On the long memory autoregressive conditional duration models

Abstract

In financial markets, transaction durations refer to the duration time between two consecutive trades. It is common that more frequent trades are expected to be followed by shorter durations between consecutive transactions, while less frequent trades are expected to be followed by longer durations. Autoregressive conditional duration (ACD) model was developed to model transaction durations, based on the assumption that the expected average duration is dependent on the past durations.

Empirically, transaction durations possess much longer memory than expected. The autocorrelation functions of durations decay slowly and are still significant after a large number of lags. Therefore, the fractionally integrated autoregressive conditional duration (FIACD) model was proposed to model this kind of long memory behavior.

The ACD model possesses short memory as the dependence of the past durations will die out exponentially. The FIACD model possesses much longer memory as the dependence of the past durations will decay hyperbolically. However, the modeling result would be misleading if the actual dependence of the past durations decays between exponential rate and hyperbolic rate. Neither of these models can truly reveal the memory properties in this case.

This thesis proposes a new duration model, named as the hyperbolic autoregressive conditional duration (HYACD) model, which combines the ACD model and the FIACD model into one. It possesses both short memory and long memory properties and allows the dependence of the past durations to decay between the exponential rate and the hyperbolic rate. It also indicates whether the dependence is close to short memory or long memory. The model is applied to the transaction data of AT&T and McDonald stocks traded on NYSE and statistically positive results are obtained when it is compared to the ACD model and the FIACD model.