Comparative Forecasting of International Air Passenger Demand: Exponential Smoothing vs. ARIMA Models

Airlines and airports use estimates of the future volume of international flight activity to plan their capacity, prices, routes, and future investments. While there is broad consensus that these forecasts are important, the challenge arises due to the difficulty in developing reliable forecast estimates. Passenger loads exhibit great seasonality and volatility from month to month making long-term forecasts much more complicated than initially perceived.In light of these issues, this research compared two traditional forecasting methodologies (exponential smoothing and ARIMA) utilizing data derived from the U.S. Bureau of Transportation Statistics T-100 International Market Data set (2010-2019) and was able to jointly analyze long-term growth and multiplicative seasonal factors. The forecast was validated using rolling-origin tests at 1, 3, and 6-months ahead as well as a 2-year holdout for evaluating the long-horizon forecast. The criteria used to assess forecast accuracy included the following measures: RMSE, MAE, MAPE, and MASE. The findings from this research are mixed, at 1, 3 and 6 month ahead forecasts both ARIMA and exponential smoothing produced very similar forecasts with neither method exhibiting superiority over the other when the forecasts are updated frequently. However, for longer-range forecasts this gap widens significantly as ARIMA forecasts drift and accumulate forecast error while exponential smoothing is more stable and achieves higher forecast accuracy with the 2-year holdout. The general conclusion drawn from this research is that what model to use when developing forecasts is driven by what level of forecasting is required. Operational forecasting at shorter time frames can be conducted with either method while longer-term forecasting is better accomplished with exponential smoothing method. This may seem intuitive; however, the evidence presented here supports the claim that the model selection based on the forecast requirements is of a greater significance than is typically reflected in most forecasting research.