Resumen
In bike-sharing systems, the inventory level is defined as the daily number of bicycles required to optimally meet the demand. Estimating these values is a major challenge for bike-sharing operators, as biased inventory levels lead to a reduced quality of service at best and a loss of customers and system failure at worst. This paper focuses on using machine learning (ML) classifiers, most notably random forest and gradient tree boosting, for estimating the inventory level from available features including historical data. However, while similar approaches adopted in the context of bike sharing assume the data to be well-balanced, this assumption is not met in the case of the inventory problem. Indeed, as the demand for bike sharing is sparse, datasets become biased toward low demand values, and systematic errors emerge. Thus, we propose to include a new iterative resampling procedure in the classification problem to deal with imbalanced datasets. The proposed model, tested on the real-world data of the Citi Bike operator in New York, allows to (i) provide upper-bound and lower-bound values for the bike-sharing inventory problem, accurately predicting both predominant and rare demand values; (ii) capture the main features that characterize the different demand classes; and (iii) work in a day-to-day framework. Finally, successful bike-sharing systems grow rapidly, opening new stations every year. In addition to changes in the mobility demand, an additional problem is that we cannot use historical information to predict inventory levels for new stations. Therefore, we test the capability of our model to predict inventory levels when historical data is not available, with a specific focus on stations that were not available for training.