The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
http://clevelandfed.org/research/workpaper/index.cfm Best available copy Working Paper 8508 THE O H I O ECONOMY: USING TIME S E R I E S CHARACTERISTICS I N FORECASTING By James G. Hoehn and James J. Balazsy, J r . Working papers of t h e Federal Reserve Bank o f Cleveland are prelimiri.ary m a t e r i a l s , c i r c u l a t e d t o s t i m u l a t e d i s c u s s i o n and c r i t i c a l comment. The views expressed h e r e i n a r e those o f t h e authors and n o t n e c e s s a r i l y o f t h e Federal Reserve Bank o f Cleveland o r the Board o f Governors o f t h e Federal Reserve System. Diane Mogren and Gordon Schlegel provided h e l p f u l programming assistance. P a r t i c u l a r thanks are due t o W i l l i a m C. Gruben o f t h e Federal Reserve Bank o f D a l l a s . The authors acknowledge h e l p f u l discussions w i t h Michael Bagshaw, John Erceg, P h i l i p I s r a i l e v i c h , and Robert Schnorbus. Kathryn Begy and Linda Shy prepared t h i s manuscript. December 1985 Federal Reserve Bank o f Cleveland http://clevelandfed.org/research/workpaper/index.cfm Best available copy Abstract T.ne s e r i e s methods are used t o determine what i n f o r m a t i o n Ohio and n a t i o r z i s t a t i s t i c s convey about the c u r r e n t and f u t u r e s t a t e o f the r e g i o n a l econory. P r o p e r t i e s o f a number of q u a r t e r l y s e r i e s measurin g aggregate economic a c t i v i t y and p r i c e s i n Ohio are described, i n c l u d i n g t h e i r growth r a t e s and v a r i a b i l i t y , c y c l i c i t y , c o r r e l a t i o n a t a moment i n time, tendency t o foreshhdow each o t h e r ' s movements, and tendency t o be foreshadowed by n a t i o n a l econom'c i n d i c a t o r s . These p r o p e r t i e s are o f i n t e r e s t b o t h f o r f o r e c a s t i n g , e i t h e r Formal o r judgmental, and f o r understanding s t r u c t u r a l c h a r a c t e r i s t i c s of the 3 h i o economy. They a r e e x t e n s i v e l y t a b u l a t e d here. I n s d d i t i o n , some methods of f o r e c a s t i n g , which e x p l o i t these time s e r i e s p r o p e r t i e s , a r e assessed i n an out- of- sample f o r e c a s t p e r i o d . The t r e a t m e n t o f these methods and means f o r comparing them i s elementary and somewhat pedogogiial f o r t h e b e n e f i t o f readers w i t h 1i t t l e p r i o r knowledge o f t i m e s e r i e s f o r e c a s t i n g methods. The nethod f o r b u i l d i n g a time s e r i e s model described i n Hoehn (1984) and applied Y - Texas w i t h considerable f o r e c a s t i n g success i s a p p l i e d , w i t h some m o d i f i c a t on, t o t h e economy o f Ohio. -a p r i o r i , i s a l s o implemented. A simple t r i c k l e - d o w n model, s p e c i f i e d Forecasts combining these methods a r e assessed. .. . . The f o r e c a s t s o f t h e m u l t i v a r i a t e models are f r e q u e n t l y found t o be b e t t e r than those o f u n i v a r i a t e autoregressions. I n some cases, they a r e s i g n i f i c a n t l y s u p e r i o r , according t o an i n d i r e c t s t a t i s t i c a l t e s t adapted f r o m Ashley, Granger, and Schmalensee (1980). The r e s u l t s show t h a t i n f o r m a t i o n can be i d e 2 t i f i e d as t o source and q u a n t i f i e d using v e r y simple r e g r e s s i o n methods. http://clevelandfed.org/research/workpaper/index.cfm Best available copy THE O H I O ECONOMY: T I M E SERIES CHARACTERISTICS The r e g i o n a l economist depends t o a l a r g e e x t e n t upon economic s t a t i s t i c s i n assessing the c u r r e n t s t a t e and l i k e l y f u t u r e course o f h i s r e g i o n . Consequently, an understanding of the p r o p e r t i e s o f the a v a i l a b l e s e r i e s can enhance h i s understanding and forecasts of the region. t h i s f e e l i s p u r e l y judgmental i n nature: One way o f a c q u i r i n g the a n a l y s t accumulates understanding by i n f o r m a l thought and observation, g e n e r a l l y over a p e r i o d of years. More formal approaches i n v o l v e b u i l d i n g models. S t r u c t u r a l models impose d e t a i l e d and somewhat i n c r e d i b l e assumptions ( " i d e n t i f y i n g r e s t r i c t i o n s " ) about economic r e l a t i o n s h i p s i n an attempt t o e x t r a c t knowledge o t h e r w i s e hidden i n the data. The time series approach a l l o w s d e s c r i p t i o n o f the d a t a w i t h o u t the requirement o f imposing extensive assumptions o r p r i o r knowledge. I t l e t s the data s e t speak f o r i t s e l f . The premise o f t h i s study i s t h a t the r e g i o n a l economist can b e t t e r understand the Ohio economy by s t u d y i n g the p r o p e r t i e s o f i m p o r t a n t Ohio time series. The r e s u l t s show t h a t i n f o r m a t i o n i s a v a i l a b l e from sources t h a t can be i d e n t i f i e d and q u a n t i f i e d through simple r e g r e s s i o n methods t h a t are widely understood. I. The Regional Forecastinq Problem Regional economic time s e r i e s e x h i b i t v a r i a t i o n from secular, c y c l i c a l , and seasonal sources. Regional f o r e c a s t e r s attempt t o assess c u r r e n t a c t i v i t y and t o p r e d i c t the f u t u r e course o f t h e r e g i o n a l economy by e x p l o i t i n g t h e i n f o r m a t i o n contained i n v a r i o u s time s e r i e s . Usually, t h i s process o f e x t r a c t i n g i n f o r m a t i o n i s q u i t e i n f o r m a l and judgmental. I n o t h e r cases, the process i n v o l v e s t h e use o f 'a formal s t a t i s t i c a l model o f some k i n d . s t u d y seeks t o provide.forma1 t o o l s f o r the Ohio f o r e c a s t e r . This http://clevelandfed.org/research/workpaper/index.cfm Best available copy Figure 1 illustrates the single series that is perhaps o f greatest interest to Ohio forecasters: (seasonally adjusted). payroll or establishment-survey employment Although it has exhibited an upward trend, its growth . . has not proceeded smoothly. The strong dependence o f Ohio on national conditions is obvious from the National Bureau of Economic Research peaks and troughs, denoted by " P s " and "T's," respectively. i If history tends to repeat itself, then the regional forecaster can benefit from knowing the trend rate o f growth, any predictable cyclical behavior, and any clues available from national data, such as the leading indicator index. Also, relations between the regional series may potentially aid in forecasting. This paper will describe these characteristics and assess their value to regional forecasters. Regional Forecasting Models 1 1 Regional forecasting models have attracted interest among government and business planners and have proliferated with the availability o f regional data. Many o f these models are of the so-called structural variety, which i nvol ve use o f detai 1 ed assumptions supposedly drawn from economic theory. Their construction reflects a primary goal of estimating the behavorial relationships (structure) corresponding t o the theory, although they are employed for forecasting as well. For some applications, involving analysis o f the effects of structural change o r o f the response o f the regional economy t o particular policies o r events, a structural model is necessary. Despite the recent proliferation o f structural models, little clear evidence exists on their ability to forecast well.' Time series models, the alternatives to structural models, are primarily designed for forecasting. Such models can be built even in contexts in which the theory o r data set required t o build a structural model is unavailable. Figure 1. http://clevelandfed.org/research/workpaper/index.cfm Best available copy ESTABLISHMENT-SURVEY EMPLOYMENT SEASONALLY ADJUSTED http://clevelandfed.org/research/workpaper/index.cfm Best available copy Most regional forecasting problems occur in such a context. The remainder of this paper is organized as follows. A general survey of some related work is presented and forecasting context and data series are Subsequent sections characterize the uni vari ate properties, described. intraregional relationships, and national-regional o r so-called trickle-down relationships. These characteristics are then used to-suggest candidate variables for inclusion in a multivariate autoregressive model (MAR) o f the Ohio economy, using a stepwise regression procedure t o select among the candidates. An priori trickle-down model is also implemented. The latter two models' forecasting ability is compared with that of univariate autoregressions in the 1979-83 out-of-sample period. 