Simultaneous Graphical Dynamic Modeling

Technical Summary: Simultaneous Graphical Dynamic Linear Models (SGDLMs)

Authors: Luke Vrotsos & Mike West (Duke University)
Date: January 31, 2026

1. The Problem: Scalability and Flexibility in Multivariate Time Series

Traditional multivariate time series models often struggle to balance three competing needs: flexibility (capturing complex, time-varying relationships), parsimony (avoiding over-parameterization), and scalability (handling high-dimensional data).

Existing Bayesian frameworks, such as time-varying vector autoregressions (TV-VARs), frequently impose structural constraints—such as strict triangularity—to remain computationally tractable. These constraints make the model dependent on a pre-chosen ordering of variables, which can bias results and limit the model's ability to capture true simultaneous dependencies. Furthermore, standard causal analysis (counterfactual forecasting) in multivariate settings is often hindered by the difficulty of modeling complex pre-intervention dependencies and handling "missing data" (the unobserved counterfactual outcomes) in a sequential, scalable way.

2. Methodology: The SGDLM Framework

The paper centers on the Simultaneous Graphical Dynamic Linear Model (SGDLM). The core structural equation is:
$$(I - \Gamma_t)y_t = \mu_t + \nu_t, \quad \nu_t \sim N(0, \Lambda_t^{-1})$$

Key Methodological Components:

• Graphical Structure: Unlike triangular models, the $q \times q$ matrix $\Gamma_t$ is sparse but allows for simultaneous (non-acyclic) relationships. The sparsity pattern defines a directed graph where $y_i$ is a child of its "parental set" $sp(i)$.
• Decouple/Recouple Strategy: To maintain scalability, the authors use a sequential Bayesian approach. They first "decouple" the multivariate system into $q$ univariate Dynamic Linear Models (DLMs). They perform analytic updates using conjugate Normal-Gamma distributions and then "recouple" them using the determinant $|I - \Gamma_t|$ to account for cross-series dependencies.
• Filtering and Forecasting: The methodology employs Importance Sampling (IS) for posterior updates and Variational Bayes (VB) to transition between time steps, avoiding the heavy computational cost of repeated batch MCMC.
• Counterfactual Forecasting (CFM): To model causal effects, the authors treat the post-intervention "counterfactual" outcomes as missing data. They develop a Monte Carlo filtering strategy that samples these unobserved outcomes from a mixture of normals, allowing for seamless sequential updating even when experimental series are parents or children of control series.
• Outcome Adaptive Model (OAM): To test the null hypothesis of "no intervention effect," they run a parallel OAM that allows for increased state/volatility flexibility (via reduced discount factors) at the moment of intervention, providing a formal statistical comparison against the CFM.

3. Key Contributions

• Theoretical Linkage (Graphical $\leftrightarrow$ Factor Models): A major theoretical advancement is proving that SGDLMs implicitly define Sparse Dynamic Factor Models. By applying Singular Value Decomposition (SVD) to the coefficient matrix $\Gamma_t$, the authors show that the graphical parental structure directly determines the sparsity patterns of the factor loadings ($L_t$) and factor scores ($S_t$).
• Model Uncertainty via Marginal Likelihood: The paper introduces a method to evaluate model marginal likelihoods in SGDLMs. This allows researchers to perform Bayesian Model Averaging (BMA) to account for uncertainty regarding the underlying graphical structure (e.g., which variables are parents of others).
• Causal Extension: The development of a robust framework for multivariate counterfactual forecasting that handles the "missingness" of counterfactuals within a sequential Bayesian filtering context.

4. Results: The OECD GDP Case Study

The authors apply the SGDLM to annual real per capita GDP log-returns for 16 OECD countries to analyze the impact of German reunification (1990).

• Structure Discovery: Using marginal likelihoods, the model identified sparse, meaningful parental structures. The SVD revealed latent factors (e.g., a "Germany-founded" factor and a "France-founded" factor) that explained significant portions of the multivariate volatility.
• Causal Inference:
  • Counterfactual vs. Observed: The CFM predicted that German GDP growth should have remained relatively stable post-1990 based on control series (Australia/New Zealand). However, the observed data showed a sharp drop.
  • OAM vs. CFM: The OAM (which allows for adaptation to change) was strongly favored over the CFM. The probability of the OAM being the correct model rose to 0.99 by 1994, providing strong evidence of a causal impact.
  • Spillover Effects: The analysis suggested that the reunification had negative spillover effects on other European economies (e.g., Austria), as evidenced by the divergence between observed and counterfactual trajectories.

5. Significance

This paper provides a unified mathematical bridge between graphical models (precision-based sparsity) and factor models (covariance-based sparsity), which were previously treated as distinct. By offering a scalable, sequential, and fully Bayesian approach, the SGDLM framework enables high-dimensional causal analysis and forecasting in fields ranging from macroeconomics to financial portfolio management, where time-varying relationships and structural uncertainty are paramount.