n t va riab le s eq u al zero. T he L eads c o lu m n s in c lu d e th e c o n te m p o ra n e o u s term s. 'D a ta are m o n th ly o b s e rva tio n s. on past m onetary information are never significant, nor are they ever significant as a group. In this vi) gtu = gt - go, period, the effect of future m oney is highly sig tripling the explanatory pow er of the esti w here Co and go are the sam ple-period means of R nificant, m ated m odels. and g, respectively. In the earlier period, there are no unam biguous Sinee the original Rozeff specification mixes d e differences among the m odels. The R 2 reveals rela viations from m eans (Rt - Co) with first differences tively equal explanatory power. T he differenced (gt - gt-i), we refer to this as the m ixed m odel. None of m odel shows a statistically significant effect of the 16 the three versions inherently makes more sense than lags of m oney growth, yet no single coefficient is the others. Our intent here is to see how sensitive the statistically significant. This m odel exhibits a high original specification is to these m inor changes. degree of autocorrelation; therefore, the F-tests should be interpreted w ith caution .15 T he m ean Table 1 provides estim ates of the original em deviation m odel also shows an apparent significant pirical specifications of the three models: the mixed effect of past m oney growth in the early period. model, given by equations 2, 3 and 4, and the modi However, w hen future terms are added to the equa fied specifications w hich we term the differenced tion, the num ber of lagged significant coefficients model and the mean deviation model. The estim ates falls to only one. As a whole, these results offer no in the table cover two subperiods, 1954-65 and 1966- clear rejection of stock m arket efficiency. The effects 77. of future m oney growth on stock returns are also robust w ith resp ect to the type of specification The results in table 1 offer no clear rejection of changes we have made. Rozeffs specification. All three models explain m ore of the variance of equity returns w hen current or 15As is well known, autocorrelation leads to a bias in the standard future m oney growth is included in the regressions. error of the regression. With negative autocorrelation, the direc In the 1966-77 tim e period, individual coefficients tion ot the bias could be positive or negative. v) RJ1 = Rt - Co, 7 FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Table 2 R eaction Function Estim ates of U nanticipated M onetary Policy (§ 1 , § 2 ) and Equity Returns (1954:7 to 1972:3)1 In d e p e n d e n t v a ria b le S“ 9“ N u m b e r of s ig n ific a n t c o e ffic ie n ts Lag (lead) M odel s p e c ific a tio n 2 3 4 16 to 1 16 to 0 16 to (9) 2 3 4 16 to 16 to 16 to 1 0 (9) F R2 DW Lags .655 .621 .946 .051 .051 .117 1.78 1.78 1.87 0 0 7 .896 .969 2.660** .068 .078 .271 1.79 1.83 1.95 0 0 0 Leads ___ 0 7 — 0 3 1See note, ta b le 1. Data are m o n th ly o b se rva tio n s. Unanticipated Monetary Base Growth and Stock Returns: Estimates Employing Monetary Reaction Functions The basic m ixed m odel is retained in this section, but two different proxies for unanticipated m onetary policy actions (gu) are tried. In these tests, we as sume that agents are rational and act as if they know the appropriate function guiding m onetary policy. Table 2 presents the results of estim ating equa tions 2, 3 and 4 using two different proxies for unan ticipated m oney growth. The first of these, denoted gi, comes from Froyen’s m onetary policy reaction function for the m onetary b ase .16 This function, w hich we assum e is used to forecast future growth rates of the m onetary base, relates the latter to past values of the Federal R eserve’s assum ed goal vari ables: the unem ploym ent rate, inflation rate, balance of paym ents and the outstanding governm ent debt held by the public. The estim ated function is used to predict the level of the m onetary base. If Mt is the prediction of the monetary base based on the estim ated reaction function, then we can define the anticipated monetary base growth rate as 17 gl.t = (Mt* - Mt*i) / Mt.i. 16See Richard T. Froyen, “A T estofthe Endogeneity of Monetary Policy "Journal of Econometrics (July 1974), pp. 175-88. ' ’Alternatively, we tried a variant of this form where g 3,t = (M; - M ul/M ,.,. The results w ere not different enough to w arrant further discussion. Digitized for 8FRASER Therefore, a first proxy for unanticipated m one tary base growth is gi'.t = gi,t - gt- The second proxy for unanticipated growth (g 2,t) is based on a sim ple third-order autoregressive process sim ilar to the specification used by Rozeff: g2,t = g2,t — gt, w here ga.t = «<> + < iig t.i + d o g t.j + «3gt-3- The results in table 2 again support the efficient m arket hypothesis. There is no clear evidence that past unanticipated m onetary base growth signifi cantly affects current stock returns using any of the proxies tested here. W hile there are num erous sig nificant lag coefficients in the g“ equation, they are not significant until leads are added, and even then the F-value is not significant. W ith regard to the effects of future m onetary base growth on current stock returns, the pattern of the results in table 2 is interesting. W hen anticipated m onetary base growth is m easured by the sim ple autoregressive specifica tion, and future “ u nan ticipated ” m onetary base growth is taken to be m oney growth that cannot be predicted with that specification, g 2 t, our results show a significant effect for these future terms. H ow ever, for the proxy constructed on the basis of the es tim ated m onetary policy reaction function,g*i t, future unanticipated m onetary base growth has no significant effect on current stock returns. MARCH 1982 FEDERAL RESERVE BANK OF ST. LOUIS Table 3 A nticipated vs. U nanticip ated M onetary B ase Growth and Equity Returns (1954:7 to 1972:3)1 N u m b e r of s ig n ific a n t Lag Anticipated variable coefficients specification gu g* F R2 DW gu g* 32 16 — .969 .078 1.83 1 — 9*1 16 — .621 .051 1.78 0 — 92 16 0 1.014 .081 1.85 1 0 91 16 0 .614 .054 1.80 0 0 92 16 6 1.001 .113 1.83 0 0 9i 16 6 .890 .102 1.85 0 1 1See no te, ta b le 1. Data are m o n th ly o b se rva tio n s. These estim ates are presented in table 3. We use the same proxies for unanticipated m oney growth and, in this case, the corresponding m easure of an ticipated m onetary base growth, as for the estim ates in table 2. The table is divided into three parts: The first two lines include only unanticipated m onetary base growth. T he second two add only the concur rent anticipation of m onetary base growth. The third pair allows up to six months lagged values of antici pations of future m onetary base growth. In each of these, unanticipated m onetary policy has the current as w ell as 16 lagged values. The results are not inconsistent with the efficient m arket hypothesis, since unanticipated m onetary base growth, current or lagged, has no significant effect on stock returns. According to equation 9, how ever, anticipated m onetary base growth should have a positive effect on stock returns, if there is a constant expected real return and if anticipated m onetary base growth affects money growth and, thereby, anticipated inflation. Our results do not show this effect and w ould seem to indicate that the expected real return on stocks is negatively affected Anticipated and Unanticipated by expected inflation that results from anticipated Monetary Base Growth and m onetary base growth. This follows since the ex Stock Returns pected real return declines w ith anticipated infla We discussed previously the Fam a version of the tion, unless there is an offsetting increase in the m odel (equation 9), w here both anticipated and un nom inal return .18 anticipated values of m onetary policy should affect 18Fama uses a general equilibrium approach and concludes that equity returns. In this section, w e again use m one real returns vary with expectations of future real economic ac tary policy reaction functions to differentiate antic tivity. He also argues that apparent correlations betw een real stock returns and expected inflation or money growth rates are ipated and unanticipated policies. The m odel tested spurious. See Eugene F. Fama, “Stock Returns, Real Activity, here is the em pirical specification of the Fam a m odel Inflation, and M oney,” American Economic Review (Septem ber 1981), pp. 545-65. given by equation 1 0 . O ne interpretation of these results is that future “ unanticipated” m onetary policy actions based on the autoregressive proxy are not in fact unanticipated. Inform ation other than past m onetary base growth — information that is available to the public and, if the reaction function specification is correct, informa tion that does affect future m oney growth — may enable the public to correctly anticipate such future m onetary base growth. Since the prediction of the reaction function already incorporates such avail able information, the public cannot forecast future unanticipated m onetary base growth as m easured by reaction function residuals; therefore, these future residuals do not affect current stock returns. O ur results then are consistent with Rozeff’s “reversed causation with correct anticipations” m odel, w here the apparent effect of future m onetary base growth on stock returns reflects the public’s correct forecasts of future m onetary base growth on the basis of cur rently available information. 9 FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Stock Returns and the Federal Funds Rate federal funds rate. The results ofthese tests are given in table 4. If the m onetary authority pegs the federal funds rate, the m oney supply becom es endogenous, and changes in the setting of the rate may be taken as an exogenous variable. In practice, the federal funds rate may change for reasons other than policy, es pecially over short intervals. C onsequently, these tests may reflect not only how efficiently the m arket absorbs infonnation about m onetary policy but also the im pact of other inform ation em bodied in m ove m ents in the federal funds rate. N evertheless, they are useful in ascertaining how changes in the federal funds rate are internalized by the m arket during a period w hen the expressed policy was to m aintain that rate w ithin a narrow range. The results of estim ating equations 2, 3 and 4 are shown in part A of the table. T hese results, using the interest rate as a m easure of m onetary policy, are less favorable to the efficient m arket hypothesis than our estim ates using m onetary aggregates. As can be seen from the first two lines of the table, lagged values of the unanticipated portion of the federal funds rate (lagged errors in forecasting the m onetary author ity’s funds rate setting) appear to affect stock returns significantly. This evidence supports the view that stock returns lag monetary policy — even though our results in the previous section w ould indicate that stock returns do not lag money growth. T he addition of current or future federal funds rate prediction errors does not increase the explanatory pow er of the equation (see estim ates of equation 4 in the table). In the m odel with m onetary aggregates, antici pated inflation was approxim ated by anticipated m onetary growth. It is less appropriate to think of anticipated changes in the federal funds rate as a proxy for anticipated inflation. However, changes in the anticipated federal funds rate that signal changes expected in financial m arkets will still provide im portant information in efficient markets. The tests in this section rem ain, therefore, as tests of m arket efficiency. T hey do, how ever, have less explicit theoretical developm ent that explains exactly how m onetary policy affects stock returns. To split m ovem ents in the federal funds rate into anticipated and unanticipated com ponents, we use the m onetary policy reaction function estim ated by Abrams, Froyen and W aud in w hich the federal