With the surge in popularity of wearable technologies a large percentage of the US population is now tracking activities such as sleep, diet and physical exercise. In this study we empirically evaluate the ability to predict metrics (e.g., weekly alcohol consumption) directly related to health outcomes from densely sampled, multi-variate time series of behavioral data. Our predictive models are based on temporal convolutional neural networks and take as input the raw historical time series of daily step counts, sleep duration, and weight scale usage sourced from an online population of thousands of digital trackers. The prediction accuracy achieved outperforms several strong baselines that use hand-engineered features and indicates that tracker data contains valuable information on individuals’ lifestyles even for behavioral aspects seemingly unrelated to the measured quantities. We believe that this insight can be applied to the design of new digital interventions and enable future largescale preventive care strategies.