Assessing the well-being of a population is of utmost importance for policy and decision makers so they can design appropriate policies and interventions to improve the quality of life of their citizens. Traditional methods to determine aggregate well-being consist of surveys which are expensive to obtain and difficult to scale. Thanks to the availability of large-scale human behavioral data, new methods to assess well-being might be possible. In this paper we describe one of such methods: MobiSenseUs, a machine-learning based system to automatically estimate geographically aggregated objective and subjective well-being measures in the UK from mobile data. We propose a comprehensive battery of features that capture different aspects of human behavior – i.e. communication patterns, mobile app usage and spatial mobility – from two sources of pseudonymized mobile data of more than one million smartphone users. We are the first to build machine-learning models to predict both objective (IMD) and subjective (SWB) indicators in the UK from these mobile features. We find that the IMD can be predicted more accurately than SWB, reaching 99% and 78% average accuracies in a binary classification task for the IMD and SWB, respectively. We analyze the most predictive features and derive implications for the design of data-driven machine-learning public health policy systems.