Semiparametric multivariate D-vine time series model

Zifeng Zhao, Peng Shi, Zhengjun Zhang

Abstract

This paper proposes a novel semiparametric multivariate D-vine time series model (mDvine) that enables the simultaneous copula-based modeling of both temporal and cross-sectional dependence for multivariate time series. To construct the mDvine, we first build a semiparametric univariate D-vine time series model (uDvine) based on a D-vine. The uDvine generalizes the existing first-order copula-based Markov chain models to Markov chains of an arbitrary-order. Building upon the uDvine, we then construct the mDvine by joining multiple uDvines via another parametric copula. As a simple and tractable model, the mDvine provides flexible models for marginal behavior of time series and can also generate sophisticated temporal and cross-sectional dependence structures. Probabilistic properties of both the uDvine and mDvine are studied in detail. Furthermore, robust and computationally efficient procedures, including a sequential model selection method and a two-stage MLE, are proposed for model estimation and inference, and their statistical properties are investigated. Numerical experiments are conducted to demonstrate the flexibility of the mDvine, and to examine the performance of the sequential model selection procedure and the two-stage MLE. Real data applications on the Australian electricity price and the Ireland wind speed data demonstrate the superior performance of the mDvine to traditional multivariate time series models.

Introduction

Modeling multivariate time series presents a significant challenge: researchers must simultaneously account for temporal dependence (the relationship between a variable's current and past values) and cross-sectional dependence (the relationship between different variables at the same point in time). Traditional approaches, such as Vector Autoregression (VAR) or Copula-based multivariate dynamic models (SCOMDY), often rely on restrictive assumptions like linearity or Gaussian innovations. While joint copula models offer more flexibility, they frequently suffer from the "curse of dimensionality" when scaled to high-dimensional data over long time horizons.

This paper introduces the mDvine framework, a semiparametric approach designed to balance flexibility and computational tractability. By utilizing D-vines to decompose complex joint distributions into simpler bivariate "pair-copulas," the model can capture non-linear relationships and tail dependencies—such as those seen during extreme market events—without the computational explosion typical of fully joint models.

Methodology: The mDvine Framework

The mDvine model is constructed in two primary stages, separating the temporal and cross-sectional modeling tasks.

1. Univariate D-vine (uDvine)

The first step involves modeling each individual time series using a uDvine. This component generalizes standard first-order copula-based Markov chains to $p$-order Markov chains. The uDvine is semiparametric: it employs nonparametric empirical distributions for the marginals and parametric copulas to describe the dependencies between time points.

The D-vine structure is particularly well-suited for time series because it naturally connects adjacent nodes in its first tree, mirroring the sequential nature of time. As shown in Figure 1, the joint distribution is decomposed into a hierarchy of pair-copulas, allowing for the modeling of dependencies between observations at different lags.

D-vine structure for a 5-node sequence, illustrating the hierarchical decomposition of dependencies into pair-copulas.

2. Multivariate D-vine (mDvine)

To handle multivariate data, multiple uDvines are linked using a $d$-dimensional parametric copula. This "hybrid" construction allows cross-sectional dependence to be modeled independently of temporal dependence. For high-dimensional datasets, the authors suggest using factor-structured or spatial-structured copulas for the cross-sectional link to maintain efficiency.

Estimation and Inference

The authors propose two key innovations for practical implementation:

  • Sequential Model Selection: A tree-by-tree procedure using the Bayesian Information Criterion (BIC) to automatically select the optimal copula families (e.g., Gaussian, Clayton, Gumbel) for each pair in the uDvine.
  • Two-stage MLE: An estimation strategy where uDvine parameters are estimated series-by-series, followed by the estimation of the cross-sectional linking copula. This approach significantly reduces the parameter space compared to full maximum likelihood estimation.

Evaluation and Results

The performance of the mDvine model was evaluated using both simulated and real-world datasets.

Simulated Data

Simulations verified that the uDvine could accurately approximate complex processes such as GARCH and GJR-GARCH (which exhibit volatility clustering) and high-order AR(9) processes. Notably, the uDvine tracked one-day-ahead Value-at-Risk (VaR) for GARCH processes effectively and matched "oracle" AR(9) performance as sample sizes increased.

Real-World Applications

  • Australian National Electricity Market (NEM): The model was applied to five years of daily maximum electricity prices across five Australian regions. The mDvine consistently outperformed VAR(1), VAR(2), and VAR(1)-DCC models, achieving lower Continuous Ranked Probability Scores (CRPS), which indicates superior density forecasting. It also successfully captured seasonal variations in regional correlations.
  • Ireland Wind Speed: This application demonstrated the model's capability in spatio-temporal modeling, further proving its versatility across different domains.

Discussion and Conclusion

The mDvine framework offers a powerful alternative to traditional multivariate time series models. Its "hybrid" nature avoids the curse of dimensionality while capturing critical features like tail dependence. One surprising insight is the model's ability to generate volatility clustering effects implicitly, suggesting that copula-based temporal modeling can serve as a robust alternative to squared-error-based approaches like GARCH.

While the model currently assumes that a series' marginal distribution depends only on its own history (ignoring potential "lead-lag" effects between different series), it provides a scalable and flexible foundation for future research. Future work may focus on relaxing this conditional independence assumption and providing rigorous proofs for the consistency of the sequential tree-by-tree selection procedure in a time-series context.

Summary updated

Abstract

This paper proposes a novel semiparametric multivariate D-vine time series model (mDvine) that enables the simultaneous copula-based modeling of both temporal and cross-sectional dependence for multivariate time series. To construct the mDvine, we first build a semiparametric univariate D-vine time series model (uDvine) based on a D-vine. The uDvine generalizes the existing first-order copula-based Markov chain models to Markov chains of an arbitrary-order. Building upon the uDvine, we then construct the mDvine by joining multiple uDvines via another parametric copula. As a simple and tractable model, the mDvine provides flexible models for marginal behavior of time series and can also generate sophisticated temporal and cross-sectional dependence structures. Probabilistic properties of both the uDvine and mDvine are studied in detail. Furthermore, robust and computationally efficient procedures, including a sequential model selection method and a two-stage MLE, are proposed for model estimation and inference, and their statistical properties are investigated. Numerical experiments are conducted to demonstrate the flexibility of the mDvine, and to examine the performance of the sequential model selection procedure and the two-stage MLE. Real data applications on the Australian electricity price and the Ireland wind speed data demonstrate the superior performance of the mDvine to traditional multivariate time series models.
BibTeX
@article{zhao2018semiparametric,
  author = {Zifeng Zhao and Peng Shi and Zhengjun Zhang},
  title = {{Semiparametric multivariate D-vine time series model}},
  journal = {Journal of Business and Economic Statistics},
  year = {2018},
}
Figure