Do Time Series Foundation Models Forecast Carbon Prices? Leakage-Free Evidence from the EU ETS and China’s Carbon Markets

Time series foundation models (TSFMs) promise accurate zero-shot forecasts, but their value for carbon markets has been examined only in the European Union Emissions Trading System (EU ETS). We evaluate Chronos-2 and TimesFM-2.5 alongside six benchmark and hybrid models in four markets (the EU ETS, China’s national ETS, and the Hubei and Guangdong pilots) at horizons of 1–60 trading days, using a rolling-origin design without information beyond the forecast origin. No model significantly outperforms the random walk in the EU ETS. TSFMs achieve the lowest point and probabilistic forecast errors in the Hubei pilot, but their point forecasts are less accurate than the random walk in the national market and, over the full test period, in Guangdong. In the national market, Chronos-2 errors reach 22.8% at the 60-day horizon and its 80% intervals cover only 62% of outcomes, whereas TimesFM-2.5 attains the lowest quantile loss despite under-coverage. The national-market and Hubei results hold over a common calendar period and for targets postdating the model weights. Across the markets and periods studied, gains are associated with a large transitory component in daily returns and losses with persistent price moves. Look-ahead bias roughly halves the apparent error of decomposition-based models. TSFMs therefore require market-specific validation before deployment.