Optimizing hotel demand forecasting through ensemble models: A Modern Portfolio Theory approach

Abstract

This study introduces Modern Portfolio Theory (MPT) as a unified framework for combining tourism-demand forecasts. By leveraging the complementarity of time-series models, machine-learning regressions, and deep-learning sequence models, we recast ensemble design as a mean–variance optimization problem. Each forecasting model is treated as an asset with an expected return (mean forecast accuracy) and risk (error variance and covariance). Using daily hotel data from a major U.S. East Coast city (2015–2019), we demonstrate that MPT-weighted ensembles outperform both individual models and traditional variance–covariance (VACO) blends, achieving MAPE values below 5.1% and reducing AvgRelMAE by up to 81.4% relative to a naïve benchmark. These findings advance tourism and hospitality forecasting by introducing a transferable efficiency frontier toolkit, enabling managers to explicitly balance forecast accuracy and volatility while making more informed pricing, staffing, and inventory decisions in volatile market environments.