Search papers, labs, and topics across Lattice.
This paper introduces Retrieval-Corrected Conformal Prediction (RCCP), a novel method for improving the calibration of prediction intervals in time series forecasting by leveraging similar past residuals for local evidence. By integrating a retrieval strategy with a conformal correction mechanism, RCCP effectively addresses the coverage errors associated with traditional calibration methods, ensuring that the final prediction intervals maintain target coverage while minimizing severe misses. The method demonstrates superior performance across standard benchmarks, achieving the lowest Winkler scores and maintaining low calibration overhead, thereby enhancing uncertainty quantification in time series predictions.
Retrieval-Corrected Conformal Prediction achieves target coverage with fewer severe misses, revolutionizing uncertainty quantification in time series forecasting.
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.
