A1244
Title: Modelling volatility with variational inference priciples
Authors: Martin Magris - ITAM - Instituto Tecnologico Autonomo de Mexico (Mexico) [presenting]
Alexandros Iosifidis - Aarhus University (Denmark)
Abstract: Variational Inference (VI) methods are gaining attention and popularity as efficient and practical approaches for performing approximate Bayesian inference in complex models. Concerning various recent VI black-box methods, it is shown that a Bayesian treatment of standard GARCH-like volatility models is immediate and of straightforward implementation. In a study involving 100 stocks and seven volatility models, first, the quality and validity of the VI approximations provided by four optimizers concerning a Monte-Carlo baseline are addressed. Then the impact of the different estimation procedures on both in-sample and out-of-sample performance metrics is discussed. Lastly, it is observed that one-step-ahead volatility forecasts obtained with the above Bayesian methods often outperform their standard likelihood-based counterparts.
