🤖 machine learning

Retrieval-Corrected Conformal Prediction for Time Series

This paper introduces Retrieval-Corrected Conformal Prediction (RCCP), a scalable method for time series uncertainty quantification that combines retrieval of similar past residuals with a scalar conformal correction to achieve target coverage and minimize prediction interval width across diverse benchmarks.

Original authors: Sangjin Jin, Kangmin Kim, Junhyeong Lee, Yongjae Lee

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Sangjin Jin, Kangmin Kim, Junhyeong Lee, Yongjae Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster. You look at the clouds, the wind, and the barometer, and you confidently predict, "It will rain at 2:00 PM." But a good forecaster knows that nature is tricky. They don't just give a single time; they give a window: "It will rain between 1:45 PM and 2:15 PM." This window is called a prediction interval. The goal is to make this window just right: not so wide that it's useless (like saying "it might rain sometime today"), and not so narrow that you get caught in the rain without an umbrella.

To build these windows, scientists use a clever trick called Conformal Prediction. Think of it like a "calibration check." Before you trust your forecast for tomorrow, you look at how wrong your past predictions were. If you were usually off by 10 minutes, you make your window 10 minutes wide on both sides. This works great if the weather is boring and predictable. But time series data—like stock prices, electricity usage, or traffic—changes constantly. The "weather" of the data shifts. Sometimes errors are tiny; sometimes they are huge. If you use a single, static rule to set your window size, you might be too safe when things are calm and dangerously risky when things get chaotic.

This is where a new study by Sangjin Jin and colleagues from the Ulsan National Institute of Science and Technology comes in. They are tackling the problem of how to build these safety windows for data that changes its mind. They propose a method called Retrieval–Corrected Conformal Prediction (RCCP). Instead of guessing the window size based on all past mistakes, or just the most recent ones, their method acts like a super-smart librarian. It hunts through history to find moments that look exactly like the current situation, grabs the mistakes made during those specific moments, and uses them to build a custom window. But here's the kicker: they realized that just grabbing similar past mistakes isn't enough to guarantee safety. So, they added a final "correction" step to fix any remaining gaps, ensuring the window is both sharp and reliable.

The Problem: The "One-Size-Fits-All" Trap

Imagine you are trying to guess how long it takes to walk to school. If you look at your entire history of walks, you might find that sometimes you run (5 minutes) and sometimes you get stuck in traffic (20 minutes). If you just take the average, you might set your window at 12 minutes. But if today is a rainy Tuesday with a heavy backpack, the "average" doesn't help. You need a window that knows it's a rainy Tuesday.

Old methods tried to solve this by looking at the most recent walks or by giving more weight to similar days. But the authors found a flaw: these methods often dilute the evidence. They might mix in a "similar" day that was actually quite different, or they might rely on complex math that gets messy. It's like trying to find a needle in a haystack by looking at the whole haystack, or by looking at just the last handful of hay you grabbed. You might miss the needle, or you might grab a piece of straw that looks like a needle but isn't.

The Solution: The "Time-Traveling Librarian"

The authors' new method, RCCP, works in two distinct steps, like a two-person team: the Retriever and the Corrector.

Step 1: The Retrieval (Finding the Twins)
First, the system looks at the current moment (say, a rainy Tuesday with a heavy backpack). It then dives into a "knowledge base" of all past moments. It doesn't just look at the last few days; it searches for any past moment that feels the same. Maybe it finds a rainy Tuesday from three years ago where you also had a heavy backpack.
Once it finds these "twins," it looks at how wrong the prediction was on those specific days. Did you get there in 5 minutes or 20? It builds a custom, asymmetric window based only on these twins. If the twins usually ran late on the right side but were early on the left, the window stretches more to the right. This is the "Retrieval" part. It's like asking your friends who have been in your exact shoes, "How long did it take you?"

