Testing the adequacy of regime-switching time series models based on aggregation operators Jozef Komornı́k1 , Magda Komornı́ková2 , and Danuša Szökeová2 1 2 Faculty of Management, Comenius University, Bratislava, Slovakia, Jozef.Komornik@fm.uniba.sk Faculty of Civil Engineering, Slovak University of Technology, Bratislava, Slovakia, (magda, szoke)@math.sk Summary. The aim of this paper is to enrich modeling procedures for aggregation operators based regime-switching time series models (that has been introduced and applied in [Bog05] and [KK05]). Namely a test of linearity (against two-regimes alternative) will be adopted to the specific form of the transition functions constructed via compositions of aggregation operators with a standard logistic shape function. An application in model building for hydrological data from Slovak rivers demonstrate promising perspectives of applicability of testing and fitting procedures developed in this article. Key words: Time series models, aggregation operators, regime-switching time series model, logistic shape function. 1 Introduction A new method of construction of regime-switching models based on utilization of aggregation operators has been recently indicated in [Bog05], [KK05]. The goal of this paper is to provide a test of linearity (against two-regimes alternative) which will be adopted to the specific form of the transition functions constructed via compositions of aggregation operator with a standard logistic shape function and to explore its usability in practical modeling (using the time series of Slovak river stream flow) 2 Theoretical background 2.1 Aggregation operators Aggregation operators and their basic property have been treated in [CMM02]. Typical continuous aggregation operators on the real line (the map Rn onto R) are - Arithmetic mean M (x1 , . . . , xn ) = 1 n n P i=1 xi ; 1188 J.Komornı́k, M.Komornı́ková, D.Szökeová - Weighted means W (x1 , . . . , xn ) = n P i=1 does not need to be symmetric); - OWA operators W ′ (x1 , . . . , xn ) = n P n P wi xi , where wi ∈ [0, 1], wi = 1 (it i=1 wi x′i , where x′i is a non-decreasing per- i=1 mutation of x1 , . . . , xn , i.e., x′1 ≤ x′2 ≤ . . . ≤ x′n . In class of OWA operators we can find also M IN (eventually M AX ) operators, corresponding to extremal cases w1 = 1 and wi = 0 otherwise, (eventually wn = 1 and wi = 0 otherwise) and all order statistics. Similarly projection to kary coordinate with wk = 1 and wi = 0 otherwise is a special weighted mean. A convenient way of producing a decreasing sequence (w1 , . . . , wn ) of generating weight coefficients is based on an increasing convex bijection ϕ of [0, 1], when we put wi = ϕ n−k+1 − ϕ n−k , k = 1, . . . , n. More about aggregation operators can n n be found, for example, in [CMM02]. 2.2 Model building using aggregation operators Outputs qt = A(yt−1 , . . . , yt−d ) of suitable aggregation operators will play the role of threshold variables for such regime-switching models. If we apply a so-called transition function on such threshold variable, we get a usual form of a regimeswitching model yt = Φ1 (B)yt (1 − F (qt )) + Φ2 (B)yt F (qt ) + εt (1) where Φ1 (B), Φ2 (B) are autoregressive polynomials of orders p1 , p2 in the shift operator B; A: Rd → R is a continuous aggregation operator for an appropriately chosen time delay d > 0; εt ’s are assumed to be a martingale difference sequence with respect to the history of the time series up to time t − 1, which is denoted as Ωt−1 = {yt−1 , yt−2 , . . . , yt−(p−1) , yt−p }, p = max(p1 , p2 ), that is, E[εt |Ωt−1 ] = 0. We also assume that the conditional variance of εt is constant, E[ε2t |Ωt−1 ] = σ 2 ; F is a so-called transition function, i.e., a non-decreasing surjective map of the values of a threshold variable qt to the interval [0, 1]: F (qt ) = h(γ(qt − c)) + 1/2, (2) where h: [−∞, ∞] → [−1/2, 1/2] is a non-decreasing shape function (an odd bijection); c = F −1 (1/2) is the threshold constant; γ is the smoothness parameter, that has a critical value γ = 0 that yields F (c) = (1/2) implying no difference between 2 regimes and thus linearity of the model (1). We will use the standard logistic shape function (shifted to h(0) = 0) h(x) = 1 1 − . 