COMPARATIVE STUDY OF THE X—11 AND BAYSEA PROCEDURES OF SEASONAL ADJUSTMENT Hirotugu Akaike and Makio Ishiguro Institute of Statistical Mathematics, Tokyo, Japan Abstract ‘A Bayesian seasonal adjustment procedure BAYSEA leads toa regression-type procedure which allows effective use of prior information ofeach particular time series. The decision on the seles- ion of the basic model is realized by minimizing an objectively defined ertrion, Numerical ests show the eelative advantage ofthe BAYSEA procedure over the X=] proce KEY WORDS Seasonal adjustment X11 BAYSEA Bayesian model Trend ‘Seasonal 1. INTRODUCTION ‘The Census method of seasonal adjustment, particularly the X11 variant of the method IL, made a significant contribu- tion tothe practice of economic data analysis. The method is 1 typical example of how an appropriate accumulation of hhuman experiences can lead to the development of an ‘extremely useful data-analytic procedure. The construction of, the provedure has often been questioned because of its lack of ‘the basic statistical model. Nevertheless, the widely spread use of the X=11 constitutes a proof that itis a good practical ‘procedure for the seasonal adjustment of economic data. Since the social impact of the seasonal adjustment of an ‘economic time series is often quite significant, it is natural ‘that the improvement of the technique has been continuously ‘contemplated. A useful reference on this subject isthe paper by Kallek (1978) who discusses the general objectives and necessary improvements of the technique of seasonal adjust- ‘ment. Also useful is the paper by Shiskin and Plewes (1978) ‘who discuss the seasonal adjustment of the U.S. unemploy- ment rate, These two papers are very informative as they explain the problems of seasonal adjustment based on the experiences of the authors, Of particular interest isthe candid a era ‘The components used to synthesize the series are described in table 1.2. The numerical values of T, Sand S, are given in table 1.3. The descriptions in terms of components of the syn- thesized series, along with the values of the accuracy meas- lures associated with the default-option X--I] and BAYSEA seasonal adjustments, are presented in table 1.41.1 X=I1 (solid) AND ARIMA MODEL-BASED (dashed) SEASONALS FOR $08 WL. (from Hillmer and Bell 1980) gee = SOS BERBRESBSESERBEEE SRRSRRESETESSSNSSRESAESSLBLES38 Figure 1.3 TREND T FROM T1506 $0." ee ‘Table 1.1 IDENTIFICATION OF SERIES USED ‘T1506 Wholesale inventries—eletrical SOSWL Wholesale sales—lumber and other building materials SISWL Wholesale sales beer, wine, and distilled scloboic beverages ‘Table 1.2 IDENTIFICATION OF COMPONENTS ‘USED TO SYNTHESIZE SERIES T ‘ARIMA model-based tend from TIS06 Ss X11 seasonal fom S03WL 3 XI seasonal from 518WL i Tug: log lig ~ided NO.0,2) b Tay log Ta, = ~ 668 log Le ‘Table 1.3 TREND AND SEASONALS: 7, 51, Sz SeeeS2S25 SGESEUYUEIEE SECTION 1 ‘Table 1.4 RELATIVE ERROR MEASURES FOR X~11 and BAYSEA ___RRMSQD (RMAD) — X-u1 BAYSEA T's, £09 (.007) 1906 008 rs! ‘09 coo?) ‘a (009 Tesh 084 (069) 51 (082) resin 1050 (039) ‘028 (023) TS 081 068) 052 (044) Tosh 028 (024) 021023) Table 1.5 AVERAGE, ABSOLUTE PERCENT ‘CHANGES FOR T $2 Ty Lag Original X-11 BAYSEA 1 11.43 42 3.64 2 v2 731 oa 3 19.00 ois 92 4 1990 1126 wR 3 won 1239 1227 6 20.0 1286 1276 7 21.27 1293, 1293 8 2197 327 B29 9 21.68 B38 Bes 0 2161 1425 14.19 u 20 1503 1490 2 16.11 1583 ‘The effect of the two different kinds of iregulars, edited- lognormal white noise and an edited-lognormal sixth-order moving average process, ar illustrated in figures 1.4 and 1.5. ‘The editing was done to remove values larger than 1.5 stand- ard deviations in. magnitude from the associated pseudo- Gaussian series. Without editing, imegulas so obtsined iad implausible looking “outliers.” (This suggests the possibility that the distributions of many economic time series have “thinner als than the lognormal.) Graphs comparing the estimated and correct seasonally adjusted values for two of the series are given in figures 1.6-1.9. Since the statisties assessing accuracy give similar results for X11 and BAYSEA with