The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
o r K in g r a p e r s e rie s Evidence on Structural Instability in Macroeconomic Times Series Relations J a m e s H . S to c k a n d M a rk W . W a ts o n 3 ■) W o rk in g P a p e rs S e rie s M a c ro e c o n o m ic Is s u e s R e s e a rc h D e p a rtm e n t F e d e ra l R e s e rv e B a n k o f C h ic a g o S e p te m b e r 1 9 9 4 (W P -9 4 -1 3) 1 iBR A S Y OCT 0 6 1994 F DR L RS R E E E A EE V 9ANK Of CHICAGO FEDERAL RESERVE BANK O F CHICAGO Evidence on Structural Instability in Macroeconomic Time Series Relations James H. Stock Kennedy School of Government, Harvard University and the NBER and Mark W. Watson Northwestern University, Chicago Federal Reserve Bank and the NBER July 1994 *We are grateful to Neil Shephard for helpful discussions and participants at the 14th Annual International Symposium on Forecasting in Stockholm, Sweden for useful comments. This research was supported by National Science Foundation grant no. SES-91-22463. Abstract An experiment is performed to assess the prevalence of instability in univariate and bivariate macroeconomic time series relations and to ascertain whether various adaptive forecasting techniques successfully handle any such instability. Formal tests for instability and out-ofsample forecasts from sixteen different models are computed using a sample of 76 representative U.S. monthly postwar macroeconomic time series, constituting 5700 bivariate forecasting relations. The tests indicate widespread instability in univariate and bivariate autoregressive models. However, adaptive forecasting models, in particular time varying parameter models, have limited success in exploiting this instability to improve upon fixed-parameter or recursive autoregressive forecasts. Keywords: forecasting; time-varying parameters; structural stability; breakpoints; recursive least squares 1 Introduction . Time series econometrics typically involves drawing inferences about the present or future using historical data. In some cases these inferences are about how the economy operates or how economic policy affects key variables. For example, much empirical work on monetary economics currently rests on inferences drawn from so-called structural vector autoregressions (VAR’s); Bemanke and Blinder (1992) and Christiano, Eichenbaum and Evans (1993) provide two recent examples. In other cases these inferences are in the form of forecasts. Both applications typically require that the model at hand be stable (that the future be like the past) for such inferences to be valid. For example, giving advice to current policy makers based on a structural VAR requires, among other things, that the historically estimated model remains relevant today. Although studies occasionally include some analysis of stability, it is often limited in scope, perhaps consisting of reestimating the model on a single subsample. The importance of stability and the current lack of systematic evidence on it therefore leads us to ask, how generic is instability in multivariate time series relations? To answer this, we undertake a two-part experiment. The first part assesses the prevalence of parameter instability in economic time series relations using a battery of recently developed tests for instability. This is done using a sample of 76 monthly time series for the postwar U.S. economy over the period 1959:1 - 1993:12 (420 observations), among which are 5700 distinct (although not independent) bivariate forecasting relations. These series are chosen to provide relations which are representative of those of interest to macroeconomists and macroeconomic forecasters. This sample is then used to compute empirical distributions of various tests for structural stability, including Nyblom’s (1989) test for parameter stability, CUSUM tests, and break point tests such as the Quandt (1960) likelihood ratio statistic. The second part of the experiment examines whether current state-of-the-art adaptive forecasting models capture the instability found by the stability tests and thereby improve upon more naive forecasts. This entails the empirical evaluation of different forecasting models - 1- which exhibit different degrees of adaptivity, ranging from no adaptivity (fixed parameter models) through moderate adaptivity (recursive least squares, rolling regression, and random walk coefficient time varying parameter (TVP) models with small coefficient evolution) to high adaptivity (TVP models with large coefficient evolution). Although work on regression models with stochastically time varying parameters (or "stochastic coefficients") dates to Cooley and Prescott (1973a, 1973b, 1976), Rosenberg (1972, 1973), and Sarris (1973), and although TVP models been applied to selected series, we know of no systematic evidence on whether these techniques might be widely useful in economic forecasting applications. * Eight univariate models are considered for each of the 76 series for a total of 608 univariate forecasting equations, and eight bivariate models are considered for each of the 5700 bivariate forecasting relations for a total of 45,600 bivariate forecasting equations. Models are compared using onemonth-ahead mean square errors (MSE’s). In the spirit of Makridakis et al. (1982) and Meese and Geweke (1984), who applied univariate forecasting techniques to large numbers of series, this part of this experiment yields a forecasting comparison suggesting which models typically do best in macroeconomic applications. The results also provide an opportunity to assess model robustness by identifying models which successfully guard against the most severe out-ofsample forecasting failures. Looking ahead to the results, the tests indicate that instability is widespread. For example, one version of the Nyblom (1989) test rejects stability (at the 10% level) in more than 70% of the 5,700 bivariate relations. This instability is more prevalent in certain classes of series, such as measures of aggregate output, than in others. However, our results also suggest that forecasting models explicitly designed for time varying parameters (rolling regressions or random walk TVP models) often fail to perform as well as traditional fixed coefficient or recursive least squares models: in 57% of the 5700 pairs, fixed-coefficient or recursive least squares forecasts have the lowest out-of-sample MSE among the sixteen competing models, while in only 10% of the pairs do bivariate TVP models have the lowest out-of-sample MSE. When they are best, the gains associated with the TVP forecasting models typically are small. - 2 - In a small fraction of the cases, the TVP and recursive least squares models perform well while the fixed coefficient models perform quite poorly, and in this sense the TVP and recursive least squares models are more robust than the fixed coefficient models. Overall, however, the TVP models fail to exploit successfully the time variation found by the stability tests. The outline of the paper is as follows. Section 2 describes the data set. The stability tests are described in section 3, and section 4 summarizes the empirical testing results. The forecasting models are described in section 5 and are evaluated empirically in section 6. Section 7 concludes. 2. The Data Set Our objective in constructing the data set was to obtain a sample of economic time series for the U.S. which is representative of the relations of primary concern to macroeconomists and macroeconomic forecasters. While one could in principle draw series at random from a large macroeconomic database, a simple random sample would oversample certain classes of heavily represented series, such as industry-specific deflators, interest rates, or financial flows. Such a sample would be representative of the monthly data which are produced but not of the forecasting relations of interest to macroeconomists. Moreover, such a sample would omit important forecasting variables which are constructed from the primary data, such as interest rate spreads. In theory, stratification could eliminate the problem of oversampling certain classes of series which are produced in detail, but would not address the issue that many of the important forecasting series are constructed by researchers and thus are not contained in standard databases. Moreover, mechanical simple or stratified sampling would produce many series with definitional changes or other internal inconsistencies. Our sample of series therefore was obtained by applying subjective judgment, using four criteria as guidelines: -3 - 1. The sample should include the main monthly economic aggregates and coincident indicators. This resulted in the inclusion of series such as industrial production, weekly hours, personal income and inventories. 2. The sample should include important leading economic indicators. This led us to include series such as monetary quantity aggregates, interest rates, interest rate spreads, stock prices, and consumer expectations. 3. The series should represent different broad classes of variables which can be expected to have quite different time series properties. 4. The series should have consistent historical definitions or, when the definitions are inconsistent (for example different base years for different segments of a real series) it should be possible to adjust the series with a simple additive or multiplicative splice. These criteria were used to select 76 monthly U.S. economic time series. Most of the raw data were obtained from the CITIBASE data base, although many series were subsequently modified (for example by creating interest rate spreads). The series can be grouped into eight categories: output and sales; employment; new orders; inventories; prices; interest rates; money and credit; and other miscellaneous series including exchange rates, government spending and taxes, and miscellaneous leading indicators. The complete list of series and their mnemonics are given in the appendix. The sample runs from 1959:1 to 1993:12. The starting date was chosen because this is the earliest date for which many of the series, in particular the monetary aggregates, are available. Four series (the series on government finance) start in 1967:6. The statistics in question were computed using the longest possible sample for which all relevant data were available. -4- Each series was screened to detect breaks and outliers due to changes in definitions or reporting practice. Most series were also transformed to induce approximate stationarity by taking either first differences or first differences of logarithms. For consistency, the stationarity transformation was in general applied to entire classes of series rather than on a case-by-case basis. For example, production, employment, prices, and money were all transformed using first differences of logarithms, and interest rates were transformed by first differencing. Some series which did not fit naturally into a broader category were analyzed on a case-by-case basis using visual inspection, a-priori reasoning, and unit root test statistics, and then transformed accordingly. The transformation for each series is listed in the appendix. It should be emphasized that many of the procedures are only slightly affected by the use of first differences vs. levels. In particular, the forecasting models in section 5 produce similar shortrun forecasts using levels or first differences (this would not be the case if many of the series were cointegrated, but there are neither theoretical nor empirical reasons to suspect widespread cointegration among these series). 