Seasonality Tests Fabio BuseTT! Research Department, Bank of Italy, 00184 Rome, Italy (busett.fabio @insedia.interbusiness.it) Andrew HARVEY Faculty of Economics and Politics, University of Cambridge, Cambridge CB3 90D, U.K. (ach34 @econ.cam.ac.uk) This aticle modifies and extends the test sigainst nonstationary stochastic seasonality proposed by Canova and Hansen, A simplified form ofthe test statistic in which the nonparametric correction for serial core lation is hased on estimates of de spectrum atthe Seasonal Frequencies is considered and shown 9 bive the same asyimpotic distribution asthe original formulation, Under the aul hypothesis, the distribution ofthe seasonality test statistics isnot affeted by the inclusion of tends, even when modifi t allow for structural breaks, or y the incision of regressors with nonseasonal unit rons. A parametric version ofthe test is proposed, and is performance is compared with hit ofthe nonparametric text using Monte Carlo experiments. A test that allows for breaks inde seasonal patter is then derived Ii chown that ity asymptotic distribution is independeat of the break point, and ics use is illustrated witha series on U.K. marriages. A general test against any form of permanent seasonality, deterministic or stochastic, i suggested and compared with a Wald test for the significance of fixed seasonal dummies. I is noted tha less constructed in similar way an be used to detect trading-day effect. An appealing feature of the Proposed test saisties i that under the null hypothesis, they’ all have asymptotic distnbutions belonging tothe Cramer-vom Mises fail. KEY WORDS: Crumé+-von Mises distribution; Locally est ivariant test: Seasonal Ireaks: Structural Tine series model Trading effets; Unobserved component 1, INTRODUCTION ‘Monthly and quarterly economic time series are often subject to seasonal movements. These seasonal patterns tend to evolve over time, and most seasonal adjustment procedures assume thal this is the case. However, if the seasonal pattern does not change, it can be modeled by a set of dummy variables. Indeed, a deterministic seasonal pattem can be removed without even ating a time series model. All that needs to be assumed is the number of times the series needs to be differenced to make it stationary (see Pierce 1978; Harvey 1989, p. 203). Adjusting y simplify the exploration of relationships esti series in this way between time series. ‘Canova and Hansen (1995), hereafter denoted by CH, pro- posed a test of the null hypothesis that the seasonal pattern is deterministic against the alternative that it evolves as a1 non= stationary stochastic process. The test includes a nonparametric correction for serial correlation and seasonal heteroscedastcity. The aim of this article is to extend the CH test in various direc tions, We show how to modify the test so as to allow for the effect of modeling breaks in the seasonal pattern. We then con= sider a different, but related testing problem—namely, testing for the presence of any kind of seasonal effects, whether de- terministic, stochastic. or both, Similar techniques can be used for detecting trading-day effects. Before describing these ex- tensions, we examine the CH test in more detail, look at some alternative formulations, and compare the performance of the ‘nonparametric and parametric tess. Section 2 shows how a nonparametric correetion for serial correlation can be set up in terms of the spectrum at seasonal frequencies. This formulation is more restrictive than the CH insofar as it does not allow for seasonal heteroscedastic- ¥. Subject to this proviso, its asymptotic distribution under the null hypothesis is the same as inthe original formulation, but its te interpretation is more transparent. In deriving asymptotic distr butions, we relax the conditions of CH by showing that the dis tribution is unaffected when a deterministic tend is included in the model; regressors with unit roots can also be included provided that they do not have seasonal unit roots Parametric versions ofthe tests against nonstationary season. ality can be based on structural time series models, Such models are set up in terms of unobserved components, stich as trends id cycles, which have a direct interpretation (see Harvey 1989; Kitagawa and Gersch 1996). The use of autoregressive models (as in Caner 1998) is less appealing in this context, because they may yield a poor approximation to the moving average terms typically found in the reduced forms of structural models (see that when (esting against the presence of «random walk compo. nent, a parametric approach will usually give tests with a higher power and more reliable size, We investigate whether this is the se for seasonality tests through a series of Monte Carlo ex periments, The results are reported in the final section, Alon ‘with casting light on the relationship between parame ‘nonparametric tests, these experiments provide information on the robustness of the nonparametric test 10 the order of differ encing. Breaks in the trend feave the asymptotic distribution unat fected if they are correctly modeled by the inclusion of dunny variables; this is proved in Section 2. Structural breaks in the seasonal pattern are also considered. Empirical results of Canova and Ghysels (1994) suggest that seasonal mean shifts are not uncommon in U.S. quarterly series. Neglecting these ‘©2003 American Statistical Association Journal of Business & Economic Statistics uly 2003, Vol. 21, No. 3 oI 10.1198/073500103288619061 420Busett and Harvey: Seasonally Tests shifts will bias the nonstationarity tests at both the O and the seasonal frequencies, Modeling breaks in the seasonal pattern will, however, affect the distribution of the seasonality test sta- tistics. Section 3 shows how to construct a test statistic against stochastic seasonality, the asymptotic distribution of which is independent of the breakpoint loc The parametric and nonparametric tests, with the breakpoint moditication, are illustrated by an application to a quarterly s ries on U.K. marriages. The point about this example is that there is an identifiable break in the seasonal pattern because of a known change in the tax laws. The modified testis trying to assess whether it is necessary to allow for stochastic season- ality once the deterministic break has been accounted for by intervention dummy variables Section 4 suggests a general test for seasonality. This takes the same form as the test against nonstationary seasonality, ex- cept that scusonal dumunies are not fitted. The asymptotic dis- tribution is given, and the performance of the test is compared with that of a Wald test for the significance of fixed seasonal dummies. Similar techniques are used to construct a test for the presence of trading-day effect ‘A unifying feature of the test statistics is that under the null hypothesis, they all have asymptotic distributions belonging to the Cramérvon Mises family. These distributions differ ac- cording to the deterministic components fitted and a degrees of freedom parameter. The same distributions arise in tests against nonstationary trends as noted by Harvey (2001), 2. TESTING AGAINST THE PRESENCE OF A NONSTATIONARY SEASONAL COMPONENT In this section we develop the trigonometric form of the test against nonstationary stochastic seasonality, show that it is lo- cally best invariant, give nonparametric corrections for serial correlation, and show how to set up a parametric test. We then use Monte Carlo simulation experiments to compare the per- formance of the parametric and nonparametric tests in small samples. 2.1. The Testing Framework Let yy be a scalar time series, let s denote the number of seasons ina year and let Z = (jj. be an (s—1) vector of trigonometric seasonal variables, that is, r= (Cos Irrjt/s, sin 2art/s)', j= 1,8", where s? = 5/2—1 is is even or [5/2] if y is odd, whereas 24/24. =(—1)! if s is even. The jth pair of trigonometric terms, selection matnx with rank & <5 — 1, and and x; ate mutually uncorrelated mean 0 disturbances with variances a? and 621. The initial value yp is assumed to be fixed ‘The aim isto test the null hypothesis that the seasonal com ponent is deterministic, Ho : 07 = 0, against the alternative that ithas unit root behavior, Hy :42 > 0. Following CH, the matrix A is used (o formulate tests at subsets ofthe seasonal frequen- cies 1... joa} Ifthe testis to be carried out at the single frequency #,, then we let A, denote the (2j— 1)th and (2j)th columns of f,-1 for < s/2 and the (2j—1)th column of f, if j =9/2: fe denotes an identity matrix of dimension k. When all of the seasonal frequencies are considered, A = /, 2.2 Locally Best Invariant Test Under Gaussianity, the locally best invariant (LBD test for the null hypothesis of deterministic seasonality for the model ()-G) can be easily obtained from the results of King and Hillier (1985) [which also show that the test is a one-sided La. ‘grange Multiplier (LM) test] and Taylor (2003a).. Specifically, let cs. r= 1, .... 