the series 7 = Sa In the AAPC values for both adjustments of this series are also presented, i table 1.5, ‘Table 1.4 shows that even when BAYSEA’s choice of models has been restricted to three clases, it sable to beter remove standard-option X11 seasonals from these series than is standard-option X—11. These seasonals are rather stable, however, and an experienced user of X—11 who krew this would probably use the nonstandard 3 %9 filter option, which does give superior performance with these series. We hhave not yet compared finely tuned adjustments by X~11 and BAYSEA, since we lack the required experience with BAYSEA.FINDLEY. Figure 1.4 7% 5 (lid) COMPARED WITH Tx $2 X I; (dashed) Py Figure 1.5 Tx S2 (solid) COMPARED WITH T X'S X Ip (dashed) To tt so te7,000 6,500 6,000 5,500 5,000 4,500 4,000 3,500 3,000 2,500 1,800 1,000 ‘SECTION 1 Figure 1.6 TX 1; (solid) COMPARED WITH THE X—11 ADJUSTMENT OF T x $2 X I; (dashed) ign th 1.7 Tx 1 (solid) COMPARED WITH THE BAYSEA ADJUSTMENT OF TX Sz X I, (dashed) Be SeFINDLEY a Figure 18 Tx 1; (solid) COMPARED WITH THE X~11 ADJUSTMENT OF T x 52 X 1; (dashed) 7,000 Ce 6,300 6,000 5,500 | Figure 1.9 T x [3 (slid) COMPARED WITH THE BAYSEA ADJUSTMENT OF 7 x S3 X I; (dashed) 7,000 ——E 6,500 6,000 s,s00 000 4,500 4,000 3,500 3,000 2,500 2,000 1,500 100 :2 2, ADJUSTMENT OF SHORT SERIES Akaike’s and Ishiguro's successful use of overlapping 4 year spans of data to accomplish adjustments by BAYSEA. Suggests that BAYSEA might be able wo adequately adjust short time series. (When enough data are available, BAYSEA uses 7 years of data to backeast starting values forthe trend and seasonal components and thereafter uses 4-year spans.) Five confidential defense unfilled orders series collected by the Census Bureau were strongly affected by the change in the beginning and ending of dates of the Federal Government's fiscal year, which occurred in 1976. Data for 51 months were available for these series, starting with October of 1976 when the first of the new fiscal years began. BAYSEA adjustments ‘were made for all five of these series. None of the series was strongly seasonal. Three were thought to have negligible seasonality. For these three, BAYSEA produced seasonal fac- tors very close 10 1.0, and the average absolute percentage changes at most lags were higher forthe adjusted series than for the unadjusted series. For the other two, some smoothing ‘occurred and it was felt by subject-area experts thatthe adjust- ‘ment was effective. Table 2.1 gives the AAPC values for the ‘most seasonal of these two series, for its X11 adjustment, and for its adjustment by BAYSEA. (The Census Bureau does ‘not recommend the use of X11 with such short series.) ‘Table 2.1 AVERAGE ABSOLUTE PERCENT CHANGES FOR A S1-MONTH, MODERATELY SEASONAL SERIES Lag: Original X-ll BAYSEA 1 331 287 2.38 2 S18 422 395 3 678 344 sul 4 8.08 673 a2 3 837 796 162 6 9.64 9.5, 65 7 10.19 939 9.69 5 1a 10:96 10.76 9 248 1198 1176 10 BB 1289 26 un B93 ba Bel BR 14.66 6 1456 3. TRANSFORMATIONS OF THE DATA BAYSEA was applied to two Census Bureau series for Which some evidence exists that a multiplicative (or log- additive) seasonal model is inappropriate. For the construction statistics series CON-BPNEI, which describes the number of building permits issued per month in the Northeast for single- family houses (figure 3.1) the value of AVABIC for an addi- tive model was 1895 as contrasted with 1909 for the los- additive model. Thus, BAYSEA prefers the additive model for tis series over the multiplicative one, which is a decision supported by visual and by some subject-matter considera- tions. The AAPC values are given for the original data of CON- SECTION 1 BPNEL, its additively adjusted series and its log-adkltive adjusted series (by BAYSEA in both cases) in table 3.1 telow. ‘Table 3.1 AVERAGE ABSOLUTE PERCENT CHANGES FOR CON-BPNEL: 1/1969-3/1980 Log: Lag Oniginat Additive active 1 22.76 $20 59 2 40.82 739 73 3 50.78 373 309 4 5534 1038 1120 5 56.84 1135 237 6 36.28 1138 1420 7 55.02 1398 1530 8 52.46 1536 1698 9 4740 1644 1787 10 39.93 17.68 wor n 2392 18.61 1990 2 au 19.74 21.02 “The