Fusing causal inference and explainable AI: dynamic mechanisms and extreme event simulation in carbon markets

Carbon market resilience to price volatility and extreme events is essential for credible climate policy. Conventional analytical models often conflate correlation with causation, weakening mechanistic interpretation and stress testing. We develop an explainable machine learning framework grounded in double or debiased causal inference to address this gap. The resulting identification clarifies effect directions and stabilizes downstream interpretability diagnostics. Using China's national Emission Trading System allowance price, which is primarily determined by power-sector compliance trading during our sample period, we isolate six dominant determinants and trace systematic shifts in their influence across seasons and market regimes. We then subject the market to thirteen counterfactual stress scenarios. Concurrent shocks to the Global Clean Energy Index and coal prices, consistent with the surge during the Russo-Ukrainian war, induce a 25.91% decline in the average allowance price and reveal a critical vulnerability. Among twelve mitigation instruments, an auction reserve price floor offers the most favorable trade-off across environmental effectiveness, economic efficiency, and risk containment. The findings provide a quantitative basis for designing robust allowance reserves and price-stabilization mechanisms.