111. A Brief Survey o f Previous Work A number o f time series approaches have been implemented t o facilitate regional forecasting. The univariate model represents the simplest approach and uses only the past history o f each regional variable t o predict its future. These models are the most straightforward t o implement, and their forecasts are often as good as--and sometimes better than--more complex models. The forecasting accuracy o f univariate models serves as an appropriate benchmark for evaluating the relative efficiency of other methods. The Box-Jenkins (1970) approach for identifying and estimating autoregressive integrated moving average (ARIMA) models is perhaps the most flexible and also the most popular framework for univariate time series model ing. Multivariate models use the history o f other variables to describe the movement in the series t o be forecast--that is, they exploit delayed interactions, o r lead-lag relations, between series. The identification and http://clevelandfed.org/research/workpaper/index.cfm Best available copy estimation of the appropriate multivariate model is problematic and is currently subject to research along different paths. The essential dilemma o f the regional multivariate model is that o f using as much information as . . possible by including as many relevant series in the equations, yet minimizing the inaccuracy due to multicollinearity and scarcity of degrees of freedom. For example, the more variables that are included, the more sources of information that are incorporated in the resulting model's forecasts, thus tending to improve accuracy. Yet, at the same time, inclusion o f more variables will increase the standard errors of the estimates o f the model's parameters, especially if variables are highly correlated, thus tending to reduce the accuracy bf forecasts. Furthermore, as the results t o fol low will illustrate, more complex models may become unstable and break down out o f the sample used'-t o specify and estimate them. for solving this dilemma is available. Unfortunate.ly, no general procedure Several recent efforts directed toward regional forecasting are o f interest. Anderson (1979) first implemented the "Bayesian approach u o f Litterman (1979) for a regional model o f the Ninth Federal Reserve District. The dilemma referred t o above is dealt with in a clever and promising way: many of parameter estimates is . variables and lags are included, but the variance . limited by the imposition o f a random walk prior distribution. The primary disadvantage o f the procedure is the bias that it introduces into estimates of parameters. The greatest practical difficulty o f the approach is the choice o f appropriate "tightness" restrictions o n the prior.' Li tterman terms the model a "vector autoregression" (VAR) because o f its (a) multivariate nature and ( b ) the absence of moving average parameters (only autoregressive parameters are present). http://clevelandfed.org/research/workpaper/index.cfm Best available copy More recently, Amirizadeh and Todd (1984) have constructed five "Bayesian VAR" models for each of five states of the Ninth Federal Reserve District. They built an elaborate structure of linkages with forecasts of the national economy. They have undertaken real-time forecasting, and plan t o publish their forecasts quarterly. Kuprianov and Luppoletti (1984) adopt a VAR approach, but without imposing priors, and implement models for the individual states o f the Fifth Federal Reserve District. The specification they employ uses six quarterly past values o f state employment and deflated personal income, plus three national variables t o forecast each of the two state variables. Hoehn, Gruben, and Fomby (1 984a, 1984b) and Hoehn ( 1 984) explore a number o f alternative methods for regional forecasting by applying them t o the state of Texas and comparing their performance in an (admittedly short, 10-quarter) out-of-sample forecast period. The Bayesian VAR generally did not perform well relative t o univariate ARIMAs, unless the VAR's prior distribution was tinkered with extensively, in which case its forecasting accuracy in some cases approached, but generally still fell short of, the univariate models. Models with many variables and no priors, using alternatively (a) other regional variables only (a closed-region model) o r (b) national variables only (a trickle-down model), also performed poorly. Using the latter two models with univariate ARIMA models t o form an unweighted combination forecast provided accuracy sometimes competitive with the ARIMAs alone. Hoehn (19841, based o n this experience with alternative models, proposed a method for-building a forecasting model and implements it for Texas. ( A more formal variant of that identification procedure, using the stepwise regression procedure, is described more fully below, where its a p p ~ r c a t i o nt o O h i o series is presented.) Essentially, "causality tests" are first used t o select a http://clevelandfed.org/research/workpaper/index.cfm Best available copy small number o f variables that are candidates for inclusion in the equations Then, combinations of variables and lag structures are used to find well-fitting and parsimonious equations. The resulting model for Texas . . provided out-of-sample forecasts consistently superior to those o f univariate ARIMAs, as measured by the criterion of the root mean square error (RMSE). For some variables and forecast horizons, the difference in forecasting accuracy between the multivariate and univariate model forecasts was significant at the .05 level. The model, while built according to strictly statistical criteria, also appeared quite reasonable in light o f intuitions about the regional economy. IV. The Forecasting Problem and the Approach The objective o f the present study is the construction o f linear forecasting equations that predict the growth rates o f Ohio variables by their own lagged growth rates and by those of each other and national series. For example, let yt(k> be the forecast of the change in the logarithm o f a regional variable Y, for period t+k, for k,O, formed at time t, when a1 1 t realizations are observed. period t growth. For example, the k=l case involves forecasting A linear forecasting equati0.n takes the general form: . . where a r and the b,, are parameters and S , t is the jth element of a vector o f q information variables available at time t. That vector, o r information set, treats each relevant lag as a distinct variable in the above equation. The forecasting equations will be used t o forecast the level of y, with particular emphasis o n the one-to-four quarter (Ockt4) horizons. The http://clevelandfed.org/research/workpaper/index.cfm Best available copy regional variables, Y, o f concern, are the following seasonally adjusted Ohio variables: (1) Payroll Employment, total (PAY ROLL) (2) Payroll Employment, Manufacturing (MFG) (3) Payroll Employment, Nonmanufacturing (4) Household-survey Employment (NONMFG) ( EMPL) (5) Civilian Labor Force (LF) (6) Personal Income ( INCOME > (7) Retail Sales (RETAIL) (8) Housing Starts (STARTS (9) Workweek in Manufacturing (HOURS) (10) Consumer Prices (PRICES) Some of these series were seasonally adjusted by the reporting agency; others were seasonally adjusted either by the data vendor o r by the authors. Some data were transformed from monthly averages to quarterly averages. The Ohio consumer price series required an elaborate method of construction from the Cleveland and Cincinnati Consumer Price Indexes. data sources and adjustments is in Appendix A. A fuller description of The series themselves, after these adjustments, but before transformation to logarithmic growth rates, are listed in Appendix B. The data series each began by at least the first quarter o f 1965 (in the format we adopt, that quarter is denoted 65QI). The working data set for initial analysis included the growth rates for 65QIV through 78QIV, o r 5 3 data points. The period from 79QI t o 83QIV (20 data points) was saved for out-of-sample analysis of models constructed during the initial analysis. http://clevelandfed.org/research/workpaper/index.cfm Best available copy V. Information Gain: A Pedaqogy of the I-Statistic The location of information available about the future course o f a given Ohio series (the identity of the S vector) will be assessed by a systematic . battery o f nested hypothesis tests. . The tests involve successive generalizations of the prediction equation to incorporate additional variables. The value of information will be measured by the improvement in the fit o f an equation as the potentially informative variable is added. The techniques and their underlying statistical basis are presented in this section. A regressor (so-called "explanatory" variable) x is informative (or contains information) about a regressand (so cal led "dependent" variable) y:to the extent that knowledge o f x conditions knowledge o f y. Formally, if ~[y-~(ylx)l~<E[y-E(y)l~then x is informative with respect t o y. An obviously useful quantitative measure o f the information value is the reduction in the condi tional variance relative to the uncondi tional variance. It is an exact measure if the loss attending an error, y-E(ylx), proportional to its square. is When scaled, o r divided, by the unconditional variance, this theoretical measure o f information value is identical to the squared correlation coefficient, r' , where th.e .relation between y and x is . . 1 i near. An a1 ternative measure, I-( 1-rL) "' , expresses the reduction of the expectation of the square root o f the error (standard deviation o f the disturbance term in the linear regression equation) relative t o the standard deviation o f y. This measure is referred to as the information gain from the , . , . It can be use o f x to condition eipectations of y and is denoted I estimated from the standard deviation of y, s,, and the standard ,error o f the regression of y o n x, s,,,: http://clevelandfed.org/research/workpaper/index.cfm Best available copy Where r e a l i z a t i o n s o f the s t a t i s t i c I a r e r e p o r t e d i n t h i s paper, they r e f l e c t m u l t i p l i c a t i o n by 100, so t h a t i n f o r m a t i o n gains are expressed as a percentage of t h e standard d e v i a t i o n . . A set o f variables, XI, xt, . . . x,, . may be assessed f o r c o l l e c t i v e i n f o r m a t i o n g a i n by c a l c u l a t i n g : where it2 is the c o r r e c t e d c o e f f i c i e n t of determi q a t i o n . More g e n e r a l l y , t h e i n f o r m a t i o n c o n t e n t of x above may be of i n t e r e s t i n c o n t e x t s i n which another v a r i a b l e , say z , o r v a r i a b l e s a r e a l s o p o t e n t i a l l y informative. This c o n t e x t i n t r o d u c e s some ambiguity, i n t h a t whether z is i n c l u d e d o r n o t w i l l a f f e c t t h e incremental r e d u c t i o n i n standard e r r o r . Hence, the i n f o r m a t i o n g a i n of x w i t h r e s p e c t t o y i s dependent on which o t h e r v a r i a b l e s a r e i n the i n f