other series for which we considered various transfor- mations was retail hardware sales. Denoting this series by y,, ‘we considered additive decompositions of y,, °° and logy. BAYSEA (easily modified to consider such power transfor- ‘mations) preferred the additive decomposition of y,, as did an ARIMA model-based likelihood analysis performed by William Bell, which suggested the use of this transformation. "The AAPC smoothness criterion favored the additive decom- position adjustments obtained from y,% and log y, over that from y;, but did not indicate a clear distinction between the adjustments obtained via y;2* and log y, 4, CONNECTIONS WITH ARIMA MODEL-BASED ‘SEASONAL ADJUSTMENT ‘Suppose a trend-seasonal-iregular decomposition into uncorrelated components of the Seasonal time series yy is given by WET tS th 1) With B denoting the backshift operator, let us make the ‘additional assumption that T, and 5; satisfy the difference equations orBNT, =U, 42 BS, = Ve 43) where or@) =~ BP and $58) = (1 eB) B + +B")43 Figure 3.1 CON-BPNEL 18,000 -7 14,000, h awe! |] 8,000 2,000 +t. ao 61 e265 of 65 65 67 68 69 and where U,, Vz and f, are mutually uncorelated white noise series whose known variances are given by 2, 77 and 1, respectively Given intial values for T, and S, and the covariances of these: inital values, the Kalman smoother can be used 10 obtain the least squares estimates of T,, 5, for 1 = ¢ = Ny given y1, .... yw (See Brotherton and Gersch 1981). ‘When c= 1 and the initial covariances are set equal to 2270, the values of S, and T, s0 obtained are those produced by ‘Aake's procedure for given values of d? and RIGID, with 2 ‘equal to 270, ic, with the constraint expressions G+ Sates. + SF dropped from his model (and with ORDER=2_ and SORDER=1). The same calculation with a equal to 2210 yields the values obtained by dropping instad the constraint ‘expressions (S; ~ $)-13°. (Obvious modifications of (4.2) and (4.3) in this procedure yield Akaike’s values for different choices of ORDER and SORDER.) ‘With both seasonal constraint terms present, itis more dif ficult to exactly relate Akaike's procedure to ARIMA models ‘George Tiao has obtained a connection with the models defined by (4.2) and (4.3), which we wish to discuss briefly. For known values of. 2 and o, the estimates of Tx, and Sys, 0f 7, and S,, obtained by fortally aplying the Feast mn mw MS 7 7 78 79 BO Al ‘squares signal extraction procedures from the theory of sta- tionary time series (Koopmans 1974, p. 148) are solutions of RB), = bsBsB)oiy, a4 and RBS, = br BYbo(B 02, as respectively, where the “shifted polynomial” RB) is given by RB) = afb BWbrB) + oFbsBWs(B) + br BNbrB YsBhs(B) Ina stuly note attached as an appendix to these comments, George Tiao shows that the estimates 7, and S,, of, and Sy obizined trom Akaike’s procedure with ORDER=2 and SORDER=1 must also satisfy (4.4) and (5), with oe = {@2 +2) -(G? + 27 ~ 4)!?} and a2 =s do} = ad? for values of rin the range deg dr ds <1 = N ~ deg orbs 48)“ where N’ denotes the length of the span ofthe observed series 1 being adjusted and deg, » bs denotes the degree ofthe product polynomial ér(B }s(B). It follows that for these ‘values, the differences AT, = Tas — Tas must satisfy R@) AT, =0 4.7) and similarly for AS, = Sas ~ Sey ‘These conditions are not terily restrictive: The solutions ‘of (4.7) need not converge to zero as ¢ becomes large and, when they do converge to 2er, the geometric rate of conver- gence can be slow. (See the comments regarding the 00ts of | R(B) given below.) Further, when N= 48, as Akaike recom- ‘mends, and when deg 4, ~ 5 = 14, as its inthis example, the 28th-order difference equation (4.7) is required to hold for only twenty ¢ values, 14 < 1 < 34 ‘Along with these qualifying remarks, it should also be men- tioned, however, that the equations (4.2) and (4.3) do correctly suggest two important features of the trends and sea- sonals produced by Akaike’s method, namely, the sensitivity to stating values mentioned by Akaike (1980) and the kind of forecasts of T, and S, obtained when, as Akaike mentions as possibilty, che y, corresponding