Volume 16, Issue 2

We develop a hybrid model of multivariate volatility that uses recurrent neural networks to capture the conditional variances of latent orthogonal factors in a GO-GARCH framework. Our approach seeks to balance model flexibility with ease of estimation and can be used to model conditional covariances of a large number of assets. The model performs favourably in comparison with relevant benchmark models in a minimum variance portfolio (MVP) scenario.

The high returns of cryptocurrencies have attracted many investors in recent years. At the same time the evolution of cryptocurrencies is characterized by extreme volatility. For investors, it is therefore key to gauge the risks related to an investment in cryptocurrencies. We provide a comparison of several GARCH and stochastic volatility models for forecasting the risk of cryptocurrencies over the out-of-sample period from 28.09.2018 to 28.02.2023. It turns out that the widely used GARCH(1,1) does not provide accurate risk predictions. In contrast, adding t -distributed innovations or allowing for regime changes improves the accuracy in both model classes. Finally, we consider a Bayesian decision-guided approach with discount learning to combine the different models and provide robust evidence that combining the model predictions leads to accurate combined risk predictions.