o r m a t i o n s e t . Even more g e n e r a l l y , t h e i n f o r m a t i o n g a i n o f a s e t o f v a r i a b l e s can be measured by t h e incremental r e d u c t i o n t h e i r i n c l u s i o n i n a m u l t i v a r i a t e l i n e a r model b r i n g s t o i t s standard e r r o r , s u b j e c t t o t h e i n c l u s i o n o f a s p e c i f i e d ( p o s s i b l y n u l l ) :st o f o t h e r i n f o r m a t i o n variables. Consider t h e f o r e c a s t i n g problem posed by t h e p r e s e n t study, i n which c u r r e n t and f u t u r e values o f y a r e t o be c o n d i t i o n e d on p a s t r e a l i z a t i o n s o f informatlon variables. The i n f o r m a t i o n g a i n from own-lags i s f i r s t assessed by p e r f o r m i n g r e g r e s s i o n ( 1 ) of y on i t s f i r s t two own-lags, i n o r d e r to o b t a i n the r e d u c t i o n i n standard e r r o r o f t h e r e g r e s s i o n e q u a t i o n r e l a t i v e t o t h e standard d e v i a t i o n . http://clevelandfed.org/research/workpaper/index.cfm Best available copy Then the i n f o r m a t i o n gain from any s p e c i f i e d candidate v a r i a b l e x can be assessed by performi ng the r e g r e s s i o n : 2 2 + Z c,x,-, + u,. . . j=1 j =1 The Granqer c a u s a l i t y t e s t (see Granger and Newbold, L19771, pp. 224-6) (2) y t = a , + 1 b:yt-, is e q u i v a l e n t t o a t e s t of whether o r n o t x i s i n f o r m a t i v e w i t h respect t o y . given past y . I t i s based on the F - s t a t i s t i c , c a l c u l a t e d from the sums o f squared e r r o r s o f regressions ( 1 ) and ( 2 ) , denoted S e and S u , r e s p e c t i v e l y : where q i s t h e number o f r e s t r i c t i o n s t e s t e d (e.g., number o f regressors i n the u n r e s t r i c t e d model. = 1 -($)"' c,=O> and k i s the The I - s t a t i s t i c i s : (n-k-l+q)"' n-k- 1 N o t i c e t h a t , a s i d e f r o m t h e adjustment f a c t o r [In-k-l+q)/(n-k-l )1 '/'--which depends u n i q u e l y upon n, k, and q--equal sums o f squared e r r o r s , which a r i s e when the ciao, b r i n g about a z e r o value f o r I. The adjustment f a c t o r e f f e c t i v e l y d e f l a t e s measured improvement i n f i t f o r t h e expenditure o f q a d d i t i o n a l degrees o f freedom i n the u n r e s t r i c t e d r e g r e s s i o n ( 2 ) . expressions r e v e a l t h e correspondence between F and I: These http://clevelandfed.org/research/workpaper/index.cfm Best available copy http://clevelandfed.org/research/workpaper/index.cfm Best available copy This relation is illustrated in Figure 2. The lower bound for I, which occurs i f F=O, is denoted L : L approaches zero as the sample size n increases. (It would be -73 percent for n=6!) L is the proportional reduction in a regression's sum o f squared errors that is expected t o occur from the inclusion of q noninformative regressors. It may also have some interpretation as a measure o f the imprecision arising from finite degrees o f freedom. Given a sample size of 53, as for the period up t o 78QIV, L=-2.04 percent. For the sample through 83QIV, n-73 and L=-1.46 percent. L is, roughly, inversely proportionate t o n; L i s o f order n - '. If F=l, the proportional reduction o f sum of squared errors of L is achieved and I is zero. As F approaches infinity (as the linear relation becomes more precise), I approaches 100 percent. These t w o properties are desirable and illustrate the usefulness of I. In the causality tests based o n the extended sample period (n=73), the critical F-values are: F.o,(2,68) and F.01(2,68) 3.13 = 4.94 which correspond t o I-statistics of: I.or(2,68) = 1 - 70 (68+2(3.13) '/' = 2.91 percent = 5.19 percent The most common criterion for inclusion of a variable in a model if the ad http://clevelandfed.org/research/workpaper/index.cfm Best available copy hoc rule that the t-statistic must exceed 2 in absolute value. This can be shown to be equivalent t o the following inequality: VI. Univariate Properties The mean and standard deviation o f each series' growth rate provide measures o f the average growth rate and its variability. Equivalently, they provide the parameter estimates for the simplest univariate model worthy of consideration, the random walk model. This model is of the form: yt = a, + e t where a, is the drift parameter and e t is a random variable with zero autocorrelation at all lags (white noise) and a constant variance u:. The random walk model serves merely to re-establish the appropriate level o f the forecast function after acquisition of a new quarterly data point. Future growth rates are revised only to the extent that the expected long-term average growth rate, a,, is revised. In particular, cyclical behavior--persistence in high o r low growth rates--is ruled out in the random walk model. u,, The mean and standard deviation, taken as estimates o f a, and respectively, are shown in table 1 , in the first t w o columns, f o r the longer sample ending 83QIV. for the 10 Ohio series. Cyclicity of growth in time series is the tendency of persistence in above- or below-average growth from one period t o the next. This persistence can be described by the correlation between.rates of change across different intervals. The series o f such correlations at various intervals is called the autocorrelation function: http://clevelandfed.org/research/workpaper/index.cfm Best available copy Table 1 Univariate Properties Sample: 65QIV - 83QIV Autocorrelation at lag 2 3 4 yt 1 - - . . Series Mean Standard Deviation PAY ROLL .0028 .0099 .57 .32 .22 .12 .58 .OO 17.7** - ,0027 .0204 .45 .22 .07 -.07 .45 .03 10.1** NONMFG .0054 .0070 .42 .32 .38 .37 .35 .19 9.6** EMPL .0028 .0130 .08 -.08 .02 .09 -.08 I NCOME .0181 ,0136 .40 -07 .06 -.05 RETAIL .0164 .0258 -.26 .09 -.03 .06 -.24 .04 2.0 STARTS -.0108 .I643 .06 .07 -.lo -.23 .05 .07 -1.0 HOURS .0001 .0107 .ll -.07 .06 -.06 .ll -.09 -0.5 PRICES .0169 .0098 .56 .46 .34 .31 .42 .24 MFG .06 Autoregressian Equation: + b l y t - , + bryt-, + e, = a + bly,-, .44 - -0.8 1 7.6** 19.1** **Significant at the .O1 level. I = [(standard deviation - . . standard error of autoregression)/tstandard devSation)l x 100. http://clevelandfed.org/research/workpaper/index.cfm Best available copy Given the sample size n, no autocorrelations are significantly different from zero (at the .05 level, two-tailed) i f they all fall between approximately With our samples of 73, the r , must exceed 0.23 in absolute -+2n-"'. value to provide strong evidence of persistence from quarter to quarter. The autocorrelation function for lags one through four is presented in columns three through six of table 1. The table reveals substantial positive persistence in growth rates for prices, payr0.11 employment and its two components, and personal income. The presence of autocorrelation in both payroll sectors implies that cyclical variation in Ohio employment is attributable to both the manufacturing and nonmanufacturing sectors. The household survey based measure of employment, EMPL, exhibited no significant autocorrelation. (1t.i~ interesting to note that all of the foregoing results regarding autocorrelations of Ohio series are consistent with those for Texas in Hoehn, Gruben, and Fomby C19841). The significant autocorrelation in the f i v e s-eries mentioned above suggests a persistence in growth rates that can be exploited by the regional forecaster. An appropriate measure of the value of information contained in the history of the series can be found by first estimating a second-order autoregression (which we denote as A R 2 > , http://clevelandfed.org/research/workpaper/index.cfm Best available copy u s i n g t h e o r d i n a r y l e a s t squares method, and then comparing t h e standard e r r o r of t h i s e q u a t i o n , s,, t o t h e standard d e v i a t i o n o f y , s,. can be expressed i n terms o f t h e i n f o r m a t i o n g a i n , [(s,-s,)/s,I . . The comparison x 100. Table 1 , i n t h e l a s t t h r e e columns, r e p o r t s t h e e s t i m a t e d r e g r e s s i o n c o e f f i c i e n t s and t h e a u t o r e g r e s s i v e i n f o r m a t i o n measure f o r each r e g i o n a l time series. R e s u l t s i n d i c a t e t h a t t h e one- quarter- ahead p r o j e c t i o n o f the consumer p r i c e measure has a standard e r r o r about o n e - f i f t h l e s s , when account i s made o f t h e l a s t two q u a r t e r l y growth r a t e s . A g a i n of 18 p e r c e n t i s found f o r p a y r o l l employment, g a i n s of about 10 p e r c e n t a r e achieved f o r the two p a y r o l l c a t e g o r i e s and 8 p e r c e n t f o r personal income. (These r e s u l t s o n l y r e f l e c t t h e e s t i m a t e d i n f o r m a t i o n v a l u e of two lagged g r o w t h r a t e s , whereas a u t o c o r r e l a t i o n f u n c t i o n s e v a l u a t e p e r s i s t e n c e a t l o n g e r l a g s as w e l l . ) VII. Intrareqional Information The v a l u e o f r e g i o n a l s e r i e s i n foreshadowing each o t h e r can be measured i n t h e f o l l o w i n g way. Regressions a r e performed t o e s t i m a t e t h e standard e r r o r o f t h e e q u a t i o n s p e c i f i e d by: where y and x , a r e two r e g i o n a l s e r i e s . I f the series x k t r u l y aids i n f o r e c a s t i n g y , t h e n t h e s t a n d a r d e r r o r o f t h i s b i v a r i a t e e q u a t i o n w i l l be lower than f o r t h e a u t o r e g r e s s i o n ( i n which c , = CL = 0 i s imposed). The j o i n t significance t e s t o r F- test for the b , provides a " causality" t e s t i n t h e sense o f Granger (Granger and Newbold, 1977, p. 225). f i r s t 10 rows, r e p o r t s r e s u l t s o f these r e g r e s s i o n s . Table 2, i n t h e The r e d u c t i o n i n t h e http://clevelandfed.org/research/workpaper/index.cfm Best available copy Table 2 Independent Variables ............................... PAYROLL MFG InformationGain' Dependent Variables---------------------------- NONMFG INCOME Reqional 2.29 MFG - . 