to the ¢ values for which Fore- casts are desired are declared to be missing values. In several cxamples we looked at the T; forecasts looked like straight- line extrapolations from the lst few 7, in the range of obser- vation, and the 5, forecasts were very close to the periodic extension of the lst twelve, values associated with the observed It is easy to generalize Tiao's argument to cover other values of ORDER and SORDER. When ORDER=1, one obtains that br(B) = 1=B. When SORDER=2, our attempts to construet b5(B) (which has degree 24) by finding the roots of ds(B s(B) met with numerical difficulties. The results obtained suggest that de(B) has double roots very close to exp(+ikw/6), k=2,3.45, along with single roots close to — 1 and exp(t 1/6), and no other roots very close tothe unit cle. In all of the cases we considered, for SORDER=1,2, all but afew ofthe roots of R (B) were close to (but not on) the unit circle, the closest roots being. near to exp(-ik'6), 40,1,2,3.4,5.6. For example, with ORDER=2, SORDER=1, and d?=19, the roots of R (B) are the seven numbers given below, along with their reciprocals, complex conjugates, and conjugate reciprocals. ‘Table 4.1 BASIC ROOTS re'® of R@) ee Tie isis Sep os ee oe os ae O) 17a eiae 2a eA ae La SECTION 1 5. CONCLUDING REMARKS ‘The number of examples we have considered is too small to support firm conclusions. Coupled with the results of Akaike (0980) and the present paper by Aksike and Ishiguro, how- ever, our results would ccm o offer atonal evidence that BAYSEA may find a place among the handfl of seasonal adjustment methods which enjoy or can expect to enjoy wide usage. Given the ease with which calendareffet variables, intervention variables, and other special trend, or seasonal or imegular compensatory adjusmens can be incorporated ito BAYSEA, this would be welcome. T would ike to thank Ted Holden fr computer assistance in the preparation ofthese comment. REFERENCES AKAIKE, H. (1977), “On Entropy Maximization Principle” in Applications of Statistics, P. R. Krishnaiah, ed., Amster- dam: North Holland, 27-41 (1980), “Seasonal Adjustment by a Bayesian Model- ing," Journal of Time Series Analysis | (1980), 1-14. BROTHERTON, T., and GERSCH, W. (1981), “A Data ‘Analytic Approach to the Smoothing Problem and Some of Its Variations,”” Proceedings of the 20th IEEE Conference ‘on Decision and Control, 1061-1069. HILLMER, S., and BELL, W. (1980), “Final Report on Research Activities: 1979-80 ASA-Census Seasonal ‘Adjustment Project,”” Statistical Research Division, U.S. Bureau of the Census. KOOPMANS, L. (1974), The Spectral Analysis of Time Series, New York: Academic Press. APPENDIX STUDY NOTES BY G. C. TIAO ON AKAIKE’S ‘SEASONAL ADJUSTMENT PROCEDURE Here we shall use ¢ for i and the backshift operator B such that By, = y,-s. Thus, lety, = 7; +S; +1, where T,, 5, and J, ate unobservable components. We show the relationships ‘between Akaike’s (1980) Bayesian minimization procedure f= BO - 1 -5F + d4s4T, ~ 2) + Ta? +S ~ Sa? FMS, +22 + Su aD ‘and the model-based procedures advocated in papers by Box, ‘Cleveland, Hillmer, Piece and TiaoFINDLEY From (A.1), + Ps46T, A044 +7 +n +T2}=0 at ng, -7- yh a -1-5) + ARS, ~ S.-12 ~ Srv) +S, + G41 +5, +105 ,12 + 5,2) + + Gen +5,-10)} =0 “Thus, we have that on, Oss, %-S %-T ay where Q)=1 + a’sy0 - BYA~F Q2=1 + dl ~B%) + 2? UBF) 1+ PUB)UEIet + (1-8) - FY) = 1+ 7UB)UFMI ~ BX ~ aF) and where F = B-!,U(B)=14+B +... +B" az?= (1 = a)? and c? = a-"d, It follows that for large Nand ¢ not close tothe end points T, = 102 - 0": - DY (a3) $= (O02 ~ N'Q1- DY "Now, let us suppose that 7, and S, follow the ARMA model orB)T, = 6r(BU, 43(B)S, = OV, (aa) 45 where {U;}, {V,) and {f,} are uncorelted ite noise processes with variances ri, and 1, respectively. Then, fiom signal extraction theory, the best estimates of 7; and S T, = R“ds(B bs (F 8B )8r(F Jordy, S, = RMbrB yor Ws(B WF oz (AS) R = ofbs(B)Os(F WB Wr(F) + o}br(B Yor(F 05(B Ost ) + br(B Ybr(F Ys (B bs(F) By comparing (A.3) with (A.5), we can simply take ds(B) (> @B)UB), or(B), = (1 ~ BY, OB) = O58) = 1, of =e ando3 = s-7d~*, Thats, the models