On The Modelling Of Heavy-Tailed Data Using Generative Adversarial Networks

Generative Adversarial Networks (GANs) were invented in 2014 and have tremendous and remarkable results in computer vision, time series, data augmentation, missing data imputation, anomaly detection, domain adaptation and semi-supervised learning. They are known to be better than other generative models due to their universal distributional learning capabilities. GANs offer competing approaches to common heavy-tailed distributions as they do not make strong distributional assumptions and can generate new plausible samples in a single step. However, it is not clear what kind of distributions they can represent. In particular, it is not clear whether and how GANs are suited to serve as heavy-tailed distribution learners in tabular data sets and this still remains an open question. In this work, we provide an in-depth overview of the use of GANs in modelling extreme events. We show that GANs can be used to enhance the data resolution of high-impact but low likelihood events and improve model robustness. We compare the popular Conditional Wasserstein GAN with a Gradient Penalty and recently proposed Extreme Value GANs i.e. ExGAN, Student’s t GAN and Pareto GAN, for modelling non-life insurance claims. Our approach is unique as it has not been previously applied for tabular data sets nor has it been previously applied for learning the distribution of heavy-tailed data sets. This realistic extreme event generation is useful in stress and scenario testing, especially in Own Risk and Solvency Assessment (ORSA). ORSA is an important tool which relies on testing an insurer’s projected future solvency on a wide range of extreme but authentic and realistic scenarios such that the insurer can manage to cope with those scenarios.

Find the Q&A here: Q&A on' Machine Learning and its Opportunities'