77, be the ordinary least squares (OLS) residuals from regressing y, on (X),Z)" and let 5 TST, e? be their sample variance, Let aj = 1 if j= 5/2 and aj = otherwise First, consider each seasonal frequency. 4. = 1... [5/21 in turn, that is A = Ay, in the model (1)-(4), Then, under Gaus: sianity, the LBI test for testing Hy: 02 =O against Hy 2 > rejects for large values of the statistic a, defined as [es om] [5/21 Note that when s is even, the test statistic at the Nyquist fre queney, 2, eum be written without the terms e, sin A,/2i, be- cause they are identically 0. A complete test against nonstationary seasonality at all fre~ quencies, (ie, A= /e-1) rejects for large values of the statistic obtained by adding up the test statistics for each individual fre- quency, namely j=l. (3) o ‘This tea is LB fora model where under the ahemtive hypoth caie, the coeficlent camesponding to cach seasonal freency Thc evelee lon ode esdea een walks with variances 07’, and if s is even, then the coefficient for fre- quency 7 is a random walk with variance 02/2. Thus Wis an identity matix unless is even, in which case he last element inthe main agonal is 12 Under ty. 2 CoMQa) and o % Comes — 1), where CVM(k) denotes a Cramér-von Mises random variable with k degresof eedomand 3173" q, = 5-1 Gaussanity othe 5422 is not needed: all that is required is that they be martingale dif ferences satisfying the conditions of Andrews (1991, p. 823) or Stock (1994, p, 2745). The proof is a special ease of Proposition 1 in the next section. 2.3. Nonparametric Correction for Serial Correlation Based on the Spectrum at Seasonal Frequencies Serial correlation in the stationary component of (1)-(4) can be treated nonparamettically by replacing the sample vari- ance, 6%, in @ with an estimator of the spectrum of f, at fre- quency 2j. We denote this spectral nonparametric test statistic by sony = Ra an 50830 + a sin] " am) v[s/2}, (7) where YS wees mfete)c0s 47 ® is the estimator of the spectral generating function, w(t.) is a weighting function or kernel, such as w(r,m) = 1 ~ |e|/ (m+ I), and felt) = TDP, everae is the sample auto- covariance of the OLS residuals at lag r. Alternative options for the kernel w(,-) have been examined by Andrews (1991, p. 821). For testing all of the seasonal frequencies, the spectral nonparametric statistic is wal oto = Yeni = Under the sumptions set out later, the asymptotic distri tions ofthe foregoing test tatstis under the null hypothesis are the same as given inthe preceding section Asnmption Al. %; a p x 1 vector of deterministic regres sors, and Dis (diagonal sealing matrix such that rt (a) tim! Sy 'XXiDz! = Oe ~ et and (b) im 7 Tee r ‘7x2 =0. where Qy is a positive definite matrix: Assumption A2. The €,'s have the structure of a linear process, ¢ = y (Les, where {sf} is a martingale difference sequence satisfying the conditions of Stock (1994, p. 2745) and V(L) = 1+ SS) yal! is a polynomial in L, the con- ventional lag operator, L'y, = y;4, k = 0,1,..., satisfy- ing (a) y(expl+i2xAj/s}) #0 for all + [5/2] and (b) DE Alva! < 00. Assumption A3, The lag truncation parameter, m, is such that, as T > 00, m—> 08 and m/T!2 + 0, Journal of Business & Economie Statistics, July 2000 Proposition 1. Let y; be generated by the model (1)-(4) un- der the assumptions A1-A3. Then, under Hy: 02 =0, when A=A FN. f5/2 1 com) 4 f Bay (0)By(r) = Coa) I when A= Iy_),co(m) CuM(s ~ 1), where By(r) = We) rWi(1), and We(r), r € [0,1], denotes a k-dimensional stan- dard Wiener process. Under Ha : 2 > 0 and when A = Aj, F= 12.25 [8/2], cylin) and calm) are O,(T/m). Proposition 1 extends the results given by CH to allow for a general specification for the deterministic trend ja;; CH gave ju: as a constant level, The limiting distribution is unchanged pro- vided that the regressors X; satisfy assumption A. In particular, contrary to what CH (p. 238) stated, the model can include car trends. Structural breaks in the trend at known points can I tadi(ce)), where da) is a dummy variable equal to 1 fort > @T with 0 0 (see Stock 1994, pp. 2797-2799) Model (1)-44) can be extended by including stochastic re- ressors with nonseasonal unit roots, and following the line of the proof of Proposition 1, it can be shown that the asymp totic critical values for the tests in the augmented model are unchanged; the proof is omitted here but is available from the authors on request, Such regressors were ruled out by CH, who stated that “the explanatory variables may he any non-trending variables that satisfy weak dependence conditions.” ‘The gen- cralization is of some practical importance, Note that the pres- ‘ence of cross-correlation between ¢, and the disturbance vector driving the stochastic regressors is not important for our testing seasonality; unless we are interested in making inferences on the coefficient vector of the regressors, there is no need to use, say, a fully modified least squares procedure instead of OLS. In summary, the inclusion of deterministic trends and sto- c’chastic integrated regressors does not affect the asymptotic dis- tribution of the seasonal test statistics. However, the inclusion of seasonal slopes does affect the distribution, just as it does for tests of seasonal unit roots, as discussed by Smith and Tay- lor (1998) and Taylor (2003a). Given the obvious parallels with stationarity tests, it is not difficult to see that the change in dis- tribution is the same as when a time trend is fitted before a sta- tionarity test statistic is constructed, ‘The critical values, once again from the Cramér-von Mises family. are as in table 2 of Nyblom and Harvey (2000). For quarterly data, the 5¢ critical value for an overall test, hased on 3 df, is .337. (There is actu- ally & typographical error in the published table, with the value printed as .332.) Dummy variables introduced to capture breaks in the sea- sonal paltern will also affect the distribution of test statistics, This is the subject of Section 3.Buserti and Harvey: Seasonality Tests 2.4 The Canova-Hansen Test Statistic ‘The test statistic proposed by CH in the framework of (1)-(4) takes the form / r alm) = Tae (Bima) Wyss), cy where S;= 1, Ze; and Q(m) is a nonparametric estimator of the “long run variance” of 261; that is, Bom = YO we mle, (10) where w(t,m) is a kernel as in (8), and P(r) = 7! 1Ziew-eZ), 8 the sample autocovariance matrix at lag 1 formed from Ze;. The main difference between (9) and co(m) is that (9) allows for seasonal heteruscedasticity as in Andrews (1991, p, 839); see CH (p. 240), Under the null hypothesis Hy :0,2 = 0, the asymptotic representation of (9) corresponds to that of Proposition 2, that is, wa(m) % Cyvank(AY), 2.5 Parametric Tests A steuctural time series model typically contains stochastic trend and seasonal components, together with an irregular. This model can be extended in various ways; for example, by in- cluding a cycle, However, for many economic time series, the Aicxibility of the stochastic trend is such that the model can still adequately capture seasonal movements if a cycle is excluded. ‘We now consider how to set up @ parametric test of whether the seasonal component in a structural time series model is stochas- tic Ifthe process generating the nonseasonal part of the model is known, then the LBI test against stochastic seasonality is con- structed from a set of “smoothing errors.” As shown in Appen= dix B, the smoothing errors are in general serially correlated, but the form of this serial correlation may be deduced from the specification of the model, thereby allowing construction of a statistic with a Cramér-von Mises distribution asymptotically, under the null hypothesis. The computation of smoothing er rors has been discussed by de Jong and Penzer (1998), but if the model contains a serially uncorrelated irregular component, then it can he shown that the smoothing errors are proportional to the optimal estimates of this component. ‘An alternative possibility is to use the T' standardized one- Mep-iheud prediction errors, the innovations, calculated. by treating nonstationary and deterministic components as having fixed initial conditions. No correction is then needed; the sta tistic is of the form (5) and has the same asymptotic distribu- tion, Calculating innovations under the assumption that the ini- tial conditions are fixed requires that the initial conditions be estimated, but a backward smoothing recursion can be avoided simply by reversing the order of the observations and caleulat- ing a set of innovations starting from the filtered estimator of the state at the end of the sample. Actually, the forward and backward innovations are not the same, and in neither case do the sums, weighted by cosA,t and sin 4,t, equal 0, so statisties 423 formed from forward and backward sums are different. Fortu- nately, the asymptotic properties are unaffected, Smoothing er- rors do not suffer from these ambiguities. For both the smoothing error and innovation formns of the test, nuisance parameters normally will have to be estimated For stationarity tests, Leybourne and McCabe (1994) argued that this is best done under the alternative using maximum like- 10d. Proceeding in this way has the compensating advantage that ecause there often will be some doubt about a suitable model specification, estimation of the unrestricted model af- fords the opportunity to check its suitability by the usual diag~ nostics and goodness-of-fit tests, Once the nuisance parameters have been estimated, the test statistic is calculated from the in- ovations obtained with o2 set to 0. The asymptotic distribution under the null is unaffected. 