for T; and S, are, respectively, a -BYt, =U, oh = sta (any +B + 48's, Yop a6 ‘To partially check the appropriateness of (A.6) in practice, note that models in (A.6) imply thatthe overall model for ¥; (1 = aB\l — BI ~ BY, = 6B), (7) where 02 0B) @(F) = R. Thus, we would argue that the ‘minimization procedure in (A.1) would be consistent with information from the data if the overall model of Y, is of the form (A.7). Note from (A.2) that (I + @2)a= (2? + 2) s0 that from Akaike's paper, the value of a is fixed. If, on the ‘other hand, the model for ¥; is vastly diferent from (A.7), then the use of (A.1) would be throvm in doubt REFERENCE, AKAIKE, H. (1980), “Seasonal Adjustment by a Bayesian Modeling,” Journal of Time Series Analysis 1, 1-14.(COMMENTS ON “COMPARATIVE STUDY OF THE X—11 AND BAYSEA PROCEDURES OF SEASONAL ADJUSTMENT” BY HIROTUGU AKAIKE AND MAKIO ISHIGURO ‘Svend Hylleberg University of Aarhus, Denmark In this interesting paper, the Census Bureau X~11 ‘method is compared with the BAYSEA seasonal adjustment ‘method developed by Akaike. The BAYSEA. procedure assumes that the observed series ¥; (or the log of ¥,) can be decomposed into three additive unobserved components: The trend-cycle component TC,, the seasonal component $,, and the imegular component J,.. When estimating the three com- ponents the objective function applied is min(Y, — TC, ~ 5)? + ds{ARTCY + (ABS? +s, re where s, 2, 8, and \ are constants chosen by the user. Ay is the difference operator (I~ Ly, L being the lag operator, ‘while a grid search is used to’ estimate d. The rationale ‘behind the term (APTC,) is that the trend-cycle component fought to be smooth, while the rationale behind the term (AAS, is tha the seasonal component should be stable. The term Bs ;) is included in order to keep the 12-month sum forthe seasonal component close to ero Obviously, several reservations can be made agninst the anbitrary citria that the tend-eycle component should be smooth and the seasonal component stable. (See Hylleberg, 1981, ch 2.) But when applying the BAYSEA procedure, it ‘becomes important to realize that one of the main objectives ‘of deseasonalizing single economic time series i to promote the prediction of tuming points. As a consequence, values of 8 less than or equal o 2 should be avoided as they imply biasing the estimated tend-eycle component towards being constant (8 = 1) o¢ linear (6 = 2). ‘Nonetheless, in applying the BAYSEA procedure to actual economic time series, Akaike and Ishiguro search over values of 8 equal to 1 and 2 ony. ‘OF course, these biases will be small if sufficiently Iow values of s are applied, but it scems preferable to apply higher degree smoothness priors, i. higher values of 8 and correspondingly higher values of s. REFERENCES HYLLEBERG, S. (1981), “Seasonality in Regressions,”” Mimeographed, Institule of Economics, University of ‘Aarhus, Denmark.RESPONSE TO DISCUSSANTS. 1H. Akaike and M, Ishiguro Institute of Statistical Mathematics, Tokyo, Japan ‘We are grateful thatthe discussant took up several impor- tant points that we did not discuss in our paper. Before answering the points, we would like to emphasize the conven tional character ofthe BAYSEA procedure. We consider it to be a simple exile rule that will allow a reasonable measure- tent of the tend and seasonal components. Thus, we do not think it sufficiently useful for forecasting. For that purpose, rmore detailed analysis and modeling of the behavior of these components are necessary. The output of BAYSEA will pro- Vide a starting point for this typeof research ‘As wo the frst point of Dr. Dagum, the frequency response analysis of BAYSEA, we must emphasize the date-adaptive ature of BAYSEA. The frequency response analysis isa pro cedure developed mainly for the analysis ofa constant near system, Inthe BAYSEA procedure, not only the parameter d but also oder parameters, such as ORDER, SORDER and RIGID, are chosen adaptively. Even when these later param cters are fixed, the adaptive choice of d gives the procedure an estenially nonlinear characteristic, For the analysis of such a system, itis more