The arXiv preprint arxiv.org/abs/2604.13253 introduces Bias-Corrected Adaptive Conformal Inference (BC-ACI), which augments Adaptive Conformal Inference (ACI) with an online exponentially weighted moving average estimate of forecast bias. BC-ACI corrects nonconformity scores prior to quantile computation and recenters prediction intervals, differing from ACI's sole adjustment of quantile thresholds. Experiments across 688 runs on two base models, four synthetic regimes, and three real datasets show 13-17% lower Winkler scores under mean and compound distribution shifts (Wilcoxon p < 0.001), with equivalent performance on stationary data (ratio 1.002x). Finite-sample analysis indicates coverage guarantees degrade gracefully with bias estimation error.

Original ACI coverage (arxiv.org/abs/2103.01261) established distribution-free guarantees under distribution shift but left unaddressed the symmetric widening around biased point forecasts, a limitation also present in conformalized quantile regression (arxiv.org/abs/1905.03222). BC-ACI synthesizes ACI with established online bias tracking via EWMA, a technique rooted in sequential learning literature, and adds a dead-zone threshold to avoid over-correction on well-calibrated series. Prior time-series conformal work (arxiv.org/abs/2208.08441) focused on coverage but rarely quantified efficiency gains under compound shifts common in energy and financial forecasting.

The preprint's 2026 submission date follows accelerated adoption of conformal methods post-2021; mainstream ML forecasting toolkits still default to uncalibrated bootstrap intervals that fail under regime change. BC-ACI fills the documented gap between theoretical coverage and practical interval tightness for multi-horizon horizons where bias compounds.