The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
The R e v ie w 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 for subscriptions, back issues, or address changes to: Research Department, Federal Reserve Bank of St. Louis, P.O. Box 442, St. Louis, Missouri 63166. Articles herein may b e reprinted provided the source is credited. Please provide the Bank’s Re search Department with a copy o f reprinted material. 2 The “Rationality” of Survey-Based Inflation Forecasts R. W. HAFER and DAVID H. RESLER nr 1 HE notion that economic agents rationally form their expectations about future economic events has emerged as a critically important hypothesis with pro found implications for macroeconomic policy. For ex ample, modern hypotheses relating to the Phillips curve emphasize that it is the departure of actual inflation from expected inflation that cause any shortrun trade-off that may exist between inflation and unemployment. Consequently, empirical tests of many macrotheoretic models require the identification not only of directly observable phenomena, such as infla tion and unemployment, but also of expectations or anticipations of these phenomena. The measurement of generally nonobservable phe nomena, such as inflation expectations, poses a diffi cult challenge in constructing empirical tests for macro models that include such variables. It is first necessary to identify an inflation expectations proxy that is consistent with the assumptions of the under lying model. As a result, tests of theories, such as the natural rate hypothesis, that employ proxy measures for inflation expectations (such as autoregressive pro cedures) are joint tests of both the underlying theory and the validity of the expectations proxy. Presumably, autoregressive procedures are used because they are less costly than opinion surveys. When survey-based data on inflation expectations are readily available, this cost argument loses some of its force. Nevertheless, it is important to determine which of the two measures is appropriate for test ing various economic theories; that is, whichever measure conforms most closely to the requirements of the underlying theory becomes the measure of choice. For instance, tests of rational expectations models should first establish that the measures of expectations conform to the criteria of rationality. This paper examines whether one particular set of survey data — the Livingston data — meets specified criteria of rationality.1 JFor examples of studies dealing with the measurement and effects of inflation expectations, see John A. Carlson, “A Study of Price Forecasts,” Annals o f Econom ics and Social Measure ment (Tune 1977), pp. 27-56; Stephen Figlewski and Paul Wachtel, “The Formation of Inflationary Expectations,” R e view o f Econom ics and Statistics (forthcoming); Rodney L. Jacobs and Robert A. Jones, “Price Expectations in the United States: 1947-1975,” American E conom ic R eview (June 1980), pp. 269-77; Edward Kane and Burton G. Malkiel, “Autore gressive and Nonautoregressive Elements in Cross-Section Forecasts of Inflation,” Econom etrica (January 1976), pp. 1-16; Donald J. Mullineaux, “On Testing for Rationality: An other Look at the Livingston Price Expectations Data,” Jour nal o f Political Econom y (April 1978), pp. 329-36; Douglas K. Pearce, “Comparing Survey and Rational Measures of Expected Inflation: Forecast Performance and Interest Rate Effects,” Journal o f Money, C redit and Banking (November 1979), pp. 447-56; James E. Pesando, “A Note on the Ration ality of the Livingston Price Expectations Data,” Journal o f Political E conom y (August 1975), pp. 849-58; and Stephen J. Tumovsky,^ “Empirical Evidence on the Formation of Price Expectations,” Journal o f the American Statistical Association (December 1975), pp. 1441-54. 3 F E D E R A L R E S E R V E B A N K O F ST . L O U I S Tests of Rational Expectations The hypothesis of rational inflation expectations, pioneered by John Muth, holds that expectations about future inflation are formed in a manner that fully reflects all currently available and relevant in formation.2 Stated somewhat differently, the observed rate of inflation differs from the expected rate of inflation only by some random error. Thus, the ration ality hypothesis can be stated algebraically as: ( 1 ) TT, = ,-t u? + U t, where Ttt is the actual rate of inflation during period t, t-iTt? *s the rate of inflation expected at time t-1 for period t, and ut is a random variable with mean zero and variance cr,V3 Expressed in this form, i.e., inflation expectations are unbiased estimates of observed inflation, the ra tionality hypothesis can be tested empirically by esti mating the equation, (2 ) TT, = Bo + B, n? + u ,, where t-i^t represents the survey-based expected in flation rate for period t made at period t-1. The notion of rational expectations, then, corresponds to the joint hypothesis that B0 = 0 and B, = 1. In addition, u. should exhibit no evidence of autocorrelation. Pesando and Figlewski and Wachtel subjected the Livingston expectations series to this test of rational ity.4 Pesando was unable to reject the joint hypothesis using consensus inflation forecasts from each survey for the periods 1959-1969 and 1962-1969. Figlewski and Wachtel, however, were able to reject the null hypothesis using a pooled time series/cross-section sample of 1,864 individual forecasts for the period 1947-1975. An additional criterion for rationality requires that inflation forecasts be efficient; in other words, the process by which inflation expectations are formed should be identical to the process that actually gener ates observed inflation. Consequently, any evidence suggesting that some of the relevant information set is not being fully (i.e., efficiently) utilized would indicate rejection of rationality. Pesando tested this notion of rationality by hypothesizing that both the expectations of inflation and inflation itself are de scribed by the history of inflation. Mathematically, 2John F. Muth, “Rational Expectations and the Theory of Price Movements,” Econom etrica (July 1961), pp. 315-35. Alternatively, equation ( 1 ) can be rewritten as ( tt, - ,-iH?) = Ut; that is, any departure of actual from expected inflation is a random variable with mean zero and variance, 05. 4Pesando, “A Note on the Rationality . . . and Figlewski and Wachtel, “The Formation of Inflationary Expectations.” 4 NOVEMBER 1980 this interpretation of rationality can be expressed as: (a) IT, = (3 ) £ 1=1 B i TT, i + (in □ (b ) ,-,H ? = 12=1 B ,' Ttt- i + LU, . Efficiency requires that Bj = B/ for all i, . . . , n. Pesando, Carlson, and Mullineaux directly tested the efficiency of the Livingston inflation forecasts by esti mating equation (3 ) and then applying an F-test to the sum of the squared residuals.5 Pesando was not able to reject the efficiency criterion at standard confi dence levels for the period 1959-1969. Carlson, using the same time period but a revised version of the Livingston data, found that the inflation forecasts do not satisfy the efficiency criterion.6 Mullineaux, on the other hand, demonstrated that the error variances of equations (3a) and (3b ) esti mated by Pesando and Carlson are not homogeneous. Consequently, the F-test used by Pesando and Carlson is inappropriate.7 Mullineaux proposed an alternative efficiency test that involves estimating the equation, (4 ) FEt = (tt, - , .it?) = b„ + Z i= l b , Tt t - i + e, where £, = [ilt _ ^i2t. The forecast error (F E t) is re gressed on past inflation rates known at the time the forecast was made.8 Efficiency requires that F E t be 5See Pesando, “A Note on the Rationality . . . This approach to testing for rationality is generally referred to as a “weakform” test because it employs only information contained in the history of inflation. It should be noted, however, that fail ure to meet the weak-form requirements of rationality sug gests that the forecast would also fail stronger forms of the test. For a discussion of weak-form and other types of tests, see John Rutledge, A M onetarist M odel o f Inflationary E xpec tations (Lexington: Lexington Books, 1974). In addition, equation (3 ) does not specify either the exact length of the lag on past inflation or the length of the period over which the inflation is observed. Pesando, Carlson, and Mullineaux each used a 5-period lag on observed 6-month inflation rates. This lag length will also be used in this paper. '■Carlson has noted that the numbers published by Livingston have been judgmentally revised. To circumvent this possible source of error, Carlson constructs a forecast series that is based on the actual responses received by Livingston. See Carlson, “A Study of Price Forecasts,” for a more detailed discussion of his construction procedures. 