Title: Hierarchical Bayes modeling of autocorrelation and intraday seasonality in financial durations
Authors: Teruo Nakatsuma - Keio University (Japan) [presenting]
Tomoki Toyabe - Keio University (Japan)
Abstract: Intraday financial transactions are irregularly spaced, and their durations exhibit positive autocorrelation and intraday seasonality. In the literature, the former is formulated as a time-dependent duration model such as the stochastic conditional duration (SCD) model while the latter is dealt with by filtering out any cyclical fluctuations in time series of durations with a spline smoothing method before the duration model is estimated. We propose a hierarchical Bayes approach to model both autocorrelation and intraday seasonality in durations simultaneously. In our new approach, the autocorrelation structure of durations is captured by the SCD model while the intraday seasonality is approximated with B-spline smoothing, and the parameters in both models are allowed to differ on each trading day. In B-spline smoothing, we incorporate the smoothness prior (the penalty on variations in the seasonality on the same trading day) as well as the similarity prior (the penalty on differences in the seasonality between consecutive trading days) in order to prevent overfitting. The resultant model is regarded as a non-linear non-Gaussian state space model of unbalanced panel data, for which a Bayesian approach is suitable. We developed an efficient Markov chain sampling scheme for the posterior analysis of the proposed model and applied it to high-frequency commodity futures transaction data in the Tokyo Commodity Exchange.