Forecasting Supply in Voronoi Regions for App-Based Taxi Hailing Services

In this paper, we deal with the problem of supply forecasting in the context of an application based taxi hailing service. We first propose a method to optimally partition the city space using a Voronoi tessellation. The generating points of the Voronoi regions are obtained as demand density cluster centers, from the taxi demand dataset. We also identify the optimal temporal resolution to use for forecasting supply in these Voronoi regions. We use a linear time-series based algorithm to forecast supply in each Voronoi region. Using this methodology for the city of Bengaluru, India, we obtained a supply forecast accuracy of about 90% for the heavily used Voronoi regions. This represents a substantial improvement in the forecast accuracy compared to similar time-series based approaches, employed over rectangular 'geohashes'.