7The Chow test used by Pesando and Carlson requires that the error terms H,, and be independently and identically dis tributed. If the error terms are not identically distributed (homogeneous variances), the Chow test is inappropriate. Mullineaux tests for variance homogeneity by using Bartlett’s test statistic and finds that the hypothesis of homogeneous variances is rejected at the five percent level of significance. See Mullineaux, “On Testing for Rationality . . . ,” pp. 331-32. 8Equation (4 ) is derived by subtracting equation (3 b ) from (3 a ). That is, b, = Bi — B| for all i. Following Mullineaux, equation ( 4 ) is estimated with a constant term (bo) instead of subsuming it into the error structure as Pesando and Carl son did. F E D E R A L R E S E R V E B A N K O F ST . L O U I S unrelated to any information known at the time ( t-1) the forecast was formed. In other words, all the in formational content of past inflation rates is fully utilized in forming expectations. Thus, the null hypoth esis is that b0 = 0 and bj = 0 for all i, . . . , n. In addition, efficiency requires that the error term be serially uncorrelated, or Cov (£ t, £i) = 0 for t / i.9 Using Carlson’s version of the Livingston data, Mullineaux was unable to reject the efficiency hypothesis for the period 1959-1969.10 Pearce, using Carlson’s data set and another test of efficiency, concluded that “the survey respondents did not efficiently use the information in the past history of the Consumer Price Index (C P I) when forming their expectations of inflation.”11 Thus, it appears that efficiency tests of the Livingston inflation expectations data are sensitive to the type of tests used, to the version of the Livingston data used, and to the time period examined. This article demonstrates that these test results are also sensitive to assumptions about the length of the forecast horizon. Therefore, it is particularly impor tant to determine the actual period over which Liv ingston respondents are making their' forecasts. The nature of this problem can be illustrated by a careful review of the survey method. The Forecast Horizon and the Forecast Error Livingston conducts his survey each spring and fall, requesting respondents to indicate their predictions about a number of economic indicators including the CPI. For example, in the spring survey they are asked to predict what the level of the CPI will be in the following December and June. Because the question naires are mailed in April and usually are returned in May, two interpretations can be made about the fore cast horizon. If, as Carlson assumes, the survey re spondents know only the April CPI, then they are implicitly predicting an 8-month rate-of-change ( April to December) and a 14-month rate-of-change (April 9It should be noted that, although the heterogeneous variance problem that plagued the Chow tests of Pesando and Carlson is alleviated here, the procedure employed does require the maintained hypothesis of independent errors. 10Mullineaux also found that for the data set used by Pesando (i.e., inflation forecasts inferred from the originally published versions of Livingston data), the hypothesis of efficiency is rejected. 11Pearce, “Comparing Survey and Rational Measures . . . p. 451. Pearce statistically analyzes the forecast errors ob tained by using either the Livingston forecasts or forecasts generated from a continuously updated moving average model [M A (1)] of the monthly CPI series. NOVEMBER 1980 to June of the following year). Alternatively, Jacobs and Jones argue that a more reasonable assumption is that the respondents actually know or have an ac curate estimate of the May CPI.12 This, of course, means that the forecast CPI implies a 7-month (or 13-month) rate of inflation. The choice of the forecast horizon can affect the results of the bias and efficiency tests, especially if the forecast is interpreted loosely as a prediction of a steady inflation. Mullineaux and Resler each made this assumption; i.e., they assume that the prediction is a constant rate-of-change for any period within a given forecast horizon.13 This assumption is often conven ient and may not be inappropriate when the investi gation focuses on the process that generates the fore cast. It may pose problems, however, when efficiency tests, such as those represented by equation (4 ), are conducted. Because the survey respondents are, in fact, fore casting an inflation rate over a 7- or 8-month horizon, it is desirable to evaluate equation (4 ) by calculating the forecast error over that time horizon. For example, F E t should be calculated by taking the difference be tween the actual rate of inflation occurring between April (or May) and December and the rate of infla tion predicted for that period. This forecast error should be regressed against lagged inflation rates known to the forecaster as of April (or May). This approach differs from Mullineaux’s procedure in which F E t was computed as of the time the next forecast was made (i.e., October). This approach seems inappro priate for evaluating the efficiency of the forecasts, especially since the forecasts exhibit expectations of accelerating inflation. The next section reevaluates the tests for bias and efficiency in light of these new tim ing assumptions. Empirical Results To investigate the importance that assumptions about the forecast horizon have on tests for bias and 12Jacobs and Jones, “Price Expectations in the United States: 1947-1975.” 13This essentially requires that inflation forecasts are linear. Thus, changes from one point to another within the fore cast horizon will not be distinguishable. If, however, infla tion expectations are not linear over different time horizons (e.g., 6 or 8 months), then the assumption of a steady rate of inflation prediction is vitiated. The fact that the 14-month forecasts are greater than the 8-month forecasts in 38 out of 40 observations from 1959-1978 suggests that the assump tion of a constant rate of inflation within the 8- or 14-month periods may not be appropriate. See Mullineaux, “On Test ing for Rationality,” fn. 3. See also, David H. Resler, “The Formation of Inflation Expectations,” this Review (April 1980), pp. 2-12. 5 F E D E R A L R E S E R V E B A N K O F ST. L O U I S efficiency (and hence rationality), the three alternative forecast horizons discussed in the preceding section are utilized in direct empirical comparisons. Based on these forecast horizons, three forecast error series are calculated and employed in the efficiency tests re ported below. To reiterate, these alternative F E t series are determined by assuming an April-October fore cast horizon (Mullineaux), a May-December fore cast horizon ( Jacobs-Jones), and an April-December forecast horizon. All tests use Carlson’s version of the Livingston data (i.e., sample average CPI forecasts from which the expected inflation rate is generated). To facilitate a comparison with previous research, the following sample periods are used: 1959-1969, 19591978, and 1959-1978 excluding the 1971-1973 period of price controls of various phases.14 To test for bias in the inflation forecasts, equation (2 ) is estimated and an F-test on the joint hypothesis that B0 = 0 and Ba = 1 is conducted for each of the alternative forecast horizons.15 The F-values calcu lated for this test are presented in table 1, and allow rejection of the null hypothesis at the 1 percent level, irrespective of the sample period chosen. This result contrasts directly with Pesando’s but is consistent with the findings of Figlewski and Wachtel, who found the Livingston data to be biased.16 An examination of the individual coefficients, B,, and B,, indicated that the joint hypothesis is rejected primarily be cause B, exceeds unity for all the sample periods. Nevertheless, the results indicate a tendency for B, to decline toward unity as more recent observations are added to the sample, suggesting that forecasters gradually adjusted to the accelerating inflation of the 1960s and early 1970s.17 Table 2 presents additional information on the accuracy of the inflation expectations series. Although the root-mean-squared error and mean error statistics 14This truncated 1959-1978 sample period was chosen to ex clude observations of forecasts errors that occurred during the period of wage and price controls. It seems reasonable that forecasters would have encountered considerably more difficulty in forecasting inflation during this period, since the controls were applied unevenly and gradually relaxed at unpredicted intervals. 