Generalised AIC for the Evaluation of Autoregressive Conditional Duration Models

Abstract

The theoretical basis of generalised versions of the Akaike Information Criterion (AIC) are reviewed and used to support their use in evaluating the autoregressive conditional duration (ACD) models that are widely used in the analysis of fine scale financial series. This provides a more solid theoretical underpinning, whilst allowing for greater generality, than the consistent information criteria discussed in our last paper (Xie and Cowpertwait, 2014). Using an Exponential and Weibull distribution of errors, the likelihood functions for two ACD models are derived and used to fit the models to IBM transaction data. The fitted models are compared using AIC and the derived generalised AIC. Both information criteria show that a Weibull distribution of errors provides the best fit to the transaction data, although the presence of autocorrelation in the residual errors indicates that further work would be needed before the models are used in practical applications.