66th ISI World Statistics Congress

66th ISI World Statistics Congress

Modelling multivariate ordinal time series

Author

DK
Dimitrios Karlis

Co-author

  • A
    Anna Nalpantidi

Conference

66th ISI World Statistics Congress

Format: WSC 2027: IPS Abstract

Keywords: pairwise

Abstract

Modeling the joint distribution of multiple ordinal time series is challenging, as it requires simultaneously accounting for autocorrelation within each series, dependence between series, and cross-correlations across time. In this paper, we propose a copula-based approach to specify joint distributions, beginning with the two-variable case and extending it to higher dimensions. To mitigate computational complexity in high dimensions, we approximate the full likelihood using composite likelihood—specifically, conditional pairwise likelihood—where each pairwise relationship is modeled via copulas. Rather than joint optimization, we maximize each bivariate model independently and combine the individual estimates using a weighted mean, with weights derived from each model’s Hessian matrix. Simulation studies demonstrate that our approach performs well across various sample sizes. Finally, we outline a forecasting framework and illustrate the practical utility of our method using a real-world application