Use of Change Point Analysis in Seasonal ARIMA Models for Forecasting Tourist Arrivals in Sri Lanka

Sri Lanka is a popular place that attracts foreign travelers, and the impact of the tourism industry has a major contribution to the Sri Lankan economy. The main objective of this study is to model the behavior and forecast tourist arrivals in Sri Lanka through a time-series approach with Change Point Analysis (CPA). Autoregressive Integrated Moving Average (ARIMA) was extended to Seasonal Autoregressive Integrated Moving Average (SARIMA) with the seasonality behavior of the tourist arrivals. The better performed models were identified using the minimum Akaike Information Criterion (AIC) while performance indicators of Mean Absolute Percentage Error (MAPE) and Normalized Root Mean Squared Error (NRMSE) were applied to evaluate the actual and fitted values. The model diagnostics were used to assess the goodness of fit of a selected model. Monthly data from January 2000 to December 2019 was used in the analysis and during this period a total of 20,217,026 tourists arrived in Sri Lanka. Moreover, there are certain decline periods of this volume mainly due to the impacts of civil war, Tsunami and many others. The findings indicate that the model ARIMA (2,1,2) (3,1,4)[3] captures the behavior well with a minimum MAPE of 0.1941 and NRMSE of 0.8800. Meanwhile, with the application of CPA (at most one change and pruned exact linear time), data was split into two separate windows, which are Window 1 (W1) from January 2000 to October 2011 and Window 2 (W2) from November 2011 to December 2019. In W1, the better model that was used in the prediction was ARIMA (1,1,1) (4,1,1)[3] with a MAPE and NRMSE of 0.1727 and 1.1190 respectively. According to the results, the better performed model (MAPE of 0.2740 and NRMSE of 0.8700) in W2, was ARIMA (0,1,1) (3,1,3)[3] and this model captured the behavior until April 2019. However, due to the Easter bomb attack in April, there was a sudden drop in the arrival of tourists in May and June 2019. Nevertheless, from this point onwards the predicted line captured the behavior of the actual values even though they did not coincide with each other. Again, in December 2019, the predicted and actual values were very close. Thus, this study will be a benefit for both the private and public sectors as it has a prominent impact on the economy of the country.