The most standard test for a theoretical model is trying to predict new data points in a time series. This general task is ubiquitous and allows different theories to be compared quantitatively. Recently, the use of deep machine learning models has shown state-of-art performances in many repetitive, yet complex, tasks, such as image classification, speech recognition, and autonomous driving. Nevertheless, simple autoregressive models still seem to be more efficient in the context of financial time series prediction. I’m interested in developing optimal architectures for time series prediction with deep machine learning models, and in characterising their efficiency for time series from different sources, ranging from finance to physics.