Carbon price forecasting based on time-lagged cross-correlation dynamic graphs and a dual-stream spatiotemporal fusion network

Carbon prices respond asynchronously to energy, financial, and international carbon-market signals, yet existing graph-based forecasting methods rarely encode these lead–lag structures as directed priors while retaining sample-level adaptability. This study proposes TLCC-DyG-STFusion, a dual-stream forecasting framework in which time-lagged cross-correlation constructs a sparse directed association prior and a learnable coefficient fuses that prior with a sample-level dynamic graph before spatiotemporal encoding. The framework was evaluated on the Hubei and Guangdong pilot markets using 68 rolling test blocks and direct 1-, 3-, and 5-step forecasts. It achieved the lowest five-seed mean root mean squared error in all six market–horizon settings. Harvey–Leybourne–Newbold-corrected Diebold–Mariano tests with Holm adjustment remained significant in 56 of 60 baseline comparisons. Configuration-level ablations favored the full model in all six settings, while single-seed hyperparameter sweeps showed relative RMSE changes from −4.65% to +2.13%. Independently recomputed prior-graph evidence showed high adjacent-window rank continuity (Spearman means of 0.858 in Hubei and 0.929 in Guangdong) while longer-span factor composition changed, supporting a stable-but-adaptive interpretation rather than a causal one. The results demonstrate broad cross-market competitiveness for the same framework while delimiting four comparisons in which adjusted significance was not established.