Does the inclusion of climate variables improve tourism demand forecasting performance?

Abstract

The aim of this study is to assess whether incorporating climate variables in econometric and combination forecasting models can improve tourism demand forecasting performance. Climate conditions are important tourism resources which can influence tourists’ decision as to when and where to travel, however, our understanding of the value of climate variables in forecasting tourism demand is limited. The current research fills this gap through empirical studies on UK’s international tourism demand. Inbound tourism demand to the UK from seven leading markets, namely, France, Germany, Irish Republic, Italy, the Netherlands, Spain and the US are studied respectively based on quarterly time series data from 1994Q1 to 2017Q4. The bounds test cointegration approach is applied to assess the long-run relationships between tourism demand and its influencing factors and to evaluate the impact of climate on tourism demand. Individual tourism demand forecasts are generated through both causal econometric and non- causal time series models, which are popular in the current tourism demand literature. Causal econometric models, which consist of the bounds test cointegration approach, the autoregressive distributed lag model (ADLM), the leading indicator (LI) model, the vector autoregressive (VAR) model, the time-varying parameter (TVP) model and the simple dynamic (SD) model, take two model specifications, which are different in identified influencing factors. Econometric models that only consider economic factors as demand determinants are named as traditional econometric models, and the others that include the climate factor as a demand determinant are called climate econometric models. Non-causal time series techniques consist of the seasonal naïve no-change model, the seasonal autoregressive integrated moving average (SARIMA) model, the exponential smoothing (ETS) model and the state space ETS model. One- to four-step-ahead out-of-sample single forecasts are generated from every individual forecasting model through the recursive forecasting procedure with the seasonal naïve no-change model serving as the benchmark. Except the naïve model, all other individual forecasting models are selected as candidate constituents for combination. For combination forecasting, the 15 selected individual models are categorized into three groups. The first group includes all individual models; the second one contains traditional econometric and time series models; and the third category consists of climate econometric and time series models. Combination is conducted for each group respectively, resulting in three sets of combination forecasts: the first set is generated through combining 15 individual models; the second and third ones are produced from integrating 9 individual models. Different combination methods are applied including the simple average (SA) method, the variance-covariance (VACO) method, the discounted mean square forecast error (DMSFE) (α = 0.85/0.90/0.95 ) methods, as well as the newly-introduced inverse-MAE and the two-stage combination approaches. Comprehensive comparisons of the predictive powers of the individual and combination forecasting approaches for seven origins and four forecasting horizons are conducted based on three accuracy measures including mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE). The results show that individual model’s forecasting performance varies greatly according to the origin market under consideration. No single model can perform the best in all cases, and in most cases, more advanced individual models forecast better than the naïve benchmark. In general, non-causal time series techniques are superior to causal econometric models. Whether including the climate factor can improve the forecasting accuracy of econometric models should be evaluated case-by-case. With respect to the forecasting ability of the combination approach, it is demonstrated that combining individual forecasts is beneficial regardless of origin country under study, forecasting horizon under consideration, accuracy measure used, combination methods applied or combination group under analysis. In all cases, there are always a portion of combination forecasts that are more accurate than the best single projections, and the worst forecasts are always produced by individual forecasting models. It means that the combination forecasting approach is superior to the individual one, as it can improve forecasting accuracy and reduce forecasting failure. Comparisons among alternative combination methods show that no single combination method can provide the best composite forecasts in all situations. The newly-introduced inverse-MAE scheme performs quite well, but the two-stage combination methods behave unsatisfactorily. Comparisons among three combination groups reveal that, generally, combining all individual models, which include traditional econometric models, climate econometric models and time series techniques produce the best combination forecasts, which means that combining econometric models with different influencing factors and introducing climate variables into combination can contribute to more accurate projections. It implies that through combining, diversity gain can be achieved not only by incorporating different modelling techniques but also by integrating different model specifications. Regarding which and how many models to combine, it is shown that individual models’ frequencies to constitute the superior combination forecasts are irrelevant to their forecasting abilities. More accurate individual forecasts do not have higher opportunities to construct superior composite projections. The number of single constituents in the best forecasts range from two to six, and for most origins, combining two individual models can bring about the most accurate projections. To the best of my knowledge, this research represents the first effort to evaluate the combination forecasting approach which consider econometric models with different explanatory variables as candidate constituents, and climate variables have been, for the first time, introduced to the combination forecasts. It proves that better combination forecasts can be obtained by integrating econometric models with different influencing factors, and the value of non-economic explanatory variables in combination forecasting deserves more attention. It is suggested that a user-friendly software for combination forecasting should be made available and combination forecasts should be included in forecasting comparisons considering the general superiority of the combination forecasting approach compared to the single forecasting method.