Coronavirus disease (COVID-19) was detected in Wuhan, China in 2019 and spread worldwide within few weeks. The COVID-19 epidemic started to gain traction in France in March 2020. Sub-national hospital admissions and deaths were then recorded daily and served as the main policy indicators. Concurrently, mobile phone positioning data have been curated to determine the frequency of users being colocalized within a given distance. Contrarily to individual tracking data, these can provide a proxy of human contact networks between subnational administrative units.Motivated by numerous studies correlating human mobility data and disease incidence, we developed predictive time series models of hospital incidence between July 2020 and April 2021. Adding human contact network analytics such as clustering coefficients, contact network strength, null links or curvature as regressors, we found that predictions can be improved substantially (more than 50%) both at the national and sub-national for up to two weeks. Our sub-national analysis also revealed the importance of spatial structure, as incidence in colocalized administrative units improved predictions. This original application of network analytics from co-localisation data to epidemic spread opens new perspectives for epidemics forecasting and public health.