How new models revolutionize hotel demand forecasting with sharper, more reliable predictions.

Hotel demand forecasting is a foundational task in revenue management, with applications in inventory control, pricing strategy, and operational planning. Traditional approaches typically treat booking data as a one-dimensional time series—e.g., modelling room demand as a function of stay date or lead-in period—thereby failing to fully exploit the two-dimensional structure inherent in hotel data.

In this work, we introduce the concept of a hotel demand time surface, a two-dimensional representation indexed by report date and stay date. We demonstrate that widely used forecasting targets—such as pickup curves, final day occupancy, or pace reports—can be interpreted as 1D slices or projections of this surface. This framing enables a more principled and unified modelling approach.

We conduct a comprehensive review of existing forecasting methods in both industry and academia, categorizing them into statistical, tree-based, and neural network approaches. Building on this foundation, we present the Cloudbeds time surface learner—a panelled ensemble model that jointly learns across report-stay pairs, capturing spatial and temporal dependencies on the demand surface.