Exploiting trading volume for volatility forecasting

In this paper, we propose a novel GARCH specification which embraces traded volume in a statistically coherent and computationally efficient way. To capture the dependence between traded volume and asset return volatility, the Mixture of Distribution Hypothesis (MDH) is used. We model the time evolution of the latent variable(return variance) ,which drives the bivariate relation in MDH, in the spirit of Generalized Autoregressive Score modelling technique. The resulting volume enhanced GARCH model has quite parsimonious yet flexible structure. Since the peculiar form of score driven models, the likelihood can be obtained in closed form and hence estimation procedure is computationally inexpensive. In an empirical application with IBM, PG and MSFT stock data, we estimate our novel model with its 3 variants. Moreover, standard GARCH models from the literature are considered for comparison. Our results highlight the superiority of volume enhanced models both in-sample estimates and out-of-sample predictions among competing models. Particularly, when we consider returns as t distributed, the gain from traded volume becomes pronounced in out-of-sample prediction study.