Information Reliability in Carbon Markets: Point-in-Time News-Augmented Forecasting and Implications for China’s Energy Transition

Abstract

Reliable carbon price forecasts support compliance planning and transition risk management, but multi-source backtests can appear accurate when observations are aligned by calendar date rather than actual availability. We develop a point-in-time framework for China’s Hubei, Guangdong, and national emissions trading markets from July 2021 to December 2025. The design explicitly recognizes that the national ETS was power-sector-dominated through 2024 and entered a multi-sector transition during 2025. Every market, energy, policy, macroeconomic, weather, and news observation receives a first-usable timestamp and a market-specific cutoff. Matched Temporal Fusion Transformer (TFT) forecasts isolate the incremental value of audited news under nested rolling-origin estimation, dependence-aware inference, and past-only conformal calibration. Relative to the same learner using the same structured information, eligible news reduces RMSE by 4.8% (95% stationary-block-bootstrap interval: 1.2–8.3%) and WIS by 3.6% (0.5–6.8%). The gain is conditional: it is strongest at one day, in Hubei and the national ETS, and during high-volatility, compliance, policy-intensive, and energy-shock states; five-day effects are imprecise. Random Forest also produces a positive matched news gain of 3.9%, indicating that the information result is not unique to TFT. National-ETS event-window exclusions leave gains of 4.6–4.9%, while excluding reform-tagged news leaves 4.4%. Rolling conformal adjustment reduces mean absolute coverage error from 8.7 to 2.1 percentage points and normalized interval width by 9.7%. A combined invalid alignment instead reports an apparent 10.9% gain. Information reliability therefore changes both measured forecast performance and its energy transition interpretation.