42 NONMFG -.53 4.51* 13.38** PRICES 4.90* 2.65 1.48 .49 3.10 1 -9 5 -.98 9.33** 3.92 3.02* 1.20 4.07* 3.84, - 1.45 9.35** 9.49** 1.33 - 1.40 1.93 -.46 .20 .01 4.26* 2.28 .30 6.25** .87 -.73 -.43 - 1.26 -.a4 - 1.25 - 1.31 - 1.03 INCOM -.I6 .01 1.61 RETAIL -.55 - 1.03 3.08* -1.06 .60 STARTS 8.16** 10.34* 2.39 2.51 - 1.29 9.45** HOURS -. 8 0 2.81 .58 - 1.43 -.72 PRICES 3.38* .51 2.25 National LEAD 19.25** 21 . 7 9 * * COIN 14.51** 22.32** PRODUCT 8.02** 14.55** USPAYROLL 9.44** 10.62** USMFG 6.61** 13.19** USHOUSEHOLD 4.11* 7.91** REALY P 4.53* 8.45** USLF HOURS 7.49** .48 LF STARTS . . PAY ROLL EMPL RETAIL - 1.22 -.88 CPI 6.93** 6.48** PPI 2.32 2.17 DEFLATOR 5.20** 3.32* 3.40* 1.87 5.08* 3.12* -.99 -.52 3.93* -1 .OO .55 -.38 1.45 2.71 3.09* - 1.20 - 1.07 -.64 11.36** - 1.22 .75 -1 - 3 1 http://clevelandfed.org/research/workpaper/index.cfm Best available copy Table 2 c o n t i n u e d , Independent Variables ............................... Dependent Variables---------------------------- INCOME :RETAIL - P AY R O LL MFG NONMFG USREALSALE 6.04** 7.96** 2.69 4.49* USSTARTS 3.43* .69 4.64* -.75 REALGNP 4.78* 9.74** 3.27* 10.80** GNP -.76 .50 -.lo 2.73 USY P 4.31* 6.69** -.67 .71* 1.42 -.94 6.89** 4.11* FUNDS .98 MOODY 7.67" Information Gain' .18 S TA RT S HOURS PRICES 6.41** .84 -.68 7 . 9 2 * * - 1 .15 .73 .70 -.80 -.28 .67 9.21** .62 1.21 7.18** -.83 .28 1.65 -.35 5.78** 3.14* -.75 -.04 3.11* -.86 5.24** 7.39** 1.69 .63 3.76, -.92 .79 9.64** 2.85 --.I0 .25 4.13* 1.25 -.76 11.55** 10.72** - 1.16 .69 2.31 * S t a t i s t i c a l l y s i g n i f i c a n t a t t h e .05 l e v e l ; g a i n exceeds 2.91 c r i t i c a l v a l u e . * * S t a t i s t i c a l l y s i g n i f i c a n t a t t h e .O1 l e v e l ; g a i n exceeds 5.19 c r i t i c a l v a l u e . ' For each c o m b i n a t i o n o f dependent and i n d e p e n d e n t v a r i a b l e s , t h e f i g u r e s i n t h e t a b l e show: 1 I = ( s t a n d a r d e r r o r o f t h e AR2 e q u a t i o n ) - ( s t a n d a r d e r r o r , o f t h e b i v a r i a t e e q u a t i o n ) x 100 s t a n d a r d e r r o r of t h e AR(2) e q u a t i o n http://clevelandfed.org/research/workpaper/index.cfm Best available copy standard error is expressed as a percent of the univariate autoregression equation's standard error. Significant evidence, at the -05 level, is found for 25 different . . causalities, or leading relations, involving regional variables. Housing starts is the only series that provided significant leading information about the total payroll employment. Housing starts and personal income appear to be the two most useful regional series: they account for 5, 4, and 4 of the significant results, respectively. These series may. however, merely reflect the same underlying forces as are more clearly revealed in national indicators. Of the two components of payroll employment, the manufacturing sector measure contains 1 eadi ng information about the nonmanufacturing sector but not vice versa. Surprisingly, the manufacturing workweek, HOURS, tended to lag behind manufacturing employment. Hours had been included in this study in the expectation that they would provide leading information on employment. The consumer price and retail sales series were the only ones for which other regional variables provided no leading information. VIII. National-Reqional Information regional series can be The value of national series in foreshadowing . . . . measured in a way analogous to the regional interactions of the previous section. Regressions are performed to estimate the standard error of the equation specified by the bivariate equation in section VII, where x, is the quarterly logarithmic growth rate of one of the 18 national variables listed in the Appendix A glossary. Rows 11-28 of table 2 report the national variable information gains. Of 180 possible relations, 89 are significant at the .05 level. Most notable is the dependence of the employment series on national economic conditions. Of the two payroll sectors, the manufacturing http://clevelandfed.org/research/workpaper/index.cfm Best available copy sector i s most dependent on the n a t i o n . This dependence conforms t o a v a i l a b l e p r i o r n o t i o n s , which tends t o c o n f i r m both the notions and the p r e s e n t methodology. Ohio payroll. employment tends t o r e f l e c t , . t o a substantial . degree, previous movements i n the n a t i o n a l leading and c o i n c i d e n t indexes, the n a t i o n a l p a y r o l l s e r i e s , and several o t h e r indicators - - even when lagged values ( a u t o r e g r e s s i o n s > o f the Ohio p a y r o l l s e r i e s i t s e l f a r e taken i n t o account. The manufacturing workweek and household- survey employment d i s p l a y a s i m i l a r dependence on p a s t n a t i o n a l c o n d i t i o n s t h a t i s s i m i l a r t o t h a t o f p a y r o l l employment. Movements i n Ohio personal income and housing s t a r t s appear t o r e f l e c t p a s t n a t i o n a l c o n d i t i o n s more than t h e i r own p a s t movements. Least dependent on p a s t n a t i o n a l conditions., s u r p r i s i n g l y , a r e Ohio r e t a i l s a l e s and consumer p r i c e s . ( I n the Texas study, r e t a i l sales and consumer p r i c e s were more s t r o n g l y r e l a t e d t o n a t i o n a l i n d i c a t o r s . ) We cannot r e j e c t t h e n o t i o n t h a t r e t a i l sales and consumer p r i c e s a r e exogenous w i t h r e s p e c t t o the o t h e r series. One o f the most u s e f u l n a t i o n a l i n d i c a t o r s i s the n a t i o n a l p a y r o l l s e r i e s , which i s s i g n i f i c a n t l y causal w i t h r e g a r d t o a l l of t h e Ohio s e r i e s except r e t a i l sales and p r i c e s . Others o f p a r t i c u l a r value a r e t h e composite i n d i c e s o f l e a d i n g and c o i n c i d e n t s e r i e s , i n d u s t r i a l p r o d u c t i o n , and manufacturing payrolls. The U.S. consumer p r i c e index and the long- run i n t e r e s t r a t e appeared t o c o n t a i n l i t t l e l e a d i n g i n f o r m a t i o n f o r t h e r e g i o n a l f o r e c a s t e r when we used d a t a through 78QIV. b u t became more i n f o r m a t i v e when t h e sample was extended. Generally, though, t h e p r i c e and i n t e r e s t r a t e s e r i e s were r e l a t i v e l y uninformative. http://clevelandfed.org/research/workpaper/index.cfm Best available copy IX. A Trickle-Down Model A simple trickle-down model was built that attempted to summarize the information from sources that actual regional forecasters are likely to be . currently placing greatest emphasis on. . In each equation for regional variables, right-hand-side variables included a constant, two own-lags, Ohio payroll employment, and one lag each o f the national leading and coincident indexes. The t w o national series' equations include two own-lags and one lag o f the other national series. The resulting model, which will be referred to as the trickle-down (TD) model, may be both too unparsimonious and not fully reflective of the information avai lable from the causality tests. O n the other hand, it embodies a rough prior notion about which series ought to b e most valuable t o the regional forecaster. Hence, it represents an interesting alternative and benchmark for a regional forecaster. It may be especially useful in combined forecasts, t o be considered later. The trickle-down model is presented in Table 3. As an illustration and an aid t o interpreting that table, the equation for payroll employment i s presented below. respect: It should be noted that this aquation is unique in one because the lagged growth o f payroll employment is the first own-lag of the equation, there is one less parameter than in the equations for the other nine regional equations. http://clevelandfed.org/research/workpaper/index.cfm Best available copy Table 3 Trickle- Down E q u a t i o n s : A l n y t = a + b , A l n y t - , + b , A l n y t - c + C , A ~ ~ L E A D +, - ~C , A ~ ~ C O I N+~ -c,AlnPAYROLL,-, ~ + e, Using Data f r o m 65QIV-83QIV Dependent Variable Parameter E s t i m a t e s (and Standard E r r o r s ) Goodness- of- Fit Measures .0004 (.0008) -.06 (.20) .36 (.I21 .18 .06) .16 (.13) -- .56 .006524 19.5 ( MFG -.0104 ( .0028) -.44 ( .24) .26 (-12) .31 (.I21 .70 (.29) .17 (.51) .55 .01369 25.3 NONMFG .0030 (.0012) .02 .20 (.I21 .03 (.I21 7.5 ( .20 .25) -005882 ( .07 .05) -31 ( . 19) .0014 (.0014> -.34 (.I31 -.31 (.12) -.I3 ( .08) .68 (.21) -. 13 .29 -010960 16.1 INCOME .0148 (.0034) -.08 (.18) .07 (.I11 .18 ( .08) .39 (.20> -.I6 (.35) -43 .O 1030 18.4 RETAIL .0197 (.0043> -.32 (.I31 .05 (.I21 .39 (.19) -. 62 .90 (.76) -06 .02498 1.2 -.0120 (.0195> -. 18 -.01 (.12> 4.16 (1.33) - 4.23 (2.93) -. 84 .16 .I510 9.0 HOURS -.0022 (.0011) -.35 (.I31 -. 08 .21 (.06) .44 -. 78 ( . I S . ) - - . (.24> -44 .007976 25.5 PRICES .0068 (.0023) .42 (.I11 .19 (.I21 -. 12 .25 (.I51 -.35 (.24) .35 .007898 0.7 .0057 (.0024) .84 (.12) .22 (.13> --- -. 