Bayesian Forecasting with Nonlinear Autoregressive Conditional Duration Models

The purpose of this paper is to provide nonlinear autoregressive conditional duration (ACD) models to forecast price duration with Bayesian method and compare the forecast performances of several ACD models altering observations. The main findings and their implications are as follows: First, though duration shock is very persistent, the nonlinear specification is effective for forecasting duration since exponential smooth transition component-ACD (EST-C-ACD) models show better performance than component-ACD (C-ACD). Second, forecast performances of most of the ACD models are not significantly different across experiments with different sample sizes. Hence, with an appropriate number of observations, extending observations to forecast intraday transactions would not be required. Third, although insignificantly different in forecast performance, Weibull-ACD (WACD) models require much more CPU time than exponential-ACD (EACD) and are therefore inefficient in forecasting intraday duration. Fourth, in all of the forecast experiments, the EST-C-EACD model shows good overall performance and better CPU time. In the presence of heterogeneous traders in a financial market, a component-model can appropriately exhibit the price process consisting of heterogeneous components. The smooth transition model could especially better exhibit a price process than others by allowing for a smooth transition between heterogeneous components.