Retrieval-Corrected Conformal Prediction for Time Series

2026-08-11

AI Summary

Traditional conformal prediction suffers from inefficient calibration in time series due to temporal dependencies and dynamic shifts in prediction errors, while existing local methods are often limited by residual weighting that dilutes critical evidence. This work proposes Retrieval-Corrected Conformal Prediction (RCCP), which constructs asymmetric prediction intervals by retrieving historical one-sided residuals similar to the current forecast and applies a scalar conformal correction to normalized retrieval errors to explicitly rectify coverage bias. By directly leveraging locally relevant evidence, RCCP provides a theoretical bound on coverage gap grounded in error stability. Experiments across multiple benchmarks and backbone models demonstrate that RCCP consistently achieves target coverage with high precision, attains the lowest Winkler scores, incurs fewer severe failures, and maintains low calibration and inference overhead, offering both strong validity and scalability.

Abstract

Conformal prediction (CP) provides distribution-free prediction intervals for fixed forecasters, but its standard calibration procedure is often inefficient for time series data, where forecast errors are temporally dependent and change across time and operating conditions. Recent time series CP methods improve local calibration using recent, weighted, or localized residuals. Yet local calibration can remain indirect, since broad residual weighting or additional adaptation procedures may dilute the evidence most relevant to the current prediction. This motivates a simple retrieval and correction strategy that selects similar past residuals as local evidence and then corrects the coverage error left by retrieval. In this paper, we propose Retrieval--Corrected Conformal Prediction (RCCP), a retrieval-augmented calibration method for time series prediction intervals. RCCP builds an asymmetric interval from retrieved one-sided residuals and calibrates its normalized retrieval error with a scalar conformal correction. Thus, retrieval provides local residual evidence, while conformal correction determines the final scale needed for coverage. We provide a coverage-gap bound based on the stability of the normalized retrieval error distribution. Across standard benchmarks and backbone forecasters, RCCP attains the target coverage in every setting and achieves the lowest Winkler scores, with fewer severe misses. RCCP also achieves low calibration and inference overhead, showing that retrieval-corrected calibration is an effective and scalable approach to uncertainty quantification in time series forecasting. Code is available at https://github.com/jinsaaang/rccp.

Problem

• conformal prediction
• time series
• prediction intervals
• local calibration
• temporal dependence

Innovation

• Conformal Prediction
• Time Series Forecasting
• Retrieval-Augmented Calibration
• Uncertainty Quantification
• Local Residual Retrieval