89 .43 .01919 12 .O (.I91 --- .0015 (.0016) .02 (.16) .22 (.12) .51 ( .09) -- -- .54 .01276 15.7 PAY ROLL EMPL STARTS LEAD COIN (.I31 - (.lo) ( .06) (.49) -- -- ( .34) (4.47) -- * I is the percent reduction o f the standard error of the trickle-down equation relative to the AR(2) regression equation.. http://clevelandfed.org/research/workpaper/index.cfm Best available copy X. The Srzpwise Regression Model Causa:'ty t e s t s performed u s i n g the sample ending 7 8 Q I V ( n o t r e p o r t e d ) served as :he p o i n t of departure f o r b u i l d i n g a m u l t i v a r i a t e autoregression model f o r :hie. The o b j e c t was t o f i n d a w e l l - f i t t i n g , y e t reasonably parsimonious, equation f o r each of the r e g i o n a l s e r i e s . each serie;, I n the equation f o r every v a r i a b l e t h a t was s i g n i f i c a n t a t t h e 0.10 l e v e l i n the c a u s a l i t y r e s t s was a candidate f o r i n c l u s i o n . The search f o r a p p r o p r i a t e equations $as made problematic by the l a r g e number of s i g n i f i c a n t causal r e l a t i o n s ~ i s c o v e r e d 3. The mccel was c o n s t r u c t e d u s i n g a single- equation method; t h a t i s , each e q u a t i o n was chosen ( i d e n t i f i e d ) and estimated i n i s o l a t i o n . More complex i d e n t i f i c a r i o n and e s t i m a t i o n procedures might be s l i g h t l y more e f f i c i e n t , though less t r a n s p a r e n t . A l e s s formal and more judgmental, b u t s i m i l a r methodology i s described i n Hoehn (1984). " automatic" and formal procedure. The p r e s e n t method employs a more The process of s e l e c t i n g t h e f i r s t equation o f the mode-'s, f o r p a y r o l l employment, i l l u s t r a t e s the p r e s e n t procedure, which i s b; red on t h e stepwise r e g r e s s i o n technique. A s u b r o u t i n e f r o m PEC (Program f o r Econometric Computation, Kim Pec, Yale U n i v e r s i t y ) was employed. T h i s program proceeds by " forward stepping," o r adding v a r i a b l e s t o t h e e q u a t i o n t h a t o b t a i n e d t - s t a t i s t i c s of 1.96 o r more i n a b s o l u t e value, and "backward stepping," o r removing v a r i a b l e s whose t - s t a t i s t i c s f e l l below one i n absolute value a f t e r o t h e r v a r i a b l e s a r e included. The backward- stepping f e a t u r e appears t o reduce the importance o f the o r d e r i n which v a r i a b l e s a r e i n c l u d e d i n the f o r w a r d steps. (As a p r i m i t i v e check, t h e o r d e r o f v a r i a b l e s was e x a c t l y reversed f o r the PAYROLL equation, b u t t h e e q u a t i o n t h e stepwise procedure s e l e c t e d was u n a f f e c t e d by t h a t reordering.) The stepwise procedure a r r i v e d a t an e q u a t i o n f o r Ohio p a y r o l l employment t h a t had ( a forced c o n s t a n t http://clevelandfed.org/research/workpaper/index.cfm Best available copy plus) the second lag of Ohio housing starts, plus one lag of the national coincident index. This equation had a standard error of .006308. In a third step, the same stepwise routine was repeated except that two own-lags were forced (that is, included regardless of their significance). This resulted in inclusion of the first lag of the national leading index and payroll employment, the second lag of national real personal income, and two lags of national housing starts. This equation, with eight parameters in all, had a standard error of .005316. Finally, the significant lags of each of the causal variables was tried to see if its inclusion would substantially reduce the standard error. In only three cases did this occur: the first lag of Ohio housing starts reduced the standard error to .005194; the first lag of national real retail sales, to .005188; and the second lag of national payroll employment, to .005288. An ad hot choice was made to tentatively include U.S. retail sales, but to exclude the other two. Last, some tinkering was done with the equation on an ad hoc basis. For the equation for payroll employment, elimination of the (insignificant) second own lag was tried, but that increased the standard error too much. The equation thus settled upon is that shown below. The stepwi se model's other equations were determined in a similar manner based on the sample ending 78QIV. Their specifications are available f-rom the authors upon request. http://clevelandfed.org/research/workpaper/index.cfm Best available copy XI. Contemporaneous Correlations The information gains described in the last three sections involve lead-lag relationships and ignore contemporaneous relationships. The latter cannot be used for forecasting the future. They are valuable, however, in estimating as-yet unreported realizations o f variables conditional on reported figures for other variables. These conditional estimates are important to real-time forecasting and monitoring of the regional economy. For example, the analyst may desire to estimate personal income for a period for which employment data are available, but a direct measure of income is not. The contemporaneous correlations between growth rates of the 10 O h i o variables and the U.S. leading and coincident indices are shown in the upper half o f table' 4. The bottom half shows correlations between residuals of the autoregressions. These residuals are nearly uncorrelated with their own past values, so that their correlations with each other, unlike those of raw growth rates, are uncontaminated by autocorrelation that can lead to spuriously significant relationships. Sample correlations have an approximate variance of n-' , so they are significant at the .OS level if they exceed approximately 2n-"' -- 0.23. Correlations among variables appear not t o be due merely t o autocorrelation. The national series, especially the coincident index, have substantial correlation with the employment and income series. The payroll employment hours, and income series generally display the highest correlations with other series. Payroll figures contain more information about current personal income than d o household-survey figures. The low correlation between manufacturing and nonmanufacturing payrolls, despite their high correlation with the U.S. coincident index, suggests that shifts between them--intersectoral technology o r preference shifts at the regional level--are http://clevelandfed.org/research/workpaper/index.cfm Best available copy important. (Lillian [19821, interprets national employment and unemployment fluctuations as arising from intersectoral shifts.) Ohio consumer prices and the labor force show li-ttle dependence on the national business cycle or o n other regional series. Housing starts and retail sales are weakly related to other series. XII. Out-of-Sample Forecasting: Univariate Models The ultimate proving-ground of any forecasting procedure is its performance outside o f the sample over which it was identified and estimated. The partition o f data available for the present study into a model-building period and an out-of-sample forecasting period was motivated by a desire t o provide evidence of the efficiency o f the forecasting model immediately. rather than after the passage o f time t o allow evidence t o accumulate. The 10-quarter period o f the Texas study appeared t o o short, because the systematic improvements o f the MAR relative t o the univariate benchmarks were generally not found t o be statistically significant. A period of 20 quarters was therefore reserved for out-of-sample forecasting in the O h i o study. period began in 7991 and ended in 8 3 Q I V . This A longer reserved period would have had the cost of unreasonably reducing the a m o u n t of data that could be used t o identify the appropriate forecasting model. The k-step-ahead forecast error for a period t forecast is et,r = yt - yt-,(k) where y is the logarithm o f the series (the level, Jon the growth rate) and y,-,(k) is the k-step-ahead forecast y, formed at time t-k (conditioned o n real i zations dated t-k and earl i er) . The criterion employed for forecast performance evaluation is the root mean square error (RMSE);4 http://clevelandfed.org/research/workpaper/index.cfm Best available copy Table 4 Contemporaneous C o r r e l a t i o n s 65QIV-83QIV C o r r e l a t i o n C o e f f i c i e n t s of G r o w t h R a t e s PAYROLL MFG .92 NONMFG .82 MFG - NONMFG EMPL - INCOME RETAIL STARTS . HOURS PRICES LEAD - . .54 EMPL LF I NCOME RETAIL STARTS HOURS PRICES LEAD COIN C o r r e l a t i o n C o e f f i c i e n t s of Residuals i n Second Order Autoregressions 65QIV-83QIV PAYROLL MFG NONMFG EMP L LF INCOME RETAIL STARTS HOURS PRICES L EAO COIN .87 MFG - NONMFG EMPL INCOME RETAIL STARTS HOURS PRICES LEAD http://clevelandfed.org/research/workpaper/index.cfm Best available copy where n i s the out- of- sample size and RMSE(k) denotes the r o o t mean square e r r o r of the k-step-ahead f o r e c a s t s . The mean e r r o r provides i n s i g h t i n t o the e x t e n t t o which the RMSE i s due t o b i a s i n o r t o variance o f the f o r e c a s t r e l a t i v e t o r e a l i z e d values. I n e v a l u a t i n g each f o r e c a s t i n g method, the model was re- estimated each q u a r t e r t o r e f l e c t a-new q u a r t e r of data. The models were n o t r e - i d e n t i f i e d each q u a r t e r , however, so t h a t the procedure does n o t f u l l y r e f l e c t the e f f i c i e n t use o f new i n f o r m a t i o n t h a t a r e a l - t i m e f o r e c a s t would make. This c o n s i d e r a t i o n i s o n l y r e l e v a n t f o r the stepwise model, because i t was the o n l y one n o t s p e c i f i e d 5 p r i o r i . An examination o f the random walk model i s p a r t i c u l a r l y i n s t r u c t i v e because o f i t s s i m p l i c i t y . Only one parameter, a, c o n s t r u c t the random walk f o r e c a s t . Since a, needs t o be estimated t o i s merely t h e average growth . between t h e l o g of the . r a t e , i t can be c a l c u l a t e d by d i v i d i n g the difference l a s t value o f t h e v a r i a b l e from t h e l o g o f i t s i n i t i a l value ( a t time p e r i o d zero) by the l e n g t h of the s e r i e s , t: a,,,, where a,,,, = t-'(yt - yo) i s t h e estimated value o f a, conditioned on d a t a a v a i l a b l e a t time t, and y i s the n a t u r a l l o g o f the v a r i a b l e . The f o r e c a s t f u n c t i o n , which associates a forecasted value o f y w i t h each k steps ahead, i s A t t+k, the e r r o r yt+,-y,(k) i s calculated. The l e v e l o f the f o r e c a s t http://clevelandfed.org/research/workpaper/index.cfm Best available copy Table 5 Out-of-Sample I n f o r m a t i o n Gains: Reduction