Attribution-State Conformal Forecasting: Residual-Independent Explanation States for Uncertainty Calibration Under Distribution Shift

Conformal prediction provides distribution-free coverage under exchangeability, but global calibration can conceal heterogeneous forecast errors. We introduce attribution-state conformal forecasting (ASCF), which learns calibration strata from residual-independent local explanation geometry. Exact TreeSHAP contributions are normalized, embedded in a square-root/Hellinger space, and clustered before residuals or test outcomes are observed. ASCF was evaluated on 1,381,005 source-level greenhouse-gas observations using separate fitting, state-learning, calibration, and test roles. At 90% nominal coverage, ASCF achieved coverage/interval-score pairs of 0.899/1.693 in the terminal temporal evaluation, 0.927/1.622 for unseen sources, and 0.929/2.294 across five leave-one-country-out folds in which ASCF was instantiable. Relative to global CP, interval-score reductions were 15.9%, 13.9%, and 18.6%. ASCF also produced lower paired interval scores than prediction-conditioned, raw-K matched, and country × gas Mondrian CP in three evaluations, with bootstrap support, reinforcing ASCF’s advantage among the residual-independent conditioning strategies evaluated. Persistence and adapted residual-aware grouping attained lower interval scores. Persistence showed stronger point-forecast accuracy, while an ASCF-plus-global-CP fallback policy extended geographic prediction to all eight countries. In a synthetic proof-of-mechanism, attribution states recovered oracle mechanisms better than matched raw-feature clustering (ARI 0.659 versus 0.040). SHAP-representation sensitivity affected state quality. Under exchangeability, a frozen state map inherits group-conditional split-conformal validity; deliberate shift experiments are interpreted empirically.