Title: The Econometrics of Financial Duration Modeling

Abstract: In this paper we discuss estimation and inference in econometric models for financial durations. By focusing on the well-known autoregressive conditional duration (ACD) models of Engle and Russell (1998), we show the unexpected result that the large sample behavior of likelihood-based estimators strongly depends on the tail behavior of the duration data, and hence on the finiteness of moments. In particular, we show that even for stationary and geometrically ergodic data, asymptotic normality breaks down when the tail index of the durations is smaller than one. In this case, the estimators are shown to be asymptotically mixed Gaussian with a non-standard rate of convergence depending on the tail index. Our results are particularly surprising when compared to the analysis of ARCH models: while ARCH and ACD likelihood functions have the same form, standard asymptotic arguments apply to ARCH models but not to ACD models. The crucial difference between the two types of models is that for ACD models, the number of observations within any given sample period is random. This feature, rather than being innocuous, requires, as we demonstrate, new, non-standard theory. Many of our results apply to general econometric models where the number of observed events over a given time span is random.