3. Description of Stability Tests The empirical analysis uses variants of three classes of tests for parameter stability: tests for random (time-varying) coefficients; tests based on cumulative forecast errors (CUSUM tests); and tests based on sequential Wald tests for a single break. For completeness, we briefly summarize these tests here, although details are available in the original references. For additional discussion of these tests see the review by Stock (1994) and for additional references to stability tests more generally see Hackl and Westlund (1989, 1991). The general model considered is, (1) yt = + “t ^ t - l + ^ t ^ xt-l + e t -5- where a t(L) and j8t(L) are p-th order lag polynomials which in general are time varying and where et is serially uncorrelated with mean zero and variance < Let k denote the total ?". number of regressors. Each test has as its null hypothesis that the parameters are constant, that is, « t(L)=a(L) and j8t(L)=j8(L). The derivation of the null distributions of the test statistics also assumes that the regressors are jointly second order stationary, along with additional technical conditions. When the discussion below refers to univariate tests, it is understood that the terms in xt_j in (1) are omitted. A. Tests for time-varving parameters The first set of tests for randomly time-varying coefficients are Nyblom’s (1989) locally most powerful tests against the alternative that the coefficients follow a random walk, where the random walk error is independent of et. Nyblom (1989) derived his statistic against the alternative that all the coefficients are stochastic, and this requires some modification since we also test subsets of coefficients. Rewrite the regression (1) as yt = 0t’zt + et in obvious notation. Suppose under the alternative that q < k linear combinations of 0t follows a random walk; that is, R0t = R0t_j + ijt, where R is a q x k matrix of constants which are either known or can be consistently estimated under the null, and the modified Nyblom statistic is L = T 9 is i.i.d. and uncorrelated with et. Then j. i n r £ t 4 - St, where St = R ^ 5 _ jZ ses, where {es} are the residuals from OLS estimation of (1), and where ^ = (RT"* £ ^ _ i Z tzt’R’)o^, where The test is evaluated for three different sets of coefficients: (2a) L ^ : test nv «t(L), 0t(L); test nv 0t(L); (2b) (2c) L test /zt, j3t, where /3t(L)=/Stb(L) and b(l) is normalized so b(l) = l. -6- For Lgjj, R is the k x k identity matrix; for Rzt = (1, xt_ j,...,x t_p)’; and for q, R is the matrix of ones and zeros such that ^ y R is the matrix such that Rzt = (1, S(L)xt_j), where 6(L)=$(L)/$(1) where $(L) is the OLS estimator of /3(L) under the null. The statistic tests for stochastic evolution of the cumulative effect of xt on the forecast of yt. Heteroskedasticity-robust variants of the Nyblom statistics were also computed by replacing P with V = RT~* X ^ = i e^ztzt’R’ (Hansen (1990)). The heteroskedasticity-robust versions of the tests in (2a), (2b) and (2c) are respectively denoted by iJjj, I_£ and L ^ g ^ . Also computed were two variants of the Breusch-Pagan (1979) LM test for random coefficients, for which the alternative hypothesis is that the coefficients are i.i.d. draws from a distribution with constant mean and finite variance. The two statistics are, (3a) 2 2 2 2 BPjji = TR from the regression of e^ onto (1, yj_j,...,yj_p,Xj_j,...,Xj_p); (3b) BP^j = TR^ from the regression of e^ onto (1, x^_j,...,Xj_p); (3c) 2 BP^g(i) = TR^ from the regression of e^ onto (1, (6(L)xt_j)^). B, Tests based on cumulative forecast errors One of the tests based on cumulative forecast errors is the maximal OLS CUSUM statistic proposed by Ploberger and Kramer (1992), which is similar to Brown, Durbin and Evans’ (1975) CUSUM statistic except that the Ploberger-Kramer (1992) statistic is computed using OLS rather than recursive residuals. Let f-p(5) = £ ^ ] e s, where [•] is the greatest lesser integer function. The Ploberger-Kramer (1992) maximal CUSUM statistic is, (4) PKsup = suPS E [0,1] I I• A related statistic is the mean square of f p (5) PKmsq = J -7- 2 The PKSU and PKmSq statistics respectively have limiting representations as the supremum and p the integral of the square of a one-dimensional Brownian bridge. C. Tests based on sequential Wald statistics The third set of tests statistics consists of functionals of the sequence of Wald test statistics, F t (5), which test the null hypothesis that the parameters are constant against the alternative that they have a single break at a fraction 8 through the sample. The break date is treated as unknown a-priori, so that the tests involve computing the sequence F^(t/T) for t= tQ ,...,tj, and then computing a functional of this sequence. Three such functionals are considered. The Quandt (1960) likelihood ratio (QLR) statistic, in Wald form, is given by (6) QLR = suP5e(50,5i)FT ^ - The mean Wald statistic (Hansen (1992), Andrews and Ploberger (1992)) is (7) MW = f ^ F T(5)d5. The Andrews-Ploberger (1992) Wald statistic is the exponential average, (8) APW = ln{ J |jexp(lAFT(5))dS}. These statistics have asymptotic representations as functionals of a k-dimensional Brownian bridge; see Andrews (1993) and Andrews and Ploberger (1992) for the details. The tests are implemented with 15% symmetric trimming (5q =1-5 j =.15). -8- 4. Stability Tests: E m pirical Evidence A. Univariate Tests The values of the univariate stability test statistics, along with summary statistics on the fraction of rejections, are given for all 76 series in table 1. The final column contains the regression F statistic testing the hypothesis that the transformed series follows an AR(0). The first panel reports summary measures of rejections for each series. The second panel reports each of the individual test statistics. For all regressions, p=6. The answer to the question of whether there is evidence of widespread instability in these univariate autoregressions evidently depends on which stability test one uses. On the one hand, 50% of the series reject at the 5% level using the QLR statistic, and similar results obtain for the APW statistic. There are also many, if fewer, rejections using the MW statistic. These results provide evidence of one-time shifts in the parameters of the univariate autogressions. While the Breusch-Pagan (1979) test often rejects, this test also has power against heteroskedasticity so it is not clear whether this indicates heteroskedasticity or time variation in the parameters. The Nyblom test has a lower rejection rate (20% at the 5% level); when an adjustment is made for heteroskedasticity, the rejection rate drops to the level of the test. However, because the random walk time variation introduces heteroskedasticity, the heteroskedasticity-robust Lr statistics seems likely to have lower power than the L statistics, so the drop in significance from the L to Lr statistics need not be interpreted as evidence against time variation in the data but rather simply as evidence that the adjusted test has lower power. The rejection rates for the PK CUSUM statistics are low, suggesting that shifts in the intercept are not a major feature in these data. The instability is more heavily concentrated in certain classes of series than others. For example, the QLR statistic rejects at the 5% level for all interest rate and inflation series. In contrast, other than the Breusch-Pagan test which could be detecting heteroskedasticity, none of the tests rejects for business failures, the government finance series, and several of the orders and inventories series. -9- B. Bivariate Tests There are 5700 bivariate forecasting relations among our 76 series so rather than present all of the test statistics we present various graphical and tabular summaries. Summary rejection rates of the bivariate tests for parameter stability are presented in table 2. The final column reports the Granger causality Wald statistic testing the hypothesis that /3(L)=0 in (1). For all regressions p= 6, and all tests have level 10%. The results for these 5700 bivariate relations are summarized in panel A of table 2. The main feature is the evidence of widespread instability in these relations, although this instability is only detected by a subset of the tests. The q statistic rejects in over 70% of the cases, and its heteroskedasticity-robust variant rejects in almost 60% of the cases. The QLR and APW statistics also reject in approximately 60% of the cases. A large fraction (58%) of cases also have significant Granger causality statistics, a result which is perhaps surprising since no apriori economic reasoning was used to select which variables should be used to forecast any particular dependent variable. As in the univariate results, the CUSUM-based tests have low rejection rates, which suggests that the instability does not arise from breaks or drift in the direction of the mean regressors. As the final two rows of panel A indicate, there is only slightly more instability among statistically significant predictive relationships (based on the Granger causality test) than among insignificant relationships. These results can be used to examine stability in relations involving those variables which commonly appear in structural VAR modeling. Industrial production, real personal income, manufacturing employment, the CPI, the PPI, the 90-day Treasury bill rate, and the commercial paper-Treasury bill spread each reject stability in at least 93% of their 75 respective bivariate relations based on either the relations based on the ^ or QLR statistics, and M l rejects in 77% of its bivariate ^ statistic, when these series are used as dependent variables (panel B). When these series appear as predictor variables (panel C), for each the QLR rejects in at least 59% of the 75 pairs. For five of the seven price series, the QLR statistic rejects stability - 10- in each of the 75 bivariate forecasting relations in which inflation is a dependent variable. When any of these five price series is instead used as a predictor, the QLR statistic again rejects in more than half the cases. If anything, it appears that instability in bivariate relations involving these key series is even more prevalent than on average across all 5700 relations. These marginal distributions provide one window on the extent of instability in these 5700 relations. However, it is possible that some of this instability is in relations which would be of little interest from a forecasting perspective because they have low overall predictive content. Exploring this possibility requires examining the joint distribution of the instability and Granger-causality test statistics. This is done graphically in figures 1-4, which are scatterplots of selected stability test statistics against the Granger causality test statistic. These figures confirm that the stability and Granger causality test statistics are only weakly correlated. In a sense, each forecasting relation can be thought of as having a temporal average level of predictive content, and deviations from that predictive relation over time are largely uncorrelated with the average predictive content. Figure 5 summarizes the estimated break dates ([TS], where S maximizes F^(5)) for the bivariate relations for the which the corresponding QLR statistics are significant at the 5 % level. Instability is concentrated around 1974-75, 1980-81, and at the endpoints (5q and in (6)). Figure 6 is a scatterplot of the QLR vs. the APW statistics. Although these are rather different functionals of the sequential Wald statistics, respectively the maximum and an exponential average, the statistics are clearly highly correlated and give quite similar inferences in these data. Evidently little is lost by considering only one or the other of these statistics. The large number of rejections when the Granger causality statistic is insignificant presents an intriguing opportunity. This is most easily seen for the QLR statistic. If there is in fact no predictive content in a bivariate relation in any subsample, then the fraction of QLR rejections will tend towards the level of the test. In contrast, a rejection by the QLR test but not by the Granger causality test suggests that the bivariate relation has predictive content in at least one - 11 - continuous subsample, even though on average over the full sample it does not (ignoring type I errors). For example, oil prices (pw561) was a useful predictor for only 23% of the other series, but 61 % of the corresponding QLR tests rejected. This raises the possibility that models which adapt to changing relations might find new, albeit transitory, forecasting relations to exploit. 