# 14-032/IV/DSF73 (2014-03-10; 2015-07-06)

Siem Jan Koopman, VU University Amsterdam; Rutger Lit, VU University Amsterdam; André Lucas, VU University Amsterdam
dynamic count data models, non-Gaussian multivariate time series models, importance sampling, numerical integration, volatility models, sports data
JEL codes:
C22, C32, C58

We introduce a dynamic statistical model for Skellam distributed random variables. The Skellam distribution can be obtained by taking differences between two Poisson distributed random variables. We treat cases where observations are measured over time and where possible serial correlation is modeled via stochastically time-varying intensities of the underlying Poisson counts. The likelihood function for our model is analytically intractable and we evaluate it via a multivariate extension of numerically accelerated importance sampling techniques. We illustrate the new model by two empirical studies and verify whether our framework can adequately handle large data sets. First, we analyze long univariate high-frequency time series of U.S. stock price changes, which evolve as discrete multiples of a fixed tick size of one dollar cent.
In a second illustration, we analyze the score differences between rival soccer teams using a large, unbalanced panel of seven seasons of weekly matches in the German Bundesliga.In both empirical studies, the new model provides interesting and non-trivial dynamics with a clear interpretation.