i n 1-Period-Ahead RMSE ( f i g u r e s i n parentheses show g a i n due t o r e d u c t i o n i n ME) RMSE(ME> of Random Walk Mode 1 PAYROLL MFG NONMFG EMP L LF INCOME RETAIL STARTS HOURS PRICES AVERAGE Reduction i n Four-Period-Ahead RMSE ( f i g u r e s i n parentheses show g a i n due t o r e d u c t i o n i n ME) RMSE(ME> o f Random Hal k Mode 1 PAYROLL MFG NONMFG EMPL LF INCOME RETAIL STARTS HOURS PRICES AVERAGE 1 I .0440 .0799 .0346 .0357 .0165 .0460 .0413 .5346 .0258 .0430 AR2IRH TDlAR2 SWlAR2 UC2lAR2 (-.0384) (-.0584) (-.0328) (-.0287) (-.0148) (-.0358) (-.0280) (-.I2561 (-.0002) ( .0265) 3.3 ( 9.1) 8.2 ( -2.5) 8.0 ( 10.3) 9.8 ( 4.0) * S i g n i f i c a n t a t t h e .05 l e v e l , according t o a t e s t adapted f r o m Ashley, Granger, and Schmalensee (see t e x t ) . http://clevelandfed.org/research/workpaper/index.cfm Best available copy function is revised upward by that error. In addition, the growth rate, or slope o f forecast function, is also revised at t+l by ( t + l > - ' times the error. The M E and RMSE for the first 10 steps ahead for the random walk model were calculated and are reported in table 5 for steps 1 and 4. Three characteristics of the results are particularly worthy o f note. First, the mean errors indicated that forecasts were typically for too-high growth, except for consumer prices (whose errors were on average positive) and the Ohio manufacturing workweek (whose forecasts were nearly unbiased). Second, the increase in RMSEs as the forecast horizon lengthens revealed that uncertainty about the series is unbounded as the horizon is extended for aF1 series, except for the workweek. In other words, only the workweek appears to have a stationary trend. (In fact, it appears to be stationary in its level.) Consequently, none of the series, except hours, should be treated in any empirical analysis as having deterministic trends; their trends are stochastic. Third, the mean absolute error accounted for most of the magnitude o f the RMSEs for all series, except the workweek, for forecasts of more than a quarter o r two ahead. What this implies is that the main source of forecast errors was the overall weakness o f the Ohio economy during most o f the 7991-83QIV period, rather than great variability in forecast accuracy from quarter to quarter. The random walk model serves as the appropriate benchmark for the autoregressive model. The out-of-sample comparison can reveal whether the autoregression found in the within-sample period not only continued to occur in the out-of-sample period, but also was sufficiently stable in its character t o be a dependable source of forecasting information. The out-of-sampl e performance of the second-order autoregressive equations generally compares http://clevelandfed.org/research/workpaper/index.cfm Best available copy f a v o r a b l y w i t h t h e random walk model. The RMSE of the AR2 was lower than f o r the random walk f o r seven of the ten r e g i o n a l v a r i a b l e s i n one-step-ahead f o r e c a s t s and f o r s i x o f the 10 i n four- step- ahead f o r e c a s t s . These comparisons, and those between the AR2 and the o t h e r f o r e c a s t i n g methods, are shown i n t a b l e 5 . I n the cases f o r which the random walk model outperforms the AR2, the d i f f e r e n c e i s modest. But some of the improvements o f the forecasts o f t h e a u t o r e g r e s s i v e equations over those o f t h e random walk a r e substantial. For example, t h e one- quarter- ahead forecasts o f PAYROLL had an RMSE o f .0089 i n the AR2 model, 26 p e r c e n t below the RMSE o f .0121 f o r the RW model. The mean e r r o r was -.0033 i n t h e AR2, compared w i t h -.0077 i n the RW model. The r e d u c t i o n i n t h e RMSE i n t h e AR2 model r e l a t i v e t o the RW model- can be a t t r i b u t e d t o r e d u c t i o n i n the a b s o l u t e value o f t h e mean e r r o r ; t h e l a t t e r r e d u c t i o n , .0044, represents 36 p e r c e n t o f the RMSE o f the RW model. The f i g u r e s i n parentheses i n t a b l e 5 i n d i c a t e t h a t t h e general improvement i n f o r e c a s t accuracy o f t h e AR2 model r e l a t i v e t o t h e RW model i s due t o r e d u c t i o n i n the absolute value of the mean e r r o r . The a u t o r e g r e s s i v e terms tended t o presage o r adapt t o c y c l i c a l movements, which tended t o e x e r t a downward i n f l u e n c e on t h e s e r i e s i n the 1979-83 p e r i o d . . . of t h e AR2 r e l a t i v e t o t h e RW The improvement i n f o r e c a s t i n g performance . . model was g r e a t e s t f o r p a y r o l l employment, i t s nonmanufacturing component, consumer p r i c e s , and personal income. The comparison was most unfavorable t o t h e AR2 model f o r the l a b o r f o r c e , household- survey employment, and housing starts. There was l i t t l e difference i n f o r e c a s t accuracy f o r r e t a i l sales. The out-of-sample r e s u l t s tend t o c o n f i r m the presence o f u s e f u l a u t o r e g r e s s i o n i n PAYROLL, MFG, and PRICES. INCOME had b o r d e r l i n e a u t o r e g r e s s i v e p r o p e r t i e s w i t h i n sample, b u t t h e out- of- sample r e s u l t s suggest moderately s t r o n g autoregression. NONMFG d i s p l a y e d no a u t o r e g r e s s i o n w i t h i n http://clevelandfed.org/research/workpaper/index.cfm Best available copy sample, but substantial autoregression out of sample. Results for the 1979-83 period confirmed the lack of autoregression in EMPL, STARTS, HOURS, and RETAIL. LF was borderline within sample, but was ultimately seen to lack useful autoregression. All these conclusions are verified by the estimation, using the sample through 1983, of the AR2 equations and their associated I-statistics, shown in the bottom half of table 2. XIII. Out-of-Sample Forecastinq: Multivariate Models The univariate autoregression results serve as the appropriate benchmark for the muitivariate models, which add terms to the autoregressiv-e equations in an attempt to capture information from other national and regional data. The out-of-sample evidence generally suggests that such information can be extracted. Table 5 displays the relative forecast performance of the trickle-down and stepwise models; their RMSEs are generally lower than those of the autoregressive model. Figure 3 depicts the performance of both the multivariate and univariate models, in their forecasts of payroll employment for forecast horizons of one to 10 quarters. The relative efficiency of . . multivariate as compared with the univariate autore7ressions do not derive particularly from reduction in the magnitude of bias, but rather more to a closer "fine-tuning" of the forecast each quarter in light of national and regional data. The payroll variable had little importance in the trickle-down model. Hence, the trickle-down model's forecasting efficiency relative to the autoregressive model can be taken as an indication of the usefulness of the lagged trickle-down relationships. In other words, those relations are sufficiently strong and stable to be useful. Figure 3. http://clevelandfed.org/research/workpaper/index.cfm Best available copy ROOT MEANS OF SQUARE ERROR OF THE PAYROLL FORECASTS - Random Walk Trickle-Do wn ommommmmmm STEPS AHEAD http://clevelandfed.org/research/workpaper/index.cfm Best available copy The t r i c k l e - d o w n model, as estimated w i t h the 1965-78 sample, suggested strong i n f o r m a t i o n gain r e l a t i v e t o the autoregressive model f o r PAYROLL, MFG, EMPL, INCOME, HOURS, and NONMFG. This strong gain c a r r i e d over t o comparisons . . o f RMSEs i n the 1979-83 p e r i o d , f o r a l l these v a r i a b l e s except NONMFG. Weaker gains i n LF and STARTS found i n 1965-78 were confirmed i n the f o r e c a s t i n g period. The absence o f gain f o r PRICES was a l s o confirmed. Finally, i n f o r m a t i o n g a i n f o r RETAIL was n o t found i n e i t h e r the 1965-78 o r the 1965-83 sample, b u t arose i n the f o r e c a s t performance comparisons. Aside from the r e s u l t s f o r RETAIL, the s h o r t and long samples and the out- of- sample f o r e c a s t simulation provide consistent r e s u l t s : i n one-period-ahead RMSES, i n f o r m a t i o n gains, I, and r e d u c t i o n s were remarkably simi l a r f o r each v a r i a b l e . The s t a t i s t i c a l s i g n i f i c a n c e of the improvement i n f o r e c a s t accuracy o f t h e TD model r e l a t i v e t o the AR2 model can be measured by t h e method proposed i n Hoehn (1984, pp. 27-81. The method i n v o l v e s an a d a p t a t i o n o f a " c a u s a l i t y " t e s t suggested by Ashley, Granger, and Schmalensee (1980). A t t h e .05 l e v e l , one-period-ahead f o r e c a s t RMSEs are s i g n i f i c a n t l y lower f o r PAYROLL, MFG, LF, and INCOME. For four- period- ahead f o r e c a s t s , the TO model i s s i g n i f i c a n t l y b e t t e r o n l y f o r MFG and LF. s i g n i f i c a n t l y worse. I n no case does the t e s t f i n d t h e TD forecasts The t e s t has some problematic i n t e r p r e t a t i o n s i n some cases, and r e s u l t s do n o t o f t e n conform t o i n t u i t i o n s , suggesting a l i m i t e d usefulness o f t h e t e s t . These ambiguities a r i s e f r o m t h e need t o make an e s s e n t i a l l y f o u r - t a i l e d t e s t u s i n g a s i n g l e F - s t a t i s t i c , u s u a l l y used f o r one- tailed tests. As a r e s u l t , t h e t e s t i s o f t e n of low power. I n f o r e c a s t i n g w i t h the stepwi se model , the exogenous n a t i o n a l v a r i a b l e s used (14 d i f f e r e n t v a r i a b l e s , n o t d i s t i n g u i s h i n g d i f f e r e n t l a g s ) were f o r e c a s t e d u s i n g second-order autoregressive equations. This may have handicapped t h e SW model somewhat i n f o r e c a s t s o f more than one q u a r t e r