Reproducible air passenger demand estimation

The availability of passenger demand estimates for air traffic routes is crucial to a plethora of application and research problems ranging from, e.g., optimization of airline fleet utilization to complex simulations of whole air transport systems. However, somewhat surprisingly, such demand estimates appear hard to come by directly or even to generate by means of published models. This is in large parts due to the widespread use of expensive proprietary data (such as airline-specific ticket prices for certain flight connections) which is typically employed both to calibrate demand estimation models as well as to evaluate such models in order to obtain demand estimates for given origin–destination airport pairs. With this work, we propose building a data set for the European air transport system from given base data and automatically extracted and processed data from external sources, all of which are (made) freely available in the public domain and thus enable reproducibility and facilitate comparability of research involving air passenger transportation. Moreover, we challenge the long-standing tradition of calibrating so-called gravity models for demand estimation by standard linear regression. For the European air transport system, using the aforementioned publicly available data, we demonstrate that machine learning models and techniques like neural networks or the “kernel trick” can significantly improve the estimation quality with respect to ordinary least-squares. In fact, our results—the best of which were obtained using our feed-forward neural network model (four hidden layers with tanh and ReLU activations)—achieve a performance at least comparable to what has been reported in earlier works that utilized non-public data. Computer code to generate air passenger demand estimates is made publicly available online along with our base data and implementation to collect and curate external data. •Origin–destination air passenger demand estimation for Europe.•Publicly available data set and code allows full reproducibility.•Comparison of several methods to calibrate gravity and other models.•Machine learning techniques significantly outperform standard linear regression.•Mixed-integer quadratic program for automated feature design and selection.