Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Forecasting Oil Volatility through Network Models with GARCH-Informed Correlation Weights

Fayçal Djebari, Kahina Mehidi, Khelifa Mazouz, Philipp Otto

TL;DR

The paper tackles high-dimensional oil-price volatility forecasting by embedding contemporaneous spillovers within a network framework whose weights are derived from MGARCH-based conditional correlations (CCC, DCC, GO). Through a rolling-window analysis of six OPEC crude prices, GARCH-informed Network Log-ARCH models achieve forecasting performance close to or better than standard DCC-GARCH while delivering substantial computational savings, thanks to a GMM-estimated, low-parameter structure and static, time-averaged networks. Key contributions include (i) introducing GARCH-informed network weights, (ii) demonstrating superior out-of-sample forecasting and robust model selection via MCS, and (iii) quantifying dramatic computational gains that enable real-time risk surveillance in interconnected energy markets. The results underscore the practical value of parsimonious network volatility models for systemic risk assessment and scalable forecasting in global oil markets, with potential extensions to time-varying networks and broader panels of producers.

Abstract

This study addresses the computational challenges of forecasting volatility in high-dimensional commodity markets. Building on the Network log-ARCH framework, we introduce a novel class of network topologies from GARCH-informed correlation weights, obtained from conditional covariance estimates of multivariate GARCH models, rather than relying on the heuristic distance measures commonly used in clustering methods. We evaluate the proposed models forecasting performance through a rolling-window exercise using a panel of OPEC members crude oil prices. The results identify network volatility models incorporating these new GARCH-informed weights as the statistically superior specifications. Remarkably, the proposed framework matches standard DCC-GARCH predictive accuracy while delivering up to 62,000-fold computational gains. By explicitly modeling contemporaneous spillovers through interpretable spatial ARCH-like lags estimated via GMM, the proposed approach offers an optimal trade-off between parsimony, interpretability, and performance. The findings establish GARCH-informed network models as robust, scalable alternatives for systemic risk measurement and volatility forecasting in interconnected financial markets.

Forecasting Oil Volatility through Network Models with GARCH-Informed Correlation Weights

TL;DR

The paper tackles high-dimensional oil-price volatility forecasting by embedding contemporaneous spillovers within a network framework whose weights are derived from MGARCH-based conditional correlations (CCC, DCC, GO). Through a rolling-window analysis of six OPEC crude prices, GARCH-informed Network Log-ARCH models achieve forecasting performance close to or better than standard DCC-GARCH while delivering substantial computational savings, thanks to a GMM-estimated, low-parameter structure and static, time-averaged networks. Key contributions include (i) introducing GARCH-informed network weights, (ii) demonstrating superior out-of-sample forecasting and robust model selection via MCS, and (iii) quantifying dramatic computational gains that enable real-time risk surveillance in interconnected energy markets. The results underscore the practical value of parsimonious network volatility models for systemic risk assessment and scalable forecasting in global oil markets, with potential extensions to time-varying networks and broader panels of producers.

Abstract

This study addresses the computational challenges of forecasting volatility in high-dimensional commodity markets. Building on the Network log-ARCH framework, we introduce a novel class of network topologies from GARCH-informed correlation weights, obtained from conditional covariance estimates of multivariate GARCH models, rather than relying on the heuristic distance measures commonly used in clustering methods. We evaluate the proposed models forecasting performance through a rolling-window exercise using a panel of OPEC members crude oil prices. The results identify network volatility models incorporating these new GARCH-informed weights as the statistically superior specifications. Remarkably, the proposed framework matches standard DCC-GARCH predictive accuracy while delivering up to 62,000-fold computational gains. By explicitly modeling contemporaneous spillovers through interpretable spatial ARCH-like lags estimated via GMM, the proposed approach offers an optimal trade-off between parsimony, interpretability, and performance. The findings establish GARCH-informed network models as robust, scalable alternatives for systemic risk measurement and volatility forecasting in interconnected financial markets.

Paper Structure

This paper contains 27 sections, 43 equations, 5 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Data, Descriptive Statistics and Preliminary Analyses
  4. Interdependence of Oil Price Returns Among OPEC Countries
  5. Overview of Time-Series Volatility Models
  6. Multivariate GARCH models
  7. Constant and Dynamic Conditional Correlation GARCH Models
  8. The Generalized Orthogonal GARCH Model
  9. Network-based Volatility models
  10. Network log-ARCH models
  11. Choice of the Network Weight Matrices and Dissimilarity Measures
  12. Results and interpretation
  13. Model Estimation
  14. Network Construction
  15. Forecasting Results
  16. ...and 12 more sections

Figures (5)

  • Figure 1: Panel (a) displays the monthly log returns and panel (b) shows the squared monthly log returns of oil prices from January 1983 to December 2024 for six selected OPEC countries: Algeria, Iran, Libya, Nigeria, Saudi Arabia, and the United Arab Emirates.
  • Figure 2: Network plots of the six OPEC oil-exporting countries based on six different distance measures. Raw data-based distances: (a) Euclidean distance, (b) Correlation-based distance. Model-based distances: (c) Piccolo-based distance, (d) CCC-GARCH, (e) DCC-GARCH, and (f) GO-GARCH. The grey colour scale of the edges is proportional to the weight degree of the connections.
  • Figure B.1: Kernel Density Estimates vs. Gaussian Benchmark
  • Figure B.2: Quantile-Quantile Plots: Standardized Returns vs. Student-$t$ Distribution
  • Figure F.1: Evolution of Network Weights for Different In-Sample Sizes ($T_0$). Note: Network plots of the six OPEC oil-exporting countries based on six different distance measures. Raw data-based distances: (a) Euclidean distance, (b) Correlation-based distance. Model-based distances: (c) Piccolo-based distance, (d) CCC-GARCH, (e) DCC-GARCH, and (f) GO-GARCH. The grey colour scale of the edges is proportional to the weight degree of the connections.