5. D escription of Forecasting M odels The second stage in this investigation is an examination of the performance of sixteen fixed and adaptive forecasting models. Eight of the forecasting models are univariate while eight are bivariate. Throughout, a (pseudo) in-sample estimation period is used for preliminary estimation of the parameters and a (pseudo) out-of-sample period is used for forecasting. The eight univariate models consist of a fixed-parameter autoregression, two autoregressions estimated by rolling regression, one autoregression estimated by recursive least squares, and four random walk TVP models. The eight multivariate models are a fixed-parameter bivariate model, two bivariate models estimated by rolling regression, one model estimated by recursive least squares, and four bivariate models with random-walk time TVP. All models are of the form (1), with the coefficients fixed or time-varying as appropriate. The bivariate models will be referred to as vector autoregressions (VAR’s), although because only one-step ahead forecasts are considered only the single equation (1) of the (yt, xt) VAR needs to be estimated. The specification of the TVP models is conventional and assumes the coefficients follow a random walk. Let 81 = (/xt, a j t,..., «pt, j8jt,..., /3pt) (where application). Then, (9) - 12- are omitted in the univariate 9 where 1^ is the k x k identity matrix, so that X is the ratio of the variance of the parameter disturbance to the variance of the regression error et. The parameters of the TVP models are O , < ", and X. For each TVP model the value of 0q is set to its OLS estimate over the inq P sample period, and <" is estimated using the in-sample data. (The out-of-sample forecasts and ? their relative performances are insensitive to choice of initial conditions because of the long insample period.) One-step ahead forecasts are then produced using period-by-period updating with the Kalman filter. We consider four TVP models that differ only in their choice of X. So that a single value of X could be applied to series measured in different units, all series were first rescaled by dividing through by their in-sample standard deviation. The eight univariate models are: (10a) AR: AR(6); n, a(L) estimated by OLS, then fixed at in-sample values for outof-sample forecasts (10b) RRA1: AR(6) estimated using rolling regression with 120 observations (10c) RRA2: AR(6) estimated using rolling regression with 240 observations (10d) RLSA: AR(6) estimated by recursive least squares (lOe) ATVP1: AR(6) estimated by TVP with X = .005 (100 ATVP2: AR(6) estimated by TVP with X = .010 (10g) ATVP3: AR(6) estimated by TVP with X = .020 (10h) ATVP4: AR(6) estimated by TVP with X = .030. The eight bivariate models are: (Ha) VAR: VAR(6); ix, cc(L), /3(L) estimated by OLS, then fixed at in-sample values for out-of-sample forecasts (11b) RRV1: VAR(6) estimated using rolling regression with 120 observations (He) RRV2: VAR(6) estimated using rolling regression with 240 observations - 13- (lid ) RLSV: (He) VTVP1: VAR(6) estimated by TVP with X = .005 (Ilf) VTVP2: VAR(6) estimated by TVP with X = .010 d ig ) VTVP3: VAR(6) estimated by TVP with X = .020 (llh ) VTVP4: VAR(6) estimated by TVP with X = .030. VAR(6) estimated by recursive least squares Two related models are univariate and bivariate TVP models with X estimated over the insample period. These models are not examined because this estimation is currently computationally impractical for the large number of forecasting relations under consideration. For the forecast comparisions, the in-sample period ends in 1978:12. This cutoff date was chosen so that the models are tested in the turbulent economic conditions of the late 1970’s and early 1980’s. For series ending in 1993:12, 180 observations remain for the out-of-sample comparison. 6. Forecasting Model Comparison: Empirical Results Comparing the various models using these data entails examining 608 univariate forecasting systems (76 variables, eight models each) and 45,600 bivariate forecasting systems (5700 bivariate forecasting relations, eight models each). All comparisons are made using out-ofsample one-month-ahead forecast MSE’s, although in principle other loss functions could be used. The term "best model" will be used to refer to the model which minimizes this out-ofsample forecast MSE, relative to some comparison group. One objective of this comparison is to see which models do best most frequently. However, because of the instability found in section 4, another objective is to ascertain which if any of the models protect the forecaster from making extreme forecast errors resulting from parameter instability. The question of which model performs best out-of-sample most frequently is examined in table 3. For each bivariate relation, MSE’s from the eight bivariate and eight univariate models - 14- were computed; for the purposes of this tabulation, the model producing the lowest out-ofsample MSE among these sixteen was then deemed the "best" model for that (yt, xt) pair. Two sets of tabulations are presented. Panel B presents the fraction of times the column model is best among the 75 bivariate relations, broken down by forecasted variable. For example, for industrial production, in 16% of the 75 bivariate pairs, RRV2 produces the smallest out-ofsample MSE; in 61% of these pairs, ATVP1 outperforms not just the other seven univariate models but also the eight bivariate models. Panel C presents analogous results, broken down by forecasted variable, except that for each forecasted variable the comparison is only among the top ten of the 75 pairs, as measured by the Bayes Information Criterion (BIC) for the in-sample OLS estimation of (1) with fixed parameters. Thus, among forecasts of IP based on the ten variables with the lowest in-sample BIC’s, in one case (10%) RRV2 has the lowest out-ofsample MSE, but in 7 cases ATVP1 outperforms the other univariate and the eight bivariate models. The first two rows of panel A respectively summarize the results of panels B and C, where the fractions are computed over all the forecasted variables. The final row of panel A presents results for bivariate relations with significant time variation, as measured by significance (at the 10% level) of the ^ statistic evaluated for the in-sample period. Several conclusions are evident from table 3. Overall, there is no clearly dominant model; no model performs best in more than 17% of the 5700 pairs. However, there is strong evidence that the adaptive models (the rolling, recursive and TVP models) outperform the two fixedparameter models. Among the set of models with predictors based on the top 10 BIC’s, 73% of the best-performing models are adaptive. However, the extreme TVP models (with X=.02 and X=.03) are rarely the top performers out of sample. Interestingly, comparing the final row of panel A with the first two rows indicates that the adaptive models perform similarly whether or not in-sample instability is detected. Consistent with the stability test evidence, the results in panels B and C show that different variables tend to be forecast best by different models. For example, exchange rates (exnwt2) have no rejections using the univariate stability tests but a moderately high fraction of rejections using the bivariate Nyblom tests, and in 88% of the 75 - 15- pairs the best forecasting model is the fixed-coefficients AR. In contrast, real personal income (gmyxp8) has widespread rejections by univariate and bivariate stability tests, and in 97% of the cases the best forecasting models are adaptive, primarily the univariate or bivariate recursive least squares models. Table 4 summarizes pairwise comparisons of the sixteen models over all 5700 bivariate relations. The ATVP3 and ATVP4 models typically have MSE’s worse than the other univariate models; the VTVP3 and VTVP4 models also typically perform worse than the other bivariate models. Among the univariate models, only RLSA and ATVP1 outperform the simple AR in more than 50% of the cases, and RLSA in turn outperforms ATVP1 in 63% of the cases. Among the bivariate models, RLSV outperforms all others at least 64% of the time. The RLSA and ATVP1 models typically outperform all bivariate models. While the univariate models often outperform the bivariate models, this is perhaps not surprising since a-priori reasoning would lead one to suspect that many of the 5700 pairs would have forecasting links which are weak at best. Table 4 and the test results from section 4 provide additional evidence of instability. While 58% of the 5700 Granger causality statistics reject at the 10% level, only 38% of the recursive VAR forecasts (the best-performing bivariate forecast) outperform the recursive AR model out of sample. Presumably some of these Granger causality rejections are Type I errors, but with 75% power against the "true" VAR’s and a 10% level test these 58% rejections correspond to 74% of the VAR’s having nonzero coefficients on the predictor variable. The much lower fraction of pairs for which bivariate techniques improve performance out of sample thus is another indication of instability in the bivariate relations. It is useful to go beyond these assessments of which model typically performs best to quantify the extent to which the various models reduce the possibility of extremely poor performance. Table 5 presents the empirical quantiles of the MSE’s of the various models. To make results comparable across series, the MSE’s are relative to the MSE for the recursive least squares AR forecast (RLSA). Panel A shows the distribution of these relative MSE’s for the - 16- univariate forecasting models. The median values all exceed 1.00, consistent with the finding in Table 4 that RLSA has lower MSE than the other forecasts more than 50% of the time. The results also suggest that RLSA is the most robust univariate forecasting model, in the sense that its worst performance is significantly better than the worst performance of the other models. For example, the minimum relative MSE for the fixed coefficent AR model is .959 while its maximum relative MSE is 1.158. Thus, at its best, the AR forecast outperforms the RLSA forecast by 4.1 %, while at its worst, the AR forecast underperforms the RLSA forecast by 13.6% (=1-(1.158)"*). Panel B presents quantiles for the 5700 bivariate forecasts. Robustness at the a ’th quantile can be determined by comparing the relative MSE at a to the inverse of the relative MSE at the (l-a )’th quantile. For example, the relative MSE of VTVP1 at <*=.001 is .600, so that in . 