funds rate is the dependent variable .19 T he fitted values from the estim ated reaction function provide a m easure of the anticipated federal funds rate (RFt). The unanticipated portion of the federal funds rate (RFU) is sim ply the actual federal funds rate minus the anticipated rate. T he m odels we estim ate using the federal funds rate as a m easure of m onetary policy again are those given by equations 2, 3, 4 and 1 0 , w here th e unanticipated (gu) or anticipated m onetary policy variables (g*) are now in terms of the In part B of the table, w e report estim ates of the m odel that allows both anticipated and unantici pated m onetary policy to affect stock prices. O ur estim ates indicate that lagged values of both unan ticipated and anticipated m onetary policy as m ea sured by the federal funds rate have significant effects on stock returns. Both here and in p artA ofthe table, all the significant coefficients on the federal funds rate variables are negative (the signs of these coefficients are not reported in the table). This accords w ith the conventional expectation that a tightening of monetary policy, as m easured by an increase in the federal funds rate setting, lowers stock prices and, hence, stock returns. In part B, as in part A of the table, how ever, the finding that past available inform ation significantly affects stock returns raises questions about m arket efficiency. This is not to say that the results in table 4 directly contradict the efficient m arket hypothesis. O ne in terpretation of these results that is potentially con sistent w ith the efficient m arket view is that the federal funds rate is a determ inant of the expected real return on stocks, which is not a constant. W ith this interpretation, the excess return on stocks w ould still be independent of past available inform ation, the condition for an efficient market. Still, the results 19The anticipated federal funds rate is a function of 1) consistent in table 4 do suggest the possibility that w hile the forecasts of future values of the unem ploym ent rate, the infla m arket efficiently absorbed data on m onetary ag tion rate and external balance variables and 2 ) lagged values of deviations of actual M l from its target values. See Richard K. gregates, information carried by observations on the Abrams, Richard Froyen and Roger N. W aud,“Monetary Policy federal funds rate was not im m ediately reflected in Reaction Functions, Consistent Expectations, and the Bums stock p rices and, h en ce, affected fu tu re stock Era "Journal of Money, Credit, and Banking (February 1980), returns. p p . 30-42. 10 FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Table 4 A nticipated vs. U nanticip ated M onetary Policy (The Federal Funds R ate) and Equity Returns (1 9 71 :7 - 1 976:6)1 A. Unanticipated Federal Funds Rate (RFU) M od e l Lag (lead) s p e c ific a tio n 2 3 4 16 to 1 16 to 0 16 to (9) N u m b e r of s ig n ific a n t c o e ffic ie n ts F 2.237* 2.097* 1.580 R2 DW Lags Leads .254 .243 .206 1.75 1.73 1.59 4 4 4 — — 0 B. Anticipated (RF*) and Unanticipated (RFU) Lead Values and Lags of the Federal Funds Rate N u m be r of s ig n ific a n t c o e ffic ie n ts Lags (leads) RFU RF* F R2 DW RFU RF* 9 9 9 9 9 — 2.661* 3.355* 3.677* 3.058* 3.150** .223 .309 .440 .332 .387 1.76 2.00 2.08 2.02 2.08 2 3 2 4 3 — 1 2 0 2 0 6 (3) 3 to (3) 1See note, ta b le 1. Data are m o n th ly o b se rva tio n s. Tests with Weekly Money Stock Data E arlier tests that split m oney growth into antici pated and unanticipated com ponents are redone using w eekly data. T he m easures of anticipated and unanticipated w eekly m oney growth are taken from Naylor .20 The tim e period for these tests is August 1974 to March 1977. As noted, the use of weekly data provides a finer test of possible lead-lag relationships b etw een money growth and stock returns. Data on the m oney supply generally were announced during our sample period on Thursday afternoons. Therefore, we as sume an injection of m onetary information occurs Thursday, w hich is new inform ation to Friday’s stock m arket transactions. By m oving to a w eekly m odel, we better capture these events. All m oney stock data used are the values originally announced on the Thursday of each w eek. T he equity returns are derived from the stock prices recorded at m arket closing on the next day, Friday. 20Naylor’s forecasts are from a 52-week autoregressive scheme. This model is re-estimated one week at a time over the entire sample period and generates one-week-ahead forecasts. For details, see John A. Naylor, “ Do Short-Term Interest Rate Ex pectations Respond to New' Information on Monetary' Growth?” Southern Economic Journal (January 1982), pp. 754-63. T able 5 presents the sum m ary data from our w eekly regression tests. The top of the table (part A) reveals that up to 16 lags and nine leads of unantic ipated m oney growth explain very little o f the vari ance in w eekly stock returns. None of the individual coefficients are statistically significant at the 5 percent level of confidence. The F-values suggest that none of the three lag specifications leads to a rejection of the null hypothesis of m arket efficiency. The bottom half of table 5 (part B) specifies past values of both anticipated (gl) and unanticipated m onetary growth (g") as determ inants of the w eekly equity returns. Adding six past weeks of anticipated m onetary growth im proves the explanatory pow er of the equation (with 16 lags of unanticipated money), doubling the R 2 to .166. The m ain contribution in statistical significance comes from the current value of gi with less added by the one week lag (t-value equal to about —1.7). The signs of the estim ated coefficients are negative, im plying an inverse rela tionship betw een anticipated m oney growth and equity returns .21 21This finding would agree with the federal funds rate results if expectations of increased monetary growth are at least partially caused by earlierbelow -targetgrow th. In this case, both higher expected money and higher federal funds rates would correlate with future falling stock returns. 11 MARCH 1982 FEDERAL RESERVE BANK OF ST. LOUIS Table 5 A nticipated and U nanticipated M onetary Growth and Equity Returns (1974:8 - 1977:3)1 A. Unanticipated Money Growth (gu) N u m b e r of s ig n ific a n t c o e ffic ie n ts M odel Lag (lead) s p e c ific a tio n F R2 DW Lags Leads 2 3 4 16 to 1 16 to 0 16 to (9) .947 .897 .890 .084 .085 .115 2.02 2.02 2.03 0 0 0 — 0 0 B. Anticipated (g*) and Unanticipated Money Growth (gu) N u m b e r of s ig n ific a n t c o e ffic ie n ts Lags gu g' F R2 DW gu g‘ 16 16 16 — 0 6 .897 1.179 1.302 .085 .115 .166 2.02 2.03 2.04 0 0 0 1 1 'S e e no te, ta b le 1. Data are w e e kly o b se rva tio n s. Overall, the results of w eekly data indicate that inform ation about m oney growth is quickly reflected in stock prices, as one w ould expect if the m arket is efficient. CONCLUSIONS The results of our study can be sum m arized as follows: Estim ates of the relationship betw een stock returns and m oney growth rates, using m onthly data, support the notion that stock m arkets are efficient. E ven from w eek to w eek, the m arket seem s to quickly u tilize th e m ost recent inform ation on m onetary aggregates. O ur estim ates of the relation ship betw een stock returns and m onetary policy actions as m easured by the federal funds rate, how ever, suggest a possible violation of the conditions for m arket efficiency. On the question of w hether stock returns lead m oney growth, our results indicate that w hen antic ipated m oney growth is a fitted value from a reaction function, future unanticipated m oney growth does not significantly affect current stock returns. But w hen future changes in m oney growth rates are based only on past m oney (using a third-order auto regressive schem e), th ey do significantly affect returns. This finding supports the hypothesis that the m arket uses inform ation other than past m oney growth rates (information em bodied in the reaction Digitized for 1FRASER 2 fu n ctio n p red ictio n ) to fo recast fu tu re m oney growth and that such anticipations affect current stock returns. This research has uncovered very little about how one can use m onetary policy information for profit in the stock market. Information about aggregates is quickly assim ilated by markets. T he m onthly esti mations show little effect of anticipated or unan ticipated aggregates (base or M l) upon stock returns. T he w eekly tests suggest that stock returns tend to fall w ithin a week after the m arket anticipates a rise in the w eek’s m onetary aggregate. T he most useful information seem s to come from the m onthly federal funds rate. We found that increases in that rate tended to lower stock returns over a six- to ninem onth period. Since the federal funds rate is an im perfect indicator of m onetary policy, this finding may say little about how m onetary policy affects stock returns. It does, how ever, reveal that for our 1971-76 sam ple period, m onths w hen the federal funds rate fell w ere followed by periods of rising stock returns. H ad m arket participants been aware of this relationship, they m ight have profited by it. Since the expressed policy of the Federal Reserve today allows the federal funds rate to float w ithin a w ide band, there is no indication that this relation ship continues. The relationship betw een m onetary growth or m ovem ents in the federal funds rate and stock returns in the post-O ctober 1979 period is a subject for future research. Monetary Policy and Short-Term Real Rates of Interest R. W. HAFER and SCOTT E. HEIN T e x t b o o k d escrip tio n s of the ch ann els of m onetary policy’s im pact on the economy usually outline a two-step procedure: “The first is that an increase in real balances generates a portfolio dis equilibrium — at the prevailing interest rate and level of income, people are holding more m oney than they want. This causes portfolio holders to attem pt to reduce their m oney holdings by buying other assets, thereby changing asset yields. In other w ords, the change in the [real] m oney supply changes [real] interest rates. The second stage of the transm ission process occurs w hen the change in interest rates affects aggregate dem and .” 1 The rational expectations literature, how ever, has raised serious questions about this description, especially the first stage w herein an increase in real m oney balances lowers expected real interest rates. Shiller, for example, draw ing from previous work in rational expectations, hypothesizes that the ex pected real interest rate is unaffected by changes in m onetary policy .2 W hile Shiller found little support for this hy pothesis, other recent em pirical work supports it. Fama, for instance, is unable to reject the hypothesis that the expected real rate on short-term financial assets was constant over m uch of the post-Accord period in the U nited States .3 This hypothesis is even •Rudiger Dornbusch and Stanley Fischer, Macroeconomics (McGraw-Hill, 1978), p. 120. 2Robert J. Shiller, “Can the Fed Control Real Interest Rates?” in Stanley Fischer, ed., Rational Expectations and Economic Policy (The University of Chicago Press, 1980), pp. 117-56. Shiller also outlined two other (non-exclusive) hypotheses: (1) the Fed can affect real rates only through unexpected policy moves and (2) Fed policies known far enough ahead of time have no effect on real rates. These hypotheses are not as stringent as the hypothesis considered in this paper. 