Adaptive Conformal Prediction with Foundation Models

Adaptive conformal prediction methods adjust prediction intervals online to maintain valid coverage under distribution shift. We study how predictor quality interacts with method effectiveness. We evaluate 21 conformal prediction methods and variants, including Conformal PID, SPCI, HopCPT, and DistMatch run from their original codebases, across four predictor classes spanning GBDT, Random Forest, the Chronos foundation model, and the MOMENT foundation model, on ten domains covering six time-series benchmarks, battery degradation, high-harmonic generation, wearable sensors, and battery cycling. Our key finding is that predictor quality governs the effectiveness of adaptive conformal prediction methods. The method-to-method spread in conditional coverage shrinks from 0.335 with GBDT to 0.150 with Chronos as predictor quality improves. Methods that exploit structured residual autocorrelation such as DistMatch and SPCI show the largest sensitivity to predictor changes, while proactive methods such as Trend, BiTrend, and Multi-Strategy FACI remain robust across predictor types. Conformal PID achieves the best conditional coverage with tree-based predictors but degrades when switching to a foundation model, whereas our Trend method improves and overtakes it under Chronos.