Enhanced demand forecasting by combining analytical models and machine learning models

Abstract

Analytical models (AM) and machine learning (ML) models are often considered to be at opposite ends of the modeling spectrum. AM are closed form expressions based on first principles which require deep domain knowledge and are difficult to construct but can extrapolate to unseen data and are data-efficient and interpretable. At the other end, ML models require little or no domain knowledge to construct, are flexible, and can provide superior accuracy in data-rich environments, but cannot extrapolate, are data-inefficient and are black boxes. We investigate how to consolidate these opposite views to obtain the best of both worlds in the context of airline demand forecasting. We leverage on an existing AM baseline and employ deep learning-based ML models as correctional multiplicative factors. This approach provides a transparent, interpretable hybrid model with a forecast accuracy outperforming both pure AM and pure ML models.