http://clevelandfed.org/research/workpaper/index.cfm Best available copy ahead. =orecasts o f two national variables, the leading index (LEAD) and the coincidert index (COIN), were both 14 percent lower for one-steps-ahead, and 19 and 9 sercent lower, respectively, for 4-periods-ahead, in the trickle-down model. Also a handicap is the maintenance of the specification of the equations throughout the period. Although the other models were not revised with regard to regressors either, their a priori specifications preclude the use of new data to revise the specifications. (Of course, the stepwise model's ccefficient values were updated each quarter.) The out-of-sample forecasting performance of the stepwise model relative to the autoregressive model confirmed a strong dependency of four regional variables t o lagged national and.regiona1 information variables: STARTS, PAYROLL, and MFG. RETAIL. EMPL, Weaker confirmation was implied for LF, HOURS, and Fina-lly, the information gain vanished for NONMFG, INCOME, and PRICES. The stepwise model significantly outperformed the AR2 model a t the .05 level, according t o the test adapted from Ashley, Granger, and Schmalensee, in the following cases. For one-period-ahead forecasts, the improvement was significant for EMPL and LF; for four-period-ahead forecasts, the improvement was significant for PAYROLL, LF, and STARTS. The properties of the errors in the T D and .SW. models were often somewhat different with regard t o bias and variance around means. consider the four-step-ahead forecasts of PAYROLL. For example, The TD and SW models had similar RMSEs, of .0358 and .0360, respectively, representing improvements of. 6.8 and 6.3 ~ e r c e n trelative t o the RMSE o f .0384 in the AR2 model. Yet the source o f erTor differed somewhat among the models, with mean errors of -.0291 in the T D mcdel and -.0205 in t h e S W model. The SW model forecasts benefited f r o m lower !.:.osolute> bias, but suffered from a larger variation in accuracy f r o m o n e quarter t o the next. A forecast that combines the forecasts of the http://clevelandfed.org/research/workpaper/index.cfm Best available copy two models is particularly promising in such a case. Giving weight to the model might seem unpromising because of its higher RMSE. SW Yet giving the SW model weight in a combined forecast will definitely reduce the magnitude of bias. This benefit must be balanced against the cost, in terms of RMSE, that results from higher variance. But unless the errors of the two forecasting models are perfectly correlated, the variance of combined forecasts will be less than the sum of the variances of the components. As it turns out, the combined, unweighted forecast (UC2 for "unweighted combination of two" forecasts) has an RMSE of -0348, lower than the TD or SW models. The contrast with the AR2 model's performance is summarized in the last two columns of table 5 , for one- and four-quarter forecasts. The UC2 forecasts do generally as well as the TD model, and better than the SW model for one-step-ahead forecasts. -,They generally do as well or better than the TD model for four-step-ahead forecasts, and better than the SW model at that forecast horizon for 8 of the 10 Ohio variables. According to the test adapted from Ashley, Granger, and Schmalensee, the improvement of the UC2 relative to the AR2 is significant at the .05 level for PAYROLL, MFG, LF, and RETAIL for forecasts one quarter ahead, but significant only for LF for the four-quarter forecasts. The improvements of the UC2 relative to the TD model do not appear substantial and are unlikely to be significant, according to casual inspection. Only small gains appear available from combining the models, as compared with giving the TD model all the weight. In the terminology of Granger and Newbold (1977, p. 2831, the TD model is conditionally efficient with respect to the alternatives considered. The importance of updating coefficients during the out-of-sample period was relatively easy to determine. Forecast performance for the TD model without updating was generally inferior to performance of the model with http://clevelandfed.org/research/workpaper/index.cfm Best available copy updating. Only f o r forecasts of HOURS, short- horizon forecasts f o r INCOME, and long- range f o r e c a s t s o f RETAIL were RMSEs lower w i t h o u t updating; i n a l l o t h e r cases u p d a t i n g was h e l p f u l . Mean e r r o r s were always lower i n absolute magnitude; u p d a t i n g had the e f f e c t of reducing p r o j e c t e d growth d u r i n g the weak c o n d i t i o n s of t h e out-of-sample p e r i o d . Generally, t h i s r e d u c t i o n accounted f o r a l l of the improvement--indeed, the means.of absolute e r r o r (MAEs) o f t e n r e f l e c t e d l e s s improvement than MEs. For example, i n one- period- ahead f o r e c a s t s of PAYROLL, updating changed the ME f r o m -.0051 t o -.0040. t o .0071. But the MAE was o n l y reduced from .0061 t o .0056; the RMSE from . k 7 8 On average, updating reduced t h e RMSEs by 4.1 percent, 5.6 p e r c e n t , and 4.4 percent, f o r one- quarter, four- quarter- , and 10- quarter- ahead f o r e c a s t s , r e s p e c t i v e l y , f o r t h e 10 r e g i o n a l v a r i a b l e s . I n t h e stepwise model, updating brought s i m i l a r b u t l e s s c o n s i s t e n t gains; the r e d u c t i o n i n b i a s was l e s s c o n s i s t e n t , b u t g e n e r a l l y s m a l l e r . f o r e c a s t s were q u i t e adversely a f f e c t e d . PRICE A more i m p o r t a n t , y e t unanswered, q u e s t i o n i s what l o s s of f o r e c a s t i n g accuracy r e s u l t e d from n o t r e s p e c i f y i n g the stepwise model each q u a r t e r i n l i g h t of new data. Some p a r t i a l evidence on t h i s q u e s t i o n c o u l d be provided by r e s p e c i f y i n g t h e equations a f t e r t h e end o f t h e out- of- sample p e r i o d . For t h e PAYROLL equation, such r e s p e c i f i c a t i o n r e s u l t e d o n l y i n t h e e x c l u s i o n o f t h e second l a g on U.S. housing s t a r t s . m i g h t be regarded as n e a r l y t h e s l i g h t e s t p o s s i b l e change. This However, we have n o t undertaken a systematic and f u l l y s a t i s f a c t o r y a n a l y s i s o f t h e b e n e f i t s o f period- by- period r e - s p e c i f i c a t i o n . Such b e n e f i t s could c o n c e i v a b l y a l t e r comparisons between t h e TD and SW models. However, we do n o t p l a c e much emphasis on such a comparison; such a comparison i s d i f f i c u l t t o i n t e r p r e t i n any case. http://clevelandfed.org/research/workpaper/index.cfm Best available copy The use of the ordinary least squares estimation procedure can be t o some degree inefficient in cases in which errors of equations estimated are correlated. Correlations in the errors o f both multivariate models were . . frequently larger than 2n-"' . Again, we have not undertaken a full and systematic study o f this issue, but have examined its implications for the PAYROLL equation. In the TD model, PAYROLL, LEAD, and COIN f o r m a system o f three variables in the equations that determine forecasts o f PAYROLL: the other regional variables' forecasts fol low recursively. Applying general ized least squares (seemingly unrelated regression) to allow for a non-diagonal disturbance variance-covariance matrix offered a potential improvement, suggested by the high correlations between residuals of ordinary least squares equations for PAYROLL and COIN (0.69) and LEAD and COIN (0.61). When compared with the 0rd.inary 1 east squares estimates, the general ized least squares method reduced the magnitude of all the coefficients of the PAYROLL equation except the one o n COIN,-,. The effects o f the equations for national variables were rather small. Forecasts of PAYROLL with the generalized least squares estimates o f the T D model were somewhat worse than for the ordinary least squares version, where the comparison is of models whose coefficients were not re-estimated each quarter. The RMSEs of the generalized least . . squares version (of the ordinary least squares version) were .0085 (.0078), .0421(.0389), and .1103(.1083), respectively. for one, four, and 10 steps ahead, This comparison may have been affected by the special characteristics of the 1979-83 period, particularly since trickle-down effects o f the national economic weakness were given less range by the generalized least squares coefficients' smaller value^.