2.6 Monte Carlo Experiments ‘This section compares the probability of rejecting the null hypothesis of constant seasonality for the nonparametric tests of Sections 2.3-2.4 and for the parametric test of Section 25 based on a correctly specified mod. The results offer guidance in assessing the effectiveness of the two approaches as well as establishing the reliability of the tests in terms of actual, as op- posed to nominal, size. The data-g on process is the basic structural model (BSM), consisting of seasonal and stationary components com- bined with a local linear trend, je. Thus Met Ziybe, &~NIDO,G2), 1 ah where He Wert Berth my ~NIDO.0,) (12) and Pra feitt, — %~MDO.e2), 3) vwith jy as in) with A ‘The relationship between this seasonal model and the owe of Harvey (1989, chap. 2) hus been discussed by Proieti (2000) “The probability of rejection depends on the seasonal signal- to-noise ratio, ge = 07 /0?, although in the tables we prefer to report the square root. Empirieal size and power of the tests are computed for @, /a- taking the values 0, .01, .025..05, .1, and .5. The results are for quarterly series of length T = 200, The empirical rejection frequencies, reported in percentages, are based on $0,000 replications and refer to tests run at the 54 significance level. Results for testing at a single frequency, 1/2 and 2 in turn, and for the joint test at both frequen are provided. The program was written in Ox using the SSF- pack set of subroutines of Koopman, Shephard, and Doornik (1999) The first set of experiments, the resulls of which are given in Tables 1, 2, and 3, is for the data-generating process (11) (13) with 22 = 0, so the trend is a random walk with constant drift, 8. The performance of the tests does not depend on the value assigned to f1and soit can be equal to 0. The critical far~ torn the nuisance parameters is the level signal-to-noise ratio, 4, = 72/2, and gy!” is set to .1 in Table 1, .5 in Table 2, and 1-0 in Table 3424 Journal of Business & Economic Statistics, July 2003 Table 1. Rejection Frequencies for a Random Walk Dritt Plus Noise Model With oy fap = <1 i= 0 oi 482~=«58]sSOSSCSSGSS*«CSCDYSCSDSOCATS Aa? «SBA Me 9 ol = as? Sw] 728 77779) Ome 549 GTA 5m 2 © 2678 aTSD ONS) MMAT S129 SAID. 24752999 2599 24.96 zE2 wom 292 2042 0 © aa? 7787407778 SOO. SOAR GTS SSG GES S249 5544 52.01 300 862792810554 «98ST 95.4] SB.IS | BOSZ MLAS B55 KSAT 85.49 B42 BSE 8045 500 9675 «10000 «9870-10000 «ET «00.00 «GRID TY 772 9589 BAI 9550 8721 94.97 ° er ee er ee oo roots ta MBT «HSK GT SS 12h 7! ROL 895 11.2 971 25 oat? HOS] AIA ©4039 ATA4 | 27H BH S149 B10 72 2654 31.49 29.10 060 $8.05 O71 7485 77-80 HHS TVOK S448 ORD 57.46 5247 S440 50.00 57.40 S247 100 00.64 911.25. UDO «9579-7579 08.92 | 70.99 71.25 TOTS ORM TRAD 7129 500 4.05 100.00 643_—=«100.00==«9.S0 «100.00 2.12 BAD BRON BSA SCZ 8920 00.00 e204 lomo sot 460958979 70061 OH 08a 4SS 3M 245 420 28s 10 10.3202 15 7U=S 1023 tH «N8OS BG] N4 1069 928 7319 527 eI Ane 2 «4088 AN ATSB 46.90 47.90 3671 HG 400 9582 24 26.80 25.04 2860 oso wat 2.08 TID. BYTE 86BH TSH T76H © TABL© 70.27 7570 7807 88:70 76.94 69.47 100 S776 «98.77 90.45.9052 WHOA SP OMA? © 96.E3. ONES 95.55 92.11 O58 G225 500 100.00 100.00 100.00 100.00 100.00 _—*100.00 9995 9959 98.90 RMB HUM MNIA MES 907 The spectral nonparametric test and the CH test are ap- ied both to levels and to first differences with the lag trun- cation parameter, m, set to 4 and 8; CH used 3 and 5 for T = 50 and 150. The parametric test statistics are computed first assuming that gy is known (with results in the columns headed “Model known”) and then with o and o? estimated by maximum likelihood under the alternative hypothesis (in columns headed “Model estimated”), In addition, the BSM ‘model is also estimated; that is, the constraint that the slope variance is 0 is not imposed (with the relevant columns headed “BSM estimated”). The parametric test results are shown for innovations, computed starting from a smoothed estimate of the initial conditions, and for smoothing errors; compare Sec- tion 2.5, ‘The main findings of Tables | a. Although the empirical sizes (rows with a /oe equal to 0) of the parametric tests with qy known are very close to the nom- inal 5%, the tests with q, estimated are somewhat oversized at frequency x. The parametric joint tests for overall seasonality are similarly oversized, with the actual sizes being around .09 in most cases, It is interesting to note that the autoregression- based! tests reported by Caner (1998, table 1) display even more oversizing. b. There appears a slight power advantage for a parametric test constructed from smoothing errors, rather than from inno- vations, when testing at a single frequency (particularly when the smaller empirical size of the parametric test is taken into account); for example, for 4 = 7/2 in Table 1, the estimated 2, and 3 are as follows: Table 2. Rejection Frequoncios for a Random Walk Drift Plus Noise Model With ay ax = .5 — Paramole tests Modo known ‘Medi estimated ‘Smoothing ‘Smoothing Joe Ianovatons —eiors” Inmovaions eters tua 0 “0 «S508 210 738 BM 928 es, 2498 20.60 0.49 0 oe | a2 25 100 9098 9eaz 07.66 500 1000 9865 100.00, den 0 499 98a ow 93941188 2 mez wes ane ‘960 sear 7290 75.27 10 0015 ahaa 8.19 500 S441 9.99 98.A3—_—100.00 er) sm 4800927 gt om 871 a5 1S 5 3825 maSN TST 0s 78818040530 100 9729982728 99.29 90 10000 100.00 0000 10000Buse and Harvey: Seasonality Tests 425 Table 3. Rejection Frequencies fora Random Walk Dit Plus Noise Model With on fox ~ 1 Paameric es (adel known ‘Model stated OSMesinaed _ —____Noparmesteg ‘Smoothing Soong ‘Snooting Seedtal Levels _Spectat: FO _CH. Levels _ CH: FD cult troovatore “eras” nnowabone “"arers” Imovatone eno” m=4_m=8 m=@ mad m=8 aad mae ° ve aa”~C«eSCS Ca 4% 450 «40g? at am 0% eae rhe tone vie Ga bane ey gat ate 225 14s ers mk meme tae fos 88) wee 158 27) tb MD TaD 980 27m slot esa Sia? a.e 27 837 dea 1281 Jet a6 AT soo 2514s 8548 81a sad fog 7am 7745 4013 499 7aa8 757 soo wooo $028 10000 gaze © 100.00 Shu o7se GEdb 9022 get GBae abla des 4s 979 asks saa ast sea aot aia a5 aes grag 17 ses ae 368 ceo 5 ws 59062354820 oe ara 287 a2 2457 ‘0 s23 mo asa sat Seah fe onan ras aa seo 9898010000 arak aD 1m Gees mes fan m6 shot ee ee 400 276 oo Bates naar atk ase na 76 cos ek St we HRND Hem) at ahaa ame ares 19 oso 7227 © 7363«««7ar2,«=«7339««7HN9«*TOTI 580 3640 70.28 G68 e722 5931 Too 570 «710 kab ORS «SONG n8D Tan T0009 SHAD OL? sae 0096 soo tooo soo) rapa 100.00 ooa 74 G25 wose moar seo ons 00.09 rejection probability is 64 versus .60 when the model is known and .78 versus.74 when the model is estimated. This advantage seems to vanish for the joint tests, . Estimating the extra parameter, q. = 2/02 in the more zeneral BSM model has no adverse effect on the performance of the parametric test, On the eontrary, there isa slight improve- size in that itis closer to the nominal; for example, in “Tuble | the empirical size of the joint test is07 when a BSM is estimated, compared with .10 when the trend isa random walk with drift 4. Regarding the nonparametric tests in first differences, only a small drop in the rejection probabilities occurs when moving from m= 4 to m=8, The seasonality test seems less sensitive 10 lg length than the test of Kwiatkowski, Phillips, Schmidt and Shin (1992); compare the comments of CH (p. 246). However, m should not be set too small, Our sim- ulations tor the joint test with m =O (unreported in the tables) showed rejection frequencies of 14.5, 13-4, and 11.0 for the models in Tables 1, 2, and 3 . The nonparametric test in levels has a much lower probit bility of rejection than the corresponding test in is differences when ql” is.5 or | (see Tables 2 and 3). This is due to the pres: ence of a so-called unattended unit root at frequency 0 when the tests are run in levels. In fact, Busetti and Taylor (2003) and Taylor (2003b) have demonstrated that, although they are wor sistent, the nonparametric ests would suffer from a great loss of power if that unit root were not removed by taking first dit ferences of the data. Indeed, these authors showed that under the null hypothesis of deterministic seasonality, the test stati ties (5), (7), and (9) all converge in probability to 0. This i also confirmed in Tables 1, 2,and 3; in Table 3, where 0; /a; is equal to 1, the joint test in levels has a size of neatly 0. f. The parametric tests show higher rejection frequencies than the nonparametric tests, but any assessment of power must be considered in terms of the larger size, Overall, the loss in power resulting from using nonparametric tests is not great For example, in Table 3, 4 = 2/2, where the