Dennis Prangle (Bristol)

A Flexible Tail Transformation and Application to Generative Models

Modelling heavy tail distributions is important in many statistical applications. Extreme value theory suggests modelling tails by the Generalised Pareto Distribution (GPD) and often uses this to model the tails separately to the body of the distribution. We aim to jointly model the tail and body by proposing a transformation which can transform Gaussian tails to GPD tails. We illustrate this approach using a particular class of generative models, normalising flows, to perform density estimation on financial data.