Step 2: The Correction (The Safety Net)
Here is the genius part. The authors realized that even if you find the perfect twins, their past mistakes might not perfectly match the current reality. Maybe the twins were slightly off, or maybe the library of twins is a bit small. If you just used the twins' mistakes as the final answer, you might still be wrong.
So, RCCP adds a Correction step. It takes the "Retrieved Window" and checks it against a separate set of data to see how much it needs to be stretched or shrunk to be 100% safe. It calculates a single "correction factor"—a magic number. If the retrieved window was too tight, the factor is 1.1 (stretch it 10%). If it was too loose, the factor is 0.9 (shrink it 10%).
This separates the job: the Retriever figures out the shape of the window (is it wide? is it lopsided?), and the Corrector figures out the scale (how big should it be to be safe?).

What They Found: Sharper, Safer Windows

The team tested this method on real-world data, including electricity usage, wind power, and solar energy. They compared RCCP to other popular methods like Split Conformal Prediction (the simple average method) and EnbPI (a method that updates based on recent errors).

The results were impressive. In their tests, RCCP managed to hit the target safety level (hitting the "target coverage") in every single setting they tried. Other methods often missed the mark, either being too wide (wasting information) or too narrow (missing the target too often).

Specifically, they looked at a score called the Winkler score, which measures how good the window is (lower is better). RCCP achieved the lowest Winkler scores across the board. This means their windows were the most efficient: they were narrow when the data was predictable and wide when it was chaotic, without sacrificing safety.

One of the most interesting findings was about "severe misses." Sometimes, a prediction interval fails, and the real value is way outside the window. The authors found that RCCP had fewer severe misses. Even when it was wrong, the real value was usually just a little bit outside the window, rather than way off. It's the difference between missing a target by an inch versus missing it by a mile.

They also checked how fast the method was. While some other methods took a long time to calculate, RCCP was surprisingly fast, taking only about 70 seconds to calibrate on the Air dataset (compared to over 600 seconds for some others). This suggests it's not just accurate, but also practical for real-time use.

Why This Matters

The paper suggests that by combining the "search for similar pasts" with a "final safety check," we can build better safety nets for the future. We don't have to guess if the future will be like the past; we can look for the specific past that matches the present, and then double-check our work.

The authors admit that their method depends on how well they can describe the "current moment" to find the right twins. If the description is bad, the retrieval might fail. But for now, this "Retrieval-Corrected" approach seems to be a powerful new tool for anyone trying to forecast the unpredictable, from stock traders to power grid managers. It turns the chaotic mess of time-series data into a story where the past helps us navigate the future, one corrected step at a time.

Technical Summary: Retrieval–Corrected Conformal Prediction for Time Series

Problem Definition

Conformal Prediction (CP) offers a distribution-free framework for constructing prediction intervals with guaranteed coverage for fixed forecasters. However, standard split CP assumes exchangeability between calibration and test scores, an assumption frequently violated in time series forecasting due to temporal dependence, heteroskedasticity, and distribution shifts.

While recent time series CP methods attempt to address this by using local calibration (e.g., weighting recent residuals or using learned similarity), they often rely on broad weighting schemes or additional adaptation procedures. These approaches can dilute the most relevant evidence for the current prediction, leading to indirect calibration. Furthermore, purely retrieval-based methods that select similar past residuals lack an inherent mechanism to guarantee coverage, as the retrieved set is finite and data-dependent. The core challenge is how to leverage specific, similar past residuals to form a sharp local interval while simultaneously correcting the residual coverage error to ensure valid statistical guarantees.