3Eugene F. Fama, “Short-Term Interest Rates as Predictors of Inflation,” American Economic Review (June 1975), pp. 269-82. stronger than Shiller’s. It holds that m onetary ac tions, as w ell as everything else, have had no system atic effect on expected real interest rates. This article re-evaluates the evidence suggesting that the expected (ex ante) real interest rate on short term financial assets is constant. Evidence is pro vided that allows us to reject this hypothesis for the 1955-79 period. Follow ing this, data are exam ined to determ ine w hether evidence supports the typical textbook description in w hich changes in expected real interest rates are associated w ith changes in real m oney growth. THE FRAMEWORK OF ANALYSIS C onsider first the relationship betw een nominal interest rates and inflation expectations em bodied w ithin the so-called Fisher relationship ,4 (1) it = rf + P f, w here it is a nom inal (or market) rate of interest (the rate m easuring how many dollars m ust be repaid in the future for a given dollar loaned today), rf is the expected real interest rate (the rate m easuring how “The belief in a positive relationship betw een expected inflation and nominal interest rates has a long history in economics. Henry Thornton recognized the relationship as early as 1811. Alfred Marshall also acknowledged the link during the latter half of the 19th century. Even so, the intensity with which Irving Fisher examined the relationship during his career has resulted in the distinction of equation 1 being dubbed the “ Fisher equation.” See Henry Thornton, “Two Speeches of Henry Thornton, esq. on the Bullion Report, May 1811,” in F. A. v. Hayek, ed.,An Enquiry into the Nature and Effects of the Paper Credit o f Great Britain (1802), (August M. Kelley, 1962), pp. 323-62; Alfred Marshall, “Remedies for Fluctuations of Gen eral Prices (1887),” in A. C. Pigou, ed., Memorials o f Alfred Marshall (Kelley & Millman, Inc., 1956) pp. 188-211; and Irving Fisher, The Theory o f In terest (Kelley & Millman, Inc., 1954), especially Chapter 2. 13 FEDERAL RESERVE BANK OF ST. LOUIS many more goods can be obtained in the future by foregoing consum ption today) and P® is expected inflation (the rate at w hich the dollar price of goods is expected to rise.) Equation 1 represents a hypothe sized equilibrium relationship. It posits thatchanges in observed nom inal rates of interest fully reflect changes in expected inflation, holding the expected real rate constant . 5 In other words, nom inal rates and expected inflation are positively related and, ceteris paribus, move on a one-to-one basis. The foundation for this equilibrium relationship is the view that investors have two possible invest m ent opportunities: they can invest either in capital goods that produce a future stream of consum ption goods or in financial assets denom inated in m onetary terms. Investm ent in capital goods is expected to produce rf percent more consum ption goods per year than the am ount of consum ption goods original ly given up to produce the capital good. To make the return on investing in the capital good com parable to the alternative investm ent (the financial asset), the value of the future stream of consum ption goods m ust be translated into dollar terms. This is accom plished by adding the expected rate of change in the dollar price of consum ption goods (Pf) to the rate of increase of consum ption goods (rf). The right-hand side of equation 1 , therefore, represents the ex pected dollar return from investing in a capital good. In equilibrium (and w ithout differential tax rates), the dollar return from investing in capital goods should equal the dollar return from investing in fi nancial assets, m easured by the nominal interest rate, it- Equation 1 thus states that an individual should not find the dollar yield on financial assets any dif ferent from the expected dollar yield on capital goods. We stress that equation 1 is an equilibrium condition: not only are the financial and capital goods m arkets hypothesized to be individually in equilibrium , but any differential in the expected real yields in these two m arkets is arbitraged away. In its present form, equation 1 cannot be exam ined em pirically because the two variables on the righthand side, the expected real rate of interest and inflation expectations, are not directly observable. W hile there are many observable nom inal interest rates on financial assets, there are no reliable aggre gate m easures of either the expected real yield on 5This equilibrium relationship also should include the crossproduct term rf Pf. Like most empirical analyses of this relation ship, we ignore this term, assuming that the magnitude of the variable is sufficiently small. Digitized for14 FRASER MARCH 1982 capital goods or the expected future inflation rate .6 IS THE EXPECTED REAL RATE OF INTEREST CONSTANT? To test the relationship specified by equation 1, one can make two assum ptions: First, assum e that the expected real interest rate is a constant, such that (2) rf = r. Second, to circum vent the problem of m easuring inflation expectations, assum e that next p eriod’s actual inflation (Pt+i) is equal to w hat is currently expected (at tim e t), plus a random disturbance ju,t+i, w here /u,t+i is independent and distributed N(0, cr2): (3) Pt+i = Pf + Mt+i- This relationship specifies that one-period-ahead inflation forecasts are unbiased; on average the actual inflation rate over the next tim e period will be the expected rate. Substituting equations 2 and 3 into 1 yields (4) it = r + Pt+i - nt+1. This equation can be arranged to test em pirically the hypothesis that today’s interest rate accurately predicts tom orrow’s inflation as follows: (5) Pt+i = - r + /3oit + Mt+i- Assuming that financial m arkets are efficient, we would expect to find /3o not to be statistically differ ent from unity and the estim ated constant term to be negative. If the estim ated coefficient (3o is not statis tically different from unity, the proposition that current interest rates fully reflect the m arket’s antici pations of the future inflation rate cannot be rejected. Similarly, if the estim ated constant term is negative, the expected real rate of return is then positive as suggested by the underlying economic theory. M ore 6Some researchers have attem pted to investigate the relationship by using directly observed inflation expectations data generated from Joseph A. Livingston’s biannual survey of economists. See, for example, William E. Gibson, “Interest Rates and Inflationary Expectations: New Evidence,” American Economic Review (December 1972), pp. 854-65; David H. Pyle, “Observed Price Expectations and Interest Rates,” Review o f Economics and Statistics (August 1972), pp. 275-80; Kajal Lahiri, “ Inflationary Expectations: T heir Form ation and Interest Rate Effects,” American Economic Review (March 1976), pp. 124-31; Thomas F. Cargill,“Anticipated Price Changes and Nominal Interest Rates in the 1950’s,” Review o f Economics and Statistics (Au gust 1976), pp. 364-67; John A. Carlson, “ Short-Term Interest Rates as Predictors of Inflation: Comment,” American Economic Review (June 1977), pp. 469-75; and Douglas K. Pearce, “Com paring Survey and Rational Measures of Expected Inflation: Forecast Performance and Interest Rate Effects,” Journal o f Money, Credit and Banking (November 1979), pp. 447-56. MARCH 1982 FEDERAL RESERVE BANK OF ST. LOUIS Table 1 Em pirical Estim ates of Equation 51 C o e ffic ie n t 1/1955-IV/1979 1/1955-IV/1959 1/1960-IV/1969 1/1970-IV/1979 Ordinary Least-Squares Estimates C o n s ta n t -0 .5 8 0 (1.46) 2.686 (3.10) —1.496 (2.65) 1.393 (1.42) 00 1.056 (13.49) -0 .0 4 1 (0.13) 1.073 (7.97) 0.840 (5.61) ■R2 0.646 -0 .0 5 5 0.616 0.439 SE 1.630 1.190 1.116 1.750 DW 1.02 1.63 1.92 1.09 Generalized Least-Squares Estimates -0 .1 2 6 (0.20) 2.584 (2.68) -1 .4 9 6 (2.65) 1.586 (1.15) 0.957 (8.00) -0 .0 0 1 (0.00) 1.073 (7.97) 0.797 (3.93) W 0.389 -0 .0 5 6 0.616 0.270 SE 1.424 1.169 1.116 1.573 DW 2.21 0.504 2.15 0.190 1.92 0.000 2.04 0.455 C o n s ta n t 0o P 1R2 re p re se n ts th e c o e ffic ie n t o f d e te rm in a tio n a d ju ste d fo r de g re e s o f fre e d o m , SE is th e re g re s sio n s ta n d a rd e rro r, DW is th e D u rb in -W a tso n test s ta tis tic and p is th e e stim a te o f th e a u to c o rre la tio n c o e ffic ie n t. A b s o lu te va lu e o f t-s ta tis tic s a p p e a r in parentheses. over, th e existence of serial correlation in the residuals w ould deny the assum ption em bodied in equation 3 and, consequently, w ould lead to a rejection of the hypothesis specified in equation 5 .7 Previous em pirical studies generally have not explicitly considered the tem poral stability of the expected real rate w ithin this framework. The con stant term in equation 5 represents the estim ate of the (negative value of the) expected real rate of return. The above theoretical foundation for this specification suggests that, in addition to being nega tive, this term is statistically tim e-invariant. Thus, a test of the tem poral stability of the constant term is also a test of the constancy of the expected real interest rate. Table 1 presents estim ates of equation 5 for vari ous periods. The inflation data used to estim ate equa’Fama tested and rejected the hypothesis that the expected real rate was linearly related to expected inflation. Critical examina tions of Fama’s results are found in Carlson, “Short-Term Inter est Rates as Predictors of Inflation” ; Douglas Joines, “Short-Term Interest Rates as Predictors of Inflation: Comment,” American Economic Review (June 1977), pp. 476-77; and Charles R. Nelson and G. William Schwert, “Short-Term Interest Rates as Predictors of Inflation: On Testing the Hypothesis that the Real Rate of Interest is Constant,” American Economic Review (June 1977), pp. 478-86. Also, see Eugene F. Fama, “Interest Rates and Inflation: The Message in the Entrails,” American Economic Review (June 1977), pp. 487-96. tion 5 are based on quarterly observations of the GNP deflator, expressed as annual rates of change .8 Since the GNP deflator provides an average m easure of prices over the quarter, the quarterly average three-m onth Treasury bill rate is used as the nominal interest rate m easure. C onsider first the results obtained by estim ating equation 5 over the full sam ple period, 1/1955IV/1979. T he constant term is negative (although not significantly different from zero), and the coef ficient on the interest rate variable is not statistically different from unity as suggested by the theory. U nfortunately, th e low D urbin-W atson statistic provides evidence of first-order serial correlation .9 8The GNP deflator is used to avoid recent problems with the con sumer price index. For a discussion of problems with this index, see Alan S. Blinder, “ The C onsum er Price Index and the M easurem ent of Recent Inflation,” Brookings Papers on Eco nomic Activity (2:1980), pp. 539-65. 