High-Dimensional Sparse Multivariate Stochastic Volatility Models

💡 Research Summary

The paper tackles the well‑known “curse of dimensionality” that hampers the practical use of multivariate stochastic volatility (MSV) models in high‑dimensional settings. While MSV models typically deliver more accurate volatility forecasts than multivariate GARCH (MGARCH) specifications, their estimation via Bayesian Markov chain Monte Carlo (MCMC) or Monte‑Carlo Likelihood (MCL) becomes computationally prohibitive as the number of series grows. To overcome this obstacle, the authors propose a novel two‑step estimation framework that relies on penalized ordinary least squares (OLS) rather than likelihood‑based or simulation‑based methods.

Model formulation
The observable vector (y_t) is modeled as (y_t = D_t \varepsilon_t) with (\varepsilon_t) i.i.d. zero‑mean and correlation matrix (\Gamma). The latent log‑volatility vector (h_t) follows a linear state equation (h_{t+1}= \mu + \Phi (h_t-\mu) + \eta_t), where (\eta_t\sim N(0,\Sigma_\eta)). By taking logs of squared returns, the authors obtain a state‑space representation that can be rewritten as a VARMA(1,1) process for the transformed series (y_t^\ell = \log(y_t^2)).

Two‑step penalized OLS procedure

1. Step 1 – Sparse VAR approximation
   The demeaned series (x_t = y_t^\ell - \mathbb{E}