Analytical ideas to improve daily demand forecasts: A case study

Abstract

With the growing popularity of app-based taxi aggregators, bike-sharing systems and supermarkets across the world, it is now important to forecast short-term (often daily) demand accurately. Imprecise forecasts generally result in daily losses due to over or under stocking. This paper proposes multiple analytical constructs for demand prediction using Capital Bikeshare’s data as an example. The aim is to provide novel and business-justified ideas on feature engineering and subsequently using these features to create different analytical constructs for the actual prediction problem. A comparison of different modeling techniques in solving the same problem is also included. The findings demonstrate that a decomposed multi-stage prediction performs better than the pure forecasting or prediction approaches. Ensembling results show that a cross-construct ensemble may perform better than the traditional multiple-learner ensembles within the same construct.