1 + e−x 2 Testing the adequacy of regime-switching models based on agops 1189 2.3 Testing for regime-switching nonlinearity For testing the linearity of (1) (or H0 : γ = 0) we adopt the approach from [FD00] to the specific form of the transition functions constructed via compositions of aggregation operator with a standard logistic shape function F ∗ (qt ) = F (qt ) − 1/2 = h(γ(qt − c)). In the reparametrized model equation the linearity can be tested by means of a Lagrange Multiplier [LM] statistic with a standard asymptotic χ2 -distribution under the null hypothesis. We denote p = max(p1 , p2 ), xt = (1, x̃′t )′ with x̃t = (yt−1 , . . . , yt−p )′ and φi = (φi,0 , φi,1 , . . . , φi,p )′ , i = 1, 2. Then we rewrite the model (1) as yt = φ′1 xt [1/2 − F ∗ (qt )] + φ′2 xt [1/2 + F ∗ (qt )] + εt or yt = 1/2(φ1 + φ2 )′ xt + (φ2 − φ1 )′ xt F ∗ (qt ) + εt = = 1/2(φ1 + φ2 )′ xt + (φ2 − φ1 )′ xt h(γ(qt − c)) + εt . In order to derive a linearity test against (1), we approximate the shape function F ∗ (qt ) with a third-order Taylor approximation around γ = 0. This results in the auxiliary regression yt = α′0 + β0′ xt + β1′ xt qt + β2′ xt qt2 + β3′ xt qt3 + et (3) where qt = A(yt−1 , . . . , yt−d ), α0 and βi = (βi,0 , βi,1 , . . . , βi,p )′ , i = 0, 1, 2, 3 are functions of the parameters φ1 , φ2 , γ, c and et = εt + (φ2 − φ1 )′ xt R3 (qt ) with R3 (qt ) the remainder term from the Taylor approximation. Under the linearity hypothesis, R3 (qt ) ≡ 0 and et = εt (see e. g. [FD00]). Consequently, this remainder term does not affect the properties of the errors under the null hypothesis and, hence, the asymptotic distribution properties. Inspection of the exact relationships shows that the null hypothesis H′0 : γ = 0 corresponds to H′′0 : β1 = β2 = β3 = 0, which can be tested by a standard LM-type test. Under the null hypothesis of linearity, the test statistic, to be denoted as LM3 , has an asymptotic χ2 distribution with 3(p + 1) degrees of freedom. The LM3 statistic based on (3) can be computed as follows: 1. Estimate the model under the null hypothesis of linearity by regressing yt on xt qti . Compute the residuals ε̂t and the sum of squared residuals SSR0 = n P ε̂2t . t=1 2. Estimate the auxiliary regression of yt on xt and xt qt , i = 1, 2, 3. Compute the residuals êt and the sum of squared residuals SSR1 = n P ê2t . t=1 3. The LM3 statistic can be computed as LM3 = n(SSR0 − SSR1 ) . SSR0 (4) 1190 J.Komornı́k, M.Komornı́ková, D.Szökeová 3 Modeling time series of monthly average Slovak rivers stream flow In illustrative examples discussed in this article are the 12 univariate time series of monthly average Slovak rivers stream flow in the period from January 1973 to December 2003. The data used for testing regime-switching nonlinearity with aggregation operators are residuals obtained after removing annual periodic components. Appropriate orders p of linear models AR(p) for considered time series have been determined by minimizing AIC and BIC information criterion (see, e.g. [FD00]). The number of variables that enter any aggregation operator (for individual time series models) is h = k − 1, where k is the first value of delay for which the values of the autocorrelation function are not significantly different from 0. In the role of aggregation operators we used Arithmetic Mean (M ), Weighted average with the generating functions x2 (W 2 ) and x3 (W 3 ) and OWA operators M IN and M AX . Usual LSTAR models with threshold variables yt−d can be considered as products of a special trivial type of aggregation operators (that map the sequence (yt−1 , . . . , yt−h ) on its single components). The following Table 1 shows the results of testing of adequacy of 2-regime models for 12 univariate time series of monthly average stream flow [m3 /s] from observation stations at Slovak rivers. The meaning of parameters k, p and d has been explained above. For each river, items in the last 7 columns of its row represent p-values for LM3 tests given by (4). The first 2 model classes are standard LSTAR models (without an explicit aggregation operator). The next 5 model classes correspond to LSTAR models with the threshold variable qt as the output from the aggregation operators M AX , M IN , M , W 2 and W 3 . Table 1. The results of testing of adequacy of 2-regime models. River k p p-value for LM3 test for LSTAR with aggregation operator for standard LSTAR with d=1 d=2 M AX M IN M Vlkyňa 4 2 0.0554 0.0018 4 1 0.0002 x Štı́tnik Boca 5 2 0.0060 0.1246 Ipel’ 3 1 0.0073 x Kysuca1 2 2 0.4746 0.9972 Kysuca2 2 1 0.1413 x Litava 15 2 0.2190 0.0005 Bebrava 4 2 0.0006 0.9173 Dobšiná 15 2 0.0013 0.2494 Krupinica 15 2 0.0555 9.4981 