Measuring the Impact of Data Anomalies on Tourism Demand Forecasts

Time series models have proven to be powerful tools for forecasting tourism demand. However, the recent Covid-19 outbreak has severely impacted the tourism industry, and models which were previously able to provide accurate forecasts may no longer be viable. This work aims to further analyse this situation by measuring the impact of data anomalies caused by the Covid-19 pandemic on the forecasting performance of different time series models. For this purpose, the monthly number of tourist Overnight Stays per region in Portugal from 2000 to 2022 is used and forecasting competitions are performed on three selected time series. These forecasting competitions contain various approaches, from simple methods to different variants of Autoregressive Integrated Moving Average and Exponential Smoothing models. The forecasting performance of the models is assessed firstly by excluding the Covid-19 pandemic from the time series and secondly by including this period. In addition, a logarithmic transformation of the forecast variable is performed as well as different types of cross-validation approaches are used. The results reveal that Autoregressive Integrated Moving Average and Exponential Smoothing models showed superior performance before the Covid-19 outbreak, but a significant loss of performance in the months thereafter. In contrast, the Naïve method produced comparatively good forecasts during these months due to its simplicity. Moreover, the Drift model applied to seasonally adjusted data was able to compete with the best models and displayed a lower deterioration in prediction accuracy following the Covid-19 outbreak. Besides, this work provides evidence that regions with a higher percentage of Portuguese tourists displayed lower declines in tourism demand during the Covid-19 period.