Hyperparameter and Feature-Choice Sensitivities: An Application to Predicting Passenger Load in Airports

This paper introduces two complementary measures of uncertainty. Hyperparameter Sensitivity (HPS) records the range of predictions produced when the hyperparameters of a single feature-construction method are varied. Feature-Choice Sensitivity (FCS) records the further uncertainty introduced by the choice between alternative modeling approaches. Using daily passenger data from a German airport, we apply these measures to three seasonality models: Repeating Radial Basis Functions (RBF), Trigonometric Functions, and Periodic Splines. Sensitivity varies sharply across the three. RBF and Trigonometric methods show low normalised HPS and produce stable predictions across hyperparameter settings. Splines reach a better point accuracy but are markedly more sensitive. Selecting a model on best-case accuracy alone can therefore underestimate operational risk. In airport demand planning, where forecast reliability shapes staffing and resource allocation, a model with lower HPS may be preferable even if its peak accuracy is slightly lower. We recommend treating sensitivity analysis as a standard step in forecasting workflows.