9Eugene F. Fama and Michael R. Gibbons, “ Inflation, Real R eturns and C apital Investm ent,” W orking Paper No. 41 (Graduate School of Business, University of Chicago, 1980), also find evidence of serially correlated disturbance terms when quarterly data are employed. In addition, in that study as well as in his “ Stock Returns, Real Activity, Inflation, and Money,” American Economic Review (Septem ber 1981), pp. 545-65, Fama drops the assumption that the expected real rate of interest is constant. Both studies estimate the inflation/interest rate rela tionship assuming that the expected real rate is a random walk. 15 FEDERAL RESERVE BANK OF ST. LOUIS This result, by itself, is enough to reject the frame work in equation 5 .10 Focusing solely on the con stancy of the expected real rate, how ever, the accom panying estim ation problem can be corrected by u sin g g en eralized least-sq uares estim atio n. T hese results appear in the low er half of table 1. The full sam ple results reported there again indi cate that next period’s rate of inflation does mirror, one-for-one, a rise in today’s interest rates. M ore over, the constant term rem ains insignificantly dif ferent from zero. Table 1 further reports estim ation results for sub periods arbitrarily truncated at the end of each decade. If the expected real rate of interest is tem porally invariant, the constant term s in these sub periods should not differ statistically. Yet, as the table im m ediately shows, they do differ significant ly across the various subperiods shown. In fact, the estim ated constant term is positive and significant in the first subperiod (late 1950s), w hile not different from zero in the last decade (1970s). It has the anticipated negative sign only in the decade of the 1960s. M oreover, the coefficient on the interest rate variable is not statistically different from zero in the late 1950s, even though theory suggests that it should equal unity. Thus, the coefficient estimates, as well as summary statistics such as the R2 and the standard errors of the equation, vary substantially across sub periods, irrespective of the estim ation technique used. T he statistical significance of the variation in the constant term (the estim ate of the ex ante real in terest rate) can be investigated by in clu ding dum m y variables for possible shifts in the intercept. Thus, equation 5 was re-estim ated with two dum m y variables: D1 equal to 1 for I/1955-IV/1959 and D2 equal to one for I/1960-IV/1969. E stim ating such an equation w ith ordinary least squares again yielded residuals that w ere significantly autocorre lated. To improve hypothesis testing, the equation was estim ated using a generalized least-squares routine to correct for assum ed first-order autocorrela tio n .T he I/1955-IV /1979 estim atio n resu lts are (absolute value of t-statistics in parentheses): (5') P,+i = 1.40 - 0.88 D1 - 1.88 D2 + 0.83 i, (1.62) (1.19) (3.46) (6.51) ~R- = 0.55 SE = 1.37 DW = 2.07 p = 0.35 10For market efficiency, past values of the disturbance, since they are past inflation forecast errors and are therefore known, should provide no additional help in assessing future inflation beyond that already incorporated in market interest rates. See Fama, “ Short-Term Interest Rates,” p. 273, for a discussion of this aspect. Digitized for16 FRASER MARCH 1982 T hese results support the previous subperiod findings: the estim ated real interest rate is signifi cantly positive only in the 1960s. The point estimates of the expected real interest rate for the 1950s, 1960s and 1970s, respectively, are —0.52, + 0.48 and —1.40. W hile the point estim ates for the 1950s an d th e 1970s are negative, they are not significantly different from zero. On the other hand, the positive point estim ate for the 1960s is significantly different from zero. Thus, the hypothesis that the expected real interest rate has been constant over the past 25 years m ust be rejected. EX POST REAL RATES: FURTHER CONSTANCY TESTS Equation 4 can be rew ritten as (6) it - Pt+i = r - /ut+1. This equation states that the ex post real rate should equal a constant (the ex ante real rate), m inus a w hite noise random error term .11 A feel for the statistical variation in the real rate can be obtained by plotting its behavior for our sam ple period. Chart 1 shows the quarterly ex post real rate for the I/1955-IV/1979 period and its m ean values for the I/1955-IV/1959 (-0.03), I/1960-IV/1969 (1.21) and I/1970-IV/1979 (—0.39) subperiods. If equation 6 holds forthe whole period, the m eans across subperiods should be equal, since the expected value of the disturbance term in each subperiod is zero. Tests for equality of the ex post real interest rate m eans across the subperiods provide another inves tigation ofthe constancy hypothesis. Such tests again lead to a rejection of this hypothesis. The t-statistie, “ This measure of the ex post real rate is somewhat different from that used by others. Many take the difference between today's interest rates and today’s inflation rate as an ex post real rate measure. Theory suggests, however, that the preferable mea sure is the difference betw een today’s interest rates and tomor row’s inflation. In the test subsequently developed and others which follow, interest rates are assumed to adjust one-for-one with inflation expectations, a hypothesis that can be rejected in equation 5'. The reader should be cautioned that there are counter theo retical arguments and some empirical evidence to suggest that the nature of the U.S. tax system has invalidated this rela tionship, with interest rates rising more than one-for-one with an increase in inflation expectations. For theoretical discus sions, see Michael R. Darby, “The Financial and Tax Effects of Monetary Policy on Interest Rates,” Economic Inquiry (June 1975), pp. 266-76; and Martin Feldstein, “ Inflation, Income Taxes, and the Rate of Interest: A Theoretical Analysis,” American Economic Review (Decem ber 1976), pp. 809-20. For empirical evidence on the matter, see John A. Carlson, “Ex pected Inflation and In terest R ates,” Economic Inquiry (October 1979), pp. 597-608. FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 C h a rt l Short-Term Ex Post Real Rate of Interest used to test w hether the m ean ex post real rate for interest to m onetary policy? After all, the textbook the latter half of the 1950s is equal to that of the description of m onetary policy’s transm ission m ech 1960s, is 3.67, sufficiently large to reject the null anism relates changes in the real rate to changes in hypothesis at the 5 p ercen t significance level. real money balances. In particular, it m aintains that Further, the t-statistie used to test the equality of an increase in real m oney balances lowers expected m ean ex post real rates in the 1960s relative to the real rates, at least tem porarily. 1970s is 4.86, again allowing rejection of the null hypothesis of constant real interest rates at the 5 The previous framework, linking ex post and ex percent level. Thus, if one accepts the propositions ante real rates, can be used to address this issue. If that interest rates move in direct proportion with inflation expectations are unbiased and financial expected inflation and that inflation expectations are m arkets are efficient, then the ex post real rate unbiased, one m ust reject the constancy of the ex (it — Pt+i) is equal to the ex ante real rate (rf), ante real interest rate over the subperiods investi m inus a random disturbance term (fit+i) capturing gated. unexpected inflation: (7) it - P t+i = rf - /xt+i. MONETARY POLICY AND THE EXPECTED REAL RATE T hese findings suggest that the real interest rate has not been constant over the past 25 years. In this light, is there any evidence that links the real rate of The typical textbook relationship can be repre sented as (8) rj = fio + f3\ (Mt/Pt) + Pi (Mt_i/Pt_i) + ... + €t, w here M is the nom inal m oney stock, P is the price 17 FEDERAL RESERVE BANK OF ST. LOUIS level and e is a random error term. This relationship represents the hypothesis that the expected real rate is related to real m oney balances. Since nothing in m acroeconom ic theory indicates how long it takes for changes in m onetary policy to have an effect, lagged real balances are included in an effort to capture em pirically the dynamics of the process. Theory does suggest, how ever, that some of the coefficients should be significantly negative. W hile it is im possible to estim ate equation 8 because of a lack of observations on r'r', equation 7 indicates that we have a close approximation in the ex post real rate. Com bining equations 7 and 8 , we get (9) it - P, + 1 = /30 + /3j (M,/ P t) + /3-2 (M u / P, i) + ... + e, - fJ-t+iEquation 9 was estim ated initially by arbitrarily trying 10 lags on real m oney balances in the relation ship. Regardless of the sam ple period considered, how ever, the only coefficients that w ere statistically different from zero in any consistent fashion were those for the contem poraneous and first-lagged real m oney balances. Thus, results including only these two variables are reported. Estim ates of equation 9 over the full sam ple peri od (I/1955-IV/1979) and most subperiods provide evidence of significant first-order autocorrelation in the residuals. Consequently, the relationship was reestim ated using a generalized least-squares tech nique to correct for this problem . The resulting fullsam ple coefficient estim ates and summary statistics are (absolute value of t-statistics in parentheses ):12 (10) it - Pl +1 = 5.00 - 0.89 (M/P), + 0.8.3 (M/P)m (1.73) (2.68) (2.48) R = 0.07 SE = 1.37 DW = 2.14 p = 0.56 F(2,97) = 4.95 W hile the variation in the ex post real rate ex plained by the equation is small, it is statistically significant. M oreover, the coefficient estim ates are consistent w ith the textbook transm ission m echa nism. An increase in real m oney balances is asso ciated w ith a statistically significant, contem po12Money (M/P) is measured (in billions of 1972 dollars) by the adjusted monetary base for all results reported here. Thus, the empirical results indicate that a $ 1 billion increase in real balances will reduce the real interest rate by 89 basis points in the current period. This decline is offset, however, by an 83 basis-point increase in the real rate in the subsequent period. We also tried the M l measure and obtained similar results. Digitized for18 FRASER MARCH 1982 raneous decline in short-term real rates during this period. Further, the results are consistent w ith the long-run policy ineffectiveness of increasing real balances to reduce real interest rates .13 The coef ficient estim ate for real m oney balances lagged one period is significantly positive and is not statistically different from the absolute value of the coefficient on contem poraneous real m oney balances. This finding indicates th at a cu rren t increase in real m oney balances w ill be associated w ith a current decline in real rates, but followed by a rise in real rates of equal size at tim e t+1. This suggests that m onetary au thorities, to the extent that they can change real balances, cannot permanently affect real rates of interest. W hile earlier evidence show ed that the ex post real rate (it — Pt+i) behaved differently across subperiods, there is little evidence to suggest that its relationship to real m oney balances has changed over the period. For exam ple, a conventional Chow test evaluating a hypothesized break in the relation ship at IV/1969 yields a calculated F-statistic of F(3,94) = 0.39, w ell below the 5 percent critical value of 2.70. Thus, the regression coefficients are not statistically different before or after IV/196914 Changes in real balances have the same statistical effect on real interest rates across the sam ple period. Finally, it is appropriate to note that the estim ated relationship im plies a positive relationship betw een the volatility in real m oney balances and the volatil ity in real interest rates. If the frequency of change in real m oney balances increases, the estim ated rela tionship im plies an increase in the frequency of change in real interest rates. The evidence pre sented here suggests that m ore stable real m oney growth, even over periods as short as a quarter, will produce a m ore stable pattern of real interest rate m ovem ents .15 I3We do not mean to suggest that monetary authorities can con trol real money balances over long periods of time. On this point, see Denis S. Karnosky, “Real Money Balances: A Mis leading Indicator of Monetary Actions,” this Review (February 1974), pp. 2-10. 14In addition, we tested the hypothesis that the variance of the error term was larger in the 1970s than in the earlier period. The calculated F-statistic (with 37 and 57 degrees of freedom, respectively) was 1.44, less that the 5 percent critical value of 1.59. Thus, the hypothesis of equal variance across these two periods cannot be rejected. 