A0436
Title: Duration modeling in the presence of zero-duration clusters
Authors: Alessandro Morelli - University of Milan (Italy) [presenting]
Massimiliano Caporin - University of Padova (Italy)
Eduardo Rossi - University of Pavia (Italy)
Abstract: The presence of groups of transactions sharing the exact same timestamps in high-frequency financial data poses a significant challenge for duration modeling. This well-known issue is addressed by proposing a novel modeling approach that leverages the predictive information contained in such groups, which is referred to as zero-duration clusters. Using millisecond-level tick data for twelve large-cap, highly liquid U.S. stocks over the last two quarters of 2024, three novel empirical regularities are documented: Durations tend to decrease before a cluster and increase afterward; larger clusters are associated with shorter subsequent durations; clusters exhibit persistence over time, especially when large. Motivated by these findings, an extension of the autoregressive conditional duration (ACD) model is proposed that jointly captures these dynamics through a limited number of additional parameters. An extensive out-of-sample evaluation against standard alternatives, including adaptations of exponential and gamma generalized autoregressive score (GAS) models within the ACD framework, demonstrates that the proposed specification consistently emerges as the only model within the model confidence set (MCS). It is concluded that while aggregating zero durations enables modeling through point processes, the informational content carried by clusters should nonetheless be exploited through an appropriate modeling approach.
