• Predicting Nominal Data in Presence of Poor Information. An Application to Air Tickets

Stacchini, Annalisa


SECS-S/03 Statistica economica


This thesis proposes an original interpretative key to a crucial problem of business travel: minimizing the flights cost, managing change and cancellation risk, choosing the optimal fare. Hitherto this topic was addressed only with reference to the revenue management of airlines, possibly because getting the information needed to model the traveler’s behavior, which, as highlighted by a qualitative study, is determined by one-off events, private life and business snags, is nearly impossible. Given that available data, collected by a travel management company for other purposes, concern mainly the tickets and flights’ characteristics, the results obtained in this work are very satisfying. For coping with the informative limitations, some original solutions were developed. First, as the literature suggests a seasonal pattern, tied to vacations, an attempt was made to insert a guess, for a vacation effect, not precise enough to be exploited within a Bayesian framework, directly in estimated models. This improved all the models’ predicting capability, but the value of a business-specific loss function, to assess predictors economically, was disappointing. Therefore, a new classification algorithm, amplifying faint signals, exploiting the whole matrix of estimated probabilities, for each prediction, was conceived. It is very flexible, as it can be applied to any matrix of probabilities, estimated by any classifier, and very effective. In fact, it improved the predictive performance of all the predictors and yielded an estimated global gain of 109,011 euros. Finally, a problem for selecting the best models for two nodes emerged, as all the candidates displayed identical forecasting performance, as assessed through the traditional measures for nominal data, based only on the predicted outcomes. Thus, a Predictive Accuracy Score is elaborated, for evaluating estimated probabilities, which are important, because, when the cost data is available, the expected value of the flights cost will be computable.




Tesi di dottorato






Stacchini, Annalisa (2016) Predicting Nominal Data in Presence of Poor Information. An Application to Air Tickets, [Dissertation thesis], Alma Mater Studiorum Università di Bologna. Dottorato di ricerca in Scienze statistiche , 28 Ciclo. DOI 10.6092/unibo/amsdottorato/7242.