empirical size of the innovation-based parametric testis broadly comparable to that of the spectral nonparametric test in first differences with m=4, the powers for a seasonal signal-to-noise ratio of 05” are around 60 and .50. The nonparametric testy perform relatively better than the results for stationarity tests (at the O frequency) would sug- gest. This is because in the experiments reported in the liter ature (€.8,, Leybourne and McCabe 1994), the process gen= crating the stationary part of the model—typically a first- order autoregression—interferes with the unit root process, This problem does not arise with the data-generating processes considered here and it would be unlikely to arise even if a first- order autoregressive process were to replace the white noise ir- ‘The second group of experiments, reported in Tables 4 and 5, is for the so-called smooth trend model; the data-generating process is given by (11)-(13) with o2 = 0. The relevant signal {o-noise ratio is now qz..The parametric model is estimated both with and without 9? set 100; the results are given inthe golurnns labeled “Model estimated” and “BSM estimated.” The nonpara- metric tests are run after taking first differences and after taking. differences a second time, Although in theory, the first differ ence operator should be applied twice to avoid the power loss induced by the presence of unattended unit roots, Table 4 indi- cates that for g,' = .1, no significant loss occurs if only first differences are taken. In fact, inthis case first differences: may be preferable, and because 4) is often smaller than .1 for eco- nomic time series, basing tests on first differences may be a good strategy The conclusions with respect to size and power that emerge from Tables 4 and 5 are similar to those reached for the ran- dom walk plus drift model of Tables 1, 2, and 3. The parametric tests are somewhat oversized but have a higher probability of rejection under the alternative. Again, there is no disadvantage to fitting a more general local linear trend model when carrying ‘out the parametric tests,426 Journal of Business & Economic Statistics, July 2003 Table 4. Rejection Frequencies for a Smooth Trend Plus Noise Moder With a oe ~.1 Parametric 056 ‘Model krown Model estimated SM estimated ‘Smoothing ‘Smoething cee Innovations "errs" Inpavaions ~erers lonovations rare 0 5477S 4225 2030s me oi 876 atk gm?) t0.06 = 087 oe ee ee oe 2713 age sta same smh 2388 1819 19.08 2238 1989 1579 1572 Oso 6058 BL mhz 7h HTB TRST 5500 5073 5160 55.52 Stas 4785 40.82, 300 8637 92500554 98.90 95.59.26, 83.07 90.75 79.83 RT BOD 7908 7HOR 500 9878 100.00 96.69 «100.00 98.86. 100.00 Boe W54s Oh G78 OMB OAT gee nx 0 51402 eos) sor 74 SBT 618 S27 Tat 587 oo 899 aes tat 960 1253 1052 10.85 960 t259 1052 5 s089 soma ao) 4188 1782.78 2100 33.17 290797 B70 817207 os) 58036074 = 7438 77.09 7490 77.95 5230 Sago 53125715 8200 5890 69:12, so) on72 6971 was 7inz 799 © 716) «7863 7102 73.90 71.560, 5009408 19000982 10000-9838 100.00 B82 035+ — 8405 80.00 ase? 99.54 84.05, Jont 0 S21 456 9a) 99 t029T0.89 419 367 3st gaz 264 BaD td 910 1965 10.08 © te tock 87a 720 Bas 828781 B1e «578 891485 O25 41154144 ards 48258858 B77 510 OD MASI 2798 NA 25.40 60 BOs mT wr3S 7B CTBT Ba 7529 7591 73.18 7593 0BB7 72K H50 400 0782 087700400880. O47 9572 G38 oR? WAT 9470 G15 500 10000 10000 100.00 1000 100.90 100.00 eae 90m a mae m0. 9981 _8a79 3. DETERMINISTIC BREAKS IN THE SEASONAL PATTERN Because the models used in the foregoing simulations do not, exhibit seasonal heteroscedastivity, itis not surprising that the spectral nonparametric test performs slightly better than the CH. test. However, once seasonal heteroscedasticity is present, the situation changes. For example, with a model consisting of a Seasonal plus white noise with variance in the four quarters of 1, 3,5, and 7, the size of a «(4) test for T = 1,000 was esti ‘mated to be .O80, and that of a w(4) test was 038, Thus the CH size is closer to the nominal .05. However, with q = .025, the In this section we consider testing against nonstationary sto- cchastic seasonality when there is a break in the seasonal pattern at time [aT a € [0, 1; that is, we replace (3) with 8 Zire + danza, «sy estimated probability of rejection was .458 for «n(4) and .662 for wa(4). Although this is rather extreme case, it does illus- trate the point that when seasonal heteroscedasticty is present, the CH test not only is more robust with respect to size, but ulso may show higher probability of rejection away from the null where d)(a) = 1 (> @T) is a break dummy variable. The model now implies that the coefficients of the seasonal terms have changed from y; when ( < [aT] to y; +0 when 1 > [aT] We focus on the nonparametric tests, although of course the same issues arise with the parametric versions. Table 8. Rojection Froquencies for a Simgoth Trend Plus Noise Model With jor =. GH Doubie FO ‘ox/ee Innovations ‘errs Innovations erors”” Innovations rors mat m8 mad m8 daw 0 400° ~«8SS ome O mw 2 29 00 yes 88598905 140193 383. 379 28 vse 05D OTH case 821 965 1548 15.15, 050 eras 7184 7558714 75.88, 9898 97.90 4087 45.89 100 D148 947978) rama Tr 716 78887827 00 $0000 9850 100.00 a1 19000 B78 G95 9462 92.59, ° 516 485 87s soar si 277 328 729 527 10 1005 ast HOT 8a? S248 1238 1023598 658 12.90 1020 02s socewati40z0 a9 3278 282 20 Bm BIA 282 os 5788 SOI2 71S TOAD HTB 77.28 S851 5254 WEE 478 Shot 5294 so =089 © 8857.28 aS BB ye7s aT ea) RIS TOT THAT 500 94409899 gag) 100.00 AR 100.00 S52 M99 GAD ISS 9552 60.99 Joint 0 50 422,97 ge) 10.10 1097 ser as) 6273 ae 2.09 01 1082 = gad ash t628 18.67 «17.08 Bot 774 203 209 G78 set s0eD 462470) 47.208 3410 BI 15.90 17 aa 28.80 oso 7096 ©8151 N62 oh aaa? «8708 7480 7228559 HRD 7202 5.50 100 9782 ©9850 wb aah? ©0940 S553 Gs mB A718 OA DOM 50010000 10000 100.00 e000 r00.00 100.00 soot ons: oa? onse oce2 90.74Busett and Harvey: Seasonality Tests It is initially assumed that the breakpoint parameter, c, is in which the breakpoint is un- known are discussed in Section 3.3. ‘When there is a break in the seasonal pattern, the nonpatam- tric test statistics «j(mm), co(m), and ca(m) of the preceding section must be constructed using the OLS residuals from re~ gressing y, on (X/, 2), dia)Zi)’. Their asymptotic representa- tions under the null hypothesis are no longer Cramér-von Mises with aj,» ~ 1, rank(A) dt, but rather they depend in a rather ‘complicated way on the breakpoint parameter a. However, a simple modification yields test statistics, the null limiting dis- tributions of which are still Cramér-von Mises but with degrees of freedom equal t 24,,2(s~ 1) and 2 rank(A). This extends to the seasonal case the modification to the LLBI test at frequency O suggested by Busetti and Harvey (2001), The parametric test can be modified along the same lines 3.1. Modified Test With Seasonal Break ‘The modified seasonal break spectral nonparametric statistic for testing against nonstationary seasonality at frequency 2, is defined as {5/2}, 5) ) + (Soesinas) « is ky = [aT] = diay) +11 = aT) *dtar), ‘a and (4.1m) are defined as in the preceding section, and the ¢}s are the OLS residuals from regressing y; on (X/, Zi. d,(a)Z))'. The corresponding statistic for the test at all frequen- cies is then w*(a; m) = S17 af (er; m), The modifications to UK mariage 427 the CH statistic, (9), and the parametric statistic are carried out ina similar way, Proposition 2, Let yy be generated by the model (1), (2) (14), (4) under the assumptions AI-A3. Then, under Hy : =0, when A= Aj, /=1,....[5/2], f(a: m) 3 Cat ay), when A= 1.1," m) 4 CvM(2s — 2). Under Hy :o2 > 0 and when A =A). j [s/2}, of (ee) and e*(crsm) are p(T Im) ‘The idea behind the construction of (15) is to combine the evidence in the two subsamples, {1 Locos} Note that @/t(S:m) = 280m breakpoint is in the middle of the sample, the tests defined by the two statistics are the same. This is important, because the latter hay properties of optimality obtained by extending the LBI/LM test t0 deal with serial correlation in the stationary component. Furthermore, for the case of testing at frequency 0, Busetti and Harvey (2001) have shown via simulation exper- iments that for a 5, the loss of power of the modified test with respect to the LBI testis not great 3.2. U.K, Marriages ‘The quarterly series of marriages registered in the United Kingdom from 19S8Q1 to 198204 was extracted from various issues of the U.K. Monthly Digest of Statistics. Iv is shown in Figure 1(@). The spectral nonparametric test statistic, e(m),cal- ‘culated from first differences, is 4.18, 2.74, and 2.1] for lags of 4, 8, and 12. This leads to a rejection of the null hypothesis as the 5% critical value for the CvM(3) distribution is 1.00. The original CH statistic, «4(m), gave smaller values: 1.78, 1.20, and 96. Estimating (11) with a random walk trend using the STAMP. 