Like in other service industries, booking cancellations impact hotel management decisions, negatively contributing to accurate forecasts. Previous research showed it is possible to develop predictive models using booking data. However, existing models did not consider high-volatile times, such as a pandemic, where mass cancellations happen. This research uses datasets from four hotels to assess in a first study how existing machine learning classification models perform under the conditions imposed by high-volatility times (COVID-19 pandemic). In a second study, this research studies how models can be improved using a sliding window training approach. Results show that existing booking cancellation models can be improved if a sliding window with nine months of training data is used, with performance increasing up to 5% points in terms of Area Under the Curve. The findings from both studies demonstrate that while pre-pandemic models remain effective, incorporating pandemic data using a sliding window approach significantly improves predictive accuracy.