1% of the cases, VTVP1 outperforms RLSA by at least 40%. At <*=.999, the relative MSE of VTVP1 is 1.146, so that in .1% of the cases, RLSA outperforms VTVP1 by more than 12.7% (=1-(1.146)~*). In this sense, in the .1% extremes, VTVP1 produces better forecasts than RLSA. The table indicates that at the . 1% quantile, all of the bivariate models dominate RSLA, while at c*=.5% and 1%, only bivariate models with small time variation (RLSV and VTVP1) dominate RLSA. Similar results obtain at c*=1% for the best 10 BIC-selected models shown in panel C. 7. Conclusions Some caveats are warranted. The stability tests are all evaluated using asymptotic critical values and some finite-sample size distortions are to be expected. However, Monte Carlo results for sequential Wald tests in Diebold and Chen (1992) suggest that these distortions are only moderate and could plausibly account for only a small fraction of the empirical rejections. Also, the quantitative results of the forecasting comparison depend on the choice of out-ofsample period. - 17 - With these caveats in mind, these results suggest some general observations about the two time series applications laid out in the introduction, structural VAR modeling and forecasting. One finding relevant to VAR modeling is that relations involving key macroeconomic variables such as industrial production, personal income, employment, prices, and interest rates are, if anything, more likely to be unstable than the average of these 5700 bivariate relations. While most VAR modeling involves more than two variables, a stable multivariate VAR implies a stable VAR for any subset of variables, so these tests can be seen as testing an implication of the hypothesis of stability of larger models. While it is possible for some (but not all) equations in a multivariate VAR to be stable despite instability of the bivariate VAR formed from two of the variables, impulse responses involve all equations in the VAR so instability in the bivaraite relations implies instability in at least some multivariate impulse responses. One practical lesson which this emphasizes is the importance of performing systematic stability analysis as part of a structural VAR modeling exercise. Given the widespread evidence of structural instability, the comparison of the forecasting models yielded some surprising results. While the fixed-parameter models occasionally worked very poorly, models with only small degrees of adaptivity performed well. In particular the univariate and bivariate recursive least squares models typically were either best or nearly best. In 38% of the cases, the recursive VAR outperformed the recursive AR out of sample; in 99% of the 5700 bivariate relations, the recursive VAR model produced an out-of-sample MSE which was at most 7.8% higher than the recursive AR model. Moreover, in 99% of the bivariate relations, the recursive VAR model produced an out-of-sample MSE which was at most 8.7% higher than the best-performing model for each bivariate relation (including the univariate models) and in 50% of the cases its MSE was within 1.4% of the best model. Of course, if the parameters are in fact constant, then the recursive estimator will be efficient relative to the fixed parameter models. One striking result is that the more adaptive models such as rolling regression or the TVP models typically failed to improve upon recursive least squares, and indeed did not even guard against extreme failures as well as did recursive least - 18 - squares. This negative finding suggests that the class of models considered here, which are the models most widely studied in adaptive forecasting, are largely unsuccessful in modeling and exploiting the instability we found in typical macroeconomic applications. - 19 - Appendix: Definitions of Series The entries for each series are the series mnemonic, the transformation code, and the definition of the series. For series obtained from CITIBASE, the CITIBASE mnemonic has been used. The transformation codes are: 0 = first difference, 1 =log first difference, 2 = level. A. Output and Sales ip 1 index of industrial production ipxmca 2 capacity util rate: manufacturing,total(% of capacity,sa)(frb) gmpy 1 personal income: total (bil$,saar) gmyxp8 1 personal income (real) less transfers rtql 1 retail trade: total (mil.87$)(s.a.) gmcq 1 personal consumption expenditure:total (bill. 1987$) ipcd 1 industrial production: durable consumer gds (1987=100,sa) ced87m 1 personal consumption expenditures:durable goods,87$ xci 1 stock-watson index of coincident indicators mt82 1 manuf. and trade sales B. Employment lpmhuadj 1 total employee hours (adjusted) lphrm 2 avg. weekly hrs. of production wkrs.: manufacturing (sa) lhell 1 index of help-wanted adv. lhnaps 1 persons at work: part time eas-slack wk,nonag (thous,sa) luinc 2 avg wkly initial claims,state unemploy.ins. ,exc p.rico(thous;sa) lhu5 1 unemploy.by duration: persons unempl.less than 5 wks (thous.,sa) lhur 0 unemployment rate: all workers, 16 years & over (%,sa) lhelx 2 employment: ratio; help-wanted ads: no. unemployed elf C. New Orders hsbp 2 housing authorized: index of new priv housing units (1967=100;sa) mdu82 1 mfg unfilled orders: durable goods industries, 82$ mpcon8 1 contracts & orders for plant & equipment in 82$(bil$,sa)! 2 mocm82 1 mfg new orders: consumer goods & material,82$(bil$,sa)! 2 mdo82 1 mfg new orders: durable goods industries,82$(bil$,sa)! 2 ivpac 2 vendor performance: % of co’s reporting slower deliveries(%,nsa) pmi 2 purchasing managers’ index (sa) pmno 2 napm new orders index (percent) D. Inventories invmt87 1 manufacturing & trade inventories:total,87$(bil$,sa) invrd 1 inventories, retail (sa) invwd 1 inventories, wholesale (sa) ivmld8 1 mfg inventories: materials & supplies, all mfg indus 87$(sa) ivm2d8 1 mfg inventories: work in process, all mfg indus 87$(sa) ivm3d8 1 mfg inventories: finished goods, all mfg industries 87$(sa) ivmtd 1 manufacturing & trade inventories: total ivmld 1 mfg inventories: materials & supplies, all mfg indus (mil$,sa) -2 0 - ivm2d 1 mfg inventories: work in process, all mfg indus (mil$,sa) ivm3d 1 mfg inventories: finished goods, all mfg industries (mil$,sa) invrd8 1 inventories, retail 87$ (sa) invwd8 1 inventories, wholesale 87$ (sa) E. Prices gmdc 1 pce,impl pr defl:pce (1987=100) punew 1 cpi-u: all items (82-84 = 100,sa) pw 1 producer price index: all commodities (82=100,nsa) pw561 1 producer price index: crude petroleum (82=100,nsa) pw561r 1 pw561/punew jocci 1 dept, of commerce commodity price index joccir 1 jocci/punew F. Interest Rates fyff 0 interest rate: federal funds (effective) (% per annum,nsa) fygm3 0 interest rate: u.s.treasury bills,sec mkt,3-mo.(% per ann,nsa) fygm6 0 interest rate: u.s.treasury bills,sec mkt,6-mo.(% per ann,nsa) fygtl 0 interest rate: u.s.treasury const maturities,l-yr.(% per ann,nsa) fybaac 0 bond yield: moody’s baa corporate (% per annum) fygtlO 0 interest rate: u.s.treasury const maturities, 10-yr.(% per ann,nsa) cp6_gm6 2 yield on 6 month commercial paper - fygm6 gl0_gl 2 fytlO - fygtl gl0_ff 2 fygtlO - fyff baa_gl0 2 fybaac - fygtlO G. Money and Credit fcbcuc 2 change in bus and consumer credit outstand.(percent,saar)(bcdl 11) fcbcucy 2 fcbcuc-annual percentage growth in GMPY delinqcr 0 delinq. rate, total install, credit cci30m 0 consumer instal.loans: delinquency rate,30 days & over, (%,sa) fmld82 1 money stock: m-1 in 1982$ (bil$,sa)(bcd 105) fm2d82 1 money stock: m-2 in 1982$(bil$,sa)(bcd 106) fmbase 1 monetary base, adj for reserve req chgs(frb of st.louis)(bil$,sa) fml 1 money stock: ml(curr,trav.cks,dem dep,other ck’able dep)(bil$,sa) fm2 1 money stock:m2(ml+o’nite rps,euro$,g/p&b/d mmmfs&sav&sm time dep(bil$, fm3 1 money stock: m3(m2+lg time dep,term rp’s&inst only mmmfs)(bil$,sa) fmbaser 1 monetary base: fmbase/punew H. Other Variables exnwt2 1 Trade weighted average nominal exhange rate fspcom 1 s&p’s common stock price index: composite (1941-43 = 10) fspcomr 1 fspcom/punew fail 1 business failures: current liabilities (mil$,nsa) failr 1 fail/punew gfosa 1 federal government outlays seasonally adjusted gfrsa 1 federal government receipts seasonally adjusted gfor 1 Real federal government outlays, gfosa/punew gfrr 1 Real federal government receipts, gfrsa/punew hhsntn 2 u. of mich. index of consumer expectations(bcd-83) -21 - Footnotes 1. Applications of adaptive forecasting include Baudin, Nadeau, and Westlund (1984), Guyton, Zhang and Foutz (1986), Engle, Brown and Stem (1988), Sessions and Chatteijee (1989), Schneider (1991), Young, Ng, Lane and Parker (1991), Zellner, Hong and Min (1991), Edlund and Sdgaard (1993), Min and Zellner (1993), and the time-varying VAR’s developed in Doan, Litterman and Sims (1984), Highfield (1986), and Sims (1982, 1993). Surveys of TVP models are provided by Chow (1984), Nichols and Pagan (1985), Engle and Watson (1987), and Harvey (1989). -22- References Andrews, D.W.K. (1993), "Tests for Parameter Instability and Structural Change with Unknown Change Point," Econometrica, 821-856. Andrews, D.W .K., I. Lee and W. Ploberger (1992), "Optimal Changepoint Tests for Normal Linear Regression," Cowles Foundation Discussion Paper #1016. Baudin, A., S. Nadeau, and A. Westlund (1984), "Estimation and Prediction Under Structural Instability: The Case of the U.S. Pulp and Paper Market," Journal o f Forecasting, 3, 63-78. Bemanke, B.S. and A.S. Blinder (1992), "The Federal Funds Rates and the Channels of Monetary Transmission," American Economic Review, 82, 901-921. Breusch, T.S. and A.R. Pagan (1979), "A Simple Test for Heteroskedasticity and Random Coefficient Variation," Econometrica, Vo. 47, pp. 1287-1294. Brown, R.L., Durbin, J. and Evans, J.M. (1975), "Techniques for Testing the Constancy of Regression Relationships over Time with Comments," Journal o f the Royal Statistical Society, Series B, 37, 149-192. Chow, G.C. (1984), "Random and Changing Coefficient Models," in Z. Griliches and M.D. Intrilligator, eds., Handbook o f Econometrics, North Holland, Amsterdam, Chapter 21. Christiano, L., M. Eichenbaum, and C. Evans (1993), "The Effects of Monetary Policy Shocks: Evidence from the Flow of Funds," manuscript, Department of Economics, Northwestern University. Cooley, T.F. and E.C. Prescott (1973a), "An Adaptive Regression Model," International Economic Review, 14, pp. 364-371. Cooley, T.F. and E.C. Prescott (1973b), "Tests of an Adaptive Regression Model," Review o f Economics and Statistics, 55, pp. 248-256. Cooley, T.F. and E.C. Prescott (1976), "Estimation in the Presence of Stochastic Parameter Variation," Econometrica, 44, pp. 167-184. Diebold, F.X. and C. Chen (1992), "Testing Structural Stability with Endogenous Break Point: A Size Comparison of Analytic and Bootstrap Procedures," manuscript, Department of Economics, University of Pennsylvania. Doan, T., R. Litterman and C. Sims (1984), "Forecasting and Conditional Projection Using Realistic Prior Distributions," Econometric Reviews, Vol. 3, 1-100. Edlund, P.