Predictive and prescriptive performance of bike-sharing demand forecasts for inventory management

Bike-sharing systems are a rapidly developing mode of transportation and provide an efficient alternative to passive, motorized personal mobility. The asymmetric nature bike demand causes the need for rebalancing stations, which is typically done during nighttime. To determine optimal starting inventory level station given day, User Dissatisfaction Function (UDF) models user pickups returns as non-homogeneous Poisson processes with piece-wise linear rates. In this paper, we devise deep generative model directly applicable in UDF by introducing variational recurrent neural network (VP-RNN) forecast future pickup return We empirically evaluate our approach against both traditional learning-based forecasting methods on real trip travel data from city New York, USA, show how outperforms benchmarks terms system efficiency satisfaction. By explicitly focusing combination decision-making algorithms methods, highlight number shortcomings literature. Crucially, more accurate predictions do not necessarily translate into better decisions. providing insights interplay between forecasts, assumptions, decisions, point out that forecasts decision should be carefully evaluated harmonized optimally control shared mobility systems.