Hotel reservation cancellations: predictive modeling and feature impact analysis

Knowing if and when a reservation will be canceled is crucial to hotels. They can try reselling rooms that are predicted to be canceled to minimize revenue losses and optimize inventory. We sought to use machine learning models to predict hotel reservation cancellations with higher f1-scores than previous studies because of their crucial impact on hotels. Using our best model of a hyper-tuned random forest model, we first used all available features in our data. We achieved test f1-scores of 0.92, for not canceled reservations and 0.83 for canceled bookings. We also performed feature importance analysis to understand the marginal contribution of each feature on the model’s predictions. As a result, we identified that the top ten features were the most important. We experimented with a tuned random forest using only the top four and top ten features and found a model trained on only the top ten features performed competitively with our best-performing model. This model achieved test f1-score of 0.91 for not canceled reservations and 0.82 for canceled bookings. The model trained on the top four most contributing features achieved a testing f1-score of 0.89 for not canceled reservations and 0.78 for canceled bookings. Compared to our best-performing model, this model, trained on the top four features, presented with a maximum 5% difference for canceled bookings. As a result of this study, a model trained on the top four features achieved acceptable trade offs in f1 scores of 0.2 and 0.4 for not canceled and canceled reservations respectively.