66 program of Koopman, Harvey, Doornik, and Shephard (2000) Figuoe 1. (a) Number of Marriages (in thousands) in the United Kingdom, 195801198204 and (b) Estimates ofthe Individual Seasonal.426 gives HH, Gy =161, and =2.69, with an equation standard error (the standard deviation of the innovations), 4, of 7.91. The parametric test statistic, con- structed from the Kalman filter innovations, is 6.96 which is a ‘much firmer rejection of the null hypothesis than was given by the nonparametric test. The reason for the rejection can be set in Figure Ia): there appears to be a break in the seasonal pat- tern at the beginning of 1969. The plot ofthe individu seasons in Figure 1(b) reveals a switeh from winter marriages to mar- riages in the spring quarter. This happened because of a change in the tax law. Up to the end of 1968, couples were allowed to claim the married persons tax allowance retrospectively for the ‘entire year in which they married. Because the tax year begins in April, this arrangement provided an incentive to marry in the first quaiter of the calendar year rather than in the spring. Adding a set of three seasonal break dummy variables, start- ing in the first quarter of 1969, to take into account a complete change in the seasonal pattern leads to the following estimates ‘of the parameters: 242, Gy = 159, and = 1.36, (9, 7) = 12.54 5.66, where OUP.) is the Box-Ljung statistic based on P residual autocorrelations but with f degrees of freedom (see Koopman et al. 2000), The ¢ statistics for the seasonal brenk dummies are =R.33, 7.58, amd 2.09, There isa large reduction in the estimate of the seasonal parameter, a, which no longer needs to allow the stochastic seasonal model to accommodute the change; the equation standard error, 6 also is eonsiderably reduced, ‘The modified seasonal break nonparametric test statistics ear ried out on the residuals obtained trom regressing first differ- ences on seasonal means and the seasonal break dummies are 2.06, 1.69, and 1.58 for m =4,8, and 12 for w* and 1.80, 1.57, ind 1.50 for the CH form, «. Thus for m= 4 and 8, the mull of a constant seasonal pattern is rejected by the «* test at the 34% level of significance because the critical value for CvM(6) is 1.69, However, the smaller values tor a lead only to a re- jection for m= 4, The comespondling parametric test statistic, calculated from the Kalman filter innovations, is 2.42, giving a stronger indication that there is still stochastic seasonality present. This is supported by the fact that estimating the model with a fixed seasonal gives a significant Box-Ljung statistic of Q(9, 8) = 22.38, whereas the fourth-order residual autocorrela- tion, r(4). is 33, and 6 3.3 Unknown Breakpoint If the breakpoint parameter @ is unknown, then the two- step strategy of Busett and Harvey (2003) for testing station arity in the presence of a structural break can be adapted in 4 straightforward manner. The idea is to estimate the break Point under the null by minimizing, over et the error vari= ance of an OLS regression of y; on (X, Z.dj(a)Zi)’, that is @ = argming T'S) e?, where e, are the OLS residuals. B. (1997) showed that under the null hypothesis of deterministi Journal of Business & Economic Statistics, July 2008 seasonality this estimator is superconsistent in the sense that it ‘converges to the true value at rate T instead of the usual rate T'2, ‘Therefore, the null asymptotic distributions of wf (@; m) and «"(@;m) are the same as if the true value @ were used, whereas under the alternative hypothesis, he statistics diverge regardless of the break date tsed due to the presence of the sea~ sonal unit roots. Hence running the seasonal break tests with an estimated breakpoint leads to an asymptotically valid proce- dure, Clearly, some loss of power with respect toa test based on a known ais to be expected. Busetti and Harvey (2003) used Monte Carlo simulation experiments to evaluate this power loss for the zero frequency stationarity tests. 4, TESTING AGAINST A PERMANENT SEASONAL COMPONENT “The test against nonstationary seasonal components takes the null hypothesis to be a model in which seasonality is determin- istic. Sometimes we may wish to test whether there is any sea~ sonality at all, irrespective of whether it iy deterministic or sto- cchastic. One strategy, implemented in the STAMP package, is to fit a structural time series model and then perform u test of ignificance on the seasonal coefficients as estimated at the end of the period. However, this has the disadvantage of not being able to indicate seasonal effects in a situation where seasonal- ity has become less pronounced over timte. This is precisely the kind of behavior noted by CH (pp. 24-50) in their analysis of USS. macroeconomic series. 4.1 Tests and Their Power ‘Our aim is to develop tests that ure powerful against the pres- tence of deterministic seasonality and/or stochastic seasonality. The stochastic seasonality is taken to be nonstationary, so that the effects are permanent in that forecasts of the seasonal pat- tern remain constant rather than dying away. We thus consider the data-generating process (1}-(4) with A=, 1 and split the seasonal component, s,. into deterministic and stochastic parts, Ps, 6) Palin, ayy and =Z mn as) Ai where yn is a (= 1) % 1 vector of fixed coefficients and ky is a vector of mean 0, serially independent disturbances with covariance matrix a2 W, independent of c;. The null hypothesis ‘of no seasonality is Ho: w=9, 0, whereas the alternative hypotheses are of deterministic season- ality, HP yw #0, 02 and stochastie seasonality, Hy =0, 02 >0.‘usetti and Harvey: Seasonality Tests We show in Proposition 3 thatthe standard Wald test on fixed, seasonal coefficients is consistent against both alternative hy- potheses. We also show that a test constructed in a similar way to the tests of Section 2, but without fitting seasonal regressors, 1s consistent against both hypotheses. More specifically let e be the OLS residuals from regressing y- on (X/.Z/)" and let e, be the residual from regressing yy on X; only. Because assump- tion AI requires that the two sets of regressors X; and Z; be or- thogonal in large samples, the X/s can be ignored in the analysis, of the Wald statistic, which may be written as ile Lea)e. F a9) where fp is the OLS estimator of yo and 6? = 7-1 2L, eP is the estimator of 2. Note that the usual form of the F statistic is our F multiplied by ( —s + 1)/2 The new statistic is 20) where each «,j but using the residuals ¢, the reverse order, that is, (Smo Jah b/2, QD where a? = T-'7 o?. Using the arguments of King and Hillier (1985) and Taylor (2003), itis easy to show that when W is specified as after formula (6), wy is the LBI test stati against /1§ for the model in (1)-(4) with yp = 0. Note that in the LBL statistics the summations run in reverse order, from 1 to T as opposed from 1 to f; in the texts of Section 1 the reverse summations can be replaced by the more usual direct summations, because iting the seasonal regressors implies that Tt Zei=0. ‘The following proposition provides the asymptotic distri- bution of F and @ under the local alternative hypotheses HP, and Hy, where HP: yo = cpt/VT.0! = 0 and HS 100,02 = A/T, where +s an s~ 1 vector of ones and ep and cs ate fixed constants, This provides the basis of a power ‘comparison between the two tests, {s/2). is defined as in Proposition | and with the summation running in Proposition 3. Let y; be generated by the model (1)-(2), (16)-18) with W-= 7, 1 and with the nonseasonal regres- sors X; satisfying assumption Al of Section 1, Let Wo,.-1("). Wi. 1(0) be independent standard Wiener processes of dimen- sion s—1.andlet A = diag(1/2, ..., 1/2, I} when sis even and A= 1/21 when y is odd, Then the following results hold: a, Under H?,, 4 Vp.s-1(0; ep) Vos-1(05 ep) ‘i 4 : Ff Voustesem)'¥ag serena where Vos—1ri¢n) = Was =) + epas AR ~ Dus re {0,1}, 429 b. Under 48 , FS V5.0: 5) Vs5-10:¢8), 8 fs. siresVsilriesddr, where Vs,-1005€5) = Wash — rte Tad {Wis 16s) ds, €[0, 1). ©. Under either HP or 9, F and a ate Op(T). Remark 1. The asymprotie distribution of F under H?, is a noncentral chi-squared distribution with s ~ 1 df and noncen- trality parameter equal to cht" Ai/o2. Remark 2. Under beth local alternatives (ie., when 3 = cot/VT and o2 the axymptotic distribution of F and & are construvied using the process Vir ep, ¢s) = Wo,.