Latent level correlation modeling of multivariate discrete-valued financial time series

Abstract

In high-frequency financial data, dynamic patterns of transaction counts in regular time intervals provide crucial insights into market microstructure, such as short-term trading activities and intermittent intensities of price oscillation. In this paper we propose a Bayesian hierarchical framework that incorporates correlated latent level and temporal effects to model multivariate count data during intraday transaction intervals. Built on the INLA method for implementation, our framework proves to be competitive with the traditional MCMC approach in terms of model inference and computational cost. We demonstrate the efficacy of our methodology by applying it to assets from three Global Industry Classification Standard (GICS) sectors, namely, healthcare, energy, and industrials. The analysis uncovers various microstructures of financial count data using our framework. Specifically, our model featuring a correlated latent effect structure adeptly captures the pattern of the empirical correlations within the count data patterns with additional statistical inference, such as assessing different associations between short-term averaged trading size as well as trading duration, the counts at different risk levels, and uncovering differential levels of uncertainty resulted from market temporal behavior and unobservable latent effects across the three sectors. We also discuss some potential applications of our framework in real-world scenarios.