15To facilitate computation of the appropriate F-statistics, equation ( 2 ) was modified slightly. Specifically, subtracting t-iTt? from each side of (2 ) produces: (2') Ut — t-illt = B„ + (B , — 1 ) t-ilT? + u,. The null hypothesis then implies that the estimated slope and intercept of equation (2 ') be jointly equal to zero. 1GPesando, “A Note on the Rationality . . . ” and Figlewski and Wachtel, “The Formation of Inflationary Expectations.” 17In studies of the process by which inflation forecasts are gen erated, more definitive evidence indicates that this process has changed over time. For more detail about this evidence, see Donald J. Mullineaux, “Inflation Expectations and Money 6 NOVEMBER 1980 Table 1 Bias Test for “Short-Run” Inflation Forecasts1 F-Values Period Forecast horizon 1959-1969 1959-1978 1959-1978- April-October 15.242 15.723 14.401 May-December 12.660 15.487 14.411 April-December 28.367 18.144 17.439 F(2,20) F(2,38) F(2,35) Critical 5% 3.49 3.25 3.27 F-values 1% 5.85 5.21 5.27 'Test based on joint hypothesis that B0 = 0 and Bi = in equation (2 ’). 1 2This period excludes the 1971-1973 price control years. vary only slightly between forecast horizons, the Theil statistics indicate that the fraction of forecast error due to bias is reduced somewhat by using the May-December horizon. It is interesting to note that, of all of the horizons examined, the April-December assumption continually yields statistics suggesting greater problems with bias than variance or covari ance in the forecasts.18 Although unbiased forecasts satisfy one criterion for rationality, it is common to find properties of bias in other non-survey-based inflation forecasts. For in stance, Lombra and Moran note that, while the Federal Reserve Board staff’s forecasts of nominal GNP are unbiased, its forecasts of GNFs real and inflation components show evidence of systematic errors.19 It is possible that inflation forecasts can show evi dence of systematic bias yet still be characterized as Growth in the United States,” American Econom ic Review (March 1980), pp. 149-161, and Resler, “The Formation of Inflation Expectations.” 18For a description of this methodology, see Henri Theil, Ap plied Econom ic Forecasting (Amsterdam: North Holland Publishing Co., 1971), pp. 26-32. 19Raymond Lombra and Michael Moran, “Policy Advice and Policy Making at the Federal Reserve,” Carnegie-Rochester C onference Series on Public Policy 13, 1980, p. 20. For evi dence that other forecasts similarly underestimate inflation and over-estimate real output, see V. Zarnowitz, “An Analysis of Annual and Multiperiod Quarterly Forecasts of Aggregate Income, Output, and the Price Level,” Journal o f Business (1 9 7 9 ), p. 133. F E D E R A L R E S E R V E B A N K O F ST. L O U I S NOVEMBER 1980 Table 2 Analysis of the “Short-Run” Forecast Errors1 Forecast horizon assumption Sample period April-October May-December April-December Theil statistics RMSE Mean error um 1959-69 1.383 0.911 1959-78 2.151 1.324 1959-78-’ 2.053 1959-69 1959-78 1959-782 Us IT 0.434 0.360 0.206 0.379 0.226 0.394 1.270 0.383 0.223 0.394 1.344 0.858 0.408 0.347 0.246 2.317 1.414 0.372 0.252 0.375 2.214 1.356 0.375 0.252 0.373 1959-69 1.307 0.934 0.513 0.344 0.143 1959-78 2.101 1.355 0.416 0.210 0.374 1959-782 1.962 1.261 0.413 0.203 0.384 'RM SE is the root-mean-squared error, Um is the Theil bias coefficient, U” the variance coefficient, and Uc the covariance coefficient. 2Omits the 1971-1973 price control years. “weakly” rational in the sense that the forecasters efficiently utilize all information contained in the his tory of inflation. To implement this efficiency test, FE, is calculated for each forecast horizon and used to estimate equation (4 ). ably from those of Mullineaux, and thev highlight the importance of specifying the time period over which FE, is calculated. If FE, is evaluated at the end of the period over which the respondents were forecasting inflation (e.g., December), the efficiency hypothesis is rejected in all but one instance. The results for the three different time periods are now discussed in greater detail. Because acceptance of the efficiency hypothesis in the present context requires that bi = 0 for all i ( i = l , . . . , n) and that the estimated relationships indicate no evidence of serial correlation, the statistics of pri mary interest are the reported F-values and the Durbin-Watson and Durbin-h statistics. The reported F-value is pertinent for testing the joint hypothesis that all the bi (i = 1, . . . ,5) are concurrently zero. Both the Durbin-Watson and Durbin-h statistics test for the presence of serial correlation. Although the DurbinWatson statistic is usually appropriate, Durbin has shown that the h statistic is more efficient when the set of independent variables includes a lagged de pendent variable.20 Because Mullineaux has interpreted equation (4 ) as containing a lagged dependent vari able, both statistics are reported. Turning first to the 1959-1969 period, the reported F-statistic for the May-December and the AprilOctober forecast horizons indicates that the efficiency criterion is satisfied. Recalling that the April-October horizon corresponds to the assumption made bv Mullineaux, these results are essentially consistent with his. The Durbin-h statistic for the April-October horizon, however, indicates the presence of negative serial correlation, even though the Durbin-Watson statistic falls within the indeterminate range.21 Since Ordinary least squares estimates of equation (4 ), using the alternative F E t series and sample periods, are presented in table 3. These results differ consider- + 0.050tt<-4 + 0.083tt, (0.25) (0.48) 20See James Durbin, “Testing for Serial Correlation in Least Squares Regression When Some of the Regressors are Lagged Dependent Variables,” E conom etrica (May 1970), pp. 410-21. 21For purposes of comparison, Mullineaux’s estimation results are presented here: (it,- t-iTrf) = -0.232 + 0.237tt,., -0.051tt,-2 + 0.251n„„ (1.91) (1.44) (0.27) (1.36) R 2 = 0.102, h = 1.89, F = 1.48. The difference between Mullineaux’s results and those in table 3 may well be due to the use of different computer algorithms. As such, the difference between the Durbin-h values may not be representative of true differences in the respective residual processes. 7 F E D E R A L R E S E R V E B A N K O F ST. L O U I S NOVEMBER 1980 Table 3 Efficiency Test Results1 CoefficientsForecast horizon b„ b, b2 Summary statistics3 b3 b, b5 R-’ D.W ./h S.E.E. F F ° (.05, .01) 1959- 1969 April-October -0.244 (0.46) 0.244 (1.48) -0.049 (0.26) 0.254 (1.37) 0.050 (0.26) 0.083 ' (0.48) 0.11 2.61/-2 .2 6 1.00 1.54 May-December -0.493 (1.02) 0.193 (1.19) 0.114 (0.61) 0.215 (1.10) 0.114 (0.58) 0.041 (0.22) 0.24 2.25/-0.91 0.92 2.36 April-December -0.345 (0.88) 0.218 (1.79) 0.051 (0.36) 0.295 (2.16) 0.019 (0.13) 0.061 (0.47) 0.38 1.85/0.43 0.74 3.52 April-October 0.695 (1.53) 0.397 (2.82) 0.003 (0.02) -0.100 (0.55) -0.261 (1.44) 0.102 (0.71) 0.20 2.14/-0.97 1.54 2.92 May-December 0.442 (1.01) 0.435 (3.29) 0.133 (0.81) -0.113 (0.64) -0.436 (2.43) 0.200 (1.40) 0.37 1.98/0.12 1.47 5.68 April-December 0.717 (1.74) 0.368 (2.88) 0.035 (0.22) -0.058 (0.35) -0.362 (2.21) 0.160 (1-23) 0.26 1.77/1.01 1.40 3.76 2.85, 4.44 19 59-1978 2.49, 3.51 19 59-1978 (Omitting 1971-73) April-October 0.649 (1.44) 0.300 (2.06) 0.062 (0.35) -0.052 (0.28) -0.257 (1.41) 0.082 (0.56) 0.14 2.21 / - 1 .22 1.51 2.16 May-December 0.414 (0.96) 0.340 (2.43) 0.175 (1.06) -0.051 (0.29) -0.412 (2.31) 0.155 (1.07) 0.34 2.01/-0.05 1.50 4.43 April-December 0.668 (1.69) 0.269 (2.11) 0.098 (0.63) -0.009 (0.06) -0.361 (2.26) 0.140 (1.10) 0.24 1.80/0.90 1.33 3.14 2.54, 3.73 iT e s t results based on equ ation ( 4 ) . -Values in parentheses represent absolute values of t-statistics. 