informative to observe the responses of the sytem under typical operating conditions, as is done in the test of audio amplifiers. Figures and 2 show the responses of X-II and BAYSEA @ 1, 1.0), without extreme value corections, tothe square wave input. The generation af the spurious seasonal component by X~11 is mach mare Significant than that of BAYSEA. Figures 3 and 4 show simi- Jar results obtained by XL1 and BAYSEA (1, 1, 1.0) for a white noise input. The spurious movement ofthe end com ponent generated by X11 clearly demonstrates the undesi- ably high sensitivity ofthe tend-eyle filter to the imegular component. These simple examples amply explin the reason ‘wy we resricted our analysis tothe empirical comparison of the 14 sets of data provided by the Bureau of the Census ‘These data ar full of completes that cannot be easly simu- lated by scale artificial iapus. Is common knowledge that the frequency response characcristc is only supplementary information in testing an audio equipment of high quality. Her, judgements by expers are stil playing a dominant role, ile, we are all in the proces of searching for a decisive characteristic in choosing a system. Maybe thsi an unending ‘process. We believe thatthe situation is much the same with Seasonal adjustment and we pay very much attention to the ‘opinions of exper inthe area of practical application. ‘We are glad to sce that Dr. Dagum i in gzeement with ws and consider that the amount of revision itself is nota decisive characteristic for the comparison of seasonal adjustment pro cedures. We believe that our numerical results have shown that BAYSEA produced snaller revisions compared ith XU, when te Hal outputs were salar BAYSEA often progced mone resonable ted estima, a i shown by the Teuls of figures 6 and 7, witout undue inoeae of he mount of revision, At leat, we canst tat BAYSEA isa frocedure tat produces results esetaly diferent fom those by XI], with eapoct to some charset ee pethas of paca! importance Coming tothe ls pont, the operational sat of BAYSEA, we must mention thatthe propram made avalabe tthe di Gussants asa prototype designed forthe ease of understand ing and modification of the poser bythe use, However, even with his program, the CPU time to roe thereat of gure 1 by out computer was 18.66 Seconds and inal the computation of covariance seqenees and pets ofthe com: ponent sets, The XI] ARIMA took 476 seconds forthe fame das, withow covariance and spectrom compton We already have a faster version of BAYSEA.! By use of this Yeti, th CPU tine forthe adjustnent ofthe sae data Was 1157 seconds, wih sii rotted cate of eta, Th te can sce thatthe computational effency He not main problem. Further improvement of computational efficiency of BAYSEA isnt que improbable. Due tothe simplicity of the srctur, anyone who intrested inthe proce can cas develop his or er own version of BAYSEA, “The panto ofthe neces of tying diferent ORDER or SORDER shows thatthe vale ofthe paraetr ds iting the boundary ofits possibe values ands usually very weal in the search for an appropiate model. However, efi Sccision shold depend on ABIC's, deste noo oper. Findley supplemented our analysis checking the pe formance of BAYSEA with respect to the tusfrmaton of tein, ndjustnent of short tie, andthe amis of yt thesied seen. We ae gla ooe that his findings ae on Siseat wth ourexpeene. The comparison vith X11 based tn avenge abslitepecentge changes wl be of partar tnerest to those wh oe familar wth ibe se ofthis sats. Censinly, we are peat stat n Dr, Fey's expe tment the BAYSEA procedreprduced results eter than, or sles equivalent to those ty Xi). The alysis through time series modeling developed by Professor George Tao is very vale, As we alfead) men toned or procedure is ey at apie ad in hat =, reponse some, sutra change of the. generating Tigre, WC, “Compo Ei fon of = getty Samal Agjsnet Pesta” Reger Sco. 28, March Tiree Sa a, 448 ‘mechanism of the original series. As is noticed by Dr. Find- ley, this characteristic may not be adequately described by stationary time series modeling. Nevertheless, this type of fnalysis might be