15An interesting investigation into the effects of monetary policy on both short- and long-term real interest rates is provided in Dean W. Hughes and Duane Weimer, “The Impact on Business Investm ent of the Federal Reserve System’s Operating Proce dures,” Federal Reserve Bank of Kansas City Economic Review (February 1982), pp. 14-25. FEDERAL RESERVE BANK OF ST. LOUIS CONCLUSION This article has provided evidence counter to the hypothesis that the expected real rate of return on short-term financial assets was constant over the period 1955-79. If such a hypothesis w ere valid, m onetary policy w ould be pow erless in affecting real econom ic activity through the conventional transm ission m echanism . W hile rejecting the con stancy hypothesis, this article also provides evi dence consistent w ith conventional m acroeconomic theory w hereby increases in real m oney balances tem porarily lower expected real rates. This effect is contem poraneous on a quarterly basis. W hile such an effect is significant, it is relatively small and MARCH 1982 is offset in the following quarter by an identical rise in ex p ected real rates. T h us, th e re is no evidence of a long-run effect running from changes in real m oney balances to changes in real interest rates. Finally, the evidence presented here suggests that more volatile short-run real m oney growth is likely to produce more volatile real interest rate fluctuations. Thus, contrary to recent claims, stable m oney growth and stable interest rates are hardly inconsistent policy objectives .16 16For another view, see Bryon Higgins, “Should the Federal Reserve Fine Tune Monetary Growth?” Federal Reserve Bank of Kansas City Economic Review (January 1982), pp. 3-16. 19 Central Banks’ Demand for Foreign Reserves Under Fixed and Floating Exchange Rates DALLAS S. BATTEN TJL HE international m onetary system has experi enced significant changes during the 1970s. The m ost dram atic of these has been the transformation from a system of pegged exchange rates to one in w hich central banks m ake no institutional com m itm ent to m aintain a particular exchange rate. D espite this change, central banks have been un willing, in general, to allow their exchange rates to be co m pletely m ark et-d eterm in ed and, co n se quently, continue to hold foreign reserves. The prim ary focus of this article is to analyze central banks’ dem and for foreign reserves in light of this institutional change. C entral banks generally are thought to hold stocks of foreign reserves so their econom ies can avoid incurring the costs of adjusting to every international im balance that w ould be transm itted to the dom estic econom y through changes in exchange rates. In par ticular, before March 1973, central banks partici pating in the Bretton Woods Agreem ent w ere com pelled to hold foreign reserves because they were com m itted to intervene in foreign currency markets w hen the value of their currencies m oved outside a predeterm ined range. It was commonly believed that the dem ise of the Bretton W oods A greem ent and the concom itant greater flexibility of exchange rates w ould reduce central b an k s’ in terven tio n in foreign currency m arkets and, consequently, reduce their dem and for foreign reserves. T hat is, since perhaps the single, m ost im portant reason for holding reserves had dim inished, central banks w ould not be expected to hold such large stocks of foreign reserves as they had under the fixed exchange rate system. In spite of this expectation, how ever, central banks have continued to m aintain sizable stocks of reserves since March The author would like to thank John Bilson, Michael Bordo and Ed Ray for their comments on an earlier draft. Digitized for 2 0FRASER 1973. This observation has led researchers to con clude that central banks have not changed appre ciably their dem and for reserves w ith the transition from a fixed to a floating exchange rate system .1 This conclusion, though potentially accurate, is founded on a framework of analysis in w hich foreign reserves are considered by central banks as a very special type of asset — one held solely to enable them to interven e in foreign currency m arkets. H ow ever, there is an alternative fram ew ork for analyzing central bank behavior that predicts that, even if all countries had adopted a com pletely cleanfloating exchange rate system in 1973, central banks w ould have continued to hold a variety of financial assets, some of which w ould have been classified as foreign reserves under the previous fixed exchange rate system. This article investigates w hich of these com peting frameworks better explains central bank behavior since March 1973. TWO MODELS OF CENTRAL BANK BEHAVIOR To analyze w hether or not central bank behavior has changed significantly since the introduction of flexible exchange rates, the dem and for reserves based on the intervention motive is com pared with an alternative one developed w ithin an asset-choice 'See, for example, Jacob A. Frenkel, “ International Reserves: Pegged Exchange Rates and Managed Float,” in Karl Brunner and Allan H. Meltzer, eds., Public Policies in Open Economies, Camegie-Rochester Conference Series on Public Policy, sup plem ent to the Journal o f Monetary Economics, Volume 9 (1978), pp. 111-40; H. Robert Heller and Mohsin S. Kahn, “The Demand for International Reserves Under Fixed and Floating Exchange Rates,” International Monetary Fund Staff Papers (December 1978), pp. 623-49; Nasser Saidi, “The Square-Root Law, Un certainty and International Reserves Under Alternative Re gimes, "Journal o f Monetary Economics (May 1981), pp. 271-90. FEDERAL RESERVE BANK OF ST. LOUIS fram ework .2 Only if the former explanation outper forms the latter for the floating period can one con clude that the changes in behavior since 1973 have been relatively m inor and inconsequential. The first m odel is the standard one based on the derived dem and for foreign reserves for purposes of intervening in foreign exchange markets. Since this m odel has appeared frequently in the literature, its characteristics are only briefly d escrib ed .3 T he second m odel is based on asset-choice behavior and has not been applied, until now, to the analysis of foreign reserve dem and. In this m odel, foreign re serves are treated as one of several assets that appear in a bank’s portfolio and are held for the general conduct of m onetary policy. The Intervention Model The central bank intervention m otive has been thoroughly investigated. E arlier studies typically have em ployed an optim izing approach in deter m ining the dem and for foreign reserves. O ne pro cedure is to find the stock of reserves at w hich the marginal costs of holding them equal the m arginal benefits of using them to in terv en e in foreign currency m arkets (i.e., the avoidance of costs asso ciated w ith the dom estic economy having to adjust to each external shock). A second procedure is con ducted in term s of w elfare m axim ization under uncertainty. In particular, a central bank’s dem and for foreign reserves is the result of its m aximizing a 2See Russell S. Boyer and David Laidler, “A Comment on the Frenkel Paper,” in Brunner and Meltzer, eds., Public Policies in Open Economies, pp. 141-43. 3Examples of this and similar models include Peter B. Clark, “Demand for International Reserves: A Cross-Country Anal ysis,” Canadian Journal o f Economics (November 1970), pp. 577-94; Peter B. Clark, “Optimum International Reserves and the Speed of Adjustment, ’’Journal o f Political Economy (March/ April 1970), pp. 356-76; T. J. Courchene and G. M. Youssef, “The D em and for International R eserves,” Journal o f Political Economy (August 1967), pp. 404-13; Jacob A. Frenkel, “The Demand for International Reserves by Developed and LessDeveloped Countries,” Economica (February 1974), pp. 14-24; Frenkel, “International Reserves: Pegged Exchange Rates and M anaged Float” ; H. Robert H eller, “Optimal International Reserves,” Economic Journal (June 1966), pp. 296-311; H eller and Khan, “The Dem and for International Reserves Under Fixed and Floating Exchange Rates” ; F. Steb Hippie, The Disturbance Approach to the Demand fo r International Reserves, Princeton Studies in International Finance No. 35 (Princeton University Press, 1974); Milton A. Iyoha, “ Demand for International Re serves in Less-Developed Countries: A Distributed Lag Speci fication,” The Review of Economics and Statistics (August 1976), pp. 351-55; Michael G. Kelly, “The Demand for International Reserves,” The American Economic Review (September 1970), pp. 655-67; and Saidi, “The Square-Root Law, Uncertainty and International Reserves.” MARCH 1982 societal welfare function which is a positive function of the expected level of real incom e and a negative function of its variability. Since the holding of for eign reserves diverts resources away from dom estic uses, the larger the stock of reserves, the lower the expected level of real income. However, if no re serves are held, the dom estic econom y w ould have to adjust to every external shock, resulting in more real income variability. Em ploying the intervention motive w ithin this framework, previous studies have identified four major determ inants of reserve dem and: the vari ability of international paym ents and receipts, the propensity to import, the opportunity cost of holding reserves and a scale variable m easuring the size of international transactions (usually the value of imports). The variability of receipts and paym ents m easures the likelihood that external disequilib rium w ill occur, in d u cin g th e cen tral bank to intervene in foreign currency markets in order to m itig ate th e im p act of th is d iseq u ilib riu m on dom estic markets. The larger the variability of a country’s receipts and paym ents, the more suscep tible is that country to external disequilibrium ; consequently, the larger is the optim al stock of reserves desired for purposes of intervention. There are two possible rationales for including the propensity to im port as a determ inant of reserve dem and. First, the average propensity to im port can be considered a m easure of the degree of openness in an economy, thus indicating the degree to which the economy is vulnerable to an external disequilib rium. A second, alternative rationale stems from the Keynesian m odel of an o p en econom y in which an external d iseq u ilib riu m could be corrected, w ithout changing the exchange rate, by a change in output in proportion to the foreign trade m ultiplier. This output cost of adjustm ent could be avoided if the central bank used its stock of foreign reserves to finance (or to sterilize) the disequilibrium . Since this output cost is directly related to the size of the foreign trade m ultiplier, and since this m ultiplier is inversely related to the m arginal propensity to im port, the output cost of not holding sufficient reserves necessary to avoid this adjustm ent and, thus, the central bank’s dem and for reserves, m ust also be inversely related to the marginal propensity to import. Because the m arginal propensity to import is difficult to m easure, most studies have substituted the average propensity as a proxy. However, if the average propensity to im port is em ployed both as a proxy for the marginal propensity and as a m easure of 21 FEDERAL RESERVE BANK OF ST. LOUIS openness, the sign of its im pact on reserve dem and is ambiguous. Since central banks do not hold an infinite stock of foreign reserves, there m ust be some cost associated with holding them . Conceptually, from society’s point of view, holding foreign reserves represents an allocation of scarce resources away from dom estic uses. Presum ably, for every dollar invested in its stock of foreign reserves (through its central bank), society foregoes a dollar of dom estic capital forma tion. C onsequently, a rate of return on dom estic capital is the appropriate m easure of the opportunity cost to society of its central bank’s stock of foreign reserves. On the margin, the optim al stock of re serves is that level at w hich the cost of holding reserves equals the m arginal benefits provided by that stock of reserves. Few studies have included explicitly a m easure of opportunity cost. Moreover, those that have included it have not found it to be em pirically significant .4 The hypothesized reason for the overall poor perform ance of this variable is the strong positive relationship betw een it and the supply of reserves. In particular, the higher the opportunity cost of holding reserves, the higher also the dom estic rate of return on financial capital which motivates capital inflows and, ceteris paribus, in creases the supply of reserves. As described below, interest rate differentials are em ployed as an attem pt to circum vent this problem . Finally, the scale variable and the dem and for foreign reserves should be positively related. In fact, if the value of international transactions is used as the scale variable, the elasticity of reserve dem and with respect to the value of international transac tions should be betw een 0.5 and 1.0 .5 An Asset-Choice Model In form ulating an asset-choice m odel of central bank behavior, foreign reserves are treated sim ply as one type of asset in a central bank’s portfolio held to enable the central bank to conduct dom estic m one tary policy. It is assum ed that the prim ary objective 4See, for example, Courchene and Youssef, “The Demand for International Reserves ’; Iyoha, “ D em and for International Reserves in Less-Developed Countries” ; Kelly, “The Demand for International Reserves” ; and Saidi, “The Square-Root Law, Uncertainty, and International Reserves.” 