^ http://clevelandfed.org/research/workpaper/index.cfm Best available copy IX. Conclusion The location of information about each o f 10 Ohio variables representing aggregate economic activity has been identified, measured, and subjected to confirming tests. Generally, the results verify two prior beliefs: (1) univariate forecasting models can be outperformed by simple multivariate models, although not consistently by a large margin, and ( 2 ) most (lagged) information other than from a variable's own past comes from national variables, and may be summarized reasonably well by the coincident and leading indices. Ohio housing starts, however, seems to contain independent leading information for other regional series such as employment. Our study is also of interest as a practical application o f statistical principles and forecasting methods in a context in which a number o f sources of information are likely t o be valuable. Conclusions in this regard may be quite sensitive t o the particular data samples employed. The two models specified priori, the univariate autoregressions and the trickle-down model, provided gains relative t o their appropriate benchmark models that were, overall, approximately equal in the 1965-78 sample and the 1979-85 out-of-sample period. In the case o f the trickle-down model, the relation between within-sample gain and out-of-sample gains in o n e period-ahead forecasts was remarkably close: the gain delivered out-of-sample approximated that o f within the sample, o n a variable-by-variable basis. The stepwise model, as might have been expected in light o f the "overfitting" problem, could not deliver out-of-sample results t o match those within the sample, nor was there much relation between them on a variable-by-variable basis. However, the stepwi se model operated under several handicaps. Its specification was not revised, as would be done by a real-time forecaster using the stepwise procedure of model construction. Second, http://clevelandfed.org/research/workpaper/index.cfm Best available copy f o r e c a s t s o f more than one-step- ahead p r o b a b l y were handicapped by t h e use o f u n i v a r i a t e a u t o r e g r e s s i o n s t o p r o v i d e paths f o r t h e exogenous n a t i o n a l variables. The p r e s e n t study adds t o t h e growing knowledge of how t o deal p r a c t i c a l l y w i t h t h e t r a d e o f f between t h e c o s t s o f i g n o r i n g i n f o r m a t i o n and t h e problems o f "overfitting." Gains o v e r u n i v a r i a t e equations have been a c h i e v e d i n t h e p r e s e n t study o f Ohio, as had been achieved f o r Texas (Hoehn, 1 9 8 4 ) . However, t h e p a r t i c u l a r methods employed a r e u n l i k e l y t o be " o p t i m a l " i n any sense." The r e s t r i c t i o n t o l i n e a r and nonseasonal models (of s e a s o n a l l y a d j u s t e d d a t a ) and t h e equal weights i n combined f o r e c a s t i n g schemes a r e a source o f inefficiency. ~ e v e r ~ h e l e s swe , contend t h a t t h e u n d e r s t a n d i n g and f o r e c a s t i rrg of an economic system, whose t r u e s t r u c t u r e i s unknown. can be enhanced by t h e s i m p l e and tr:ansparent t i m e - s e r i e s methods employed. S t r u c t u r a l models i n such a c o n t e x t m i g h t b e s t be c o n s t r u c t e d a f t e r t h e s t y l i z e d f a c t s o f t h e t i m e s e r i e s a r e uncovered. http://clevelandfed.org/research/workpaper/index.cfm Best available copy Footnotes 1 . Strictly speaking, a structural model cannot forecast in the same fashion as a time series model. The former is always "incomplete" in the sense that it forecasts the endogenous variables conditional on arbitrarily specified 'values of the exogenous variables, which the forecaster~must provide. While thi s condi tional nature of structural forecasting al'lows for interesting simulations of "what if" questions, it complicates forecast construction and performance evaluation in the more realistic case for which no future values o f any variables are known when forecasts are made. This difference between structural and time series models makes this relative forecast performance difficult to assess. (See, for example, Granger and Newbold (1977, pp. 289-302). 2. These issues are more fully addressed in Hoehn, Gruben, and Fomby (1984a>, pp. 34-49. 3 . Ohio series displayed more frequently significant dependence on lagged national series than did Texas series, in conformance with prior beliefs. Also, this study of Ohio examined 19 national variables whereas only 14 wereexamined in the Texas study. In the latter, only 21 out of 92, or 24 percent, o f the possible trickle-down causal relations were significant at the .05 level (see pp. 26-27, Hoehn, Gruben, and Fomby, 1984b). The proportion for this Ohio study was 47 percent. However, the comparison may be affected b y . the longer sample for Ohio. 4. The RMSE is an exact criterion for comparison of alternative forecast generating mechanisms if the loss function is proportional to the square of forecasting errors (see Granger and Newbold, 1977, pp. 279-280). 5. However, other studies have also suggested that the gains from accounting for contemporaneous correlations in errors in the estimation of linear forecasting models may be slight. Unpublished results by Hoehn for "VARsU of the Texas economy showed generally inferior forecasts for seven regional series, with updating. 6. Granger and Newbold (1977, pp. 268-9) offer . some reasons why optimal . forecasts are practically never available. . . http://clevelandfed.org/research/workpaper/index.cfm Best available copy References Amirizadeh, Hossain, and Richard M. Todd. "More Growth Ahead f o r Ninth D i s t r i c t S t a t e s , " Q u a r t e r l y Review, Federal Reserve Bank o f Minneapolis, F a l l 1984, pp. 8-17. . . Anderson, Paul A . "Help f o r the Regional Forecaster: Vector Autoregression," Q u a r t e r l y Review, Federal Reserve Bank o f Minneapolis. v o l . 3 , no. 2 (Summer 1979). pp. 2-7. Ashley, Richard A . , C.W.J. Granger, and R. Schmalensee. " A d v e r t i s i n g and Aggregate Consumption: An Analysis of C a u s a l i t y , " ~conometiica, v o l . 48, no. 5, ( J u l y 1980), pp. 1149-67. Box, G.E.P., and G.M. Jenkins. Time Series Analysis: C o n t r o l . San Francisco: Holden-Day, 1976. Granger, C.W.J., and Paul Newbold. York: Academic Press, 1977 Forecastinq and Forecastinq Economic Time Series. New Hoehn, James G., and '~ames J . Balazsy, J r . "The Ohio Economy: A Time Series Analysis," Economic Review,. Quarter 111, 1985, Federal Reserve Bank o f Cleveland. Hoehn, James'.G. " A Regional Forecasting Procedure A p p l i e d t o Texas," Working Paper No. 8402, Federal Reserve Bank o f Cleveland, September 1984. Hoehn, James G., Wi-lliam C. Gruben, and Thomas B. Fomby. "Time Series Models o f the Texas Economy: A Comparison," Economic Review, Federal Reserve Bank o f D a l l a s (May 1984a1, pp,. 11-23. . "Some Time Series Methods o f Forecasting the Texas Economy,"~orkinq Paper No. 8402, Federal Reserve Bank o f D a l l a s , A p r i l 1984b. Kuprianov, A n a t o l i , and W i l l i a m L u p o l e t t i . "The Economic Outlook f o r F i f t h D i s t r i c t States i n 1984: Forecasts from Vector Autoregression Models," Economic Review, Federal Reserve Bank o f Richmond, v o l . 7011 (JanuaryIFebruary 19841, pp. 12-23. L i 1ien, David M. " Sectoral S h i f t s and Cycl i c a l Unemployment, " Journal o f P o l i t i c a l Economy, v o l . 90, no. 4 (August 19821, pp. 777-93. Litterman, Robert 8. "Techniques o f Forecasting Using Vector Autoregressions," Working Paper No. 115, Federal Reserve Bank o f Minneapolis, 1979. Nelson. Charles R. " A Benchmark f o r t h e Accuracv o f Econometric - f o i e c a s t s o f GNP," Business Economics, v o l . i 9 , no. 3 ( A p r i l 19841, pp. 52-58. , and Charles I. Plosser. "Trends and Random Walks i n Macroeconomic Time Series: Some Evidence and I m p l i c a t i o n s , " Journal o f Monetary Economics, v o l . 10, no. 2 (September 19821, pp. 139-62. - 35 Appendix A: http://clevelandfed.org/research/workpaper/index.cfm Best available copy Glossary of Variables Regional variables* Mnemon i c Description EMPL - Total civilian employment (household survey), in thousands, Bureau o f Labor Statistics (BLS). Seasonally adjusted by Chase Econometrics (Chase). Transformed from monthly values t o quarterly averages by Hoehn and Balazsy (HB). HOURS Average weekly hours in manufacturing. BLS. Seasonally adjusted by Chase. Transformed from monthly t o quarterly by HB. I NCOME Personal income at annual rates, in billions o f current dollars. Commerce Department. From Data Resources, Inc. (DRI). LF - Labor force, in thousands. BLS. Seasonally adjusted by Chase. Transformed from monthly t o quarterly by HB. MFG - Employment in manufacturing, in thousands. BLS. Seasonally adjusted by Chase. Transformed from monthly t o quarterly by HB. PAY ROLL Total nonagricultural employment: total private and. government, in thousands. Seasonally adjusted by Chase. Transformed from monthly t o quarterly by HB. PRICES Constructed average for consumer prices for Ohio. Constructed f r o m bi-monthly series for Cleveland CPI and Cincinnati CPI, BLS. S e e special description of construction method, below. RETAIL Total retail sales, in millions of current dollars. Bureau o f Census. Seasonal ly adjusted by Chase. Transformed from monthly t o quarterly by HB. STARTS Total private housing starts, in thousands o f units, at annual rates, Bureau of Census. Special note o f PRICES The consumer price index for O h i o (PRICES) was constructed in the following manner. First, t h e seasonal adjustment factors f o r each month for the U.S. CPI was determined by dividing the U.S. CPI, not seasonally adjusted, by the U.S. CPI, seasonally adjusted. This factor was used t o seasonally adjust values for the (bimonthly) Cleveland and Cincinnati CPIs. From these seasonally adjusted bimonthly figures, quarterly averages were constructed for each city. The average used the available months within each quarter (one o r two) rather than interpolated values. Then the quarterly city values were averaged. http://clevelandfed.org/research/workpaper/index.cfm Best available copy National Variables* Mnemon ic Description COIN Coincidental Indicators Composite Index with Trend Adjustment. CP I Consumer Price Index (Revised) - A11 Items. DEFLATOR Gross National Product Implicit Price Deflator. FUNDS Effective Rate on Federal Funds. GNP Gross National Product LEAD Leading Indicators Composite Index with Trend Adjustment. MOODY Yield on Moody's Industrial Corporate Bonds. PPI ~ r o d i c e rPrice Index PRODUCT Total Industri a1 Production Index. REALGNP Gross National Product in 1972 Dollars. REALYP Personal Income in 1972 Dollars. USHOUSEHOLD Nonagricultural Employment (Household Survey). USLF Civilian Labor Force. USMFG Manufacturing Employment. USPAYROLL Nonagricultural Establishments Employment. USREALSALE Total Retail Sales in 1972 Dollars. USSTARTS Total Private Housing Starts Including Farm. USYP Personal Income. - Finished Goods. EHHEA *The source of a1 1 the national variables is Data Resources Inc. All variables, except the two interest rates, are seasonally adjusted. http://clevelandfed.org/research/workpaper/index.cfm Best available copy