3R- is the coefficient of determination corrected for degrees of freedom; D.W. is the Durbin-Watson statistic; h is the Durbin-h statistic; S.E.E. is the standard error of the equation; F is the calculated F-value to test the joint hypothesis that all bi (i = 1, . . . , 5 ) equal zero; and F ° represents the relevant critical F-value. efficiency requires no serial correlation among the residuals, the hypothesis of efficiency for the AprilOctober horizon remains unresolved. Unlike these two forecast horizons, however, the results based on using the April-December assumption clearly permit rejec tion of the efficiency hypothesis.22 In contrast to the results for the 1959-1969 period, the hypothesis of efficiency is unambiguously rejected at the 5 percent level for each forecast horizon ex amined during the entire 1959-1978 sample period. The hypothesis is also rejected at the 1 percent level for the May-December and April-December horizon 22It should be recalled that the April-December forecast hori zon does not require the special assumptions necessary to construct the competing forecast error series. We know that Livingston supplies the April CPI to the survey recipients and specifically asks for their D ecem ber CPI forecast. periods. Based on these test results, the period from 1959-1978 does not appear to be one in which Livingston forecasters, on average, efficiently utilized the information contained in the history of observed inflation rates. Similarly, when the period of wage price controls is excluded, the efficiency criterion is not satisfied if the forecast error is calculated at the end of the fore cast period (e.g., in December). For instance, when the forecast error is measured at the end of the pe riod over which the forecast is made, the F-test per mits a rejection of the efficiency hypothesis at the 5 percent level.23 The efficiency hypothesis is not re23The efficiency hypothesis cannot be rejected, however, at the 1 percent level when the 8-month (April-December) fore cast horizon is employed. NOVEMBER F E D E R A L R E S E R V E B A N K O F ST . L O U I S jected only when the forecast error is evaluated in October (as in Mullineaux). 1980 _________________________ Table 4 Efficiency of the 12-Month Forecasts Most previous analyses of the Livingston inflation forecasts focus on the short-run (8-month) forecasts. Because the respondents are asked at each survey date to predict the level of the CPI for the following De cember and June, the forecasts embody both an 8month and a 14-month (long-run) prediction of the inflation rate. This section examines the rationality of the 14-month forecasts. Bias Test for 14-Month Inflation Forecasts1 F-Values Sample period June forecast December forecast Critical F (.05, .01) 1959-1969 16.130 20.800 4.26, 8.02 1959-1978 9.533 5.188 3.55, 6.01 10.599 4.592 3.63, 6.23 1959-19782 The methodology used here slightly modifies the approach used for the 8-month forecasts. Specifically, the lagged inflation rates in equation (4 ) are now interpreted as occurring over 12-month periods ( again, observed in either April or October). This assump tion requires that the estimation of these equations for the 14-month forecasts be modified. Because the forecasts are made at 6-month intervals, this new interpretation means that the first lagged term in equation (4) contains information that over laps from the previous period, if all available observa tions are included in the estimation procedure. Such overlapping observations may introduce serial corre lation into the equation.24 To avoid this problem, separate estimations of equations (2) and (4 ) are made for each semiannual observation of the 14month forecast; that is, each sample period is split into two data sets, one consisting only of the June forecasts and the other consisting only of the Decem ber forecasts. With these modifications, equations (2) and (4 ) are estimated for the three time periods used in the previous section. The analysis first examines the 14-month forecasts for bias. F-statistics were computed from the regres sions of equation (2) for each semiannual forecast series over each sample period. These F-values, re ported in table 4, again indicate that the forecasts are biased. Table 5 provides the statistics for Theil’s analysis of the forecast errors. These results also show that .33-54 percent of the forecast error is due to bias. Nevertheless, as with the “short-run” forecasts, the portion due to bias declines as new data are added. The efficiency test is then applied to the 14-month forecast errors. The forecast errors are consistently 24Introduction of serial correlation tends to bias the efficiency test toward rejecting the null hypothesis. Recall that an additional criterion for efficiency is that the estimation be free of autocorrelation. 'Test based on joint hypothesis that B» ~ 0 and B, — 1 in equation (2 ’). 2Oniits the 1971-1973 price control years. measured as of the end of the period over which the forecast was made. The F-statistics and the DurbinWatson statistics for these equations are reported in table 6.-5 In contrast to the 8-month ( April-December) inflation forecasts, the results for the 14-month forecasts do not permit rejection of the efficiency hypothesis. Because halving the sample period severely reduces the degrees of freedom, these results should be interpreted with considerable caution. Nevertheless, the F-statistics suggest that the errors in the 14-month forecasts are not correlated with observations of past inflation available at the time the forecast was made. The Durbin-Watson statistics, however, indicate that the hypothesis of no serial correlation can neither be rejected nor accepted. Thus it appears that, based on the F-test, the 14-month forecasts comply with the efficiency criterion. These contrasting results for the 8-month and 14month forecast horizons cast some doubt on the find ings that the Livingston forecasts are not formed efficiently. This disparity may indicate that forecasters are better able to anticipate longer-term movements in economic variables, such as inflation, relative to explaining the short-term vagaries of the time series. For instance, if the actual rate of inflation is accelerat ing within the 14-month period, the forecaster may be able to forecast efficiently the overall rate of change but not be able to forecast the rate within shorter sub-periods. 25The Durbin-h statistic is not appropriate for small samples (n < 3 0 ). On this point, see J. Johnston, Econom etric M eth ods, 2nd ed. (New York: McGraw-Hill, 1971). 9 F E D E R A L R E S E R V E B AN K O F ST. LO UI S NOVEMBER 1980 Table 5 Analysis of 14-Month Forecast Errors1 Forecast horizon assumption Sample period June December Mean error RMSE Theil statistics ---------------------------------------Um U* Uc 1959-69 1.120 0.824 0.540 0.337 0.123 1959-78 1.964 1.298 0.436 0.208 0.356 1959-782 2.022 1.383 0.468 0.222 0.310 1959-69 1.182 0.782 0.438 0.474 0.088 1959-78 2.085 1.190 0.326 0.198 0.477 1959-782 1.976 1.133 0.329 0.194 0.477 1RMSE is the root-mean-squared enror; Um is the Theil bias coefficient; Us the variance coefficient; and Uc the covariance coefficient. 2Omits the 1971-1973 price control years. Table 6 Efficiency Test Results: 14-Month Forecasts1 Sample period June forecasts D.W. 1959-69 11 0.426 December forecasts N 1.38 F D.W. F *(.05, .01) 1.344 2.25 5.05, 10.97 1959-78 20 1.049 1.88 2.029 1.87 2.96, 4.69 1959-782 18 0.875 1.34 1.993 2.30 3.11, 5.06 *N is the respective sample size; F is the calculated F-statistic; D.W. is the Durbin-Watson test statistic; and F * represents the relevant critical F-value. 2Omits the 1971-1973 price control years. Summary This paper has reexamined the rationality of the inflation forecasts contained in the Livingston survey data by emphasizing that the inflation forecast error should be calculated in a manner consistent with the forecast horizon used by the survey respondents. Specifically, empirical tests for bias and efficiency of the forecasts were employed to determine the effect that changes in the assumption about the forecast horizon have on the conclusions of previous investiga tions. The test for bias indicated that, regardless of the forecast horizon or the sample period used, the Livingston forecasts exhibited characteristics of bias. The “efficiency” test suggested by Mullineaux was also employed. These test results indicate that over 10 the period, 1959-1969, only one forecast horizon ( April-December) could be judged unambiguously inefficient. When the 1959-1978 period is examined, however, the results for each forecast horizon allow rejection of the efficiency hypothesis. When the period of wage-price controls is deleted from this sample period, only the April-October forecast horizon is judged efficient. These findings imply that conclusions regarding the forecast efficiency (and, therefore, rationality) of the Livingston inflation expectations are sensitive to the period over which the forecast error is evaluated. Because the survey respondents are asked specifically to predict the level of the CPI for the following June or December, it seems appropriate that tests of F E D E R A L R E S E R V E BAN K O F ST. L OUI S efficiency be formulated to measure the forecast error only after the actual value of the predicted CPI be comes known. When this approach is used in con junction with the assumption of either a May-December or April-December forecast horizon, the results indicate that the forecasters did not efficiently use the information available at the time of the survey in five out of six samples. This conclusion contrasts sharply with that reached when the forecast error is calcu lated at the