-il1 Db epAh = net esac! AS {Wy a6dds.r © 0.1 stead of either Vp.) (rcp) oF Vg.y—105 ¢s). Remark 3. A modification of w would be to replace o th ? that is, to fit seasonal dummies when calculating the denom- inator of the statistics in (21). This makes no difference to the asymptotic distribution under the null and the local alternative hypotheses Although the asymptotic distribution of cw under the null hy- pothesis differs trom that of w, it still belongs to the Cramér~ von Mises family, The 5% critical values for 1, 2, and 3 df— the latter appropriate for a full test on quarterly data—are 1.65 2.63, and 3.46 (see table | of Nyblom 1989, p. 227). The 5% critical value for 11 df (kindly supplied by J. Nyblom) is 9.03. For the reasons given in subsection 2.3, the asymptotic distrib- tution is unaffected by the inclusion of u constant or a constant and a time trend. ‘The asymptotic distributions of F and « under the local al temnatives H?, und H® , can be used to compare the power per- formance of the two tests, This is done in Table 6. Specifically, fora quarterly model, s = 4, we have generated 50,000 replica: tions of the limiting random variables defined in Proposition 3 by replacing the continuous-time Wiener processes Wo, Wi 4-1 by their diserete counterparts (dividing the unit 0 1,000 parts), We have also considered the limiting behavior the «test, invariant to the presence of deterministic seasonal: ity: its asymptotic distribution against 1° ; was given by Taylor (2003u). Note that under both the fixed and local alternative H? and #12, the asymprotic power ofthis tes is equal to its nomi- nal size. “Thus Table 6 reports, for a qu totic power of the three tests at the nominal 5% signi level across a range of values for the parameters ep and cs (with 2 set equal to 1), As expected, the Wald test is more powerful lunder the local alternative of deterministic seasonality, whereas xachieves the highest power under pure stochastic seasonality being the LBI test for this case. For example, for ey = 2 and 5 = 0, the asymptotic power of the Wald test is 652, as op posed to .559 for the fest hased on a, In contrast, for cs 4 and p= 0, the power of the go testis 668, whereas that of the Walk test is 627, Finally, note that under pure stochastic seasonality, the power of the « test of Section 2 is considerably lower that of e.Journal of Business & Economic Statistics, July 2003 Table 6, Loca! Asymploe Power Against Determine anor Pure Stochastic Seasonality of he Fw ond Toes Ga0 op=5 qp=10 Gpa18 epa80 ep 25 o-30 op -35 quo F490 7001986 a0 72 6b21 ae of oa 0028 G im 7a wor seo Sean 7st ots or ta Sis is “ies tee tee "tao tas cae co-1 F984 1298 24524817 88a gee 98.75 (183 tae. 330 Seon Seen ae beet Seo coe ‘son “coo “60s 600 G08) “e008 8.08 o4se 2019 see S351 6058300 oes. 10 focr Stas Some | stes ste res rys eat jo tow fos 10s © toae=tose «tose toae teat 4774 5435 Bk T4134 OHTA Stas 508 seas eeue | rest wove rae ee its tats ters fonts fas Tos Toa sere e097 rai 7312 Tage 5.26 0.16 tees em voto © 700 on seo oa? ost tse wo moze tome oak asad 1479 7350 7752 060 adOT.-«76G 106.96 foro apes ome © ose © boas «oer fie fies fae ates fee ates ater ses9 e982 9027 0091.87 «BOT hts. S72 see soon seve «| akoe ake sz 87 S Ge eer! aet cet cone gatt cant gat so=10 F 9809 0521 0528 054005709805. 08.45—— 08.00 Sas Gros rosary avar ars avgk rr ee The local power of the modified test, using 4? rather than a? is, ay suggested in Remark 3, the same as that of w. However, in practice it may well have a higher power than w against de- terninistic seasonality. This is because when o =O, the probs ability limit of 0 exceeds that of 6”; indeed, because 0” > 6”, the modified statistic will always be greater than or equal to ‘There is a parallel with the test on jo in that using 0” instead of 6? would give the LM statistic. ‘A second modification is also in order. As formulated in (21), the fest is LBI against stochastic seasonality with yy = 0. In practice, we are more concerned with seasonal patterns that di- ‘minish over time. Thus our recommendation is to use the for- ward summation just as in the « test of Seetion 2, because this ‘would be the LBI test if the data were gencrated backward start ‘Taking these points into consideration, our is constructed using SIE] lyeng 7 ing with yr 22) When ¢; is serially correlates be modified as in Section 2.3. Ifthe spectrum is computed using the residuals after fitting the seasonal regressors, then the statistic is denoted by wt(m). The test can be extended to deal with both serial correlation and heteroscedasticity by making the amendment of 9) The Wald test can be carried out by fiting a roxil that is unrestricted except insofar as the seasonal component is taken the a test e to be nonstochastic; that i, 9 is vet to 0. Alternatively, a non- parametric es can be setup ising a nonparametric covariance ‘matrix estimator, as was done by Andrews (1991). This is es- sentially the same correction as in (9). To be specific, F(m) = TH1O-' QmO-' jn, ay where @=7~! C7, 2,2). If there is no need to guard against heteroscedasticity, the modifications are made simply using es- timates of the spectrum as for *(m). 4.2. A Diminishing Seasonal Pattern in Spanish Interest Rates. Asan example, we consider the logarithm of 3-momth money smarket interest rate in Spain for the period 1977Q1-20014; the source is the Bank of International Settlements macroeco- nomic series database. The series is depicted in Figure 2(a). It is difficult deteet a seasonal pattern from a casual glance at the graph, and one would not normally expect such a pattern to be present in an interest rate series; however, the Functioning ‘of the interbank loans market may imply some seasonality (see, eg. Hamilton 1996) Fitting the BSM to the series gives scusonal component, as shown in Figure 2(b); the slope variance is estimated to be 0, and the estimate of the (fixed) slope is small and insiguificunt, We have used logarithms of the data only because the diag nostics are better, ifthe raw series is used, then the resulting seasonal pattern is similar ‘The chi-squared statistic for the seasonals ut the end of the series is only .09, which is clearly not significant, because the 5% critical value for a x? is 7.81. However, the graph shows a fairly strong seasonal pattern until the mid-1980s, The question 1s whether the pattern as a whole isin any sense significant Setting the seasonal variance to 0 and reestimating the BSM. szivey a Wald statistic of 4.76, with a p value of .19, This is stillBusetti and Harvey: Seasonality Tests (a) 35 sa] 2s 20) Ms lw ©) os ies 431 = atte ate — Tren 2000 Figure 2. (a) Logarithm of he 3-Momh Spanish Interest Rate, 197701-200104 and (b) Estimate of the Seasonal Component. not significant, If the seties is differenced and a nonparametric ‘Wald test, (23), is computed using the Newey-West covariance ‘matrix estimator with three lags, then a similar p value (.17) is ‘oblained. On the other hand, the spectral nonparametric statisti ‘computed using forward summations takes the values 3.83 and 3.01 for m = 3 and 6, rising to 4.64 and 3.89 for w*(m), the preterred torm in which the spectrum is estimated after fitting seasonal regressors. Because the 5% critical value is 3,46, this test provides a firm rejection of the hypothesis that there is no seasonality in the series. Finally, for m = 3 and 6, the «(m) statistic of Section 2 takes, the values 1,17 and 1,02 (against a 5% critical value of 1.00). This confirms the presence of stochastic seasonality. 43. Seasonal Adjustment ‘The foregoing tests can be applied to a seasonally adjusted series to check whether the adjustment has been effective. This assumes that the adjustment has been done by means of mov= ing averages, rather than by regres It dummies have been used, then the «test statistics have the asymptotic distributions of Section 2 4.4 Detection of Trading-Day Effects Cleveland and Devlin (1980) showed that peaks at certain frequencies in the estimated spectra of monthly time series in- dicate the presence of truding-day effects, Specifically, there is a peak at a frequency of .348 x 2x radians, with the possibility of subsidiary peaks at 432 x 25 and .304 x 2or radians, An ‘option in the output of the X-12-ARIMA program provides a comparison of the estimates of these frequencies with the ad- Jacent frequencies (see Soukup and Findley 2000). However, there is no formal test. One possibility is to construct paramet- Fie OF nonparametric Statistics analogous to ay and w, 50 as to carry out tests for permanent eyclical effects at one or all of the three trading-day frequencies. Assuming that no (deterministic) trading-day model has been fitted, the asymptotic distributions Under the null will be CvMo, with a 5% critical value of 2.63 for a test at a single frequency and 5.68 for a test at all three frequencies. As an example, we took the irregular component, obtained from X12-ARIMA, of series sObS6ym, U.S. Retail Sales of Children’s, Family, and Miscellaneous Appare!, as supplied by the Bureau of the Census. Because the process followed by this, irregular component cannot be derived, we decided to use the nonparametric test. The test statistic with 10 lags was 7.03 for the single main frequency and 8.21 for all three frequencies. Both give a clear rejection of the null hypothesis that there is no trading-day effect 5. CONCLUSION The seasonality test statistic proposed by CH may be sim- plified so that a nonparametric correction for serial correlation is based on estimating the spectrum of the series at the rele- vant seasonal frequency or frequencies. This test statistic then has a very straightforward interpretation. As might be expected, Monte Carlo experiments show a slight gain in power over the original CH test for homoxcedastic series, but a size distortion and lower power when there is seasonal heteroscedasticity If a model is fitted, then a parametric seasonality test may be based on the innovations or smouthing errors, but Monte Carlo experiments show that