Navigating uncertainty: enhancing hotel cancellation predictions with adaptive machine learning

Accurately predicting hotel booking cancellations is critical for hotel management, especially during volatile periods such as the COVID-19 pandemic. Prior work demonstrated that machine-learning (ML) models perform well on historical data, yet few studies test robustness under severe disruption. We evaluate ML classifiers trained on pre-pandemic data from four hotels and assess their adaptability to pandemic conditions (Study One). We then examine whether adding pandemic observations via a dynamic sliding-window approach improves accuracy (Study Two). Pre-pandemic models exhibit reasonable discrimination, but including pandemic-period data can raise the Area Under the Curve (AUC) by up to 5% points. A nine-month training window balances stability and responsiveness, capturing rapid shifts in booking patterns and customer behavior. Feature importance also changes: Lead time and other drivers show altered effects during the pandemic, underscoring the need for continuously updated models. Anchored in concept-drift theory, we interpret the pandemic as an abrupt shift in the cancellation decision boundary and show that sliding-window retraining together with interpretable diagnostics (e.g., the Lead time crossover threshold) provides a theoretically grounded blueprint for prediction under distributional change. Our results advocate scheduled retraining and lightweight drift diagnostics to sustain forecast accuracy and managerial actionability. For hotel managers and technology providers, the proposed approach supports proactive cancellation management, more reliable forecasting, and resilient operations in volatile markets, demonstrating the robustness of adaptive ML under conditions of extreme market volatility. The study advances theoretical understanding and practical applications by operationalizing concept-drift management in revenue-critical settings.