time the forecasts are made ( i.e., April or October). Finally, evidence about the bias and efficiency of the 14-month forecasts indicates that these longer forecasts are efficient, even though, like the 8-month forecasts, they are apparently biased. Although the apparent disparity in the efficiency tests between the “short-” and ‘long-run” forecasts is somewhat puz zling, it suggests that the forecasters are more efficient NOVEMBER 1980 at predicting longer term inflation trends than short term movements in the series. The evidence presented here indicates that Carlson’s sample average forecasts of the rate of CPI inflation in the Livingston data do not conform to two criteria of rationality. Consequently, the use of these data in empirical investigations of rational expectations mod els appears to have serious limitations. In addition, the observation that these survey-based inflation ex pectations fail to conform to rationality criteria sug gests that adjustments in expectations evolve slowly. This further implies that, even if inflation forecasts are ultimately rational, fully anticipated short-run monetary policy actions may have important economic effects since inflation expectations adapt slowly. These and other possible implications of the apparent non rationality of survey-based expectations deserve fur ther study. We would like to thank Don Mullineaux and Doug Pearce for their helpful com ments on an earlier draft of this paper. Their contributions in no way imply complete agreement with the opinions expressed herein. 11 Monetary Aggregates as Monetary Indicators KEITH M. CARLSON and SCOTT E. HEIN T 1 HE monetary aggregates are being relied upon rently being taken by monetary authorities. more and more as indicators of the thrust of mone Early this year, the Federal Reserve Board an tary policy actions on aggregate economic activity.1 nounced a redefinition of the monetary aggregates. In To be useful as a monetary indicator, a monetary ag some cases, the differences between the old and new gregate should satisfy at least two criteria. First, it money measures are quite substantial. While the re must be sensitive to policy actions taken by the lationship between the old monetary aggregates and Federal Reserve — such as open market operations economic activity has received much attention in the and changes in reserve requirements, the discount economic literature, the usefulness of the new mone rate, and Regulation Q ceilings; it must not be sensi tary aggregates as monetary indicators has yet to be tive to influences other than Federal Reserve actions. examined in detail. This article reports some results If the monetary aggregate is responsive to nonpolicy bearing on this issue. forces, it will provide erroneous signals as to the thrust of monetary policy.2 The analysis focuses primarily on the relationship of the new MIA, M1B, and M2 measures to economic Second, a monetary aggregate should be both con activity. To provide historical continuity, the results sistently and predictably related to the pace of eco are compared with those derived from analyses of the nomic activity. If it is not, changes in the monetary old M l, M2, and M3 aggregates. aggregate will not “indicate” what will happen to aggregate economic activity as a result of actions cur THE NEW MONETARY AGGREGATES 1For a general discussion of monetary indicators, see Albert E. Burger, “The Implementation Problem of Monetary Policy,” this R eview (March 1971), pp. 20-30. 2This criterion explains why many argue against the use of market interest rates as monetary indicators. See Albert E. Burger, “The Implementation Problem . . . ,” where he argues that market interest rates are poor monetary indicators be cause they are sensitive to nonpolicy impulses, such as factors that affect the demand for credit. 12 Components of the new MIA, M1B, and M2 mone tary aggregates are listed in table l.3 MIA is identical 3For a detailed description of the new see R. W. Hafer, “The New Monetary view (February 1980), pp. 25-32; or “The Redefined Monetary Aggregates,” letin (February 1980), pp. 97-114. monetary aggregates, Aggregates,” this R e Thomas D. Simpson, F ederal Reserve Bul F E D E R A L R E S E R V E B A N K O F ST . L O U I S to old M l, except that it excludes demand deposits due to foreign commercial banks and official institu tions. The new M1B aggregate, a broader transaction measure, is equal to M l A, except that it includes newly developed interest-bearing transaction deposits. These latter deposits include negotiable order of with drawal (NOW ) accounts, automatic transfer system deposit (ATS) accounts, and credit union share drafts. NOW accounts were legalized in certain New England states early in the 1970s, and such legaliza tion will extend nationwide as of December 31, 1980.4 Commercial banks have been permitted to offer indi vidual ATS accounts since November 1, 1978. Chart 1 presents compounded annual rates of change of old M l, MIA, and M1B for the period II/ 1959 through IV/1979.5 The chart shows that the ex clusion of demand deposits held by foreign commer cial banks and institutions has had little effect on the growth rates of the monetary aggregates. Growth rates of new MIA closely resemble those of old M l. Furthermore, the growth rates of MIA and M1B differ little prior to early 1974 and, although M1B growth usually exceeds that of MIA over the period 1/1974 through III/1978, the disparity between these aggregates is quite small. It is only after the nation wide introduction of ATS accounts in late 1978 that the growth rates of these new aggregates show any marked divergence. While the new MIA and M1B measures are similar in scope to old M l, the new M2 measure is quite different from old M2. In fact, the new M2 measure is more closely related to the old M3, which included savings and small time deposits of thrift institutions; old M2 did not include such deposits. Because the monetary aggregates are no longer differentiated on the basis of institutional considerations, old M2 does not have a counterpart among the new measures. As shown in table 1, there is essentially only one component of the old M3 measure — large time de posits (other than large negotiable CDs) at commer cial banks and thrift institutions — that is not included 4For a description of the New England experience with NOW accounts, as well as a discussion of how their legalization will affect other parts of the country, see William N. Cox III, “NOW Accounts: Applying the Northeast’s Experience to the Southeast,” Econom ic R eview of the Federal Reserve Bank of Atlanta (September/October 1980), pp. 4-10; and Patrick J. Lawler, “NOW Accounts in the Southwest: A Break for Con sumers, an Entry from S&L’s, and a Test for Banks,” V oice o f the F ederal R eserve Bank o f Dallas (October 1980), pp. 5The historical series for the new monetary aggregates begins in 1/1959. NOVEMBER 1980 Table 1 The New Monetary Aggregates Component M1A M1B M2 Currency in circulation X X X Demand deposits at commercial banks and thrift institutions, exclusive of deposits due to foreign commercial banks and official institutions X X X X X NOW and ATS accounts and credit union share drafts Overnight RPs X Savings deposits at commercial banks and thrift institutions X Small time deposits (less than $100,000) at commercial banks and thrift institutions X Overnight Eurodollar deposits issued by Caribbean branches of member banks and held by U.S. nonbank residents X Money market mutual fund shares X in the new M2 measure. On the other hand, a num ber of the changes that have been made make new M2 even more comprehensive than old M3. In addi tion to the interest-bearing transaction deposits in cluded in M1B, the new M2 measure also includes overnight BPs at commercial banks, money market mutual funds, and overnight Eurodollar deposits is sued by Caribbean branches of member banks and held by U.S. nonbank residents.6 Chart 2 depicts the compounded annual rates of change of new M2, old M2, and old M3. Growth rates of the new M2 and old M3 aggregates were similar from the 11/1959 through 11/1973 period; growth rates of old M2, on the other hand, generally were much slower than these aggregates. The similarity in the growth rates of old M3 and new M2 breaks down in late 1973, however, when overnight BPs, money mar ket mutual funds, and the overnight Eurodollar de posit component of new M2 became increasingly popular. 6Timothy Q. Cook and Jeremy G. Duffield, “Short-Term In vestment Pools,” E conom ic Review of the Federal Reserve Bank of Richmond ( September/October 1980), pp. 3-23. The authors have recently argued that there are many other in vestment pools, similar to money market mutual funds, which should be included in the new M2 measure. 