they have similar properties. If the main reason for fiting a model isto investigate seasonality, then a basic structural time series model consisting of stochastic ‘tend, seasonal, and irregular components usually will be ade- quate. However, itis worth noting that the innovations test can be implemented for any structural time series model, including432 those that do not have a time invariant structure. Nonparamet- ric tests require a decision about lag truncation, but our Monte Carlo experiments show that in samples of size 200 the rejec- tion probabilities do not fall by very much when the lag length is increased from 4 to 8, Nonparametric tests are also dependent on decisions regarding differencing, but an important practical finding to emerge trom the Monte Carlo experiments is that for most economic time series, taking first differences is likely to be a good strategy. The Monte Carlo experiments indicate higher probabilities of rejection from parametric tests, but any assessment of power must be offset against the higher size. For quarterly data, there is a tendeney for parametric tests to be oversized at frequency zr, and this carries over to the joint test. The actual sizes of the joint tests at the 5% level of significance in a sample of size 200 are nearly 10%. On the other hand, with a lag length of 4, the actual size of the corresponding nonparametric tests never exceeded 5.5%. Parumettic tests are attractive within the context of a model-building exercise, but ifthe sole focus is on ‘esting for stochastic seasonality, then there is no overwhelming case for preferring them to nonparametric texts, If there are breaks in the seasonal pattern, then the season, ality test may be modified so as to have an asymptotic distri bution that is independent of the position of the break points under the null hypothesis that the seasonal pattern is determini tic, The U.K. marriages example yields much greater values for the parametric tests both with and without the seasonal ‘break dummy variables. In the modeled break case, the para metric text indicates a rejection of the null hypothesis, whereas the conclusions from the nonparametric tests are ambivalent. ‘Although a fixed seasonal component is normally estimated under the null hypothesis, there may be situations in which the researcher wishes to carry out a general test against a perma- nent seasonal component, regardless of whether itis determi istic and stochastic, We propose the use of test statisties that have the same form as the tests against nonstationary seasonal- ity except insofar as no fixed seasonal effects are removed when the residuals used (o construct the partial sums are formed. The asymptotic critical values are easily obtained. We compared the test with a Wald test carried out to determine the joint signifi- cance of a set of seasonal dummies asstimed to be constant over time. This test can also be carried out nonparamettically. When only deterministic seasonality is present, an analysis of local power shows the Wald test to be more powerful than our modi- fication of the test against nonstationary seasonality, but not by very much, In the not uncommon situation when the seasonal pattern is diminishing over time, the modified test against non- stationary seasonality is slightly more attractive. Indeed, in the ‘example of Spanish interest rates, this test shows a clear rejec~ tion of the null hypothesis of no permanent seasonality, whereas the Wald test does not reject. Tests against permanent seasonal effects can also be used to detect trading-day effects by exploit- ing the fact that these give rise to eycles at known frequencies, ‘An appealing feature of the proposed test statistics is that ‘under the null hypothesis, they all have asymptotic distributions belonging to the Cramér-von Mises family. Thus they provide an integrated approach to testing a wide range of hypotheses that arise in the context of seasonal time series, Journal of Business & Economic Statistics, July 2008 ACKNOWLEDGMENTS: We thank Robert Taylor and Bruce Hansen for helpful com- ‘ments on the issues raised in this article. David Findley kindly supplied the dats used in the section on trading-day effects. An carlier version of the article was presented at a conference on Seasonality at the University of the Algarve in October 2000. APPENDIX A: ASYMPTOTIC REPRESENTATIONS, FOR THE SPECTRAL NONPARAMETRIC TESTS Proof of Proposition 1 AI-A2, we have that under Ho, 72 Dr(i ) and T# (fy ~ yo) are Opi) and asymptotically orthogonal. {In particular, From assump TG» — mS NOG"), (Ady where G isan (51) x (1) diagonal matrix whose elements ane proportional to the spectral-zenerating function evaluate at the seasonal frequencies 2, )~ 1.....1s/2k; when sis even, G 1 1 1 tige( 5201). 520) 15802) whereas when sis odd, the last two diagonal elements of @ are both equal to 4g(Aj./2))- Note that asymptotic orthogonality is a direct consequence of assumption A2(b), whereas the result (A.1) follows mainly ‘rom the central limit theorem of Brillinger (1975, thn namely, for j= |. ! TAS eos a : Tr} Desinay (A2) whereas when s is even, 7? 22 eycos daj2t- NO, e(d/2)). In addition, the limiting random vectors of (A.2) are also inde- pendent across j. Furthermore, a functional central limit the rem also holds: that is, the partial sums of ¢,-c0s 2 and est yt weakly converge to independent Wiener processes (see Chun and Wei 1988), Now write the OLS jdduals as = 81 — X(B ~ B)— Ziv — yw) [8/2], consider the (normalized) partial sum tym. Let se re 10,‘Busett and Harvey: Seasonally Tests We then have that, under Ho, wm L. EV sper — TSO eaXDz!DrT (A - BY Ser) = wm TS Zi? ho i where the last expression is due t0 the orthogonality relation Fimpsac TO! OE sect =0 for J Ah. Then, using the functional central limit theorem of Chan and, Wei (1988) and the continuous mapping theorem, we have that under Hy, (aye) $8.17) > By, 10,1) (Aa) where Be(r) = Wels) — rWiCl) is a k-dimensional standard Brownian bridge, W)(r) is a k-dimensional standard Wiener process, ancl => denotes weak convergence: furthermore 5)-r(7) is asymptotically independent of Sp,r(r) for jh. As under Ho and by assumption A3, (4: an) 2 w(iy). it then follows by the continuous mapping theorem that the null limiting distrib- tutions of «j(m) when A = Aj. j= 1... [5/2] and wm) whea A=y-1 are CuM(a)) and CoM(s ~ 1). Under Hy: 42> 0, when A = A; itis easily seen that the partial sum S).7(r) is Op¢T4), whereas assumption A3 implies that @(2j; n) is Optm) (see Stock 1994, pp. 2797-2799), Thus both cj(m) and aotm) are Op T/m), Proot of Proposition 2 Consider the two subsamples (1...., [a7], oP] + 1 7), and Let $8) = [aT]! DHT yey, r€ (0, UJ, and Sr) Lie 10) a7} SF Sse 7 10, 1, be the partial sum processes inthe ist utd Secon subsamples. Using functional central limit theorem as in (A.3), we have that under Ho, Cag) 4 (Shp SPrhe)') > (BLO. BRU. re (01), where (7) and H2)(7) are independent k-dimensional stan- Soone=torr Sel sie()(5)) +la-arr! trace(!r($)s'-() ). sales 433 it then follows by the functional central limit theorem of Chan and Wei (1988), the continuous mapping theorem, and the addi- livity property of independent Cramér—von Mises random vari- ables that cf (ce; m) CyM(2aj) and a*(e: m) CoMQs 2) under Ho, Then, from the same arguments as in Proposi- tion 2, under Hy : 02 > 0, when A =A), both a (rm) and coe) are OT fan, APPENDIX 8: LBI TEST When ¢; in (1) is generalized so as to be a linear station- ary process, possibly consisting of more than one component, its 1 > T matrix covariance matrix will be denoted V = 22 Vy. where o isa variance parameter. If V, is known, then it follows from King and Hillier (1985) thatthe LBI testis ofthe form (5) with the OLS residuals replaced by the elements of the Tx 1 vector, V5'2, where @ is the 7 x T vector of generalized least squares residuals. The LBI test against stochastic seasonality at all frequencies is constructed similarly. Ie; contains & white noise component with variance 0? then itis straightforward to show that V, '2 is equal to the smoothed estimator of the vector of the white noise series. More generally, when multiplied by 7 iL becomes the 7 x | vector of smoothing errors, denoted by . The smoothing errors are produced as a byprod- uct ofthe smoother applied to the state-space forma af the model (ee de Jong and Penzer 1998; Hurvey und Streibel 1997) With V, known, an exact test eun be earriod out using nu merical inversion to construct critical values or probability val ues. However, V_ normally will depend on unknown parame- ters, 0 there are good reasons for wishing to use a statistic with 4 known asymptotic distribution. Ifthe test statistic is formed from smoothing errors. then it is necessary t0 take into. ac~ count their serial correlation, Following a argument similar to that used to give (7), the denominator needs an estimator of the spectral generating function of V—!e, This is equal 10 1/ge(. where g(a) isthe speetral-generating function of ¢, The para metric test statistic is, therefore, se) Faded where uj is the ith smoothing error. The