-O. and H.T. S6gaard (1993), "Fixed versus Time-Varying Transfer Functions for Modelling Business Cycles," Journal o f Forecasting, 12, 345-364. Engle, R.F. and M.W. Watson (1987), "The Kalman Filter Model: Applications to Forecasting and Rational Expectations Models," in Advances in Econometrics: Fifth World Congress o f the Econometric Society, T. Bewley (ed.), Cambridge University Press, Cambridge, U.K. -23- Engle, R.F., S. Brown and G. Stem (1988), "A Comparison of Adaptive Structural Forecasting Methods for Electricity Sales," Journal o f Forecasting, Vol. 7, 149-172. Guyton, D.A., N-F Zhang, and R.V. Foutz (1986), "A Random Parameter Process for Modeling and Forecasting Time Series," Journal o f Time Series Analysis, 7, 105-115. Hackl, P. and A. Westlund (1989), "Statistical Analysis of ’Structural Change’: An Annotated Bibliography," Empirical Economics, 143, 167-192. Hackl, P. and A.H. Westlund (eds) (1991), Economic Structural Change: Analysis and Forecasting, Springer-Verlag, Berlin. Hansen, B.E. (1990), "Lagrange Multiplier Tests of Parameter Instability in Non-linear Models," manuscript, Department of Economics, University of Rochester. Hansen, B.E. (1992), "Tests for Parameter Instability in Regressions with 1(1) Processes," Journal o f Business and Economic Statistics, 10, 321-336. Harvey, A.C. (1989), Forecasting, Structural Time Series Models, and the Kalman Filter, Cambridge: Cambridge University Press. Highfield, R.A. (1986), "Forecasting with Bayesian State Space Models," Unpublished Ph. D. thesis, University of Chicago, Graduate School of Business. Makridakis, S., A. Anderson, R. Carbonne, R. Fildes, M. Hibon, R. Lewandowski, J. Newton, E. Parzen, and R. Winkler (1982), "The Accuracy of Extrapolation (Time Series) Methods: Results of a Forecasting Competition," Journal o f Forecasting, 1, 111-153. Meese, R. and J. Geweke (1984), "A Comparison of Autoregressive Univariate Forecasting Procedures for Macroeconomic Time Series," Journal o f Business and Economic Statistics, 2, 191-200. Min, C. and A. Zellner (1993), "Bayesian and non-Bayesian Methods for Combining Models and Forecasts with Applications to Forecasting International Growth Rates," Journal o f Econometrics, 56, pp. 89-118. Nichols, D.F. and A.R. Pagan, (1985), "Varying Coefficient Regression," in E.J. Hannan, P.R. Krishnaiah and M.M. Rao, eds., Handbook o f Statistics, Vol. 5, North Holland, Amsterdam, pp. 413-449. Nyblom, J. (1989), "Testing for the Constancy of Parameters Over Time," Journal o f the American Statistical Association, 84, 223-230. Ploberger, W. and W. Kramer (1992), "The CUSUM test with OLS Residuals," Econometrica, 60, 271-286. Quandt, R.E. (1960), "Tests of the Hypothesis that a Linear Regression System Obeys Two Separate Regimes," Journal o f the American Statistical Association, 55, 324-330. -24- Rosenberg, B. (1972), "The Estimation of Stationary Stochastic Regression Parameters ReExamined," J o u r n a l o f th e A m e r ic a n S t a t is t ic a l A s s o c ia t io n , 67, pp. 650-654. Rosenberg, B. (1973), "The Analysis of a Cross-Section of Time Series by Stochastically Convergent Parameter Regression," A n n a ls o f E c o n o m ic a n d S o c ia l M e a s u re m e n t, 2, 461484. Sarris, A.H. (1973), "A Bayesian Approach to Estimation of Time Varying Regression Coefficients," A n n a ls o f E c o n o m ic a n d S o c ia l M e a s u re m e n t, 2, 501-523. Schneider, W. (1991), "Stability Analysis Using Kalman Filtering, Scoring, EM, and an Adaptive E M Method," in Hackl, P. and A.H. Westlund (eds), E c o n o m ic S t r u c t u r a l C h a n g e : A n a ly s is a n d F o r e c a s t in g , Springer-Verlag, Berlin. Sessions, D.N. and S. Chatteijee (1989), "The Combining of Forecasts Using Recursive Techniques with Non-stationary Weights," J o u r n a l o f F o r e c a s t in g , 8,239-251. Sims, C.A. (1982), "Policy Analysis with Econometric Models," B r o o k in g s A c t iv it y , 1 107-164. , P a p e r s o n E c o n o m ic Sims, C.A. (1993), "A Nine-Variable Probabilistic Macroeconomic Forecasting Model," in J.H. Stock and M.W. Watson ( ds.), B u s in e s s C y c le s , I n d ic a to r s a n d F o r e c a s t in g , University of e Chicago Press for the NBER, 179-204. Stock, J.H. (1994), "Unit Roots, Structural Breaks, and Trends," in R.F. Engle and D. McFadden (eds), T h e H a n d b o o k o f E c o n o m e t r ic s , V o l. I V . Amsterdam: Elsevier. Young, P.C., C.N. Ng, K. Lane, and D. Parker (1991), "Recursive Forecasting, Smoothing and Seasonal Adjustment of Non-stationary Environmental Data," J o u r n a l o f F o r e c a s t in g , 10, 57-89. Zellner, A., C. Hong and C. Min (1991), "Forecasting Turning Points in International Output Growth Rates Using Bayesian Exponentially Weighted Autoregression, Time-Varying Parameter, and Pooling Techniques," J o u r n a l o f E c o n o m e t r ic s , 49, pp. 275-304. - 2 5 - Table 1 Univariate Tests for Stability A. Summary: Percent rejections over all series T e s t ----- Test Size Lall Lall PKsup PKmsq 10% 5% 1% 28.9 19.7 9.2 10.5 5.3 0.0 21.1 10.5 0.0 11.8 6.6 0.0 BPall QLR 72.4 65.8 55.3 53.9 50.0 30.3 MW APW 34.2 25.0 14.5 52.6 42.1 27.6 F 98.7 97.4 96.1 B. Results for Individual Series Series Sample PK sup Lall PK msq BP all QLR MW APW F A. Output and Sales ip 59:2 93:12 1.72* 1.06 0.83 0.19 18.61*** 21.94** 8.74 6.70 ipxmca 59:1 93:12 1 .8 8 * 1.17 0.81 0.13 27.73*** 19.87* 9.30 7.05* 14.22*** 2450.44*** gmpy 59:2 93:12 3.58*** 1.14 1.23* 0.45** 145.24*** 56.81*** 20.98*** 25.51*** gmyxp8 59:2 93:12 5.46*** 1.54 1 .0 2 0.25 133.62*** 70.15*** 30.76*** 32.03*** rtql 59:2 93:12 0.98 0.72 0.60 0.04 43.37*** 16.66 gmcq 59:2 93:12 1.83* 1.58 1.23* 0.37* 47.51*** 24.47** ipcd 59:2 93:12 1.50 1.09 0 .6 6 0.09 87.19*** 23.61** ced87m 59:2 93:10 3.02*** 1.73* 0.54 0.05 55.30*** 46.60*** xci 59:3 93:12 1 .6 8 1.20 0.87 0 .2 1 23.47*** 18.33 9.63 5.90 20.08*** mt82 59:2 93:12 1.32 1.17 0 .6 8 0.07 10.85* 13.93 8 .00 4.95 2.06* 6.70 12.17** 9.77 20.44*** 3.04*** 3.95*** 5.20 4.68*** 8.50** 2.63** 7.90** 18.47*** 1.76 5.37*** B . Employment lpmhuadj 59:2 93:12 1.48 1.26 0.56 0.07 15.84** 17.53 9.20 5.92 6.27*** lphrm 59:1 93:12 0.93 1.33 1.30* 0.24 27.65*** 20.92* 5.61 5.10 317.38*** lhell 59:2 93:12 1.85* 1.70* 0.44 0.03 13.72** 30.11*** lhnaps 59:2 93:12 1.26 1.08 0.56 0.05 32.22*** 11.96 luinc 59:1 93:12 1.41 1.17 1.04 0.24 56.08*** 22.91** lhu5 59:2 93:12 0.99 0.96 0.90 0.16 10.89* 12.36 lhur 59:2 93:12 1.25 1.10 0.78 0.04 18.23*** 15.85 8 .2 2 5.13 lhelx 59:1 93:12 1 .1 2 0.76 0.89 0.17 62.00*** 25.58** 8.36 8 21.81*** 11.30* 11.52*** 8.29 4.60 10.07 7.77* 5.64 3.68 .2 0 ** 18.43*** 3.07*** 1 2 2 0 .6 6 *** 14.76*** 9.68*** 6275.90*** C. Hew Orders hsbp 59:1 93:12 1.45 1.13 0.62 0.06 mdu82 59:2 93:12 2.23** 1.44 0.79 0.17 mpcon8 59:2 93: 9 0.58 0.42 0.58 0.05 24.72*** mocm82 59:2 93: 9 1.03 0 .8 6 0.69 0.08 11.31* mdo82 59:2 93: 9 1.47 1.39 0.69 0 .1 0 ivpac 59:1 93:12 1.31 1.11 0.89 0 .1 2 pmi 59:1 93:12 0.98 0.90 0.87 0.16 7.72 11.49 5.99 3.74 504.26*** prano 59:1 93:12 1 .0 1 0.85 0.65 0.08 4.76 12.32 6.25 4.15 225.84*** 3.54 21.95** 10.96* 7.57* 24.38** 14.08** 9.83** 1005.56*** 55.09*** 9.66 4.14 2.59 17.85*** 15.50 7.13 4.99 3.15*** 3.61 28.64*** 9.95 9.50** 15.17** 22.27** 8.84 7.73* 8.26*** 690.03*** D. Inventories invmt87 59:2 93: 9 1 .1 2 0.98 0.92 0.24 4.90 12.60 6.26 4.09 invrd 59:1 93: 9 1.33 1.24 0.87 0.24 9.04 20.34* 7.62 5.87 invwd 59:1 93: 9 1.82* 1.32 ivmld8 59:2 93: 9 3.46*** 2 ivm2 d8 59:2 93: 9 0.97 ivm3d8 59:2 93: 9 1.55 11.41* .8 8 *** 4.47*** 1 .1 2 0.23 25.33*** 2 2 1.13 0.23 38.93*** 38.58*** 1.01 0.97 0.27 6 .0 0 13.66 5.41 4.35 25.43*** 1.52 1.33** 0.59** 4.35 19.44 9.89 6.44 7.67*** .2 0 ** .1 2 ** 2 0 19.54*** 8.28** 14.39*** 12.63*** 19.83*** T a b le Series Sample Lall Lall PK sup 1 , PK msq c o n tin u e d BPall i i 59:1 93: 9 0.98 0.91 1.03 0.23 15.26** 12.44 ivrald 59:1 93: 9 3.70*** 2.08** 0.94 0 .2 0 60.32*** 36.31*** ivmtd m QLR 6 .2 0 21.71*** APW 3.66 14.87*** F 71.58*** 55.00*** 13.28 5.06 3.84 51.62*** 1 0 .1 1 19.43 9.47 6.30 24.58*** 0.06 14.87** 15.40 8.32 5.69 4.92*** 0.16 23.71*** 19.54 10.37* 7.50* 3.30*** ivm2 d 59:1 93: 9 0.99 0.99 1.01 0 .2 1 ivm3d 59:1 93: 9 1.45 1.37 1.48** 0.37* invrd8 59:2 93: 9 1.38 1.22 0 .6 8 invwd8 59:2 93: 9 1.77* 1.23 1.00 7.18 E. Prices gmdc 59:2 93:12 1.40 1.05 0.77 0.18 45.15*** 14.54*** 18.49*** 42.31*** punew 59:2 93:12 2.44*** 1.42 0.73 0.16 18.06*** 44.09*** 2 0 .6 6 *** 18.10*** 73.50*** 0 .66 0.95 0.23 113.95*** 93.39*** 19.78*** 41.02*** 17.65*** 0.71 1.27* 0.28 54.89*** 56.30*** 13.08** 24.13*** 11.06*** pw 59:2 93:12 2.35** pw561 59:2 93:12 2 .2 2 ** 4.28 pw561r 59:2 93:12 1.91** 0.64 1.15 0.23 47.17*** 49.08*** 10.61* 20.37*** 10.24*** jocci 59:2 93:11 1.09 0.92 0.61 0.08 22.54*** 25.88** 7.23 9.22** 21.14*** joccir 59:2 93:11 0.98 0.92 0.82 0.07 16.85*** 22.50** 6.45 7.61* 19.25*** F. Interest Rates fyff 59:2 93:12 1.15 0.49 1.18 0.14 32.68*** 39.82*** 15.13*** 15.05*** fygm3 59:2 93:12 1.15 0.54 1.41** 0 .2 1 71.15*** 34.74*** 9.70 12.45*** 19.94*** fygm6 59:2 93:12 1.32 0.57 1.36** 0 .2 1 73.86*** 32.45*** 10.50* 11.49*** 19.75*** fygtl 59:2 93:12 1.05 0.50 1.36** 0 .2 0 87.81*** 30.78*** 9.03 10.53*** 23.86*** 27.23*** 11.53** fybaac 59:2 93:12 0.76 0.56 1.42** 0.27 74.80*** 27.13*** 6.25 8.80** fygtlO 59:2 93:12 0.83 0.50 1.41** 0.24 46.68*** 25.64** 6.90 8.16** cp6 _gm6 59:1 93:12 0.65 0.39 0.76 0.07 145.60*** 34.04*** 5.25 11.38*** 17.91*** 218.52*** glO_gl 59:1 93:12 1.18 1.16 1.27* 0.28 26.07*** glO_ff 59:1 93:12 1.45 0.45 0.65 0.06 114.79*** 49.81*** 11.56** 19.43*** 216.13*** baa__glO 59:1 93:12 1.51 1.59 1.14 0.37* 13.04** 26.78*** 12.54** 9.64** 936.64*** 22.48** 9.37 7.23* 921.55*** G. Money and Credit fcbcuc 59:1 92:11 0.95 0.80 1.05 0.25 fcbcucy 59:1 92:11 1.61 1.65 1.10 0.52** 1.52 1.45 0.98 0.16 1.19 1.03 0.99 0.17 delinqcr 59:2 93: cci30m 6 59:2 93: 9 fmld82 59:2 93: 9 1.93** 1.65 fm2d82 59:2 93: 9 2.08** 2 fmbase 59:2 93:12 1.79* fml 59:2 93:12 15.11 5.70 4.65 149.90*** 17.79 8.69 6.14 15.71*** 6.89 12.49 8.97 4.74 5.27 19.09 7.52 6.13 5.33*** 6.4 7 *** 6 .2 1 29.15*** 1.23* 0.27 26.96*** 12.79** 9.46** 23.05*** 0.97 0.24 1.59 26.43** 14.98*** 10.08*** 50.49*** 1.87* 1.31* 0.58** 6 .1 1 19.19 10.60* 1.42 1.33 1.18 0.62** 15.38 8.51 .0 0 ** 24.52*** 39.42*** 7.18* 15.69*** 4.94 21.47*** fm2 59:2 93:12 1.06 1.07 1.32** 0.37* 3.25 18.13 fm3 59:2 93:12 2.44*** 1.80* 1.11 0.33 3.98 24.97** 15.56*** 10.17*** 141.76*** frabaser 59:2 93:12 H. Other Variables 1.91** 1.53 1.30* 0.23 3.83 29.64*** 13.12** 12.54*** 19.72*** exnwt2 59:2 93:12 0.90 0.67 0.99 0 .1 1 15.07** 16.86 6.63 5.46 9.33*** fspcomr 59:2 93:12 1.14 1.04 0.99 0.16 18.60*** 14.72 7.42 5.45 7.49*** 6.91 5.68 73.20*** fspcom 59:2 93:12 1.15 1.07 0.96 0.18 15.32** 14.99 7.36 5.47 6.93*** fail 59:2 93:12 1.25 1.06 0.63 0.06 11.06* 16.56 7.92 5.91 36.23*** 36.49*** failr 59:2 93:12 1.25 1.05 0.62 0.05 11.53* 16.56 7.91 5.93 gfosa 67:8 93:10 1.0 1 0.85 0.91 0 .2 2 43.70*** 11.07 6.43 3.78 4.29*** gfrsa 67:8 93:10 1.00 1.12 0.89 0.18 10.07 5.85 3.34 11.94*** gfor 67:8 93:10 0.81 0 .6 8 0.49 0.04 10.21 5.07 3.27 4.88*** gfrr 67:8 93:10 1.04 1.20 0.55 0.06 10.72 5.82 3.31 11.46*** hhsntn 59:1 93:12 1.95** 2.03** 0.84 0.15 4.16 33.18*** 5.30 41.74*** 47.34*** 20.23*** 21.00*** 1077.94*** * ** ^'' kkk Notes: Tests are significant at the : 10% . 