helpful when our procedure encounters some particular typeof difficulty. Dr. W. S, Cleveland’s suggestion of robustfication of SECTION 1 BAYSEA against outliers seems very natural and interesting, ‘The tentative procedure of outlier correction in the present version of BAYSEA is based on a Bayesian modeling. Although it works fairly well itis extremely time consuming, ‘We hope to investigate the possiblity of implementing acom- pputation by more efficient procedures of outlier correction.ona wrnoaea sess wu wNrorgD =n 88 ASALTAVM SAMVAOS FHL OL VASAVE JO ASNOdSA “7 Bnd ASAL AVA SAVAOS TAL OL 11-X JO ASNOdSA “1 ambitonset ox wernoaua -ww0se26COMMENTS ON “COMPARATIVE STUDY OF THE X-11 AND BAYSEA. PROCEDURES OF SEASONAL ADJUSTMENT” BY H. AKAIKE AND M. ISHIGURO ‘Adi Raveh Akaike and Ishiguro formulated the decomposition of a given time series into trend, seasonal, and irregular com- ponents as a minimization problem. They assume the additive structure ¥; = T; + S; + 1; where ¥; denotes the observation at time i and 7, 5), and f; denote the trend, seasonal and inregular component, respectively. ‘They expect the two systematic components, trend and sea- sonal, to have atleast locally smooth behavior in linear terms and a stable yearly pattem, respectively. This expectation is represented by requiring that (9°) — 2a + TP +S + 24S, +51 + iia) + 5.98 be small. The last term within the brackets is added to keep the 12 months sum of the seasonal variations close to ze, Here $ and Z are propesiy chosen constants. The authors also expect thatthe systematic par T; + S, will not deviate signift- cantly from the original observation ¥,. This suggests the minimization of m-1-SP wo Consideration of the above two quantities leads to the minimization of 0) ~ 1) — SP + 8% ~ 1 + Ta) +S ~ Sah + 2G +...+ SF] @ ‘where d is a properly chosen constant, Estimates of the two systematic components, 7; and S, can be obtained by minim- izing (2) using the standard procedure of least squares. Of ‘course, the three parameters (weights) d2, S® and Z must be specified before the procedure becomes operational. This problem is solved as a problem of statistical mode! selection using AIC. The very same procedure was suggested recently bby Schlicht (1981). His method minimizes the equation (2) for ‘monthly series. His a, f, andy play the same role as °S°, 4, and d°Z* in Akaike’s and Ishiguro paper. Schlicht has derived a unique solution to the minimization problem. He claims thatthe appropriate choice of parameter values will depend on the shape ofthe time series, but he did not show how to choose the desired values of the parameters ‘Thus, if the seasonal patter is nearly fixed over the whole period of observation, « very high B (not a. as is mistakenly claimed by Schlicht) would be appropriate, High values of are required to keep the seasonally adjusted data (SAD) inthe same scale a the original data. In other words, the sum of the ‘original series and the SAD for any 12 consecutive observa- tions will be about the same. Procedures to chock that the pat- tem is constant and to choose the parameter values are not given by Schlict. Both methods trade off among the four components ofthe overall loss function by means of the three weight parameters ‘Thus, for example, when B increases to infinity, the moving Seasonality becomes fixed. When B = 0, the moving seasonality behaves like an ifegular component with every consecutive 12 values adding up to 12 approximaely, depending on the value of y. Both procedures try to estimate as smooth a trend as possi- ble in terms of local linearity. The part of the loss function that relates to trend is based on squares of the second differ ‘ences an is provided by minimizing uve) = San ~ ana > 0 ° where AT) = 7, ~ 7)-..IN(T) = 0 if and only if the trend is perfectly linear. LIV() = 0 (relatively, close to ze), if the tend is locally inar, namely, if there ae very fw turn- jing poins and between turing point the trend is Hear. Linearity conditions forthe trend are: 7) ~ 7, = T)- ~ T.