Prognostic models in survival analysis are aimed at understanding the relationship between patients' covariates and the distribution of survival time. Traditionally, semi-parametric models, such as the Cox model and the Additive Hazards model, have been assumed. In this talk I will derive a measure of explained variation under the Additive Hazards model showing its properties. Moreover I will describe the development of a new flexible method for survival prediction: DeepHazard, a neural network for time-varying risks. I will show its performance on popular real datasets.