1% levels. All tests 5%, and were performed for AR(6) models including a constant term. See the appendix for series definitions and the text for descriptions of the tests. Table 2 Bivariate Tests for Stability Percent of Tests Significant at 10% Level A. Summary of All Regressions Test Statistic PK ‘L . Lall sup PK msq BP all BP* BP /HI) QLR MW APW GC Combined 23.3 70.2 18.8 10.7 59.6 18.8 18.8 15.3 66.3 2 2 .6 23.8 60.2 37.9 59.3 GC significant 25.5 74.1 2 1 .0 13.0 62.9 2 1 .0 19.7 16.3 6 6 .6 27.6 30.3 61.8 42.0 60.5 100.0 GC insignif. 2 0 .2 64.8 15.6 7.4 55.0 15.6 17.6 13.8 65.9 15.7 14.7 57.9 32.2 57.6 0 .0 B. Percent rejections, listed by variable being forecasted Test Statistic Series r Lall \ . 0 V/3U) PK Lall sup PK msq BP all % BP QLR MW APW GC 66.7 0(1) A. Output and Sales ip ipxmca 49.3 93.3 17.3 5.3 54.7 17.3 13.3 25.3 22.7 25.3 6 8 .0 52.0 62.7 50.7 94.7 21.3 8 .0 61.3 21.3 1.3 5.3 1 0 0.0 16.0 18.7 60.0 48.0 62.7 54.7 22.7 62.7 93.3 1 0 0 .0 18.7 12.0 100.0 100.0 100.0 100.0 100.0 100.0 gmpy 1 0 0 .0 100.0 22.7 14.7 gmyxp8 98.7 69.3 29.3 1 0 0 .0 1 0 0.0 25.3 22.7 97.3 25.3 16.0 40.0 1 0 0 .0 18.7 16.0 rtql 0 .0 2 0 .0 13.3 0 .0 37.3 13.3 4.0 1.3 100 .0 12.0 14.7 gmcq 46.7 6 8 .0 24.0 28.0 89.3 24.0 77.3 81.3 1 0 0 .0 10.7 9.3 81.3 70.7 77.3 57.3 ipcd 17.3 80.0 17.3 14.7 82.7 17.3 13.3 16.0 100 .0 24.0 32.0 86.7 52.0 86.7 78.7 17.3 29.3 ced87ra 93.3 xci 37.3 90.7 8 .0 66.7 mt82 100.0 98.7 17.3 5.3 68.0 12.0 9.3 2.7 68.0 12.0 9.3 30.7 9.3 30.7 56.0 60.0 2.7 4.0 1 0 0 .0 14.7 16.0 18.7 26.7 100 .0 32.0 32.0 52.0 49.3 54.7 64.0 5.3 5.3 41.3 21.3 33.3 1 2.0 10.7 14.7 70.7 1 0 0.0 1 0 0.0 100.0 72.0 B . Employment Ipmhuadj 24.0 97.3 18.7 22.7 94.7 18.7 1.3 2.7 44.0 21.3 22.7 32.0 38.7 34.7 73.3 lphrm 13.3 52.0 22.7 29.3 60.0 22.7 56.0 34.7 98.7 17.3 34.7 38.7 16.0 32.0 78.7 lhel 53.3 98.7 10.7 42.7 97.3 10.7 2.7 2.7 45.3 2 0 .0 26.7 98.7 72.0 97.3 74.7 lhnaps 16.0 80.0 13.3 10.7 84.0 13.3 0 .0 1.3 1 0 0 .0 28.0 30.7 42.7 45.3 46.7 77.3 100 .0 luinc 1.3 80.0 14.7 1.3 46.7 14.7 9.3 10.7 lhu5 2.7 62.7 17.3 1.3 65.3 17.3 13.3 16.0 lhur 10.7 76.0 10.7 5.3 62.7 10.7 8 .0 5.3 4.0 70.7 18.7 6.7 52.0 18.7 0 .0 1.3 8 8 .0 1 2.0 4.0 66.7 12.0 5.3 2.7 94.7 16.0 10.7 85.3 0 .0 0 .0 13.3 24.0 lhelx 34.7 45.3 58.7 30.7 58.7 74.7 28.0 18.7 22.7 40.0 6.7 33.3 73.3 94.7 37.3 52.0 34.7 28.0 37.3 72.0 13.3 13.3 92.0 24.0 84.0 56.0 22.7 45.3 24.0 37.3 48.0 22.7 64.0 86.7 84.0 40.0 1 0 0 .0 C. New Orders hsbp mdu82 5.3 65.3 90.7 6.7 6.7 mpcon8 0 .0 13.3 10.7 0 .0 14.7 10.7 4.0 mocm82 0 .0 24.0 13.3 0 .0 24.0 13.3 10.7 mdo82 2 0 .0 54.7 14.7 14.7 ivpac 13.3 80.0 24.0 9.3 2.7 12.0 8 .0 22.7 1.3 1.3 1.3 65.3 33.3 17.3 22.7 8 .0 9.3 9.3 78.7 10 0 .0 68.0 14.7 6.7 9.3 13.3 22.7 36.0 90.7 41.3 78.7 77.3 72.0 24.0 5.3 0 .0 62.7 17.3 10.7 77.3 28.0 74.7 42.7 2 0 .0 77.3 29.3 84.0 pcni 1.3 49.3 1 2.0 4.0 30.7 12.0 17.3 16.0 34.7 33.3 25.3 21.3 6.7 pmno 8 .0 65.3 16.0 4.0 53.3 16.0 16.0 2 0 .0 14.7 25.3 22.7 25.3 10.7 17.3 D. Inventories invmt87 1 2 .0 81.3 8 .0 4.0 64.0 8 .0 2.7 6.7 22.7 24.0 37.3 16.0 33.3 68 .0 invrd 21.3 69.3 21.3 10.7 73.3 21.3 24.0 30.7 1 2 .0 5.3 4.0 66.7 28.0 64.0 66.7 invwd 2 0 .0 97.3 13.3 2.7 57.3 13.3 17.3 6.7 1 0 0 .0 28.0 32.0 65.3 44.0 62.7 58.7 ivmld8 94.7 97.3 8 .0 38.7 97.3 8 .0 22.7 6.7 10 0 .0 22.7 14.7 97.3 94.7 94.7 69.3 ivm2 d8 0 .0 46.7 6.7 0 .0 64.0 6.7 8 .0 6.7 6.7 6.7 4.0 12.0 8 .0 12.0 52.0 ivm3d8 17.3 96.0 16.0 14.7 97.3 16.0 1.3 8 .0 56.0 42.7 45.3 61.3 60.0 6 8 .0 10.7 58.4 T a b le 2 , c o n tin u e d Test Statistic Series Lall Lm ,0 ( D V^U) PK PK sup msq ivmtd 13.3 49.3 16.0 0.0 28.0 16.0 8 .0 10.7 ivmld 94.7 97.3 14.7 45.3 97.3 14.7 8 .0 8 .0 BP all 64.0 100.0 QLR BP % 0 MW APW (1 ) GC 36.0 41.3 33.3 18.7 33.3 72.0 2 0 .0 14.7 96.0 97.3 94.7 66.7 ivm2 d 2.7 40.0 5.3 0.0 56.0 5.3 2.7 4.0 4.0 5.3 1.3 13.3 5.3 10.7 60.0 ivm3d 10.7 76.0 18.7 9.3 73.3 18.7 58.7 25.3 16.0 10.7 13.3 46.7 30.7 42.7 78.7 invrd8 1 2 .0 6 8 .0 10.7 5.3 77.3 10.7 2.7 1.3 46.7 8 .0 8 .0 40.0 24.0 45.3 69.3 invwd8 10.7 94.7 10.7 2.7 66.7 10.7 1.3 1.3 98.7 14.7 17.3 49.3 29.3 52.0 46.7 10.7 52.0 9.3 5.3 41.3 9.3 1.3 0.0 6.7 E. Prices grodc 25.3 29.3 100.0 100.0 100.0 36.0 17.3 100.0 64.0 100.0 32.0 100.0 34.7 100.0 26.7 32.0 68.0 2 0.0 20.0 12.0 8 .0 52.0 8 .0 42.7 52.0 42.7 5.3 100.0 12.0 50.7 6.7 36.0 50.7 4.0 4.0 100.0 12.0 17.3 0.0 97.3 13.3 16.0 93.3 17.3 92.0 41.3 82.7 17.3 14.7 76.0 12.0 66.7 58.7 62.7 100.0 2 0 .0 pw561 50.7 1 0 0 .0 pw561r 38.7 50.7 16.0 1.3 53.3 16.0 0.0 52.0 4.0 100.0 100.0 100.0 0.0 pw 14.7 14.7 0.0 56.0 14.7 4.0 1.3 69.3 92.0 1.3 24.0 joccir 100.0 10.7 100.0 4.0 100.0 24.0 86.7 jocci 100.0 22.7 8 .0 82.7 punew 98.7 5.3 18.7 F. Interest Rates fyff 10.7 76.0 42.7 2.7 36.0 42.7 32.0 5.3 100.0 41.3 48.0 fygm3 10.7 77.3 46.7 2.7 38.7 46.7 77.3 5.3 100.0 49.3 48.0 98.7 41.3 97.3 70.7 fygm6 12.0 89.3 46.7 2.7 36.0 46.7 72.0 5.3 100.0 42.7 45.3 94.7 45.3 94.7 66.7 4.0 100.0 41.3 42.7 100.0 41.3 37.3 100.0 94.7 60.0 8 .0 89.3 46.7 89.3 10.7 82.7 48.0 10.7 97.3 40.0 22.7 62.7 62.7 6 8 .0 42.7 1.3 29.3 42.7 73.3 32.0 30.7 2.7 30.7 30.7 77.3 fygtlO 41.3 41.3 0.0 24.0 41.3 81.3 6.7 100.0 32.0 36.0 cp6 _gm6 1.3 28.0 22.7 1.3 10.7 22.7 4.0 1.3 100.0 49.3 46.7 100.0 glO_gl 0.0 41.3 26.7 1.3 16.0 26.7 42.7 6.7 100.0 38.7 41.3 6 8 .0 100.0 100.0 9.3 66.7 26.7 0.0 10.7 26.7 0.0 0.0 baa_glO 4.0 89.3 33.3 10.7 82.7 33.3 13.3 45.3 44.0 10.7 21.3 2 0 .0 10.7 8 .0 82.7 2.7 0.0 4.0 65.3 22.7 9.3 0.0 0.0 glO_ff 100.0 93.3 fygtl fybaac 12.0 53.3 100.0 38.7 38.7 49.3 37.3 42.7 78.7 13.3 16.0 25.3 16.0 14.7 40.0 100.0 60.0 49.3 77.3 62.7 33.3 12.0 30.7 62.7 34.7 21.3 34.7 60.0 9.3 32.0 33.3 30.7 60.0 G . Money and Credit fcbcuc 18.7 delinqcr 10.7 5.3 10.7 26.7 84.0 5.3 fcbcucy 10.7 12.0 78.7 10.7 56.0 18.7 100.0 100.0 cci30m 6.7 70.7 22.7 10.7 61.3 22.7 5.3 fmld82 49.3 92.0 29.3 29.3 80.0 29.3 33.3 8 .0 100.0 5.3 2.7 100.0 8 .0 9.3 10.7 64.0 14.7 60.0 69.3 34.7 18.7 92.0 80.0 93.3 78.7 fm2d82 66.7 30.7 62.7 30.7 13.3 26.7 14.7 29.3 26.7 93.3 93.3 93.3 70.7 fmbase 28.0 92.0 12.0 18.7 86.7 12.0 54.7 81.3 21.3 28.0 33.3 26.7 44.0 37.3 54.7 fral 16.0 77.3 24.0 14.7 58.7 24.0 24.0 78.7 32.0 25.3 42.7 25.3 36.0 56.0 fm2 9.3 54.7 25.3 16.0 62.7 25.3 72.0 64.0 18.7 13.3 45.3 26.7 42.7 69.3 £ro3 fmbaser 89.3 100.0 14.7 37.3 57.3 13.3 2 0 .0 2 0 .0 100.0 44.0 100.0 100.0 14.7 14.7 4.0 44.0 17.3 22.7 14.7 84.0 93.3 42.7 54.7 13.3 45.3 13.3 16.0 25.3 24.0 89.3 73.3 96.0 68 .0 2 0.0 2.7 0.0 0.0 0.0 0.0 0.0 21.3 100.0 100.0 H . Other Variables exnwt2 2.7 45.3 4.0 33.3 fspcomr 0.0 0.0 56.0 2.7 1.3 58.7 2.7 4.0 52.0 2.7 1.3 48.0 2.7 1.3 fspcom fail 2.7 6 8 .0 9.3 4.0 58.7 9.3 failr 2.7 6 8 .0 10.7 4.0 57.3 10.7 gfosa 2.7 2 0 .0 17.3 2.7 33.3 17.3 0.0 0.0 0.0 gfrsa 0.0 24.0 4.0 0.0 61.3 4.0 2.7 2.7 gfor 1.3 13.3 10.7 1.3 21.3 10.7 2.7 0.0 69.3 2.7 0.0 0.0 0.0 0.0 30.7 49.3 74.7 30.7 2.7 9.3 gfrr hhsntn 0.0 41.3 36.0 100.0 1.3 64.0 28.0 24.0 34.7 10.7 33.3 82.7 21.3 2 0 .0 16.0 8 .0 16.0 48.0 56.0 22.7 14.7 16.0 5.3 16.0 46.7 42.7 26.7 2 0 .0 38.7 4.0 37.3 16.0 45.3 29.3 2 0 .0 38.7 4.0 38.7 14.7 100.0 26.7 100.0 25.3 100.0 8 .0 8 .0 4.0 9.3 46.7 30.7 36.0 2.7 0.0 4.0 45.3 12.0 5.3 10.7 9.3 2.7 8 .0 42.7 29.3 34.7 1.3 0.0 4.0 48.0 28.0 29.3 100.0 97.3 100.0 54.7 T a b le 2 , c o n tin u e d C. Percent rejections, listed by variable used as predictor ---------Series all L L M D .0C all Test Statistic ---- L1 flLr fl/1xPK H,P n,p(l) sup PK msq BP all BP. BP /3(1) QLR v MW APW GC A. Output and Sales ip 33.3 89.3 34.7 16.0 78.7 34.7 16.0 12.0 70.7 22.7 17.3 72.0 49.3 69.3 76.0 ipxmca 40.0 85.3 18.7 25.3 81.3 18.7 16.0 21.3 73.3 49.3 44.0 6 8 .0 56.0 69.3 82.7 gmpy 28.0 80.0 14.7 17.3 85.3 14.7 9.3 12.0 58.7 4.0 10.7 57.3 38.7 56.0 61.3 grayxp8 38.7 72.0 13.3 28.0 69.3 13.3 16.0 8 .0 52.0 2.7 1.3 64.0 49.3 66.7 52.0 rtql 14.7 6 8 .0 2 0 .0 4.0 53.3 2 0 .0 2 0 .0 10.7 62.7 12.0 25.3 58.7 25.3 60.0 66.7 gmcq 22.7 72.0 17.3 6.7 61.3 17.3 2 0 .0 13.3 65.3 16.0 21.3 56.0 32.0 56.0 72.0 ipcd 22.7 73.3 30.7 8 .0 62.7 30.7 14.7 8 .0 78.7 25.3 30.7 61.3 30.7 56.0 45.3 ced87m 2 0 .0 69.3 2 0 .0 5.3 54.7 2 0 .0 17.3 9.3 62.7 13.3 24.0 53.3 30.7 53.3 49.3 xci 40.0 90.7 29.3 16.0 82.7 29.3 2 0 .0 12.0 76.0 25.3 13.3 73.3 48.0 6 8 .0 88 .0 mt82 37.3 8 8 .0 36.0 16.0 70.7 36.0 16.0 13.3 66.7 26.7 34.7 6 8 .0 46.7 69.3 81.3 B . Employment lpmhuadj 29.3 80.0 14.7 16.0 65.3 14.7 22.7 16.0 64.0 4.0 10.7 6 8 .0 42.7 66.7 72.0 lphrm 21.3 73.3 26.7 12.0 62.7 26.7 29.3 29.3 70.7 40.0 41.3 73.3 46.7 73.3 62.7 lhel 32.0 78.7 24.0 16.0 65.3 24.0 22.7 18.7 62.7 21.3 38.7 62.7 42.7 65.3 92.0 lhnaps 29.3 78.7 45.3 13.3 74.7 45.3 17.3 14.7 61.3 10.7 30.7 6 8 .0 42.7 6 8 .0 66.7 luinc 2 0 .0 73.3 13.3 6.7 60.0 13.3 6.7 8 .0 70.7 54.7 42.7 69.3 40.0 70.7 8 8.0 lhu5 24.0 82.7 34.7 77.3 34.7 21.3 10.7 65.3 5.3 18.7 58.7 38.7 58.7 49.3 lhur 24.0 82.7 24.0 14.7 72.0 24.0 32.0 24.0 69.3 16.0 29.3 66.7 42.7 65.3 81.3 lhelx 33.3 80.0 53.3 22.7 82.7 53.3 16.0 24.0 6 8 .0 33.3 2 0 .0 80.0 65.3 80.0 85.3 12.0 C. New Orders hsbp 13.3 60.0 9.3 5.3 44.0 9.3 25.3 13.3 66.7 33.3 25.3 56.0 30.7 56.0 85.3 mdu82 21.3 84.0 29.3 10.7 74.7 29.3 18.7 16.0 62.7 13.3 13.3 49.3 33.3 48.0 57.3 mpcon8 17.3 6 8 .0 10.7 0 .0 56.0 10.7 2 0 .0 16.0 60.0 1.3 13.3 56.0 26.7 49.3 53.3 mocm82 30.7 77.3 29.3 17.3 70.7 29.3 13.3 10.7 65.3 16.0 29.3 48.0 36.0 46.7 77.3 mdo82 26.7 78.7 28.0 10.7 65.3 28.0 13.3 10.7 69.3 18.7 21.3 60.0 37.3 53.3 80.0 ivpac 12.0 70.7 13.3 4.0 54.7 13.3 17.3 12.0 65.3 21.3 2 0 .0 61.3 34.7 58.7 77.3 13.3 9.3 18.7 13.3 70.7 42.7 29.3 64.0 40.0 62.7 85.3 16.0 72.0 41.3 41.3 6 8 .0 41.3 66.7 81.3 prai 30.7 84.0 13.3 14.7 pnxno 36.0 77.3 16.0 16.0 64.0 65.3 16.0 D. Inventories invmt87 22.7 74.7 14.7 8 .0 58.7 14.7 16.0 13.3 65.3 5.3 9.3 54.7 28.0 53.3 32.0 invrd 26.7 6 8 .0 10.7 12.0 62.7 10.7 26.7 10.7 66.7 10.7 10.7 57.3 28.0 54.7 45.3 38.7 invwd 22.7 69.3 5.3 1 2.0 54.7 5.3 6.7 8 .0 66.7 26.7 30.7 53.3 30.7 54.7 ivmld8 26.7 70.7 16.0 10.7 62.7 16.0 14.7 17.3 70.7 28.0 30.7 66.7 32.0 65.3 42.7 ivm2 d8 2 0 .0 73.3 16.0 12.0 65.3 16.0 8 .0 14.7 6 8 .0 8 .0 6.7 52.0 21.3 50.7 38.7 ivm3d8 28.0 78.7 13.3 1 2.0 73.3 13.3 34.7 37.3 73.3 25.3 21.3 6 8 .0 45.3 72.0 48.0 ivmtd 26.7 74.7 5.3 17.3 62.7 5.3 14.7 17.3 64.0 25.3 28.0 62.7 36.0 58.7 6 8 .0 ivmld 30.7 74.7 14.7 14.7 69.3 14.7 8 .0 14.7 73.3 37.3 36.0 62.7 40.0 61.3 66.7 ivm2 d 22.7 74.7 14.7 16.0 66.7 14.7 10.7 18.7 6 8 .0 22.7 25.3 49.3 24.0 45.3 65.3 32.0 66.7 48.0 66.7 56.0 12 .0 60.0 24.0 60.0 38.7 13.3 50.7 28.0 48.0 12.0 ivm3d 29.3 85.3 9.3 21.3 70.7 9.3 29.3 24.0 70.7 37.3 invrd8 22.7 72.0 17.3 10.7 65.3 17.3 24.0 13.3 61.3 9.3 invwd8 18.7 66.7 17.3 9.3 53.3 17.3 13.3 8 .0 6 8 .0 2 0 .0 T a b le 2 , c o n tin u e d Test Statistic Series Lall Lnf .i 1 ) r r Lall *V.0U> PK sup PK msq BP all % BP 0 (1 ) QLR MW APW GC E. Prices gmdc 28.0 78.7 12 .0 22.7 74.7 12.0 29.3 25.3 74.7 46.7 42.7 50.7 36.0 50.7 50.7 punew 2 0 .0 69.3 10.7 13.3 60.0 10.7 26.7 26.7 70.7 46.7 37.3 62.7 40.0 61.3 70.7 pw 18.7 52.0 0.0 9.3 45.3 17.3 pw561 8 .0 38.7 0.0 18.7 21.3 64.0 24.0 28.0 61.3 33.3 60.0 64.0 1.3 28.0 17.3 17.3 10.7 62.7 5.3 9.3 61.3 41.3 61.3 22.7 9.3 45.3 17.3 0.0 29.3 17.3 18.7 10.7 62.7 5.3 9.3 52.0 40.0 52.0 20 .0 jocci 24.0 66.7 21.3 9.3 60.0 21.3 17.3 14.7 61.3 17.3 18.7 54.7 34.7 56.0 73.3 joccir 22.7 66.7 18.7 5.3 57.3 18.7 16.0 13.3 72.0 32.0 38.7 54.7 34.7 54.7 80.0 pw561r F. Interest Rates fyff 16.0 41.3 4.0 1.3 2 0 .0 4.0 13.3 8 .0 54.7 17.3 29.3 69.3 37.3 66.7 70.7 fygm3 14.7 42.7 1.3 1.3 21.3 1.3 10.7 8 .0 61.3 25.3 30.7 77.3 44.0 74.7 58.7 fygm6 17.3 53.3 1.3 2.7 26.7 1.3 13.3 8 .0 60.0 24.0 38.7 76.0 40.0 76.0 68.0 fygtl 16.0 50.7 2.7 1.3 29.3 2.7 12.0 8 .0 65.3 25.3 34.7 73.3 41.3 77.3 77.3 fybaac 13.3 56.0 14.7 2.7 38.7 14.7 24.0 21.3 70.7 29.3 30.7 65.3 45.3 64.0 76.0 fygtlO 13.3 54.7 6.7 2.7 33.3 6.7 16.0 12.0 76.0 29.3 33.3 69.3 