Methodology: Retrieval–Corrected Conformal Prediction (RCCP)

The authors propose RCCP, a framework that decouples the construction of local residual evidence from the final coverage calibration. The method operates in two distinct stages:

Retrieval-Augmented Local Interval Construction:
- Knowledge Base: RCCP maintains a time-ordered knowledge base of past prediction contexts. Each entry consists of a retrieval key (derived from input windows, forecaster states, or embeddings) and the realized one-sided residuals ( $e^+$ for positive errors, $e^-$ for negative errors).
- Retrieval: For a new prediction context, the system retrieves the $K$ most similar past contexts using a distance metric (e.g., Euclidean distance).
- Asymmetric Interval: Instead of assuming symmetric errors, RCCP computes weighted quantiles separately for the positive and negative retrieved residuals. This forms an initial, asymmetric local interval $\tilde{C}_t = [\hat{Y}_t - \tilde{R}^-_t, \hat{Y}_t + \tilde{R}^+_t]$ . This step adapts the interval shape to the local error distribution without training a new model.
Scalar Conformal Correction:
- Normalization: The retrieved interval is not used directly as the final output. Instead, RCCP evaluates the "normalized retrieval error" on a calibration set. For each calibration point, it computes a scalar score $B_j$ representing the ratio of the actual residual to the retrieved bound (e.g., $(Y_j - \hat{Y}_j)^+ / \tilde{R}^+_j$ ).
- Correction Factor: A scalar correction factor $\hat{c}$ is estimated as the empirical $(1-\alpha)$ -quantile of these normalized scores.
- Final Interval: The final prediction interval is obtained by scaling the retrieved asymmetric bounds by $\hat{c}$ : $C^{RCCP}_t = [\hat{Y}_t - \hat{c}\tilde{R}^-_t, \hat{Y}_t + \hat{c}\tilde{R}^+_t]$ .

This separation allows retrieval to determine the local shape and scale of the interval, while the scalar correction ensures the final coverage meets the target level.

Key Contributions

Direct Residual Retrieval: The paper applies residual retrieval as a direct alternative to broad weighting or complex adaptation procedures, selecting evidence from similar past contexts to define local intervals.
RCCP Framework: It introduces a novel framework combining asymmetric retrieved intervals with a scalar conformal correction factor.
Theoretical Guarantees: The authors provide a coverage-gap bound based on the stability of the normalized retrieval error distribution. They establish that RCCP achieves asymptotic coverage if the distribution of normalized retrieval errors is stable between calibration and test sets and the empirical correction is consistent.
Empirical Performance: Extensive experiments demonstrate that RCCP achieves target coverage across various benchmarks and backbone forecasters (LSTM, Transformer) while minimizing Winkler scores and reducing severe misses compared to state-of-the-art baselines.

Experimental Results

The authors evaluated RCCP on four benchmark datasets (Air, Solar, Electricity, Wind) using LSTM and Transformer backbones, comparing against Split CP (SCP), adaptive methods (EnbPI, SPCI), and localized methods (NexCP, HopCPT, ResCP).

Coverage and Efficiency: RCCP consistently attained the target coverage (e.g., 90% for $\alpha=0.1$ ) across all settings, avoiding the severe undercoverage observed in some baselines. It achieved the lowest Winkler scores, indicating a superior trade-off between interval width and coverage violations.
Miss Severity: Analysis of miscoverage showed that when RCCP intervals failed to cover the true value, the deviation was typically smaller than in other methods. This suggests RCCP effectively reallocates interval width to high-error regions rather than uniformly widening intervals.
Adaptivity: RCCP demonstrated high width adaptivity, significantly widening intervals for high-error deciles (e.g., a width ratio of 20.86 on the Solar dataset for the highest error decile vs. the lowest), whereas other methods showed less sensitivity to realized error magnitude.
Computational Cost: RCCP maintained low calibration and inference overhead, performing more efficiently than complex learned local methods like HopCPT while offering better coverage guarantees.

Significance and Claims

The paper claims that RCCP offers a simple, scalable, and effective approach to uncertainty quantification in time series forecasting. By separating the retrieval of local evidence from the conformal correction, the method avoids the "dilution" of evidence common in weighted approaches and the lack of calibration in pure retrieval methods. The authors assert that this two-stage process provides locally informative intervals that are rigorously calibrated, achieving a better balance between sharpness and reliability than existing time series CP methods. The approach is presented as a practical solution that requires no additional training of the base forecaster and operates with minimal computational overhead.