arxiv: 2605.05497 · v2 · submitted 2026-05-06 · 💻 cs.LG

Recognition: no theorem link

Online Localized Conformal Prediction

Yuheng Lai , Garvesh Raskutti

Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:46 UTC · model grok-4.3

classification 💻 cs.LG

keywords online conformal predictionlocalized calibrationadaptive prediction intervalsonline convex optimizationcoverage guaranteesheterogeneous covariatestime series

0 comments

The pith

Localizing calibration to similar covariates produces narrower valid online prediction sets.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses conformal prediction in online and time-series settings where data exchangeability fails. Global calibration methods like adaptive conformal inference achieve long-run coverage but produce inefficiently wide sets under covariate heterogeneity. The authors introduce online localized conformal prediction that restricts calibration to nearby covariates and adds an online convex optimization step to select the localization bandwidth automatically. This combination is shown to preserve long-run coverage while yielding narrower intervals, with supporting theory and experiments on simulated and real data.

Core claim

Online Localized Conformal Prediction (OLCP) pairs online adaptation of the conformal threshold with localization of nonconformity scores to covariate-similar points. OLCP-Hedge further casts bandwidth choice as an online expert aggregation task solved by constrained convex optimization. Both procedures come with long-run coverage guarantees and, in simulations and real-data experiments, deliver valid coverage with narrower prediction sets than global baselines.

What carries the argument

Covariate-dependent localization of nonconformity scores together with online convex optimization for automatic bandwidth hedging.

If this is right

Long-run coverage remains valid even without exchangeability.
Prediction sets become narrower by using only locally relevant calibration data.
Bandwidth sensitivity is reduced by treating localization radius as an online expert problem.
The methods outperform standard adaptive conformal inference on heterogeneous data streams.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The localization idea could extend to streaming settings such as online reinforcement learning where state similarity matters.
Dynamic or learned localization radii might further improve efficiency beyond fixed-bandwidth hedging.
Narrower valid sets would directly reduce over-conservatism in sequential forecasting and control tasks.

Load-bearing premise

The underlying online data process must allow the localized scores and online updates to produce long-run coverage at the nominal level.

What would settle it

A long online sequence in which the empirical coverage rate of the localized method falls materially below the nominal level while the global baseline does not.

Figures

Figures reproduced from arXiv: 2605.05497 by Garvesh Raskutti, Yuheng Lai.

**Figure 1.** Figure 1: Diagnostics for simulation. Left panels show Scenario B conditional coverage and average size across Xt. Right panels show Scenario C rolling coverage and rolling average size with window size 100; the vertical dashed line marks the change point. Shaded bands show mean ± one standard deviation across repetitions. OLCP (DtACI) curves are partially hidden behind OLCP-Hedge (ACI). • A: Stationary: Yt = 0.5Yt−… view at source ↗

**Figure 2.** Figure 2: ELEC2 rolling diagnostics. Top: rolling coverage using a one-week window (48 × 7 half-hourly observations). Bottom: rolling average interval size using the same window. The dashed line marks target coverage 0.90. 22 view at source ↗

**Figure 3.** Figure 3: ILINet rolling diagnostics. Top: rolling coverage over weekly test observations. Bottom: rolling average interval size. The horizontal dashed line marks target coverage 0.90 while the vertical line marks the start of COVID. 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Coverage ETF volatility: rolling coverage (window=20 trading days) 2020 2021 2022 2023 2024 2025 2026 Calendar time 2 4 6 8 10 12 Average size (%) ETF volati… view at source ↗

