A0673
Title: Learning the predictive density of mixed-causal ARMA processes for portfolio optimization
Authors: Arthur Thomas - Paris Dauphine University - PSL (France) [presenting]
Abstract: Asset price bubbles have become increasingly common in financial markets around the world, and risk management during these extreme events is a challenging task. In fact, standard financial econometric models (ARMA-GARCH) do a poor job of capturing the non-linear characteristics of speculative bubbles. At the same time, mixed-causal ARMA processes are known to capture their dynamics well. However, the limited knowledge of the predictive density of mixed-causal processes, especially during explosive bubble events, complicates their forecasting ability and thus limits their use in practical applications. Recognising the lack of closed-form formulae for the conditional prediction density, except in exceptional univariate cases, simulation-based and sample-based methods have been proposed in the literature. However, these methods can be computationally expensive for multivariate processes, rely on distributional assumptions for the error term, and do not accurately capture the dynamics during explosive episodes. It is shown that K nearest neighbours and random forest learning methods are promising for this task. First, in a simple univariate case, the tested approaches provide an interesting approximation to the true theoretical predictive densities compared to standard approaches available in the literature, then this approach is extended to the multivariate cases and an application to portfolio optimization is proposed.