42.7 61.3 78.7 cp6 _gm6 25.3 70.7 17.3 8 .0 44.0 17.3 22.7 17.3 66.7 38.7 36.0 6 8 .0 46.7 70.7 76.0 glO_gl 30.7 73.3 34.7 13.3 56.0 34.7 25.3 36.0 65.3 26.7 24.0 6 8 .0 57.3 72.0 89.3 glO_ff 12.0 52.0 8 .0 1.3 25.3 8.0 18.7 17.3 69.3 42.7 37.3 56.0 28.0 53.3 68 .0 baa_glO 22.7 72.0 8 .0 8 .0 60.0 8.0 17.3 2 0 .0 81.3 61.3 34.7 73.3 53.3 74.7 81.3 G . Money and Credit fcbcuc 21.3 66.7 21.3 12.0 50.7 21.3 6.7 6.7 64.0 18.7 17.3 58.7 30.7 54.7 60.0 fcbcucy 25.3 65.3 26.7 13.3 65.3 26.7 9.3 5.3 64.0 9.3 9.3 62.7 32.0 57.3 21.3 delinqcr 16.0 61.3 22.7 6.7 50.7 22.7 17.3 12.0 58.7 13.3 21.3 46.7 25.3 48.0 26.7 1.3 53.3 22.7 2 0.0 13.3 58.7 16.0 25.3 54.7 37.3 56.0 36.0 65.3 cci30m 13.3 66.7 22.7 fmld82 36.0 81.3 46.7 28.0 76.0 46.7 30.7 24.0 70.7 33.3 36.0 66.7 50.7 65.3 fm2d82 22.7 70.7 9.3 9.3 57.3 9.3 24.0 2 0 .0 64.0 2.7 17.3 49.3 30.7 50.7 65.3 fmbase 26.7 80.0 33.3 16.0 72.0 33.3 29.3 26.7 72.0 21.3 13.3 56.0 46.7 60.0 28.0 fml 26.7 78.7 29.3 12.0 69.3 29.3 38.7 32.0 6 8.0 25.3 32.0 61.3 53.3 61.3 53.3 fm2 18.7 72.0 1.3 12.0 69.3 1.3 9.3 10.7 61.3 1.3 13.3 58.7 45.3 60.0 68.0 fm3 30.7 82.7 16.0 17.3 76.0 16.0 5.3 5.3 68.0 22.7 21.3 6 8 .0 54.7 6 8 .0 42.7 fmbaser 26.7 77.3 37.3 18.7 72.0 37.3 28.0 24.0 69.3 32.0 26.7 57.3 41.3 60.0 56.0 H. Other Variables exnwt2 2 0.0 72.0 24.0 2.7 58.7 24.0 24.0 10.7 66.7 25.3 2 0 .0 58.7 38.7 57.3 29.3 fspcomr 28.0 76.0 18.7 9.3 65.3 18.7 24.0 10.7 62.7 22.7 24.0 50.7 37.3 49.3 66.7 fspcom 26.7 76.0 17.3 9.3 64.0 17.3 22.7 13.3 64.0 2 0 .0 25.3 50.7 38.7 50.7 65.3 fail 18.7 62.7 18.7 4.0 62.7 18.7 18.7 12.0 65.3 14.7 9.3 49.3 28.0 44.0 2 0 .0 failr 18.7 64.0 18.7 2.7 61.3 18.7 18.7 12.0 65.3 14.7 9.3 49.3 26.7 45.3 18.7 gfosa 18.7 70.7 2 0 .0 9.3 6 8 .0 20 .0 24.0 16.0 60.0 26.7 13.3 42.7 33.3 45.3 45.3 gfrsa 16.0 42.7 14.7 9.3 53.3 14.7 2 0 .0 12.0 60.0 4.0 8 .0 46.7 21.3 40.0 18.7 gfor 2 0 .0 73.3 29.3 10.7 69.3 29.3 25.3 17.3 57.3 18.7 12.0 44.0 32.0 46.7 41.3 gfrr 16.0 45.3 14.7 10.7 50.7 14.7 21.3 14.7 60.0 4.0 9.3 44.0 21.3 41.3 16.0 hhsntn 14.7 62.7 13.3 5.3 42.7 13.3 21.3 12.0 72.0 50.7 29.3 58.7 29.3 56.0 49.3 Notes: All statistics are based on regression (1) with 6 lags. See the appendix for series definitions and the text for descriptions of the tests. Table 3 Best Out-of-Sample Forecasting Models: Percentage of cases in which model was best out-of-sample A. AR KRAI RRA2 RLSA ATVP1 9 ATVP2 Summary ATVP3 ATVP4 VAR RRV1 RRV2 RLSV VTVP1 VTVP2 VTVP3 VTVP4 All Models 17 4 8 15 4 4 0 13 1 5 12 6 3 1 0 10 BIC Sel. 15 2 6 14 10 3 3 0 12 2 8 12 8 6 1 0 Stab, rej. 13 5 8 14 10 4 5 0 10 1 6 12 8 4 1 0 B. Best models among all bivariate pairs, by variable being forecasted AR RRA1 RRA2 RLSA ATVP1 ATVP2 ATVP3 ATVP4 VAR RRV1 RRV2 RLSV VTVP1 VTVP2 VTVP3 VTVP4 A. Output and Sales ip ipxmca gmpy gmyxp8 0 0 0 0 61 0 0 0 1 0 16 7 0 0 0 61 0 0 0 0 1 1 3 15 11 20 1 0 3 0 0 0 0 4 1 44 1 0 0 0 76 0 0 0 0 0 0 0 0 0 0 47 0 0 0 0 3 5 9 0 0 7 1 0 rtql 71 0 0 0 0 0 0 0 20 0 0 8 1 0 0 0 gracq 0 0 73 0 0 0 0 0 0 0 21 1 1 3 0 0 17 16 9 0 0 0 1 0 0 0 ipcd 0 0 0 0 49 0 0 0 3 0 5 ced87m 0 0 56 0 0 0 0 0 0 0 43 xci 0 0 0 68 0 0 0 0 0 0 9 12 7 4 0 0 mt82 0 0 0 65 0 0 0 0 11 0 0 20 4 0 0 0 B . Employment lpmhuadj 0 61 0 0 0 0 0 0 0 39 0 0 0 0 0 0 0 21 lhel 0 40 0 0 0 0 0 0 lhnaps 0 0 0 0 51 0 0 0 0 0 51 0 0 0 0 0 12 52 0 0 0 0 0 0 0 12 lphrm luinc lhu5 13 5 1 15 4 0 0 0 7 15 15 4 0 0 0 3 0 16 37 4 0 0 0 3 1 7 3 0 0 19 17 0 0 1 9 20 12 24 1 0 0 5 0 0 lhur 0 0 0 59 0 0 0 0 4 1 0 16 19 1 0 0 lhelx 0 0 0 69 0 0 0 0 0 0 1 16 11 3 0 0 hsbp 0 0 67 0 0 0 0 0 0 0 5 20 mdu82 0 0 79 0 0 0 0 0 0 0 mpcon8 60 0 0 0 0 0 0 0 20 0 4 12 mocm82 51 0 0 0 0 0 0 0 28 0 0 16 C. New Orders 20 0 5 1 1 0 0 1 0 0 3 1 0 0 5 0 0 0 mdo82 0 0 0 0 48 0 0 0 4 7 0 25 15 1 0 0 ivpac 0 0 0 87 0 0 0 0 1 1 1 5 4 0 0 0 pmi pmno 0 0 0 65 0 0 0 0 9 0 0 21 3 1 0 0 55 0 0 0 0 0 0 0 17 0 4 19 4 1 0 0 D. Inventories invmt87 0 0 0 45 0 0 0 0 invrd 0 0 0 47 0 0 0 0 invwd 5 20 1 11 23 13 1 0 0 0 1 23 8 1 0 0 0 0 0 72 0 0 0 0 3 0 0 24 1 0 0 0 ivmld8 0 63 0 0 0 0 0 0 0 4 9 3 0 15 7 0 ivm2 d8 56 0 0 0 0 0 0 0 17 0 3 16 8 0 0 0 ivm3d8 0 0 0 0 41 0 0 0 0 3 5 0 0 13 12 25 T a b le AR RRA1 RRA2 RLSA ATVP1 ATVP2 3 , ATVP3 c o n tin u e d ATVP4 VAR RRV1 RRV2 12 RLSV 3 VTVP1 VTVP2 VTVP3 VTVP4 0 3 4 0 9 16 4 0 0 0 5 27 25 7 4 0 0 16 0 16 13 7 0 0 0 0 3 0 0 8 13 1 0 0 0 0 16 0 0 1 3 4 1 1 85 0 0 0 3 0 0 3 9 0 ivrald 0 68 0 0 0 0 0 0 0 11 ivm2 d 0 0 0 44 0 0 0 0 27 0 ivm3d 0 0 0 0 32 0 0 0 0 invrd8 0 0 0 48 0 0 0 0 invwd8 0 0 0 0 75 0 0 0 0 0 0 0 73 0 E. Prices gmdc punew 0 0 0 0 0 pw 0 0 0 0 0 0 93 0 0 0 0 0 1 0 5 0 0 0 0 0 0 75 0 0 0 0 0 0 0 0 pw561 25 0 0 pw561r 36 0 0 0 0 0 0 0 64 0 0 0 0 0 0 jocci 65 0 0 0 0 0 0 0 15 0 0 16 4 0 0 0 0 0 0 64 0 0 0 0 11 0 0 23 3 0 0 0 joccir F. Interest Rates fyff 33 0 0 0 0 0 0 0 57 0 3 7 0 0 0 0 fygm3 49 0 0 0 0 0 0 0 48 0 0 1 0 1 0 0 fygm6 43 0 0 0 0 0 0 0 39 0 1 16 0 1 0 0 fygtl 47 0 0 0 0 0 0 0 39 0 1 12 1 0 0 0 fybaac 59 0 0 0 0 0 0 0 41 0 0 0 0 0 0 0 0 fygtlO 51 0 0 0 0 0 0 0 47 0 0 3 0 0 0 cp6 _gm6 91 0 0 0 0 0 0 0 9 0 0 0 0 0 0 0 glO_gl 63 0 0 0 0 0 0 0 37 0 0 0 0 0 0 0 glO_ff 0 0 44 0 0 0 0 0 36 0 8 9 3 0 0 0 baa_glO 0 65 0 0 0 0 0 0 17 1 12 3 0 1 0 0 0 G. Money and Credit fcbcuc 0 0 0 55 0 0 0 0 12 0 0 23 0 0 fcbcucy 0 0 0 0 0 56 0 0 0 0 0 3 29 11 0 1 delinqcr 0 0 0 52 0 0 0 0 0 1 0 44 3 0 0 0 cci30m 33 0 0 0 0 0 0 0 37 0 0 19 11 0 0 0 fmld82 0 0 0 0 40 0 0 0 1 0 5 28 21 4 0 0 8 17 7 0 19 0 0 0 fm2d82 0 0 fmbase 0 0 11 0 0 0 63 0 0 0 1 1 0 , 0 53 0 0 0 0 0 5 23 3 fml 61 0 0 0 0 0 0 0 21 0 0 11 7 0 0 0 fm2 0 0 0 0 0 49 0 0 16 0 0 1 5 28 0 0 fm3 0 0 0 0 0 0 93 0 0 0 0 0 1 fmbaser 0 0 0 0 0 57 0 0 0 0 3 5 13 3 21 3 0 0 0 H. Other Variables exnwt2 fspcomr 88 0 0 0 0 0 0 0 11 0 1 0 0 0 0 0 0 0 0 0 79 0 0 0 0 0 3 15 4 0 0 0 fspcom 0 0 0 0 84 0 0 0 0 0 1 13 1 0 0 0 fail 0 0 0 0 73 0 0 0 15 0 0 8 4 0 0 0 0 failr 75 0 0 0 0 0 0 0 16 0 0 7 3 0 0 gfosa 0 0 0 68 0 0 0 0 0 3 0 29 0 0 0 0 gfrsa 0 0 87 0 0 0 0 0 0 0 7 7 0 0 0 0 gfor 0 0 0 76 0 0 0 0 0 1 1 gfrr 0 0 84 0 0 0 0 0 0 0 11 hhsntn 0 0 55 0 0 0 0 0 0 7 19 21 4 20 0 0 0 0 1 0 0 0 0 0 0 0 T a b le C. 3 , c o n tin u e d Best models among those bivariate pairs with the 10-lowest in-sample BIC, by variable being forecasted AR KRAI RRA2 RLSA ATVP1 ATVP2 ATVP3 ATVPA VAR RRV1 RRV2 RLSV VTVP1 VTVP2 VTVP3 VTVPA A. Output and Sales ip ipxmca gmpy gmyxp8 rtql 0 0 0 0 70 0 0 0 0 0 10 0 0 20 0 0 0 0 0 70 0 0 0 0 0 0 0 0 30 0 0 0 60 0 0 0 0 0 0 0 A0 0 0 0 0 0 0 0 0 0 0 20 0 0 0 0 20 0 0 60 0 0 0 0 AO 0 0 0 0 0 0 0 20 0 0 A0 0 0 0 0 0 gmcq 0 0 20 0 0 0 0 0 0 0 70 ipcd 0 0 0 0 80 0 0 0 0 0 10 ced87m 0 0 10 0 0 0 0 0 0 0 90 0 0 10 0 10 0 0 0 0 0 0 0 0 0 xci 0 0 0 60 0 0 0 0 0 0 0 10 20 10 0 0 mt82 0 0 0 60 0 0 0 0 20 0 0 20 0 0 0 0 0 B . Employment lpmhuadj 0 20 0 0 0 0 0 0 0 30 0 0 50 0 0 lphrm 0 0 0 0 0 0 0 0 0 0 10 20 A0 30 0 0 lhel 0 30 0 0 0 0 0 0 0 0 0 10 50 10 0 0 30 lhnaps 0 0 0 0 A0 0 0 0 0 0 0 0 10 luinc 0 0 50 0 0 0 0 0 10 0 30 0 0 10 20 0 0 0 lhu5 AO 0 0 0 0 0 0 0 0 0 0 10 30 20 0 0 lhur 0 0 0 20 0 0 0 0 0 10 0 50 20 0 0 0 lhelx 0 0 0 60 0 0 0 0 0 0 10 10 20 0 0 0 C. New Orders hsbp 0 0 80 0 0 0 0 0 0 0 0 10 0 10 0 0 mdu82 0 0 70 0 0 0 0 0 0 0 30 0 0 0 0 0 mpcon8 30 0 0 0 0 0 0 0 30 0 10 30 0 0 0 0 mocm82 AO 0 0 0 0 0 0 0 A0 0 0 20 0 0 0 0 0 mdo82 0 0 0 0 30 0 0 0 0 20 0 30 20 0 0 ivpac 0 0 0 100 0 0 0 0 0 0 0 0 0 0 0 0 pcni 0 0 0 50 0 0 0 0 10 0 0 30 0 10 0 0 AO 0 0 0 0 0 0 0 0 0 10 30 10 10 0 0 0 porno D. Inventories invmt87 0 0 0 20 0 0 0 0 0 10 20 20 20 10 0 invrd 0 0 0 50 0 0 0 0 0 0 0 20 20 10 0 0 invwd 0 0 0 80 0 0 0 0 0 0 0 20 0 0 0 0 ivmld8 0 20 0 0 0 0 0 0 0 10 20 0 0 A0 10 0 ivm2 d8 90 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 10 0 0 10 10 0 0 0 10 20 10 0 0 0 0 ivm3d8 0 0 0 0 70 0 0 0 0 ivmtd 0 0 0 20 0 0 0 0 A0 ivmld 0 20 0 0 0 0 0 0 0 20 30 0 0 20 10 ivm2 d 0 0 0 80 0 0 0 0 0 0 0 10 10 0 0 0 ivm3d 0 0 0 0 0 0 0 0 0 10 60 20 0 10 0 0 0 0 0 0 0 0 50 20 10 0 0 0 90 0 0 0 0 0 0 0 10 0 0 0 invrd8 0 0 0 20 invwd8 0 0 0 0 E. Prices gmdc 0 0 0 0 0 90 0 0 0 0 0 0 0 10 0 0 punew 0 0 0 0 0 0 90 0 0 0 0 0 0 0 10 0 pw 0 0 0 0 0 0 80 0 0 0 0 0 10 0 10 0 pw561 20 0 0 0 0 0 0 0 80 0 0 0 0 0 0 0 pw561r 30 0 0 0 0 0 0 0 70 0 0 0 0 0 0 0 jocci 90 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 joccir 0 0 0 70 0 0 0 0 0 0 0 20 10 0 0 0 T a b le AR RRAl RRA2 RLSA ATVPl ATVP2 3 , ATVP3 c o n tin u e d ATVP4 VAR RRV1 RRV2 RLSV VTVP1 VTVP2 VTVP3 VTVP4 F. Interest Rates fyff 0 0 0 0 0 0 0 0 90 0 0 10 0 0 0 0 fygm3 30 0 0 0 0 0 0 0 60 0 0 10 0 0 0 0 fygm6 30 0 0 0 0 0 0 0 50 0 0 20 0 0 0 0 fygtl 50 0 0 0 0 0 0 0 30 0 0 10 10 0 0 0 fybaac 50 0 0 0 0 0 0 0 50 0 0 0 0 0 0 0 fygtlO 40 0 0 0 0 0 0 0 50 0 0 10 0 0 0 0 cp6 _gm6 gl0 _gl 100 80 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 20 0 0 0 0 0 0 0 gl0 _ff 0 0 0 0 0 0 0 0 40 0 50 0 10 0 0 0 baa_gl0 0 50 0 0 0 0 0 0 40 0 10 0 0 0 0 0 G. Money and Credit fcbcuc 0 0 0 30 0 0 0 0 20 0 0 20 30 0 0 0 fcbcucy 0 0 0 0 0 0 0 0 0 0 0 10 50 40 0 0 delinqcr 0 0 0 80 0 0 0 0 0 0 0 20 0 0 0 0 cci30m 20 0 0 0 0 0 0 0 40 0 0 20 20 0 0 0 fmld82 0 0 0 0 40 0 0 0 0 0 0 50 10 0 0 fm2d82 0 0 0 0 0 50 0 0 0 0 0 0 0 fmbase 0 0 0 0 60 0 0 0 0 0 0 30 10 fml 50 0 0 0 0 0 0 0 40 0 0 10 0 fm2 0 0 0 0 0 40 0 0 0 0 0 0 20 0 40 0 10 0 0 0 0 0 0 40 0 0 fm3 0 0 0 0 0 0 90 0 0 0 0 0 0 10 0 0 fmbaser 0 0 0 0 0 20 0 0 0 0 20 30 10 20 0 0 H. Other Variables exnwt2 100 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 fspcomr 0 0 0 0 100 0 0 0 0 0 0 0 0 0 0 0 100 fspcom 0 0 0 0 fail 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 90 0 0 0 0 0 0 10 0 0 0 0 failr 90 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 gfosa 0 0 0 70 0 0 0 0 0 0 0 30 0 0 0 0 gfrsa 0 0 70 0 0 0 0 0 0 0 0 30 0 0 0 0 gfor 0 0 0 80 0 0 0 0 0 0 0 20 0 0 0 0 gfrr 0 0 70 0 0 0 0 0 0 0 10 10 10 0 0 0 hhsntn 0 0 60 0 0 0 0 0 0 0 20 20 0 0 0 0 Notes: See the appendix for series descriptions. All models included six lags plus a constant. The first row of panel A shows results for all 76x76 models. Entries in the second row are the corresponding fraction, except that the set of bivariate relations is restricted from 75 to 10 for each forecasted variable, where the 10 predictors are chosen to be those with the lowest in-sample BIC for the forecasted variable at hand. Entries in the third row are for the set of bivariate relations restricted to be those for which the L n statistic is significant at the 10% level, when calculated through 1978:12. Panel*a shows detailed results for all model for each variable, and panel C shows detailed results for the 10 best fitting (insample) models for each variable. The in-sample period was from the later of 1959:1 or the first data for which data are available, to 1978:12, and the out-of-sample period is from 1979:1 through the earlier of the final date for the series or 1993:12. Table 4 Comparsion of out-of-sample forecasts among all 5700 bivariate bivariate combinations Percentage of times that row forecast MSE is less than column forecast MSE Model AR RRA1 RRA2 RLSA AR ATVP1 ATVP2 ATVP3 ATVP4 — 66 55 36 45 53 79 87 RRA1 34 — 26 12 20 30 72 RRA2 45 74 — 20 34 62 RLSA 64 88 80 — 63 ATVP1 55 80 66 37 ATVP2 47 70 38 18 ATVP3 21 28 16 11 ATVP4 13 11 VAR 35 RRV1 VAR RRVl RRV2 RLSV VTVPl VTVP2 65 79 60 50 54 63 79 88 89 53 80 48 36 41 58 82 91 84 93 61 85 65 51 57 70 85 93 82 89 93 69 88 72 62 68 76 88 94 — 86 93 93 66 86 67 56 63 77 89 95 14 — 93 97 59 78 55 45 48 67 88 94 VTVP3 VTVP4 7 7 — 100 44 60 33 27 27 36 73 90 0 — 34 42 20 16 15 19 46 77 45 68 80 20 61 87 7 7 7 3 47 39 31 34 41 56 66 — 65 42 26 33 21 20 15 12 14 22 40 58 35 — 9 7 6 