13 F E D E R A L R E S E R V E B A N K O F ST . L O U I S Chart 1 Compounded A n n u a l Rates of Change of M l(o ld ), M I A and M1B 14 L a te st d a ta p lo tte d : M l (o ld ) - 4 th q u a r te r 1 979; O th e rs-2 n d q u a rte r 1980 C h a rt 2 Compounded A n n u a l Rates of Change of L a te st d a ta p lo tte d : M 2(new )-2nd q u a rte r 1 98 0; O th e rs -4 th q u a rte r 1979 NOVEMBER 1980 F E D E R A L R E S E R V E B A N K O F ST . L O U I S NOVEMBER 1980 C h a rt 3 C om pounded A n n u a l Rates of C hange of M 2 ( n e w ) and M1B Percent 1959 6 0 Percent il 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 1980 S h a d e d a re a s re p re s e n t p e r io d s d u r in g w h ic h th e th re e -m o n th tre a s u ry b ill ra te w a s a t le a s t 100 b a s is p o in ts a b o v e th e c e ilin g ra te o n c o m m e rc ia l b a n k s a vin g s d e p o s its . L a te s t d a ta p lo tte d : 2nd q u a r te r 1 980 Finally, chart 3 presents the compounded annual rates of change of the new M1B and M2 aggregates. This chart illustrates the differential growth rates of narrow versus broad money measures.7 Note the dif ference in average growth rates; new M2 growth is usually above that of M1B. The average growth rate of new M2 over the 11/1959 through IV/1979 period is 8.4 percent, compared to 5.0 percent for M1B. The differential between the two growth rates sometimes varies. The chart indicates a definite pat tern in the relative growth rates. Over the periods II/1959-IV/1965, III/1970-I/1973, and I/1975-I/1978, growth rates of new M2 are substantially above those of M1B. In the intervening periods, the differential between growth rates of these two aggregates is very small. Historical experience indicates that the growth rate of the broad money stock measure is sensitive to the differential between market interest rates and Regulation Q ceilings. This is clearly indicated by the 7M1A is excluded for simplification purposes; prior to late 1978, quarterly growth rates of MIA were very similar to those of M1B (see chart 1). Further, while only the new aggregates are shown, old M l and M2 display a similar pattern. shaded areas in chart 3, which depict periods of two quarters or more during which the three-month treas ury bill rate was at least 100 basis points above the ceiling rate on commercial bank savings deposits.8 Redefining this broader monetary aggregate has not made it insensitive to nonpolicy influences. Nonpolicy factors that affect the supply or demand for credit and, as a result, change market interest rates will clearly influence the growth of new M2 just as they affected the growth of old M2 and M3. The sensitivity of new M2 to such nonpolicy factors thus reduces its usefulness as an indicator of monetary policy actions. THE RELATIONSHIP BETW EEN ECONOMIC ACTIVITY AND THE MONETARY AGGREGATES The relationship between economic activity and the new monetary aggregates is investigated with 8The chart indicates that the most recent period of disinterme diation, IV/1977-II/1980, has not had the same effect in re ducing new M2 growth relative to M1B as observed in pre vious periods of disintermediation. However, at least part of this phenomenon is explained by the rapid growth of over night RPs and Eurodollar deposit holdings and, more recently, by money market mutual funds. F E D E R A L R E S E R V E B A N K O F ST. L O U I S reference to nominal GNP. Nominal GNP is chosen because this is the apparent channel by which mone tary policy variables directly affect the economy.9 The general form of the relationship to be estimated is: (1 ) Yt — C + 2 mi M u 1=0 + Z e, E t l + |at I=0 where Y is the compounded annual growth rate of nominal GNP, M is the compounded annual growth rate of the given monetary aggregate, E is the com pounded annual growth rate of high-employment ex penditures, and p is a random error term.10 This re lationship is estimated using the new MIA, M1B, and M2 aggregates and the old M l, M2, and M3 measures. The relationships are estimated with the ordinary least squares estimation technique. The investigation subjects the six different relation ships to a number of statistical tests. The strategy is first to find the optimal lag structure for the different relationships over the sample period, III/1962 through III/1977. After investigating the in-sample stability of the relationships and the likelihood of simultaneous equation bias problems, these estimated relationships are then used to project nominal GNP over the post sample period, IV/1977 through IV/1979, to deter mine which relationship would have yielded the most accurate forecasts for this period. This period was chosen because of the divergent growth rates for the various aggregates, as shown in the preceding charts. Sample Period Relationships The first concern in estimating the general relation ship given in equation (1) is to determine the ap propriate values of f and g, the number of lags on the monetary and fiscal variables. Lag values of 0, 4, and 8 were considered for each of the six relation ships. Interestingly enough, F-tests for each of the equations indicated that the appropriate lag value was 9See Milton Friedman, “A Theoretical Framework for Mone tary Analysis,” in Milton F riedm an s Monetary Fram ew ork: A D ebate with His Critics, ed. R. J. Gordon (University of Chicago Press, 1974), pp. 1-63; and Charles R. Nelson, “Re cursive Structure in U.S. Income, Prices, and Output,” Journal o f Political Econom y (December 1979), pp. 1307-27. 10This relationship is similar to the original Andersen-Jordan equation. Such a relationship has been estimated more re cently by Keith M. Carlson, “Money, Inflation, and Eco nomic Growth: Some Updated Reduced Form Results and Their Implications,” this Review (April 1980), pp. 13-19. Usually, the relationship is estimated assuming that the lag coefficients lie along a polynomial of a given degree. No such constraints are imposed here. 16 NOVEMBER 1980 4 for each of the separate monetary aggregates, as well as for the fiscal variable. Table 2 provides the sample period coefficient esti mates and summary statistics for the six different equations, where the relationships are estimated with ordinary least squares and four lags on the fiscal and monetary variables are assumed. There is very little difference between the sample period fit provided by the various aggregates. In all cases, the standard error of the estimating equation (S E E ) is less than onethird the size of average GNP growth over the sample period (9.61 percent). While the pattern of the distributed lag effects of both the fiscal and monetary variables is similar across equations, the size of the coefficients is clearly de pendent on the comprehensiveness of the monetary aggregate employed. In general, the more comprehen sive the aggregate, the smaller the size of any lagged monetary coefficient. The sum of the money coeffi cients is close to 1.0 for both MIA and M1B.11 On the other hand, the sum of the money coefficients for new M2 is close to 0.7. Begardless of the aggregate used, the sum of the high-employment expenditures coeffi cients is close to zero. Stability Tests A question to be considered with these estimation results is whether the relationships reported in table 2 are structurally stable (i.e. whether the regression coefficients change significantly with time). The hy pothesis of structural stability was investigated with the use of the Chow test. The formal hypothesis tested is whether the regression coefficients estimated for the III/1967 through IV/1969 sample period differ significantly from those obtained for the same equa tion in the 1/1970 through III/1977 period. The null hypothesis is that the coefficients are equal in each of these periods. The midpoint of the sample was chosen as the breakpoint because it maximizes the power of the test.12 Table 3 lists the F-statistics for each of the various equations. None of the cases considered provide eviu The results reported for the narrow aggregates are similar to those found by Keith M. Carlson, “Money, Inflation and Economic Growth . . . ,” where a third degree polynomial with tail constraints was employed in the estimation. 