Scholarlink

<p>Carbon emission trading systems exhibit pronounced nonstationarity and policy-related volatility, which complicates reliable forecasting and risk-oriented monitoring. Many econometric and deep learning models assume near-synchronous effects and therefore struggle to represent time-lagged responses between carbon prices and external drivers such as energy markets and policy uncertainty. We propose a Dynamic Graph-based Carbon Price Network that uses Dynamic Time Warping to construct window-wise interaction graphs capturing lead–lag dependence among drivers, and combines graph convolution with temporal convolution to learn evolving spatiotemporal dynamics. Predictive uncertainty is obtained by Monte Carlo dropout at inference to form prediction intervals. Experiments on China’s pilot markets in Hubei, Shanghai, and Guangdong under time-ordered evaluation show improved accuracy and stable behavior during volatile periods and regime transitions. Perturbation-based attribution and hierarchical feature removal further indicate that policy uncertainty provides informative predictive signals alongside macro-financial indicators, international benchmarks, and energy cost proxies, while lower-ranked variables contribute complementary information under turbulence. These findings support uncertainty-aware monitoring of market stability and time-lagged shock transmission in emerging carbon markets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lead–lag Aware Carbon Price Forecasting with Dynamic Graphs and Uncertainty Intervals

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Ke Ren,
  • Weiyu Zhang,
  • Haoxiang Chang,
  • Yang Zhou,
  • Chengyao Jin,
  • Ye Zhang

摘要

Carbon emission trading systems exhibit pronounced nonstationarity and policy-related volatility, which complicates reliable forecasting and risk-oriented monitoring. Many econometric and deep learning models assume near-synchronous effects and therefore struggle to represent time-lagged responses between carbon prices and external drivers such as energy markets and policy uncertainty. We propose a Dynamic Graph-based Carbon Price Network that uses Dynamic Time Warping to construct window-wise interaction graphs capturing lead–lag dependence among drivers, and combines graph convolution with temporal convolution to learn evolving spatiotemporal dynamics. Predictive uncertainty is obtained by Monte Carlo dropout at inference to form prediction intervals. Experiments on China’s pilot markets in Hubei, Shanghai, and Guangdong under time-ordered evaluation show improved accuracy and stable behavior during volatile periods and regime transitions. Perturbation-based attribution and hierarchical feature removal further indicate that policy uncertainty provides informative predictive signals alongside macro-financial indicators, international benchmarks, and energy cost proxies, while lower-ranked variables contribute complementary information under turbulence. These findings support uncertainty-aware monitoring of market stability and time-lagged shock transmission in emerging carbon markets.