Loganathan Nanthakumar, Thirunaukarasu Subramaniam & Mori Kogid376Table 2 Regression Results and Diagnostic Tests for SARIMAModelsModelsSARIMA(1,1,1)Season-ISeason-IISeason-IIIAR(1)MA(1)AIC valueSARIMA(1,1,2)Season-ISeason-IISeason-IIIAR(1)MA(1)MA(2)AIC valueARIMA(2,1,2)Season-ISeason-IISeason-IIIAR(1)AR(2)MA(1)MA(2)AIC valueCoefficient0.08 (0.24)-0.03 (0.64)0.08 (0.21)0.74 (0.00)*-0.99 (0.00)*-0.680.09 (0.21)-0.02 (0.68)0.09 (0.22)0.81 (0.00)*-1.13 (0.00)*0.14 (0.45)-0.640.08 (0.39)-0.02 (0.68)0.08 (0.40)-0.14 (0.56)0.56 (0.00)*-0.24 (0.38)-0.74 (0.01)*-0.68ARCH-LMTest (a)3.91(0.05)2.21(0.14)1.11(0.29)H o : No serial H o :correlation (b) NormalityTest (c)0.93(0.40)1.12(0.33)1.48(0.24)23.63(0.00)*31.99(0.00)*42.14(0.00)*Note: (a), (b) and (c) indicates Autoregressive conditional heteroscedasticity (ARCH) LMtest, Breusch-Godfrey (BG) serial correlation test and Jarque-Bera (JB) normalitytest. Figures in ( ) indicates probability values. Aster<strong>is</strong>ks (*) denote stat<strong>is</strong>ticallysignificant at 1% significance levels. Season-1 (April-June), Season-2 (July-September), Season-3 (October-December)The flow of ACF and PACF shows the autoregressive effects withfirst difference of the h<strong>is</strong>torical data. Table 2 indicates AIC values and thedec<strong>is</strong>ion to select the most suitable model <strong>is</strong> by comparing the value ofAIC to the ARIMA models used in th<strong>is</strong> study. The smaller the value ofAIC, the better <strong>is</strong> the fit in the ARIMA model used. ThereforeARIMA(2,1,2) <strong>is</strong> relevant because the value of AIC <strong>is</strong> smaller comparedto ARIMA models. Besides that, diagnostics tests are applied in th<strong>is</strong> studyto determine whether the estimated models deviate <strong>from</strong> the assumptions
TOURISMOS: AN INTERNATIONAL MULTIDISCIPLINARY JOURNAL OF TOURISMVolume 7, Number 1, Spring-Summer 2012, pp. 367-381UDC: 338.48+640(050)of standard linear regression model. Further, we tested for autoregressiveconditional heteroscedasticity (ARCH), serial correlation using Breusch-Godfrey (BG) test; and normality using Jarque-Bera (JB) test. Ascorrelogram of squared residuals <strong>from</strong> ARMA(2,2) shows autocorrelationpattern in squared residuals which could be attributed to volatilityclustering, to test the presence of ARCH effect, we compute ARCH-LMtest.The results in Table 2 do not indicate any ARCH effects in all modelsestimated. The Breusch-Godfrey (BG) test of serial correlation indicatesthat serial correlation hypothes<strong>is</strong> cannot be rejected in all three SARIMAmodels. Meanwhile, normality test using Jarque-Bera (JB) indicates thatnormality in the errors <strong>is</strong> rejected in all SARIMA models. Th<strong>is</strong> indicatesthat all three models are not normally d<strong>is</strong>tributed because of someseasonal effects. One interesting results that can be derived <strong>from</strong> Table 2<strong>is</strong> that, although seasonal effects have been taken into account in everySARIMA models, the results does not d<strong>is</strong>play any significant seasonaldummy effects, either positively or negatively.To choose suitable SARIMA model for th<strong>is</strong> study, we use theinequality coefficients technique. The inequality coefficient ofSARIMA(2,1,2) model are marginally smaller than those of theSARIMA(1,1,1) and SARIMA(1,1,2). Therefore ARIMA(2,1,2) <strong>is</strong> thebest SARIMA model for th<strong>is</strong> study. For one-step-ahead <strong>forecasting</strong> which<strong>is</strong> applied in th<strong>is</strong> study, the MA(2) model <strong>is</strong> found to be the best becausethe values of RMSE and MAPE are the lowest as in Table 3 below.Table 3 Summary of Forecast Evolutions of SARIMA ModelsSARIMA ModelsInequality Coefficient SARIMA(1,1,1)SARIMA(1,1,2)SARIMA(2,1,2)RMSE 0.17 0.17 0.16MAE 0.13 0.13 0.12MAPE 302.39 311.97 292.88Theil Coefficient 0.71 0.70 0.68It <strong>is</strong> clear that the dummy variables which represent seasonalityeffects do not give any clear evidence for ASEAN tour<strong>is</strong>t arrivals toMalaysia. Therefore, in order to forecast ASEAN tour<strong>is</strong>t arrivals usingone-period-ahead approach, th<strong>is</strong> study used only ARIMA(2,1,2) modelwithout any seasonal effect to identify the autoregressive (AR) andmoving average (MA) effects. The ARIMA(2,1,2) model that had been377