12See John U. Farley, Melvin Hinich, and Timothy W. Mc Guire “Some Comparisons of Tests for a Shift in the Slopes of a Multivariate Linear Time Series Model,” Journal of Econom etrics (Volume 3, No. 3, 1975), pp. 297-318. F E D E R A L R E S E R V E BAN K O F ST. LO UI S NOVEMBER 1980 Table 2 Relationships Between GNP and The Monetary Aggregates Yt = C + I mi Mt-i + Z ei Et-i + jit 1=0 1=0 (Sample Period, III/1962-III/1977; absolute value of t-statistic in parenthesis) Monetary aggregates Old New M1 M2 M3 M1A M1B M2 C 2.94 (2.00) 0.56 (0.32) 0.90 (0.54) 2.35 (1.56) 2.20 (1.49) 1.30 (0.85) m0 0.58 (2.97) 0.38 (2.03) 0.16 (0.81) 0.61 (3.20) 0.60 (3.18) 0.13 (0.74) mi 0.02 ( 0.10) 0.14 (0.59) 0.26 (0.90) 0.03 (0.13) 0.05 (0.24) 0.39 (1.67) m2 0.20 (0.83) 0.12 (0.52) 0.08 (0.27) 0.32 (1-36) 0.31 (1.35) -0.07 (0.28) m» 0.56 (2.28) 0.43 (1.82) 0.49 ( 1.68) 0.36 (1.53) 0.38 (1.62) 0.44 (1.85) m< -0.54 (2.67) -0.19 (0.99) -0.30 (1.48) -0.35 (1.75) -0.35 (1.79) - 0.22 (1.28) e0 0.04 (0.89) 0.05 (0.98) 0.08 (1.74) 0.05 (1.18) 0.05 (1.19) 0.08 (1.95) e, 0.12 (2.67) 0.10 (2.15) 0.13 (2.80) 0.12 (2.55) 0.12 (2.59) 0.12 (2.80) 62 -0.07 (1.54) -0.07 (1.54) -0.07 (1.61) -0.08 (1.81) -0.08 (1.82) -0.08 ( 1.88) es 0.02 (0.42) 0.01 (0.23) 0.02 (0.34) 0.02 (0.35) 0.01 (0.31) 0.02 (0.43) e< XX1 ( 0.02) - 0.02 (0.43) - 0.01 (0.28) 0.01 (0.13) xxi (0.04) - 0.01 (0.23) Ra 0.49 0.46 0.49 0.48 0.49 0.51 SEE 2.96 3.05 2.95 2.98 2.98 2.90 DW 1.78 1.87 2.00 1.82 1.89 2.03 Mess than 0.005 Table 3 Calculated F-Statistics for Test of Break In Relationships (111/1962-IV/1969 vs. 1/1970-111/1977) Monetary aggregates Old F (1 1,39)1 New M1 M2 M3 M1A M1B M2 1.52 1.11 0.64 1.64 1.62 0.61 1The 5 percent critical level is 2.05; the 10 percent critical level is 1.73. r 17 NOVEMBER F E D E R A L R E S E R V E BANK O F ST LOUIS dence to reject the null hypothesis at traditional levels of significance.13 Simultaneous Equation Bias Tests A further question with regard to the estimation results reported in table 2 is whether or not they are subject to significant simultaneous equation bias. Equations such as those reported in table 2 can be estimated reliably with ordinary least squares methods only if the independent variables are exogenous. A major criticism of equations of this type is that the monetary aggregates are not exogenous with respect to GNP.14 Sims has recently suggested a test to examine whether the independent variables in a distributed lag relationship, such as equation (1 ), can be said to be statistically exogenous.15 The test procedure in volves adding leading values of the independent vari ables to the basic distributed lag equation. If the regression coefficients of the leading values of the 13A break in the relationship in 1/1974 was also considered. With the exception of the new M2 relationship, there is evi dence, at traditional levels of significance, to suggest a break in all the relationships. With regard to the inability to reject the stability of the new M2 relationship, it should be noted that none of the separate subperiod money coefficients dif fered from zero. The fact that all other equations break is evidence of the specification error. There appear to be two likely candidates for omitted variables. First, none of the relationships include a variable to capture the impact of the oil shock which oc curred near 1974. Second, there is no variable to capture a shift in money demand if, as many argue, money demand shifted in 1974. (For example, see Stephen M. Goldfeld, “The Case of the Missing Money,” Brookings Papers on Econom ic Activity (3 :1 9 7 6 ), pp. 683-730.) Since we are primarily concerned with the coefficients on the money variables, either of these specification errors will cause a problem only to the extent that the excluded vari able is correlated with the independent variables. It is only when such correlation exists that the estimated coefficients will be biased. Regardless of whether either or both of the above specification errors exist, it is unlikely that this bias problem will result. Both of the suggested specification errors resulted because shock variables were excluded. For evi dence of the “shock” view of money demand, see R. W. Hafer and Scott E. Hein, “The Dynamics and Estimation of Short-Run Money Demand,” this Review (March 1980), pp. 26-35. By definition, these shock variables should not be correlated with the included independent variables. The out-of-sample simulation results to be reported later in this aper indicate that there is little evidence of a significant ias in these simulations. ,4See Frank de Leeuw and John Kalchbrenner, “Monetary and Fiscal Actions: A Test of Their Relative Importance in Eco nomic Stabilization — Comment,” this R eview ( April 1969), pp. 6-11. 15Christopher A. Sims “Exogeneity and Causal Ordering in Macroeconomic Models,” in New M ethods in Business Cycle R esearch: Proceedings from a C onference (Federal Reserve Bank of Minneapolis, 1977), pp. 23-44. Digitized for18 FRASER 1980 independent variable are not different from zero, the null hypothesis of exogeneity is supported. On the other hand, statistical significance of leading coeffi cients suggests that simultaneous equation bias prob lems would result if the equation were estimated with ordinary least squares. To test for the presence of simultaneous equation bias, four leads on both the fiscal and monetary vari ables were added to the basic equation as follows: (2 ) Y, ~ C + £ in , M ,-, + 1=0 4 Z e , E, , + Zm! M „, 1=0 1 =1 . + Z e( E „ i + Ht. i-i Since the Sims test depends crucially on the statistical significance of regression coefficients, every effort was made to assure the absence of serially correlated error terms. This was accomplished by following Sims’ re commendation of filtering the data prior to estimation. In most cases, the filter employed was the first order linear filter (1-K L ), where L is the lag operation and K is a constant. The value of K was determined by iterating over values from 0 to 1, at intervals of 0.1. The first value of K which yielded no evidence of a relationship between the contemporaneous residual and residuals lagged, first two and then four periods, was chosen as the appropriate value.10 This search procedure removed the problem of serially correlated disturbances in all relationships ex cept that using old M l. In this case, the fourth lagged residual always remained statistically significant in an autoregressive error structure in the residuals. Thus, in the case of old M l, the filter employed was (1-KL4). Table 4 lists the F-statistics testing the null hypo theses; (1 ) mf = 0 for i = 1, 2, 3, 4; (2 ) e,' = 0 for i = 1, 2, 3, 4; and (3 ) m| = e[ = 0 for i = 1, 2, 3, 4. In none of the cases considered were F-statistics large enough to reject the null hypothesis at the 5 percent level, thus suggesting the absence of any simultaneous equation bias problems associated with the estimation results reported in table 2.17 16A similar search procedure was employed by Yash P. Mehra and David E. Spencer, “The St. Louis Equation and Re verse Causation: The Evidence Reexamined,” Southern E co nomic Journal (April 1979), pp. 1104-20. 17This conclusion is somewhat different than that obtained by Mehra and Spencer, “The St. Louis Equation. . . .” In esti mating a relationship similar to equation ( 1 ) , they found evidence of simultaneous equation bias problems. However, their study differed in three important ways. First, the only FEDERAL. R E S E R V E BAN K O F ST. LO UI S NOVEMBER 1980 Table 4 F-Statistics for Simultaneous Equation Bias Tests Monetary aggregates Null hypothesis1 (degrees of freedom) Old New M1 M2 M3 m,' = 0 (4,42)2 1.62 0.57 e,' = 0 (4,42)2 0.22 0.47 m,' = e,' = 0.94 0.10 0 (8,42)3 Value of K M1A M1B M2 0.85 0.23 0.25 2.20 0.12 0.14 0.13 0.27 0.37 0.44 0.23 0.23 1.18 0.10 0.10 0.20 0.20 0.10 *i = 1, 2, 3, 4 in all cases. 2The 5 percent and 10 percent critical values for F (4 ,4 2 ) are 2.60 and 2.09, respectively. 