A data-driven dynamic demand hotspots forecasting framework for on-demand meal delivery platforms

Abstract

Speed and reliability are the keys to high quality on-demand meal delivery service. The rebalancing of couriers locations according to future demand remains an operational challenge in the industry. This study proposes an adaptive framework to identify and predict the near-future demand hotspots, utilizing the semi real-time predictive information as input. This framework provides demand insights to assist meal delivery platforms in making operational forward-looking resourcedemand rebalancing decisions in real time, such as fleet management and demand management. To generate fast and accurate demand forecasting, we incorporate time series features and datadriven machine learning methods to create an adaptive forecasting approach. We create a dynamic demand hotspot clustering algorithm which takes predictive and geographic information as input. In the case study, our predictive forecasting model outperforms the time series and deep learning benchmarks in deterministic forecasting. The hotspots clustering performance is improved by using probabilistic predictive input.