Generative Adversarial Networks (GANs) were invented in 2014 and have generated more interest since then. GANs are useful for learning the structure of the data without explicitly postulating the model. They are better than other generative models, used for data augmentation, boosting predictive models, attention prediction, anomaly detection, domain adaptation, privacy preservation, missing data imputation and discriminative modelling. GANs have been successful on images, music, text, speech and tabular data sets. However, GAN applications have been lacking in actuarial science. This work offers to provide practical applications of GANs in actuarial science. This presentation covers the following aims and objectives:
• Overview of generative models and why GANs are of better quality than other generative models
• An overview of the GAN architecture with practical insurance, banking and healthcare examples, trained using Python for:
• Data augmentation;
• Alleviating class imbalance; and
• Boost predictive models
• Cover some challenges with GANs, including recent advances and scope for the future and actuarial science.
Overall, we show a significant superiority of GANs for predictive models and stochastic simulations compared to current actuarial approaches. This work has recently been submitted for an MSc in Data Science at the University of the Witwatersrand in 2019.