test statistic has the same asymptotic distribution as (5), namely, CvM(2). This re- inains true when parameters in Vq are estimated (ef. Leybourne and MeCabe 1994; Saikkonen and Luukkonen 1993). ‘The foregoing correction can be made even if the model con- tains stochastic trend, The smoothing error series is stationary, and although it is not (strictly) invertible, the noninvertibility affects only the zero frequency, and the “quasi” Spectral Gen. erating Function (gf) of the nonscasonal part of the model can be inverted at 2). Thus for the special case of (11) in which the trend is a random walk, &-(2,) in (B.1) is replaced by &} +201 — eos? “|, o 0, Lemma C.1. (a) Under H¥ + 09 = 0,0, rot TED een f Witten aim fat re (0,1, (b) Under #99 =0,02 T'S Dex; Vwi 5 0. rt iat Proof of the Lemma To avoid unnecessary complications in the notations, we as- sume that 7; = T/s is an integer. Let” = (7—1)/s +1, where T= 1,....7. For part (i), first note that 1 LV Rw if isan integer ti ht YE B+ Rw= Vi al L Fut otherwise, sub Lo Note that we ean write, fort _[% iftejes-1 a=[Z ifj=0 where j = rmods. Then, using the decomposition above, we have that under the local alternative H° ,, for r¢ [0, 1] rh y 2am im Ly 2 1 — o ADE YS 2airhe + bu +o) * uclTyr] Piacoa 1 Laat By a standard functional central limit theorem, we obtain that ler Hy. YTB, + Ryu) + op. Trek TB SOMiv AIM, re (0,1) Journal of Business & Economic Statistics, July 2008 with Wi.s1(0) being an s ~ 1 dimensional standard Wiener process, whereas, forall j= 1,....8 ‘Thus, by an application of the continuous mapping theorem, for re(0.1}, r ' ry 26 «Seca f Wis1dr, wy iat where A= 5°01, 7:7; is the diagonal matrix defined in the statement ofthe proposition, For part (b), first note that Fl Then, proceeding similarly as before, we have that 7 : FP Das / cy rt te DP r XZ} (By + Ryu) +0p(1) (C2) z 1 VDF Xyada + op( Pr top(1), ‘where the second equality uses (C.1) and = ZR in ia 1,....T, is an independent sequence, hecause Lis made as a weighted sum of s nonoverlapping disturbances x, = ,..., 7. Thus part (b) of the lemma is proved by applying to (C.3) a law of large numbers for inde- pendent heteroscedastic sequences, (3) Proof of Proposition 3 Let yy be generated by the model (1)-(2),(16)-(18) with, 1-13 tnd let 6 and Jy be the OLS estimators of # and yo from regressing y; on (Xj, Z;)'; and let be the OLS estimator from regressing y; on X), Then (fzf) =a fu-y ” where r 1 TS DF XX Ds! TH! DF! xz a-| a tT So2xp;' Tag a a oi A Tr Spr'x (+23) b=Buse and Harvey: Seasonality Tests and 1 a= (1 Sp xg 43m ein) ‘The OLS residuals for the two regressions can be written as e642 ms — XB B)— Zyl} cay and 8) +7) Ki + Ziv — XB — BY 4 cs) By the orthogonality of the regressors Z, and X; of assump- tion AI, it can be straightforwardly shown that = f+ ep(). Furthermore, by the functional central limit theorem of Chan and Wei (1988), we have the following weak convergence re- sult oN IT? SS 25 Wo l=) i where Wo,.-1(r) and A ure defined in the statement of the proposition. Note that rel01), (C6) r ae Taste Consider the local alternative HP; : yo = ent/ VT, 0; = 0. We have that f= B+ op(1), UAT jo Wo) + enor A he 'p.-03 ep), and ¢; = € + op(1). Then F % Vp..-10;cp)'Vns-1 (0:0): note that because Wo, (1) isa standard s— I dimensional nor- ‘mal, the asyraptotic distribution of F is a noncentral chi-squared distribution with s— 1 degrees of freedom und noncentrality pa rameter equal to cho, °V At. Furthermore, from (€.6) and (C.5), r TED 2, > Wor 1 A eo 1AT — ne ara -1 A Vos-tep), re {0,1}, and because ¢ =F ap(1.02% fl ps1: €D)’Vp.s-1x (rep) dr. Under the fixed alternative H?, using similar arguments itis, not difficult to show that yp is Op(1), 6? and a? are O,(1 LDL, Ze, is Op(7). Thus both F and w are p(T). Consider the local alternative A ,. By (C.6) and using part (a) of the lemma, we have that and 1 d - MTHS Wo. wn tesatad f Wistar b = Vs.s-105 ¢5). 495 As under Hf pe, =; + op(1), it easily follows that F Vix. (05 €8)'¥5,s-1(0: 5) Using similar arguments, we also have that r ate! 2 Ze, = Wo, “th Wie adds = Vsoles). re 101) Tye Bad Thus way oe) 4 f Vs.s-1075 65) V5.9-1drs 5) dr. lo Finally, under the fixed alternative H}, from part (b) of the Jemma, we first obtain that both A — 6 and BB are op( 1). ‘Then, as Ze, is Op(F 4) and jy is Op(T4), and 6? and are Op(1). it easily Follows that both the Wald statistic F and the statistic w are Op(7).. Received December 2000, Revised November 2002, REFERENCES Andrews, D.W.K. (1991), Heteroscedasticty and Autocorrelation Con Covariance Matrix Estimation” Econometrica, $9, 817-858 Bai, J. (1997), "Estimation ofa Change Poi in Molliple Regtesion Models” Review of Economics and Statistics, 551-563 Biillinges, D.R. (1978), Time Series” Date Analysis and Theory, New York ol, Rinehart and Winston Bust, F, and Harvey, A.C. (2001), “Testing forthe Presence ofa Random ‘Walkin Series Wih Structural Breaks." Journal of Time Series Analysis, 2. 17-156. (2003), “Further Comments on tationaity Tessin Series With Struc tural Breaks at Unknown Poin.” Joumal of Time Series Analysis, 24, 137 140, use, Fad Taylor, A.M. , (2008), "Testing Apainst Stochastic Teen aad ‘Seasonality in dhe Presence of Unancaded Breaks and Unit Rots” Joaymal of Econometrics, forbcoming. Caner, M. (1998), “A Locally Optimal Seasonal Unit Root Test” Journal of ‘Business ane! Economie Stites 16, 349-386 Canova, F and Gysls (1994), “Changes in Seasonal Paters: Are They Cyclical?" Jounal of Economic Dynamies and Conte, 18, 48-117] Canova, F, and Hansen, BE. (1998), "Are Seasonal Pallems Constant Over Time? A Tes for Seasonal Stability” Journal af Business and Economic St istics, 2200-9, (Chan, NH and Wei, CZ. (1988), “Limiting Distebutions of Least Squares Fstimates of Unsable Autoregressive Processes” The Annals of Statics 16, 367-40 Cleveland, 5, and Devlin, $1. 1980), "Calendar Effects in Monthy Time Serie: Detection by Spectrum and Graphical Merb” Journal ofthe Amer. ‘ean Statistical Association, 75, 187-496, de Jong, P, and Penver (1998), "Diagnosing Shacks nTime Series" Journal ofthe American Suistcat Assocation, 93, 790-806, Hamition, 1D (1996), “The Daily Market for Federal Funds” Jounal of Po Iie Econemy, 104, 26-36 lavey, A. C. (1989), Forecasting, Sirctural Time Series Meaels andthe ‘Kalman Filter, Cambridge, UK. Cambridge University Pres (2001), “Testing in Unabserved! Components Models.” Journal of Forecasting. 2.1 9.496 Harvey, A.C. and Strebel, M (1997), “Testing for Nonsationary Unobserved “Components.” mimeo, King, M.L.. and Hiller, G. 1. (1985), “Locally Best Invariant Tests of the rior Covariance Matix it the Linear Regression Move,” Journal of the Royal Stance Soctey, Sot. 47, 98-102 Kitagawa. G. and Gersch W, (1996), Smoothness Priore Analysis of Time Se: ‘ies, Beli: Springer Ver Koopman, $.J, Harvey, A.C.,Deomik, JA. and Shephart, N. (2000), AMP “60 Structural Tine Series Anotssis Modeller and Predictor, London: Tim herlake Consultants Koopman, 3 Shephard, N. ann! Doon, J. (1999), “Satiseal Algerithns for Moxels in Sine Space Using SsPack 32." Econometrics Joumal,2.113- 166. Kiathowsh, D Philips, PC. B, Schini, P. and Shin, Y, (1992), “Testing the Null Hypothesis ot Staionurty Aguinst the Alternative of a Unit ROO: How Sure Are We Thae Beonomic Time Series Have & Unit Rost? Jounal of Econemetnies, 4 159-178 Leybourme, §. J, and McCabe, BPM. (1994), "A Consistent Test for a Unit Root” Journal of Business and Economic Statistics, 12, 157-166. [Nyblom, 1. (1989), “Testing forthe Constancy of Parameters Over Tim" Jnr ‘nal of the American Statistical Assocation, 84, 222-230 Nyblom, J, and Harvey, A.C. (2000), “Tess of Common Stochastic Tends,” ‘Econometric Theor. 1, 176-198. Journal of Business & Economic Statistics, July 2008 Pirce, D_A. (1978), "Seasonal Adjustment When Roth Deterministic and Sto chaste Seasonality Are Present," in Seasonal Anassis of Economie Time Series, ed, A, Zelines, Washington, DC: Bureau of the Census, pp. 242- 268, Proiet T2000), “Compariag Seasonal Components for Structaral Time Se= ‘es Models" temationel Journal of Forecasting. 16, 247-260, Saikkonen, P, and Luukkonca, R, (199), “Testing fora Moning Average Unit Root in Autoregressive Interated Moving Average Miles onal of the American Sisal Assocation, 88, 590-01 South, RJ, and Taylor, A-M. R. (1998), “Additional Chtical Values and AS- ‘ypu Representations for Seasonal Unit Root Tess." Journal of Beano Invtres, 85, 269-288 Soukup. KJ. und Findley D.F.(2000), “On the Spectrum Diagnosties Usod by X-I2-ARIMA to Indicae the Presence of Trading-Day Etfects Afr Model- ing or Adjustrent.”in Proceedings of He Business and Economic Stites Section, American Stisteu! Assocation, pp. [44-149 Stook,J-H. (1994), "Unit Roots, Structural Breaks and Trends in Handbook of earomerries 4 cds. R. F Engle and D. L, MeFadden, Amsterdam: Elsevier Science, pp, 2739-280. ‘Taylor, AMR (20034), "Locally Optional Tests Agalnst Seasonal Unit Rvs, urna of Time Series Anahsis ontcoming. 20035), “Robust Stationarity Tests in Seasonal Time Series Provesses" Jour of Business and Fonomic Statsties, fortes