Forecasting High‐Frequency Trade Durations: A Regime‐Switching Approach With Flexible Hazard Functions

What the paper says

This paper introduces a threshold stochastic conditional duration (TSCD) model to capture regime‐switching behaviour in both the observed duration process and the latent process. Additionally, the model captures complex market dynamics using roller‐coaster‐shaped hazard functions derived from the extended generalised inverse Gaussian (EGIG) distribution. The model parameters are estimated using the simulation‐based maximum likelihood method implemented via the sampling importance resampling algorithm, and validated through a simulation study. The empirical analysis employs trade duration data from the Apple incorporated and Tesla incorporated stocks. The results demonstrate that the TSCD model incorporated with EGIG distribution effectively captures distinct regime‐switching behaviours in trade durations, characterised by different parameter sets across regimes in the in‐sample analysis, yielding the highest log‐likelihood value and the lowest Akaike information criterion and Bayesian information criterion scores amongst all benchmark models. For the out‐of‐sample forecasts, the TSCD EGIG model consistently achieved the strongest performance, ranking first in most of the evaluated loss functions (mean squared forecast error and quasi‐likelihood). The Diebold–Mariano test further provides robust evidence of significant differences in the relative predictive performance of the models. Time‐at‐risk forecasts across various risk levels are computed and evaluated using the Kupiec likelihood ratio test. Lastly, density forecasts are assessed through the probability integral transform technique.