Forecasting Carbon Price Volatility in China's Ets a Policy-augmented Deep Learning Frame-work Inte-grating Innovation-driven Policies and Green Investment Thresholds

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a novel methodology combining causal inference and deep learning for carbon price volatility, useful for further research in carbon market dynamics and policy evaluation.

🏢実務担当者:Offers a forecasting model and insights into key volatility drivers, aiding risk management and investment decisions in carbon markets.

🏛政策担当者:Demonstrates that consistent innovation policies and green investment thresholds can reduce carbon price volatility, providing evidence-based guidance for market design and stabilization.

Abstract The stability of carbon pricing mechanisms is fundamental to achieving global climate targets, yet carbon price volatility in emerging markets like China's Emissions Trading Scheme (ETS) poses significant challenges to market efficiency and investment certainty. Existing forecasting models predominantly focus on statistical patterns while neglecting the profound influence of policy interventions and firm-level behavioral dynamics. This study addresses this critical gap by developing a comprehensive analytical framework that integrates causal inference with advanced machine learning to both explain and pre-dict carbon price volatility. We make four primary contributions. First, grounded in sig-naling theory, we construct a dynamic Innovation-Driven Policy (IDP) Index through sys-tematic textual analysis of 5,820 government policy documents spanning 2013–2023, quantifying the intensity and consistency of green innovation signals. Second, drawing from real options theory, we empirically identify a statistically significant Green Invest-ment Threshold (GIT) at 28.5 billion RMB monthly green bond issuance using Hansen's threshold regression model, revealing distinct market volatility regimes. Third, we devel-op a novel hybrid forecasting architecture, the VMD-Attention-GRU (VAG) model, which synergistically combines Variational Mode Decomposition (VMD) for noise reduction, an attention mechanism for dynamic feature weighting, and Gated Recurrent Units (GRU) optimized via the Sparrow Search Algorithm (SSA). Fourth, we employ SHAP (SHapley Additive exPlanations) analysis to achieve unprecedented model interpretability, quanti-fying the marginal contribution of each factor to volatility forecasts. Utilizing comprehen-sive data from China's pilot and national ETS markets (2014–2023, N = 2,452), our causal analysis demonstrates that innovation-driven policies significantly dampen long-term volatility (cumulative abnormal volatility reduction of 8.5% following major policy an-nouncements, p < 0.05), while the identified investment threshold creates a structural break in volatility persistence (autoregressive coefficient drops from 0.45 to 0.28 above threshold). The VAG model achieves superior forecasting performance, reducing RMSE by 27.8% compared to standalone GRU and by 14.3% compared to VMD-SSA-GRU without policy variables (test set RMSE = 0.0078, MAPE = 17.2%). SHAP analysis confirms that policy sig-nals and investment regime indicators rank as the second and third most influential pre-dictors after market memory effects, validating our theoretical framework. These findings provide actionable insights for policymakers to design stabilizing interventions and for market participants to navigate the complexities of carbon markets in the context of global sustainability transitions.

• openaire https://doi.org/10.21203/rs.3.rs-8326151/v1first seen 2026-05-14 21:18:05