5See E rnst B altensperger, “The Precautionary D em and for Reserves,” The American Economic Review (March 1974), pp. 205-10; and J. H. G. Olivera, “The Square-Root Law of Precau tionary Reserves,” Journal o f Political Economy (September/ October 1971), pp. 1095-1104. 2 MARCH 1982 of m onetary policy is to provide an econom ic en vironm ent conducive to the stable, noninflationary growth of real output. To this end, the central bank affects the level of commercial bank reserves (and, subsequently, the m oney supply) through activity in governm ent securities and foreign currency m arkets and by making loans directly to the banking sector. C onsequently, to conduct m onetary policy ad e quately, its portfolio should contain at least three assets: foreign reserves, governm ent securities and claims on com mercial banks. A central bank typically confronts two types of econom ic phenom ena — expected and unexpected — to w hich it makes policy responses. In light of this, the specific m odeling of the portfolio decision making process of a central bank involves separating its assets into two categories: com m itted and un com m itted assets. In response to its anticipations of prospective events, a central bank commits a portion of its portfolio so that it can pursue its m onetary policy objective w ithin this “expected” econom ic environm ent. However, since a central bank also is faced with unanticipated econom ic events to which it may wish to respond, it m ust hold additional reserves to enable it to respond to these “unexpected” occurrences (or shocks) as well. T hese “precautionary” reserves may or may not be used for the conduct of m onetary policy in any specific period, w hile the com m itted portion, is, by definition, fully involved in the monetary control process. C onsequently, a central bank is concerned only with the yield (cost) on the potentially idle, precautionary portion. That is, a central bank’s dem and for the assets that form the com m itted com ponent is hypothesized to be insensi tive to their relative yields, w hereas the com position of the precautionary (or uncom m itted) reserve com ponent is hypothesized to be sensitive to changes in relative asset yields. To formalize this discussion of central bank b e havior, assum e that a central bank (subject to certain constraints) desires to m axim ize its “ ability” to respond to unanticipated events. It accom plishes this by maximizing the uncom m itted portion of its portfolio .6 This can be sum m arized with the fol6A model assuming a wealth-maximizing objective of the U.S. Federal Reserve System has been shown to be a better predictor of Fed behavior than the traditional model of the Fed as an automaton reacting only to political pressures. See Mark Toma, “Inflationary Bias of the Federal Reserve System: A Bureaucratic Perspective,” unpublished manuscript (California State Uni versity, Northridge, 1981). Consequently, applying a similar assumption to other central banks is not without precedent. FEDERAL RESERVE BANK OF ST. LOUIS lowing objective function: (1) F(xi,...,xn) = II (xk - y k) k=l w here Xk = asset k’s maturity value at the end of the time period, 7k = the committed or required value of asset k, xk — -yk = the uncommitted or precautionary value of asset k, /8k = asset k’s share of the uncommitted portfolio, and 2 /3k = k=l The resulting system of asset-dem and equations is as follows :8 £k / (3) xk = n + Vk [TA j= i k = 1, ., n It is clear from equation 3 that a central bank’s dem and for each asset in its portfolio has two primary com ponents. The first is the required or com m itted portion (yk), which is determ ined regardless of yields. The second, or precautionary, com ponent is the 7A11 assets are assumed to mature in one period, but longer-lived assets could be included without a substantive change in the analysis. Also, since the issue investigated here is a central bank’s allocation of a given portfolio among various assets, the determination of the size of the portfolio in any time period (TAt) is not considered. For some insights into this question, see Toma, “ Inflationary Bias of the Federal Reserve System.” 8More formally, the system of demand equations represented by equation 3 is derived by setting up the Lagrangean function and maximizing it with respect to each asset as follows: ft ft (x2 - 72 ) (Xj +1 “ 7j + l) F 1 + rk = the yield on asset k w ithin the period, = the present value of the assets in the portfolio.7 n — = (x i - y i) fix, ft+l ft -1 - & - (xj.i (x„ - 7n) - 7j-i) ft-1 fti Xvj = 0 = ft (xj - 7j) 1 F - Avj = 0 or 1 (l') L = n (xk - n ) k=l (2') 1, (2) TA = 1 vk xk, k= l i'k TA rem ainder of its b alan ce sh eet (TA — £ yj vj)> j= l w hich the bank allocates to the various assets (in proportions denoted by /3k) according to relative yields in a m anner that m aximizes its objective function .9 ft (Xj - 7j) which the central bank maximizes subject to the following balance sheet accounting constraint: where vk MARCH 1982 X (TA 2 k=l Vk Xk) vj (3') - = f t (x i-7 j) Solving (3') for /3j yields: (4') 13, = Kv>(xJ - y Q . F Since 2 y8k = 1, k (5') 2 ft = p 2 vj (xj - 7j) = 1 j J K = p (2 Vj Xj - 2 Vj y j ) j j X = p (TA —2 Vj -yj) from (2) in the text or j (6') “ = TA - 2 Vj yj = T-(xj - yj) from (3’). j Solving (6’) for Xj yields: (7') xj = yj + 7 J (TA - 2 Vj yj) J which is the system represented by (3) in the text. It can be shown that the own-price elasticity of demand for asset j is ( S ') f xv - - 1 + yj(i-ft) and that the Allen partial elasticity of substitution between assets i and j is (9') xj - y s Xj - y j n xk For (xk — yk) > 0, all assets are Hicksian substitutes. 9The value of yk is determ ined by those variables that influence each country’s monetary policy decisions (e.g., economic activ ity, unemployment, inflation). Certainly, interest rates may be included in this group of determinants. However, since yk is estimated, the hypothesized interest insensitivity of a portion of a central bank’s portfolio can be easily tested. Specifically, if yk is statistically significant, the hypothesis that a central bank holds a portion of its portfolio for reasons other than relative yields cannot be rejected. Also, the hypothesis that any part of the portfolio is sensitive to changes in interest rates can be tested by testing the statistical significance of /3k• 23 FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 ESTIMATION OF THE MODELS The Intervention Model The rationale for this is that central banks hold most of their foreign reserves in the form of U.S. dollars. Instead of holding idle balances of dollars, central banks typically invest their reserves in some short T he functional form of central bank dem and for term asset in order to m aintain a relatively high d e foreign reserves for the purpose of exchange m arket gree of liquidity; hence, the ratio (or log difference) intervention is a familiar o ne :10 m easures the net foregone yield. C onsequently, an (4) In Rit = a0 + a; In Mit + a2 In mit + a3 In crit appropriate yield on invested foreign reserves is a short-term in terest rate on d ollar-d en om in ated + a4 In rit + uit, assets .12 where Rit = the sum of country i’s holdings of gold, The sam ple em ployed consists of seven coun convertible foreign exchange, SDRs and reserve position in the IM F at the end of tries for the tim e period 1/1964 to IV/1979.13 The time period t, countries in clu d ed are D enm ark, F rance, W est G erm any, Japan, th e N eth erlan ds, N orw ay and Mit = imports of i during t, Sweden. The U nited States is not included because m it = i’s average propensity to import during it is considered to be the prim ary supplier of foreign t (Mit/GD Pit), reserves. The data set consists of a pooling of crosso"it = the trend-adjusted variance of i’s stock of section and tim e-series observations. foreign reserves in t, The possibilities that country-specific variation rit = i’s opportunity cost of holding foreign may be present and that a lagged adjustm ent process reserves during t, may exist are provided for in the following assum ed autoregressive error structure: uit = error term. (All variables denom inated in domestic currency units are converted into U.S. dollars using the end-of-period ex change rate.) The use of imports as a scale variable and the average propensity to im port as an indicator of openness have been discussed above. The trend-adjusted vari ance of country i’s stock of foreign reserves is a proxy for the variability of international receipts and ex penditures. It is calculated using a m ethod sim ilar to Frenkel’s .11 The m easure of opportunity cost em ployed is the ratio of the discount rate in each country to the threemonth Eurodollar deposit rate. For a given portfolio of assets, the discount rate represents a m easure of the foregone earnings of central banks as a result of holding assets in the form of foreign reserves; the three-m onth Eurodollar deposit rate is a m easure of the income earned from invested foreign reserves. 10See, for example, Frenkel, “International Reserves”; and Heller and Khan, “The Demand for International Reserves Under Fixed and Floating Exchange Rates.” 11 Frenkel, “International Reserves,” p. 136. O ur measure of vari ability is actually Frenkel’s divided by the num ber of degrees of freedom (14 in this case); i.e., t -1 , CTit = 2 (Rim - Rim-1 m =t-15, V i m ) 2/ 14, where r)im is the slope of a linear time-trend equation esti mated over the period t-15 to t-1. Digitized for24 FRASER (5) uit = pi uit_i + eit, where pf = autocorrelation param eter for country i, eit = white noise random error. Including a separate autocorrelation param eter for each country captures the country-specific variation 12The discount rate is em ployed because, even though it is not market-determined, its movement closely parallels market rates in the countries in the sample. Also, since most of the central banks studied use interest rates as a mechanism of monetary control, the discount rate reflects conditions in the respective credit markets. Government securities markets are not suffi ciently developed in all of the countries to be able to use an interest rate from that market. The Eurodollar deposit rate is used as the yield on foreign reserve stocks even though other currencies are held as foreign reserves and even though some central banks have refrained generally from investing in the Eurodollar market directly. The justifications for this are: (a) the U.S. dollar is still the major reserve currency, comprising 66 to 75 percent of the foreign reserves held by central banks, (b) some central banks do invest directly in the Eurodollar market while others invest indirectly using the Bank for International Settlements as an intermediary and (c) the major alternative to the Eurodollar market is the market for U.S. Treasury bills. However, since the three-month Eurodollar rate and the threemonth Treasury bill rate move very closely together, they yield virtually identical results when em ployed individually in the estimation of both the intervention and the asset-choice models. Finally, the ratio has been criticized as simply a proxy forthe forward discount orprem ium on the currencies included. However, when the covered ratio is substituted for the un covered one, no significant qualitative changes occur. 