Scholarlink

<p>Accurate forecasting of CO<sub>2</sub> allowance prices remains difficult because carbon markets differ substantially in maturity, liquidity, and volatility across jurisdictions. This study compares five deep learning architectures, Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Network (TCN), under a strict one-step-ahead rolling-window design across seven carbon markets: the EU, China, Australia, New Zealand, Korea, California, and the Regional Greenhouse Gas Initiative (RGGI), USA. Using technical, financial, and macroeconomic indicators, this study evaluates model performance under common tuning and out-of-sample scoring rules across multiple machine learning models with various technical and macroeconomic indicators. The results show that forecasting performance is more market-dependent than architecture-dependent: tuned LSTM, GRU, and TCN models are consistently competitive, technical indicators contain most of the short-run predictive signal, and the 10-day window usually delivers slightly better accuracy than the 20-day window. These findings provide an empirical cross-market benchmark for climate-finance forecasting and suggest that short-run forecasting systems can support market trend monitoring and financial planning. Through comprehensive hyperparameter optimization and cross-market comparison, this research demonstrates the use of deep learning with hyperparameter tuning in the context of forecasting accuracy and interpretability.</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

Comparative market dynamics using deep learning architectures for forecasting CO2 allowance prices across multiple jurisdictions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月18日

查看全文
  • Jeonghoe Lee,
  • Jim H. Yang,
  • Yewen Li

摘要

Accurate forecasting of CO2 allowance prices remains difficult because carbon markets differ substantially in maturity, liquidity, and volatility across jurisdictions. This study compares five deep learning architectures, Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Network (TCN), under a strict one-step-ahead rolling-window design across seven carbon markets: the EU, China, Australia, New Zealand, Korea, California, and the Regional Greenhouse Gas Initiative (RGGI), USA. Using technical, financial, and macroeconomic indicators, this study evaluates model performance under common tuning and out-of-sample scoring rules across multiple machine learning models with various technical and macroeconomic indicators. The results show that forecasting performance is more market-dependent than architecture-dependent: tuned LSTM, GRU, and TCN models are consistently competitive, technical indicators contain most of the short-run predictive signal, and the 10-day window usually delivers slightly better accuracy than the 20-day window. These findings provide an empirical cross-market benchmark for climate-finance forecasting and suggest that short-run forecasting systems can support market trend monitoring and financial planning. Through comprehensive hyperparameter optimization and cross-market comparison, this research demonstrates the use of deep learning with hyperparameter tuning in the context of forecasting accuracy and interpretability.