Hron1 5 2 0.0011 0.9913 Hron2 6 1 0.0014 x 0.0001 0.0033 0.0039 0.0239 0.7756 0.1613 0.1303 0.0065 0.0056 0.0004 0.0019 0.0079 0.0119 0.4339 0.2048 0.4751 0.4144 0.6591 0.0261 0.4222 0.0059 0.6796 0.2586 0.6757 0.0005 0.0452 0.0535 0.0494 0.8558 0.3093 0.0007 0.0642 0.0004 0.0015 0.0011 0.4191 W2 W3 0.0004 0.0013 0.0024 0.0007 0.0046 0.0019 0.0133 0.0071 0.5313 0.5168 0.1729 0.1532 0.0004 0.0067 0.0366 0.0330 0.0001 0.00001 0.0051 0.0098 0.0002 0.0001 0.0346 0.0102 From results in Table 1 we see that in case of data from Kysuca1 and Kysuca2 the p-value greatly exceed standard threshold 0.05, which implies that 1-regime models Testing the adequacy of regime-switching models based on agops 1191 are superior to 2-regimes alternatives. In case of data from Štitnik, Bebrava and Hron2 the minimal p-values for LM3 test are attained for standard LSTAR models. For data from other observation stations the minimum p-values for LM3 test are attained for aggregation operators M AX (in case of Vlkyňa and Krupinica), W 3 (in case of Boca, Ipel’, Hron1, Dobšiná) and W 2 (in case of Litava). The operators M IN and M do not provide minimum p-value for any of the investigated river station data. The next Table 2 shows an overview of the models characteristics for 10 observation stations, where the adequacy of regime-switching models has been demonstrated by the previous tests. Based on computed p-value, each river flow stream was modeled by two classes of models, namely by the standard LSTAR with threshold variable yt−d (with delay d determined by the smaller p-value from columns 4 and 5) and by the LSTAR model with threshold variable qt = A(yt−1 , . . . , yt−h ), where the type of aggregation operator A was determined by the minimal p-value from columns 6 – 10. In each class we have estimated the parameters for models with p1 , p2 ≤ 3. The best model of the class is determined by information criteria AIC and sBIC modified for regime-switching models (see [LCh03]). The criteria RM SE = 1 m m P (yi − Fi )2 and M AE = i=1 1 m m P |yi − Fi | (where m is the number i=1 of predictions, yi is an observed value and Fi a prediction, i = 1, . . . , m) are measures of out-of-sample fitting (for the last year data left out from the model building sample and used for a subsequent testing, see [FD00]). Table 2. The measures of out-of-sample fitting for 12 predictions. River Vlkyňa Štı́tnik Ipel’ Boca Hron1 Dobšiná Litava Bebrava Krupinica Hron2 Standard LSTAR (p1 , p2 , d) RM SE M AE (2,2,2) (1,1,1) (2,1,1) (2,2,1) (1,2,1) (2,2,1) (2,1,2) (2,1,1) (2,2,1) (1,2,1) 5.068 0.695 1.089 0.960 18.501 0.523 0.473 0.405 0.685 24.08 3.424 0.637 0.985 0.771 15.598 0.457 0.428 0.332 0.533 21.33 Aggregation operator Type RM SE M AE M AX W3 W3 W3 W3 W3 W2 M AX M AX M AX 4.835 0.654 1.093 0.954 18.184 0.438 0.419 0.541 0.590 23.67 3.178 0.496 0.917 0.729 15.704 0.358 0.323 0.452 0.532 21.23 The previous results demonstrate that regime-switching models based on aggregation operators provide clearly better fit in 8 of 10 cases and comparable fit in 2 cases. 1192 J.Komornı́k, M.Komornı́ková, D.Szökeová 4 Conclusion The above analysis presents an enriched procedure for regime-switching time series models based on exploiting of aggregation operators and their adequacy testing. The outcomes of modeling hydrological data yield very promising results. Acknowledgement The research summarized in this paper was partly supported by the Grants APVT20-003204, VEGA 1/2032/05 and VEGA 1/1145/04. References [Bog05] Bognár, T.: Time series analysis applied in geodesy and geodynamics. PhD thesis, Bratislava (2005) [BKK04] Bognár, T., Komornı́ková, M., Komornı́k, J.: New STAR models of time series and their application in finance. In: Proc. 16th Symposium COMPSTAT 2004, Prague. Physica-Verlag, A Springer Company, 713 – 720 (2004) [CMM02] Calvo, T., Mesiar, R., Mayor, G. (Eds.): Aggregation Operators. Studies in Fuzziness and Soft Computing, 31, Physica-Verlag, Heidelberg (2002) [FD00] Franses, P.H., Dijk, D.: Non-linear time series models in empirical finance. Cambridge University Press (2000) [KK05] Komornı́ková, M., Komornı́k, J.: Application of Aggregation Operators in Regime-Switching Models for Exchange Rates. In: Proc. EUSFLAT-LFA 2005, 1297 – 1300, Barcelona (2005) [LCh03] Liew, V.K., Chong, T.T.: Effect of STAR and TAR types nonlinearities on order selection criteria, http://econwpa.wustl.edu:80/eps/em/papers/0307/ 0307005.pdf (2003)