13The sample period extends to IV/1980 for Japan, W est Germany and the Netherlands. Gross domestic product data were not available for the other countries in the sample for this extended period. FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Table 1 Estim ation of Intervention M odel 1/1964-11/1973 111/1973-IV/19791 P a ra m e te r E stim a te S ta n d a rd e rro r ao ai a2 a3 a4 - .5 2 5 .883* - .6 1 4 * .064* - .0 7 3 .420 .066 .072 .023 .060 E stim a te S ta n d a rd e rro r 1.102* .644* - .2 8 9 * .113* - .2 1 7 * .503 .069 .070 .038 .055 C o u n try P P D e n m a rk F rance G erm any Japan N e th e rla n d s N orw ay S w eden .94 .94 .83 .92 .92 .90 .96 .89 .78 .93 .93 .54 .88 .74 N RM SE R 2 b e tw een a ctu a l va lu e s and p re d ic te d values = 259 = .118 = .96 N RM SE R 2 b e tw e en a ctual values and p re d ic te d va lu e s = 187 = .141 = .87 1The sa m p le p e rio d e xte n d s to IV/1980 fo r Japan, W est G e rm a n y and th e N e th e rla n d s. "S ig n ific a n tly d iffe re n t fro m ze ro at th e 5 p e rc e n t level. and also provides a m eans of introducing dynam ic the Sm ithsonian Agreem ent, that is, betw een the behavior into the m odel .14 second and third quarters of 1973.16 Finally, the date ofthe switch from fixed to floating The results obtained from estim ating the solution exchange rates m ust be identified. Since the data are of equations 4 and 5 over the two tim e periods in pooled, it is extrem ely difficult to identify the break dicated above are reported in table 1. Several dif point as occurring at a specific point in time. It is ferences in the estim ated relationships for the two likely the switch occurred over different intervals for periods are apparent. First, the im port elasticity (ai) each country an aly zed .15 E xperim entation w ith in the fixed exchange rate period is significantly various breakpoints around the March 1973 collapse larger than that in the floating rate period. In fact, the of the Sm ithsonian Agreem ent yielded no single im port elasticity in the fixed period is not statistically quarter as the most likely break point for all of the different from one, w hich indicates that central bank countries in the sample. C onsequently, the break is holdings of foreign reserves do not exhibit econ simply assum ed to coincide with the actual failure of omies of scale during that period. Second, the mag nitude of the response to changes in variability (as) is 14For further explanation, see John F. O. Bilson and Jacob A. Frenkel, “Dynamic Adjustment and the Demand for Interna tional Reserves,” NBFR Working Paper No. 407 (November 1979), pp. 1-4; and H eller and Khan, “The Demand for Inter national Reserves Under Fixed and Floating Exchange Rates,” p. 631. As pointed out by H eller and Khan, when equation 5 is substituted into equation 4, the result is observationally equiv alent to an adaptive-expectations or an error-learning process. 15This is supported by Frenkel, “International Reserves,” pp. 122-25; and Saidi, “The Square-Root Law, Uncertainty and International Reserves,” pp. 280-83. 16This choice is generally supported by Frenkel, “ International Reserves,” pp. 124-25, and by Heller and Khan, “The Demand for International Reserves Under Fixed and Floating Exchange Rates,” pp. 637-39. The selection of the break point is also constrained by the necessity to choose the same break point for each model so that the performance can be compared over identical sample periods. Also, for each model, the hypothesis that the estimated parameters before this point are equal to those after this point is rejected at the 5 percent confidence level. 25 FEDERAL RESERVE BANK OF ST. LOUIS larger under floating than under fixed rates. This is som ewhat paradoxical since one m ight expect that the increased exchange rate flexibility during the floating rate period would serve as a buffer and, consequently, reduce central banks’ response to changes in variability .17 Third, the sensitivity of central banks’ reserve holdings to interest rate changes under fixed rates (a.4) is insignificant, a result sim ilar to that of other studies .18 Alternatively, under floating rates, central banks are found to respond in a significant and con ceptually consistent m anner to changes in interest rates. W hen com pared with those of previous studies, these results suggest that an interest rate differential is a better m easure of the opportunity cost of holding reserves. Finally, a com parison of the intercepts (ao) suggests that central banks are holding larger stocks of foreign reserves, on average, in the floating rate period than they did in the fixed rate period, indi cating that they have actually added to their stocks during the floating period. The Asset-Choice Model To estim ate the system of asset-dem and equations represented by equation 3, it was assum ed that norm ally distributed random errors enter additively with zero m ean and constant variance. As a result of introducing a random com ponent in this m anner, the sum of the error term s across all equations in the system m ust equal zero if the system is to be con sistent .19 This restriction on the error structure, by introducing linear dependence across equations, has at least two im portant im plications for estim ation. First, single-equation estim ating techniques are in appropriate. Efficient estim ation requires the use of a system technique. Second, the covariance matrix of the entire system is singular. Because of this, a full17Frenkel, “International Reserves,” p. 120, also obtained this result; however, Saidi, “The Square-Root Law, Uncertainty and International Reserves,” p. 285, found smaller responses to changes in variability in the floating-rate period. 18See footnote 4. 19For the system of asset-demand equations represented by equa tion 3 to be consistent, the value of the estimated portfolio must equal the value of the actual portfolio. This condition implies that the error terms across all n asset-demand equations must sum to zero. That is, the error terms across equations are linearly dependent and thus, by definition, correlated. It could also be argued that, for this analysis, the demands for assets are cor related regardless of the consistency condition. In particular, if the impact of foreign exchange market intervention upon the domestic money supply is sterilized (e.g., through an offsetting sale or purchase of government securities), then foreign ex change holdings and government security holdings are neces sarily negatively correlated. Digitized for26 FRASER MARCH 1982 inform ation technique cannot be em ployed on the entire system of n asset-dem and equations sim ul taneously because the inversion of this covariance matrix is required during the estim ation process. Consequently, only n-1 equations can be estim ated sim ultaneously.20 The countries and tim e periods em ployed here are identical to those used in estim ating the intervention m odel. The assets of the central banks of these countries are aggregated into three categories: for eign reserves, claims on governm ent and claim s on com mercial banks. The interest rates used for these asset groups are the three-m onth Eurodollar deposit rate (for foreign reserves), short-term governm ent bond yield in country i (for claims on governm ent) and the discount rate in i (for claims on commercial banks). The three-m onth E urodollar rate is used here for the same reason it was used in the estim ation of the intervention m odel. Also, a dynam ic specifi cation is em ployed to capture lagged adjustm ent of the com m itted param eters (yu) by allow ing them to vary over time. This dynam ic feature is introduced into the system by assum ing that the com m itted level of each asset is a function of the total holding of that asset during the previous tim e period as follow s: (6) yv t = 0k Xkt-i, w ith 0 =£ 9\r s; 1 for all k. T h e p aram e ter 0^ reflects a p ro p ortio n al relatio n sh ip b e tw e e n th e co m m itted level o f asset k in th e cu rre n t p erio d to th e total h o ld in g o f th a t a ss e t in th e p re c e d in g p e rio d . F inally, th e d ate of th e sw itch from fixed to floating exchange rates is th e sam e as in estim atin g th e in terv en tio n m odel. Substituting equation 6 into equation 3 and recog nizing that n = 3 in this case, the resulting system of asset-dem and equations is as follows: (7.1) x iit = Hi x u m + —'lBi- t (T' A it - . 3X , Oj x jin vjit)/ + ulit J=1 (7 .2 ) X2it (7 .3 ) = 0o X2 iM (T A it v 2it ' x3jt = 03 x 3it-i + — + — V 3 it (T A it - . 1 Oj Xjit.i Vjit) + u 2it s Oj Xjit.i Vjit) + u3it J=1 . J=1 / 1 20Robert A. Poliak and Terence J. Wales, “Estimation of the Linear Expenditure System,” Econometrica (October 1969), pp. 611-28. They prove that if a full-information, maximumlikelihood estimation procedure is employed, the estimated param eters are invariant to w hichever n -1 equations are included. FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Table 2 Estim ation of A sset-C hoice M odel 1/1964-11/1973 P a ra m e te r2 01 tt2 t>3 01 02 03 E stim a te III/1973-IV/19791 S ta n d a rd e rro r .987* .893* 1.005* .536* .326* .138* .022 .032 .017 .048 .035 .041 RM SE o f e q u a tio n 7.1 = .697 R2 b etw een a ctu a l and p re d ic te d va lues fo r e q u a tio n 7.1 = .98 E stim ate .983* 1.001* .897* .261* .213* .526* S ta n d a rd e rro r .011 .018 .049 .038 .047 .058 RM SE o f e q u a tio n 7.1 = 1.404 R2 b e tw een a ctu a l and p re d ic te d va lues fo r e q u a tio n 7.1 = .99 'T h e sa m p le p e rio d e xte n d s to IV/1980 fo r Ja p a n, W est G erm any and th e N e th e rla n d s. 2The s u b s c rip ts 1, 2 a nd 3 re fe r to th e th re e asset c a te g o rie s, fo re ig n reserves, cla im s on g o v e rn m en t and cla im s on c o m m e rc ia l banks, resp e ctive ly. 'S ig n ific a n tly d iffe re n t fro m ze ro at th e 5 p e rc e n t level. other hand, the estim ated com m itted param eters for claims on governm ent (6 2 ) and for claims on com m ercial banks (6 3 ) have changed significantly with 1 v jit the change in regim es .22 Furtherm ore, the p er 1 + r,jit centage of their discretionary portfolio that central ijit = the yield on asset j in country i from be banks held in the form of foreign reserves (/3i) fell ginning to end of time period t, significantly from the fixed to the floating period. TAit = the value of i’s portfolio at beginning The sensitivity of the dem and for foreign reserves to of period t, changes in interest rates (as m easured by the abso lute value of the price elasticity of dem and) also fell = error term. .jit from .563 in the fixed rate period to .289 in the Table 2 presents the results of estim ating the above floating rate period. N onetheless, the fact that this system om itting equation 7.3 .21 A full-information, percentage is statistically significant in both periods m axim um -likelihood technique is used to obtain indicates that reserve holdings are at least partially efficient estim ates. sensitive to changes in interest rates. All param eter estim ates are statistically significant Taken together, the changes in 91 and /Si over the and w ithin conceptually acceptable ranges of values. two periods shed some light on why H eller and Khan As before, differences betw een tim e periods, but consistently overpredict central bank dem and for also across assets, are readily apparent. In particular, foreign reserves during the floating period .23 In their the estim ated com m itted param eter for foreign re m odel, central banks hold foreign reserves solely to serves (fli) is relatively constant across tim e periods, intervene in foreign exchange m arkets. A lterna indicating that central banks have not altered the tively, in the asset-choice m odel, intervention is com m itted portion of their foreign reserves in the simply one of several m otives (where the com m itted move from fixed to floating exchange rates. On the param eter m easures the dem and for reserves for where jit the value of country i’s holding of asset j at the end of time period t, 21Except for ^3 and its variance, all parameters and their variances are estimated directly. Since 2 f3y = 1, 03 = 1 — 0i — 02 and k Var ((83) = Var (/3i) + Var ((82) + 2 Cov (/3i, 02 ). The same results as those reported were obtained when either equation 7.1 or 7.2 (instead of 7.3) was deleted. 