Operationalizing SHAP for Imbalanced No-Show Prediction in Hospitality

Hotel revenue management relies on accurately forecasting capacity. When a guest books a room, the property commits a perishable asset. If the guest fails to arrive without formally cancelling, the hotel absorbs a complete financial loss for that given night. Identifying these "no-show" bookings represents a difficult classification problem due to the rarity of the behaviour. In this dataset of 87,396 validated records, only 1.58% of reservations resulted in a realized no-show. Standard machine learning models generally fail when confronted with a 62-to-1 class imbalance. They achieve high raw accuracy by predicting that every guest will arrive, which provides zero operational utility. This paper details an XGBoost classification pipeline explicitly optimized to identify the minority class. By adjusting the class weights (scale_pos_weight = 62.477), the model accepted lower absolute precision to achieve a 0.60 recall rate on the minority class, successfully identifying 60% of all no-shows in the test set. The final model returned a ROC-AUC of 0.7406 and an F2-score of 0.1516, reflecting the heavy weighting toward recall. This research then subsequently applied SHAP (SHapley Additive exPlanations) to interpret the algorithmic decision-making process. Permutation importance tests and SHAP matrices confirmed that the absence of a parking space request was the primary predictor of a no-show. The resulting framework provides front-desk teams with clear, explainable thresholds for implementing targeted deposit policies.