Carbon emission pricing prediction based on autoregressive history and external signals: a multi-frequency deep time series network based on MF-VAR-LSTM

• Hanlong Xie,
• Tianhuiyu Zou,
• Fangli Fu,
• Yanji Piao,

Vol. 19, No. 11, November 30, 2025
10.3837/tiis.2025.11.005, Download Paper (Free):

• Carbon Market Price Prediction
• Policy Sentiment Analysis
• Bidirectional long short-term memory network
• Deep Learning

Abstract

In order to solve the problem of multi-source heterogeneous data integration and nonlinear modeling in China carbon market price prediction, this study proposes a hybrid prediction framework (MF-VAR-LSTM) that integrates mixed frequency VAR model and bidirectional LSTM. To address this, an intelligent mixed-frequency data processor is constructed to integrate three types of data from 2022 to 2023: daily trading data of China’s carbon market, monthly nighttime light satellite data, and unstructured policy texts. Notably, data from 2022 to 2023 is adopted because the 2024 nighttime light data has not been fully released. This processor effectively resolves the issues of data frequency differences and structural heterogeneity. The model combines the VAR model’s ability to capture linear cointegration relationships between variables with the LSTM network’s deep learning advantage in modeling nonlinear time series features. It also introduces a residual learning mechanism to dynamically fuse the outputs of these two components. The results show that the prediction performance of the model is significantly better than that of the traditional method, with the R² of the test set reaching 0.8545 and the RMSE of 0.7412, which is significantly better than that of the single VAR and LSTM models. The introduction of policy sentiment index and nighttime light data reduces the prediction error and verifies the effectiveness of multi-source information fusion. The characteristic analysis shows that historical price, policy intensity and trading activity are the core predictors, and the policy signal presents a 1-2 days leading trend at the market turning point. The innovation of the research is reflected in the dynamic weight fusion architecture, the heterogeneous frequency data alignment algorithm and the policy sentiment quantification system, which provides a high-precision and highly explanatory analysis tool for carbon market risk management and policy evaluation. Future research can be extended to multi-regional linkage forecasting and real-time online learning frameworks to further support the realization of the "dual carbon" goal.

Statistics

Cite this article