-for all | = 3, ..., N,or AT, = AT, 0¢ M7, = 0. ‘Another point of view which i slighty diffrent is that the tend will be as smoath as possible in tems of local monotn- ‘ct. The conitions for weak monotonicity are i, -TWa Tj) = Ofor all i= 3, ..., AT/AT;-. = 0 The author is on leave from Hebrew University Jerusalem, to Stanford University, Calif and currently is an ASA Junior Research Fellow at the Bureau of the Census st2 JATAT,. | for al 1 ATIAT;-1 ‘Thus, smooth trend can be measured by a coefficient of Tocal monotonicity given in (4) below. San an. MON@) = Wy = 5 ———_ S lan - at] ug = 1 if and only ifthe trend is perfectly weak monotone with either positive or negative slope. If very few turning points exist (relative to the length of the series) and between them the tend is monotone then y2a = 1, and we call it local monotonicity. The least monotone trend would be obtained for series that their slope change rapidly. Thus for the follow- ing series: a,b, ..., bwhere a 6, the coefficient a =I In tems of monotonicity, a smooth trend is achieved as wy is increased and thus max jis desired. To combine this oss function with the overall loss function, let us consider the ‘minimization ofthe quantity 0% - 7, ~ SP +4 [tar ar,-, ~ aTar al? + BS, ~SiP + CG +... +S uP 5) where A, B,C, P are properly chosen constants (usually P = Tor 2 can be used). The solution can be obtained by mumeri- cal algorithms for minimization such as in Zangwill (1967). 1. is obvious that LIN (T) = 0 yields MON (T) = 1, but not vice versa, In other words, a trend can be very smooth in mono- ‘one terms but a the same time may be far from being linear SECTION 1 ‘one, as in figure 1, below. By adopting this minimization for- ‘mula, a smoother trend should be obtained in terms of local monotonicity. By minimizing (2) for a discontinuous and polytone (namely, local monotone) trend, one could obtain a Jocal linear trend as an estimate although such does not ex for the data. In figure 2, original series and trend estimation are exhibited for wholesale inventories of grocery stores inthe United States for January 1967 to July 1980. In 2a and 2b, estimates of global monotone (nearly linear) trend and local ‘monotone (polytone of order 3) wend are obtained, respec- tively. It seems that the estimation in 2b represents the inrin- sic (hidden) local trend more accurately, especially around the ‘uring points. REFERENCES SCHLICHT, E, (1981), ““A Seasonal Adjustment Principle ‘and a Seasonal Adjustment Method Derived From This Principle,"” Journal of the American Statistical Associa- tion, vol. 76, No. 374, 374-378. ZANGWILL, W. (1967), ‘*Minimizing a Function Without Calculating Derivatives,” Computer Journal 10, 293-296. Figure 1. A LOCAL POLYTONICITY SERIES WHICH IS NOT SMOOTH IN THE LOCAL LINEAR- ITY SENSE Figure 2. WHOLESALE INVENTORIES OF GROCERY STORES: JANUARY 1967 TO JULY 1980 2a, Global monotone (nearly linear) trend 2b, A local monotone trend (a polytone curve of order 3) Mitions Mitions 7,500, 87,500, s7,000}- $7,000 86,500}. ‘$6,500 eee em Trend estimation 5,500}- $5,500 A se 0007- Original series $5,000 Oriana seri 4,500}- $4,500 4 4,000} A ees ‘Trend estimation ey 83,500}- $3,500 $3,000 ‘$3,000 4 $2,500 $2,500 4 eae Source: Bureau ofthe Cansut $2,000 Source: Buren ofthe Canaus”| S505 95 G7 ao 61 73 6 OT ToOTAT ISB aBTa7§— $™°°5-95Sy ao Gi 7 OH 7 100 121100145 157 Months MonthsREPLY TO DR. RAVEH’S COMMENT HL Akaike and M. Ishiguro We are grateful to Dr. Raveh for pointing out the distine- tion between the Schlicht’s constrained least squares and our ‘Bayesian approach. As was mentioned in Dr. Dagum’s diseas- sion, the constrained least squares approach his a long his- tory. Is our use of ABIC, or the likelihood of the Bayesian ‘model, that mace BAYSEA a viable altemative to the X11 procedure. ‘The concept of local monotonicity is interesting. However, the minimization of (5) will allow irregular variations ofthe trend in monotone pases and only curb the behavior at turn- ing points, This seems to be somewhat in contraction to Dr. [Raveh’s final statement. It may also be mentioned that the application of the Bayesian modeling to (5) may not be quite feasible a a practical procedure