Forecasting Transaction Rates: The Autoregressive Conditional Duration Model

This paper will propose a new statistical model for the analysis of data that does not arrive in equal time intervals such as financial transactions data, telephone calls, or sales data on commodities that are tracked electronically. In contrast to fixed interval analysis, the model treats the time between observation arrivals as a stochastic time varying process and therefore is in the spirit of the models of time deformation initially proposed by Tauchen and Pitts (1983), Clark (1973) and more recently discussed by Stock (1988), Lamoureux and Lastrapes (1992), Muller et al. (1990) and Ghysels and Jasiak (1994) but does not require auxiliary data or assumptions on the causes of time flow. Strong evidence is provided for duration clustering beyond a deterministic component for the financial transactions data analyzed. We will show that a very simple version of the model can successfully account for the significant autocorrelations in the observed durations between trades of IBM stock on the consolidated market. A simple transformation of the duration data allows us to include volume in the model.