3The 5 percent and 10 percent critical values for F (8 ,4 2 ) are 2.17 and 1.82, respectively. Two qualifications to this conclusion are required. These qualifications concern the regressions employ ing old M l and new M2. While the F-statistics re ported in table 4 do not allow the rejection of the null hypothesis at the 5 percent level, there were individual lead money coefficients in these two cases that were different from zero at certain levels of sig nificance; thus, there is some evidence to reject the null hypothesis at lower significance levels. For ex ample, in the case of old M l, the regression coefficient on the one-quarter lead of money was 0.64. The tstatistic associated with this individual coefficient was 2.32, indicating that the estimate was statistically dif ferent from zero at the 5 percent level. In this regard, there is some evidence of “reverse causation” — an increase in economic activity “causing” an increase in future money growth.18 This result generates some concern about the regression estimates reported for the equation using old M l in table 2. It is interesting to note that the redefinitions of the monetary aggregates, although not directly concerned with this simultaneity problem, have done much to resolve it. None of the individual leading money coefficients were close to being statistically different from zero when the MIA aggregate was employed. Together, these findings suggest that the simultaneous equation bias, to the extent it exists, is due to the inclusion of demand deposits held by foreign institu tions or commercial banks. In the case of new M2, the coefficient on the money variable led two quarters was -0.50; and its abso lute t-statistic of 1.83 was significantly different from zero at the 10 percent level. In addition, the joint hypothesis that all leading money coefficients are zero had to be rejected at the 10 percent level. This again suggests the possibility of a simultaneous equation bias problem. However, it is important to recognize that the problem does not appear to be a result of a ;positive association between current economic activ ity and future money growth, as traditionally sug gested. Rather, in this case, this regression coefficient suggests that current economic activity is negatively associated with new M2 growth two quarters in the future.19 This negative relationship should not come as a sur prise in light of the evidence of the impact of disin termediation on new M2 growth. An increase in eco nomic activity, by causing market interest rates to rise above Regulation Q ceilings, will be associated, other things being equal, with a reduction in future new M2 growth. In summary, it appears that the redefinitions of the monetary variable they consider is the monetary base. Sec ond, they include high-employment receipts, as well as highemployment expenditures, in their relationship. Finally, they focus on a diiferent time period (I/1952-1V/1974). 18More formally, if one were willing to use the 25 percent significance level, the null hypothesis that the leading M l coefficients are equal to zero must be rejected. 19In this regard, it is to be noted that when old M3 is used, the coefficient on money variable led two quarters is also negative. However, the coefficient is not different from zero even at the 10 percent level. Thus, it appears that including overnight RPs, overnight Eurodollars, and money market mutual funds in new M2 has compounded the simultaneity problem. 19 F E D E R A L R E S E R V E B A N K O F ST . L O U I S NOVEMBER 1980 Table 5 Summary Measures of Out-of-Sample (IV/1977-IV/1979) GNP Predictions Monetary aggregates Old New M1 M2 M3 -2.17 -2.67 -2.60 4.05 4.55 4.62 0.29 0.34 0.32 due to variation 0.29 0.36 0.31 0.20 0.47 0.17 due to covariation 0.42 0.29 0.38 0.64 0.53 0.56 Mean error Root mean squared error M1A M1B M2 -0.02 -2.57 4.04 3.68 5.03 0.16 0.00 0.26 -1.61 Fraction of error1 due to bias lNeed not sum to unity as a result of rounding. monetary aggregates have removed possible problems associated with simultaneity as far as the narrow transaction aggregates are concerned. However, there still remains a question concerning simultaneity with regard to the more comprehensive measure. Prediction Results How well do the relationships presented in table 2 simulate nominal GNP over the IV/1977 through IV/1979 period? Table 5 indicates that the equation using the new M1B aggregate performs the best in simulating GNP growth over this period, regardless of the criteria considered. The strength of this equa tion is most evident in the lack of bias in the pre dictions. The other aggregates underpredict GNP growth over this period, on average, by approximately 2.5 percent. In comparison, the average prediction error for M1B is a trivial -0.02 percent. It is also appropriate to note that the bias in pre diction errors is smaller for new MIA than for old M l. Removing demand deposits held by foreign com mercial banks and institutions did not reduce the variance of forecast error; it did, however, reduce the average error and the bias in the forecast. The fact that the more comprehensive monetary aggregates (old M2, old M3, and new M 2), which include savings deposits subjected to Regulation Q ceilings, underpredict GNP growth by more than the transaction aggregates is again consistent with the Digitized for2 0FRASER view that disintermediation has adversely affected the growth of these deposits. The whole period from IV/1977 through IV/1979 has been characterized by market interest rates well above Regulation Q ceilings. This has led to a relative slowing in the growth of these regulated deposits. As a result, equations using these aggregates have underpredicted economic ac tivity since IV/1977. SUMMARY The monetary aggregates were redefined early this year. The purpose of this article was to examine these new aggregates in terms of their usefulness as mone tary policy indicators. Two criteria for judging the usefulness of the monetary aggregates as indicators were suggested. First, to serve as an indicator, the aggregate should reflect the policy actions of the monetary authority and not be highly sensitive to nonpolicy influences. Second, the aggregate should be consistently and predictably related to economic activity. Although the first criterion was not considered for mally, examination of the rates of change of the new monetary aggregates indicated that redefining M2 did not remove the influence of nonpolicy forces. In par ticular, the movement of market interest rates relative to Regulation Q ceilings has had an adverse effect on new M2 growth (relative to the narrowly de fined aggregates), as it did with the old M2 and M3 aggregates. F E D E R A L R E S E R V E BANK O F ST. LO UI S The second criterion was examined more extensively by regressing nominal GNP growth on the growth of the various monetary aggregates and a fiscal variable ( growth rates of high-employment expenditures). These relationships were checked for structural sta bility, simultaneous equation bias, and out-of-sample prediction accuracy. Of the new monetary aggregates, only M2 showed any evidence of simultaneous equa tion bias. This problem is felt to be closely related to the impact of Regulation Q ceilings. In out-ofsample simulations, M1B performed better than any of the other new aggregates analyzed, indicating that it had a closer relationship to economic activity than did the other new aggregates. In light of the criteria suggested for judging the usefulness of the new monetary aggregates as mone NOVEMBER 1980 tary indicators, M1B was thus found to best satisfy these requirements. It appears to be relatively insen sitive to nonpolicy influences ( a characteristic it shares with M IA ), and it is more predictably and consistently related to movements of nominal GNP than MIA or new M2. On the other hand, new M2 was found to be par ticularly unreliable as a monetary indicator. Growth in this aggregate was found to be sensitive to non policy forces. While proposed actions under the Finan cial Institutions Deregulation and Monetary Control Act of 1980 should eventually resolve this type of problem, new M2 growth will likely remain a poor monetary indicator in the seven-year transition period, especially in light of the absence of any reliable his torical relationship with economic activity. 21 EIGHTH F E D E R A L .. R E S E R V E DI S T R I C T v ■ / Ol B e d fo rd Q S e y m o u r^ p * - fV in c e n n e s ^ M a d is o n j^ ^ * ^ O W ashington C e n tr a lia o M o u n t V e rn o n M u rp h y s b o r o Q M a r io n V D a n v ille ^ o* J /E N S B O R O O E liz a b e th to w n C a rb o n d a le . V C a p e G ir a r d e a u ^ f- V M a d is o n v ille q ^ \i\£ O P o p la r B lu ff o v IFAYETTEV1LLE-SPRINGDALE ^ P a ra g o u ld o<L.__■ # J o n e s b o ro / <§> ^ D y e rs b u rg O Q H u m b o ld t B ly th e v i o G,as9°w M u rra y O O U n io n C ity J ack s o n U SM ITH CMA R K S V IL L E - H Q P K IN S V IL L E q M a y fie ld / ^ B o w lin g G re e n flHHi S ik e s t o n ^ B - p a d u c a h Legend © H e a d O f f ic e ® B r a n c h O f f ic e s o f th e FRB, St. Louis o f th e FRB, St. Louis I S ta n d a rd M e tr o p o lita n S ta tis tic a l A re a s O P laces H P laces o f 4 0 -5 0 ,0 0 0 (§) P la c e s o f 3 0 -4 0 ,0 0 0 ® ov e r 5 0 ,0 0 0 R u s s e llv ille Q Conw ay F o rre s t C ity o LIT' C o r in th O W e s t H e le n a .O S tu ttg a rt <§> H o t S p rin g o Helena^ LUFF O x fo r d .^ C la r k s d a le O T u p e lo • P laces o f 10 -2 0 ,0 0 0 O P laces o f 2 0 -3 0 ,0 0 0 P o p u la tio n is b a s e d on th e 1970 census. / • > C am den \R K A N A o < — S ta te B o u n d a rie s C le v e la n d ° ^ • ^ G r e e n v ille M a g n o lia o El D o ra d o .*§> District States — D is tric t B o u n d a ry G re e n w o o d Colum busjl S t a r k v ille Q ® ' 32 64 961 S ca le in M ile s F e d e ra l R eserve B a n k o f St. L o u is DECEMBER 1975