A Poisson-Based Distribution Learning Framework for Short-Term Prediction of Food Delivery Demand Ranges

The COVID-19 pandemic has caused a dramatic change in the demand composition of restaurants and, at the same time, catalyzed on-demand food delivery (OFD) services - such as DoorDash, Grubhub, and Uber Eats - to a large extent. With massive amounts of data on customers, drivers, and merchants, OFD platforms can achieve higher efficiency with better strategic and operational decisions; these include dynamic pricing, order bundling and dispatching, and driver relocation. Some of these decisions, and especially proactive decisions in real time, rely on accurate and reliable short-term predictions of demand ranges or distributions. In this paper, we develop a Poisson-based distribution prediction (PDP) framework equipped with a double-hurdle mechanism to forecast the range and distribution of potential customer demand. Specifically, a multi-objective function is designed to learn the likelihood of zero demand and approximate true demand and label distribution. An uncertainty-based multi-task learning technique is further employed to dynamically assign weights to different objective functions. The proposed model, evaluated by numerical experiments based on a real-world dataset collected from an OFD platform in Singapore, is shown to outperform several benchmarks by achieving more reliable demand range forecasting.