22Even though 63 in the fixed period and 62 in the floating period are greater than 1 (the conceptual limit of each), neither is significantly greater than 1 in a statistical sense. 23H eller and Khan, “The Demand for International Reserves Under Fixed and Floating Exchange Rates,” pp. 639-43. 27 FEDERAL RESERVE BANK OF ST. LOUIS MARCH 1982 Table 3 Table 4 Partial Elasticities of Substitution R esidu al-V arian ce Estim ates Assets 1/1964-11/1973 F o re ig n reserves and cla im s on g o v e rn m e n t .116 F o re ig n reserves and c la im s on c o m m e rc ia l banks .031 C la im s on g o v e rn m e n t and c la im s on c o m m e rc ia l banks .065 111/1973-IV/19791 .028 Level L o g -le ve l .076 .094 'T h e s a m p le p e rio d e xte n d s to IV/1980 fo r Japan, W est G e rm a n y and th e N e th e rla n d s. purposes of intervention ).24 Even though this rela tionship appears to be relatively stable across tim e (in the asset-choice model), overlooking the sig nificant decline in the percentage of the precau tionary portfolio held in the form of foreign reserves by basing predictions on a interven tio n m odel should lead to an overprediction of reserve dem and, ceteris paribus. O ne final question remains to be answ ered: Are the assets in central banks’ portfolios close substi tutes for each other? To answ er this question, partial elasticities of substitution are calculated for each of the asset pairs over each tim e period. Since these elasticities are functions, inter alia, of the com m itted and uncom m itted levels of each asset, the elasticities reported are evaluated using the m ean holdings of the relevant assets (table 3). Given the relatively high estim ated values of the com m itted param eters, it is not too surprising to find that none of the assets are close substitutes. Predictive Abilities o f the Two Models T he ultim ate test of a structural m odel is how well it predicts behavior. This section compares the pre dictive abilities of the two m odels described above. 24One may infer that, since the asset-choice model does not ex plicitly contain explanatory variables that represent the in tervention motive, it is fundamentally mis specified. However, the estimation of the asset-choice model clearly indicates that the foreign reserve demands of central banks are sensitive to yields on other assets in their portfolio. Since the intervention model ignores these explanatory variables, it is also funda mentally misspecified. Consequently, future research should be directed at combining the features of both of these models to specify correctly a central bank’s demand for foreign reserves. Digitized for 28 FRASER 1/1964-11/1973 111/1973-IV/19791 .788 .0139 4.577 .0199 .486 .0144 1.970 .0167 Intervention model Asset-choice model (E q u a tio n 7.1) Level L o g -le ve l 1The sa m p le p e rio d e xte n d s to IV/1980 fo r Ja p a n, W est G e rm a n y and th e N e th e rla n d s. Two m ethods of com parison are em ployed: T he first is the residual-variance criterion d evelop ed by Theil 25 The use of the residual-variance criterion involves calculating a residual-variance estim ate (error sum of squares divided by degrees of freedom) for each m odel and selecting the m odel w ith the sm allest residual variance .26 Since the intervention m odel is estim ated in log-level form and the assetchoice m odel is not, the residual-variance estim ates from the two m odels are not directly com parable. To m ake these estim ates com parable, eith er the re siduals of the estim ated intervention m odel have to be transform ed from logarithms to levels or the re siduals of the estim ated asset-choice m odel have to be transform ed from levels to logarithm s .27 Table 4 presents the results of both of these transform ations. Except for the logarithm ic specification estim ated over the fixed rate period, the asset-choice m odel appears to outperform the intervention model. T hese results, how ever, m ust be qualified. The residual-variance m ethod presupposes that one of the specifications is the correct one, a som ewhat presum ptuous supposition. Also, in this case the two 25Henri Theil, Principles o f Econometrics (John Wiley and Sons, Inc., 1971), pp. 543-45, 553-54. 26The selection of the specification with the smallest residual variance is justified by the following proposition: if the correct specification has uncorrelated disturbances with zero mean and constant variance and if the explanatory variables are non stochastic, the residual-variance estim ator of the correct specification has an expectation that is never larger than that of an incorrect specification. See Theil, Principles o f Econo metrics, p. 543. 27This transformation is accomplished by converting the actual and the predicted values from the level (logarithmic) specifi cation into logarithms (anti-logs), calculating the sum of squared deviations of the predicted value from the actual, then adjust ing for degrees of freedom. FEDERAL RESERVE BANK OF ST. LOUIS m odels com pared are non-nested; that is, the m odels have separate sets of explanatory variables such that one m odel cannot be obtained from the other. Con sequently, the conventional use of summary statis tics and F-tests to discrim inate am ong alternatives can be m isleading and even inappropriate .28 The second m ethod is an extension of the Cox test developed by Pesaran and D eaton .29 This procedure for testing non-nested hypotheses is not subject to either of the above qualifications necessary for in te rp re tin g the resu lts of th e resid u al-v arian ce m ethod. In particular, Pesaran and D eaton’s pro cedure does not em ploy a single m aintained (null) hypothesis. (No m odel is considered a priori to be the correct one.) The alternative m odels are anal yzed one at a time. O ne by one, each is assum ed to be the correct one.) The alternative m odels are ana lyzed one at a time. O ne by one, each is assum ed to be the correct model (null hypothesis); the alternative has been observed. T he notion of absolute goodness of fit plays no role in this procedure. In fact, the possibility exists that all com peting m odels may be rejected. This is not the case for conventional testing procedures .30 The test statistics calculated w ith the intervention m odel and equation 7.1 of the asset-choice m odel, respectively, as the null hypothesis are reported in table 5 .31 U nder the null hypothesis, this test statistic is asym ptotically distributed as a norm al random variable with zero m ean and unit variance. The rp28See M. H. Pesaran, “On the General Problem of Model Selec tion,” The Review o f Economic Studies (April 1974), pp. 153-71. 29D. R. Cox, “Tests of Separate Families of Hypotheses,” in Proceedings o f the Fourth Berkeley Symposium on Mathe matical Statistics and Probability (University of California Press, 1961), pp. 105-123; M. H. Pesaran and A. S. Deaton, “Test ing Non-Nested Nonlinear Regression M odels,” Econometrica (May 1978), pp. 677-94. 30A necessary condition for the use of this test is that both models explain the same dependent variable. In this case, the first equation of the asset-choice model explains the quantity of reserves dem anded while the intervention model explains the logarithm of the quantity of reserves demanded. Consequently, to perform the Cox test, the anti-log of the intervention model (i.e., a non-linear, Cobb-Douglas-type function) is estimated using a maximum-likelihood procedure. The resulting pre dicted values and e stimated parameters are essentially identical to those obtained from a least-squares estimation of the loglinear functional form. 31The test statistic (C) is defined as: where To = s I n --------;-------------------------------------------- , °o + “ g(0Ao)]'[f(<£o - g(<?>Ao)] MARCH 1982 Table 5 Statistics for Testing H ypotheses Involving N on-N ested M odels Ho: In te rv e n tio n m o d e l H a : A sse t-ch o ice m od e l (E q u a tio n 7.1) T est sta tis tic Period 1/1964-11/1973 -1 8 .3 6 * 111/1973-IV/19791 -2 2 .6 2 * Ho: A sse t-ch o ice m o d e l (E q u a tio n 7.1) H a : In te rve n tio n m odel T est s ta tis tic P e riod 1/1964-11/1973 - 0 .8 5 111/1973-IV/19791 - 0 .7 5 1T he sa m p le p e rio d e x te n d s to IV /1980 f o r Ja p a n , W est G e rm a n y and th e N e th e rla n d s. ‘ S ta tis tic a lly d iffe re n t fro m z e ro at th e 5 p e rc e n t level. suits are unam biguous. W hen confronted w ith the "data and the asset-choice m odel as an alternative, the intervention m odel m ust be rejected. Alternatively, the asset-choice m odel cannot be rejected. This conclusion is invariant across sam ple periods. W hile the rejection of the intervention m odel for the float ing rate period is not unexpected, it is certainly interesting that this m odel is also rejected for the fixed rate period. This result confirms that the assetchoice m odel provides a m ore general explanation of central banks’ dem and for reserves than does the intervention m odel. SUMMARY AND CONCLUSION The purpose of this article has been to compare central bank behavior before and after the m ove m ent to floating exchange rates w ithin the fram e work of two alternative m odels of a central bank’s dem and for foreign reserves. In the first m odel, sample size, estimated variance of the model under H0, estimated variance of the model under Ha, f($o) - g(cf>Ao) = the residuals from an auxiliary esti Var (T„) mation of the model under H a using the predicted values from the model u n d e r Ho as th e d e p e n d e n t variable, the variance of To as defined in Pesaran and Deaton, “Testing NonN e ste d N o n lin e a r R e g re ss ion M odels,” p. 687. 29 FEDERAL RESERVE BANK OF ST. LOUIS foreign reserves are treated as a special type of asset, one dem anded solely to enable a central bank to intervene in foreign currency markets. T he second m odel considers foreign reserves to be the same as — and also to be held for the same reasons as — any other asset w ithin a central bank’s portfolio. The estim ation of the asset-choice m odel as an alternative to the intervention m odel yielded several interesting results. First, a central bank’s dem and for foreign reserves is sensitive to relative changes in the yields of the assets in the portfolio. Second, central banks consider foreign reserves as substi tutes to other assets in their portfolio. Third, the decrease in the percentage of the uncom m itted portfolio com posed of foreign reserves is identified as a possible reason for the usual overprediction of reserve dem and by the intervention m odel in the 30 MARCH 1982 floating rate period. Finally, and m ost im portantly, the asset-choice m odel consistently outperform s the intervention m odel. Since the testing procedure em ployed could lead to the rejection of both m odels, the fact that the asset-choice m odel cannot b e rejected in eith er sam ple period is an extrem ely robust result. The im plication is sim ply that, regardless of exchange rate regim e, central banks hold foreign reserves for a w ide variety of purposes — not ju st for intervention in foreign exchange m arkets. C onsequently, the in v estig atio n of w h eth er or n ot cen tral b an k s’ general behavior has changed with the m ovem ent to a system of floating exchange rates w ithin the framework of the intervention m odel appears to be m isdirected. Investigation should focus on the argum ents, instead of the param eters, w ithin the dem and function.