RRV2 40 52 35 28 33 45 67 80 58 91 — 22 37 63 89 96 RLSV 50 64 49 38 44 55 73 84 74 93 78 — 64 78 91 96 VTVP1 46 59 43 32 37 52 73 85 67 94 63 36 — 86 95 98 VTVP2 37 42 30 24 23 33 64 81 55 80 37 22 14 — 98 99 VTVP3 21 18 15 12 11 12 27 54 32 39 11 9 5 2 — 100 VTVP4 12 9 7 6 10 23 20 13 4 2 1 0 — Notes: See the notes to table 3 5 6 4 Table 5 Selected Quantiles of Distributions of Mean Square Forecast Errors Relative to MSE for the AR Recursive Least Squares (RLSA) Forecast A. Univariate Forecasts ----------- Percentile -----------Model Min 0.050 0.250 0.500 0.750 0.950 Max 1.158 AR 0.959 0.972 0.998 1.003 1.016 1.032 RLA1 0.967 0.987 1.006 1.017 1.049 1.076 RLA2 0.977 0.991 1.004 1.022 1.010 1.023 1.076 1.002 1.004 1.025 1.035 1.007 1.014 1.089 1.107 ATVP1 0.988 0.992 ATVP2 0.976 0.987 1.001 1.000 1.002 ATVP3 0.969 0.984 1.014 1.024 1.039 1.233 1.269 ATVP4 0.973 0.989 1.029 1.040 1.063 1.368 1.413 B. Bivariate Forecasts (All) ------------------------------- Percentile -------------------------------------Model Min 0.001 0.005 0.010 0.050 0.250 0.500 0.750 0.950 0.990 0.995 0.999 Max 2.278 VAR 0.512 0.607 0.914 0.931 0.960 0.994 1.018 1.048 1.145 1.276 1.338 1.570 RLV1 0.493 0.612 0.920 0.934 0.979 1.018 1.037 1.057 1.093 1.131 1.145 1.200 1.392 RLV2 0.491 0.602 0.906 0.922 0.964 0.998 1.013 1.029 1.059 1.093 1.105 1.159 1.268 RLSV 0.486 0.598 0.903 0.919 0.961 0.992 1.006 1.019 1.047 1.078 1.094 1.152 1.263 VTVP1 0.491 0.600 0.903 0.921 0.961 0.995 1.009 1.024 1.050 1.076 1.096 1.146 1.274 VTVP2 0.499 0.602 0.905 0.923 0.964 1.001 1.020 1.038 1.094 1.123 1.133 1.183 1.320 VTVP3 0.517 0.615 0.917 0.935 0.978 1.020 1.048 1.074 1.220 1.280 1.289 1.329 1.454 VTVP4 0.534 0.639 0.934 0.951 0.996 1.043 1.077 1.114 1.331 1.428 1.446 1.485 1.574 C. Best 10 bivariate models as selected using in-sample BIC ---------------------------- Percentile ------------------------Model Min 0.005 0.010 0.050 0.250 0.500 0.750 0.950 0.990 0.995 Max VAR 0.512 0 .6 8 6 0.901 0.940 0.980 1.017 1.066 1.195 1.519 1.947 2.278 RLV1 0.493 0.712 0.889 0.936 1.003 1.031 1.052 1 .101 1.175 1.240 1.392 RLV2 0.491 0 .6 8 8 0.875 0.928 0.981 1.009 1.031 1.071 1.157 1.193 1.268 RLSV 0.486 0.689 0.870 0.922 0.977 1.002 1.024 1.061 1.146 1.194 1.263 VTVP1 0.491 0.701 0.873 0.922 0.978 1.008 1.027 1.059 1.130 1.149 1.274 VTVP2 0.499 0.724 0.884 0.926 0.983 1.018 1.042 1.096 1.157 1.201 1.320 VTVP3 0.517 0.772 0.904 0.939 1.001 1.043 1.077 1 .222 1.306 1.368 1.454 VTVP4 0.534 0.809 0.926 0.958 1.022 1.070 1.112 1.331 1.437 1.517 1.574 Note: Mean square forecast errors are relative to mean square error of the recursive least squares AR forecast. See the note to Table 3 for a description of the models. G r a n g e r —C a u s a l i t y S t a t i s t i c Figure 1 Scatterplot of Nyblom (1989) s a i t c vs. Granger causality F s a i t c ttsi -ttsi S o lid lin e s d e n o te 10% c r it ic a l v a lu e s G r a n g e r —C a u s a l i t y S t a t i s t i c 0 .2 0.4 0 .6 0 .8 1 .0 1.2 1.4 1.6 1.8 2.0 PKsup Figure 2 Scatterplot of Ploberger-Kramer (1992) P K sup s a i t c vs. Granger causality F s a i t c ttsi -ttsi S o lid lin e s d e n o te 10% c r it ic a l v a lu e s o ro CM G r a n g e r —C a u s a l i t y S t a t i s t i c IN CM 00 m CM a> co m o 0 40 80 120 160 200 QLR Figure 3 Scatterplot of Quandt (1960) Q L R s a i t c vs. Granger causality F-s a i t c ttsi ttsi S o lid lin e s d e n o te 10% c r it ic a l v a lu e s 240 o n rCJ N G r a n g e r —C a u s a l i t y S t a t i s t i c CJ N (N C O m V — < N <> 7 to to o 0 10 20 30 40 50 60 70 80 MW Figure 4 Scatterplot o f MW statistic vs. Granger causality F-statistic S o lid lin e s d e n o te 10% c r it ic a l v a lu e s 90 100 C 0 .08 0 .06 0 .0 4 0.0 2 0.00 Frequency Relative 1964 1968 1972 1976 1980 1984 Year Figure 5 Histogram of break-dates from Q L R s a i t c ttsi 1988 APW o QLR Figure 6 Scatterplot of Andrews-Ploberger (1992) APW statistic vs. QLR statistic Working Paper Series A s ries of research studies on regional economic issues relating to the Seventh Federal e Reserve D s r c ,and on financial and economic t p c . itit ois REGIONAL ECONOMIC ISSUES Estimating Monthly Regional Value Added by Combining Regional Input With National Production Data WP-92-8 P h ilip R. Israilevich and Kenneth N . Kuttner Local Impact of Foreign Trade Zone WP-92-9 D a v id D . Weiss Trends and Prospects for Rural Manufacturing WP-92-12 W illiam A . Testa State and Local Government Spending--The Balance Between Investment and Consumption WP-92-14 Richard H . Mattoon Forecasting with Regional Input-Output Tables WP-92-20 P.R . Israilevich, R . M ahidhara, and G J D . Hew ings A Primer on Global Auto Markets WP-93-1 P au l D . Ballew and Robert H . Schnorbus Industry Approaches t Environmental Policy o in the Great Lakes Region W P-93-8 D a v id R . Allardice, Richard H . Mattoon and William A . Testa The Midwest Stock Price Index-Leading Indicator of Regional Economic Activity W P-93-9 W illiam A . Strauss Lean Manufacturing and the Decision to Vertically Integrate Some Empirical Evidence From the U.S. Automobile Industry WP-94-1 Thomas H . K lie r Domestic Consumption Patterns and the Midwest Economy WP-94-4 Robert Schnorbus and P au l Ballew 1 W orking p aper series continued ISSUES IN FINANCIAL REGULATION Incentive Conflict in Deposit-Institution Regulation: Evidence from Australia W P -9 2 -5 Edw ard J. Kane and George G . Kaufman Capital Adequacy and the Growth of U.S. Banks WP-92-11 H erbert Baer and John M cElravey Bank Contagion: Theory and Evidence WP-92-13 George G. Kaufman Trading Activity, Progarm Trading and the Volatility of Stock Returns WP-92-16 James T. Moser Preferred Sources of Market Discipline: Depositors v . s Subordinated Debt Holders WP-92-21 D ouglas D . Evan o ff An Investigation of Returns Conditional on Trading Performance WP-92-24 James T. M oser and Jaclcy C. So The Effect of Capital on Portfolio Risk a Life Insurance Companies t WP-92-29 E lijah Brew er I II , Thomas H . Mondschean, and P hilip E . Strahan A Framework forEstimating the Value and Inte e tRate Risk of Retail Bank Deposits rs WP-92-30 D a vid E . Hutchison, George G . Pennacchi Capital Shocks and Bank Growth-1973 t 1991 o WP-92-31 H erbert L . B aer and John N . M cElravey The Impact of S&L Failures and Regulatory Changes on the C D Market 1987-1991 WP-92-33 E lijah Brew er and Thomas H . Mondschean Junk Bond Holdings, Premium Tax Offsets, and Risk Exposure a Life Insurance Companies t WP-93-3 E lijah Brewer I I I and Thomas H . Mondschean 2 W alk in g p aper series continued Stock Margins and the Conditional Probability ofPrice Reversals W P -9 3 -S P a u l Kofman and James T. M oser I There Lif(f)e After DTB? s Competitive Aspects of Cross Listed Futures Contracts on Synchronous Markets WP-93-11 P a u l Kofman, Ton y Bouwman and James T. M oser Opportunity Cost and Prudentiality: A RepresentativeAgent Model of Futures Clearinghouse Behavior WP-93-18 H erbert L . Baer, Virginia G . France and James T. M oser The Ownership Structure of Japanese Financial I s i u i n ntttos W P-93-19 Hesna Genay Origins of the Modem Exchange Clearinghouse: A History ofEarly Clearing and Settlement Methods atFutures Exchanges WP-94-3 James T. M oser The Effect of Bank-Held Derivatives on Credit Accessibility WP-94-5 E lija h Brew er I I I , Bernadette A . Minton and James T. M oser Small Business Investment Companies: Financial Characteristics and Investments WP-94-10 E lija h B rew er I I I and Hesna Genay MACROECONOMIC ISSUES An Examination of Change i Energy Dependence and Efficiency n in the Six Largest Energy Using Countries-1970-1988 WP-92-2 Ja c k L . H ervey Does the Federal Reserve Affect Asset Prices? WP-92-3 Vefa Tarhan Investment and Market Imperfections i the U.S. Manufacturing Sector n WP-92-4 Paula R. Worthington Business Cycle Durations and Postwar Stabilization of the U.S. Economy WP-92-6 M ark W. Watson 3 W orking p aper series continued A Procedure forPredicting Recessions with Leading I d n icators: Econometric Issues and Recent Performance W P -9 2 -7 Jantes H . Stock and M ark W. Watson Production and Inventory Control a the General Motors Corporation t During the 1920s and 1930s WP-92-10 A n il K . Kashyap and D a vid W. W ilcox Liquidity Effects, Monetary Policy and the Business Cycle W P-92-15 Lawrence J. Christiano and M artin Eichenbaum Monetary Policy and External Finance: Interpreting the Behavior ofFinancial Hows and I t r s Rate Spreads neet WP-92-17 Kenneth N . Kuttner Testing Long Run Neutrality WP-92-18 Robert G. K ing and M ark W. Watson A Policymaker’ Guide to Indicators of Economic Activity s Charles Evans, Steven Strongin, and Francesca Eugeni WP-92-19 Barriers t Trade and Union Wage Dynamics o WP-92-22 Ellen R. Rissman Wage Growth and Sectoral S i t : Phillips Curve Redux hfs WP-92-23 Ellen R. Rissman Excess Volatility and The Smoothing of I t r s Rates: neet An Application Using Money Announcements WP-92-25 Steven Strongin Market Structure, Technology and the Cyclicality of Output WP-92-26 Bruce Petersen and Steven Strongin The Identification of Monetary Policy Disturbances: Explaining the Liquidity Puzzle WP-92-27 Steven Strongin Earnings Losses and Displaced Workers WP-92-28 Louis S. Jacobson , Robert J. LaLonde, and D aniel G. Sullivan 4 W orking pap er series continued Some Empirical Evidence of the Effects on Monetary Policy Shocks on Exchange Rates WP-92-32 M artin Eichenbaum and Charles Evans An Unobserved-Components Model of Constant-Inflation Potential Output WP-93-2 Kenneth N . Kuttner Investment, Cash How, and Sunk Costs WP-93-4 Paula R. Worthington Lessons from the Japanese Main Bank System for Financial System Reform in Poland WP-93-6 Takeo H oshi, A n il Kashyap, and G a ry Loveman Credit Conditions and the Cyclical Behavior of Inventories WP-93-7 A n il K . Kashyap, Owen A . Lam ont and Jerem y C. Stein Labor Productivity During the Great Depression WP-93-1o M ichael D . B ordo and Charles L . Evans Monetary Policy Shocks and Productivity Measures in the G-7 Countries WP-93-12 Charles L . Evans and Fernando Santos Consumer Confidence and Economic Ructuations WP-93-13 John G . Matsusaka and A rg ia M . Sbordone Vector Autoregressions and Cointegration WP-93-14 M ark I . Watson V Testing for Cointegration When Some of the Cointegrating Vectors Are Known WP-93-15 M ichael T. K . H orvath and M ark W. Watson Technical Change, Diffusion, and Productivity WP-93-16 Jeffrey R. Campbell Economic Activity and the Short-Term Credit Markets: An Analysis of Prices and Quantities Benjamin M . Friedm an and Kenneth N. Kuttner WP-93-17 W alk in g p aper series continued Cyclical Productivity i a Model of Labor Hoarding n W P -9 3 -2 0 A rg ia M . Sbordone The Effects of Monetary Policy Shocks: Evidence from the Flow of Funds WP-94-2 Lawrence J. Christiano, Martin Eichenbaum and Charles Evans Algorithms for Solving Dynamic Models with Occasionally Binding Constraints WP-94-6 Lawrence J .Christiano and Jonas D M . Fisher Identification and the Effects of Monetary Policy Shocks WP-94-7 Lawrence J .Christiano, M artin Eichenbaum and Charles L. Evans Small Sample Bias i G M M Estimation of Covariance Structures n WP-94-8 Joseph G. Altonji and Lewis M . Segal Interpreting the Procyclical Productivity of Manufacturing Sectors: External Effects of Labor Hoarding? WP-94-9 A rgia M. Sbordone Evidence on Structural I s a i i y i Macroeconomic Time Series Relations ntblt n WP-94-13 James H . Stock and M ark W. Watson 6