Abstract: Despite some negative press, deep neural networks are a useful tool for statistical risk modelling. We present deep extremal regression models - neural networks designed to target descriptions of the conditional tails, such as extreme quantiles or tail indices. By combining deep neural networks with asymptotically-justified models from extreme value theory, we create grey box models which permit high-dimensional inference on extremes while retaining some of the interpretability of traditional statistical models. This talk covers a few examples of deep extremal regression models, which we use for extreme quantile regression, full (semi-parametric) conditional density estimation, and modelling of multivariate extremes via an angular-radial approach. Real data applications include US and Mediterranean wildfire extremes, and UK Metocean storms.