**Figure 4.** Figure 4: ETF volatility rolling diagnostics. Top: rolling coverage over daily test observations. Bottom: rolling average interval size, reported in percentage points of absolute log return. The dashed line marks target coverage 0.90. Across the rolling diagnostics, SPCI is consistently much smaller but also persistently below the coverage target, indicating that its residual autoregression is not sufficiently calib… view at source ↗

read the original abstract

Conformal prediction is a framework that provides valid uncertainty quantification for general models with exchangeable data. However, in the online learning and time-series settings, exchangeability is not satisfied. Existing online conformal methods, such as adaptive conformal inference (ACI), can achieve long-run validity, yet they remain inefficient under covariate heterogeneity because they rely on global calibration. We propose \emph{Online Localized Conformal Prediction (OLCP)}, which combines online adaptation with covariate-dependent localization to better reflect heterogeneity. To reduce sensitivity to the localization bandwidth, we further develop \emph{OLCP-Hedge}, which performs bandwidth selection as an online expert aggregation problem using a constrained online convex optimization framework. Importantly, we provide coverage guarantees for both algorithms and demonstrate through simulations and real-data experiments that the proposed methods attain valid long-run coverage with narrower prediction sets than existing baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper pairs covariate-localized calibration with online hedge aggregation for bandwidth in non-exchangeable settings, but the coverage claim rests on local sample size growth that the abstract leaves underspecified.

read the letter

The new piece here is taking standard online conformal methods like ACI and restricting the calibration set to a covariate neighborhood whose size is controlled by a bandwidth that itself gets updated online via hedge-style expert aggregation. That combination is not in the prior work they cite, and it directly targets the inefficiency of global calibration when the data distribution varies across regions. The experiments back this up with narrower sets on both simulated heterogeneous streams and real datasets while still hitting the target long-run coverage rate. That part is useful and cleanly presented. The coverage argument itself follows the usual online convex optimization route on the miscoverage indicator, which is fine as far as it goes. The soft spot is exactly the one the stress-test flags. Localization shrinks the effective sample at each step, so in low-density covariate regions the local quantile can stay high-variance even after many observations arrive globally. Without explicit conditions on local density or mixing rates, the regret bound does not automatically translate into long-run coverage at level alpha. The abstract asserts the guarantee without listing those conditions, which leaves the central claim less secured than it needs to be. A referee would want to see either a clear density assumption or a modified analysis that accounts for regions where the local count grows too slowly. This is the kind of paper that belongs in a conformal-prediction or online-learning venue. Readers already working on adaptive uncertainty quantification will get immediate value from the algorithm and the empirical comparison. It is worth sending to peer review because the algorithmic idea is straightforward to implement and the experiments show a practical gain, even though the theory will probably need tightening on the local-sample-size issue.

Referee Report

3 major / 2 minor

Summary. The paper proposes Online Localized Conformal Prediction (OLCP), which augments online conformal methods with covariate-dependent localization to handle heterogeneity, and OLCP-Hedge, which treats bandwidth selection as an online expert aggregation problem solved via constrained online convex optimization. It asserts coverage guarantees for both algorithms and reports that simulations and real-data experiments show valid long-run coverage together with narrower prediction sets than global baselines such as adaptive conformal inference.

Significance. If the stated coverage guarantees hold, the work would meaningfully extend online conformal prediction to heterogeneous streaming settings by replacing global calibration with localized, adaptively tuned neighborhoods. The framing of bandwidth selection as an online convex optimization problem with expert aggregation is a clean technical contribution that could reduce manual tuning. Empirical demonstrations of narrower sets while preserving coverage would be practically useful, provided the theoretical conditions are made explicit and verifiable.

major comments (3)

[§3] §3 (Coverage Guarantees): The long-run coverage claim for OLCP rests on sublinear regret of the online update applied to a local empirical quantile. Under covariate heterogeneity the effective local sample size is controlled by the bandwidth and the local density; the manuscript does not state density or mixing conditions that would guarantee the local sample size grows sufficiently fast for the quantile to concentrate. Without such conditions the regret argument does not automatically translate into long-run coverage at level α.
[§4.1] §4.1 (OLCP-Hedge Algorithm): The constrained online convex optimization formulation for bandwidth selection is presented, but the regret bound is stated only with respect to the expert loss; it is not shown that the selected bandwidth sequence preserves the coverage property of the underlying OLCP procedure when the local sample size is small. A concrete bound linking the Hedge regret to the deviation of the local quantile would be needed to support the joint guarantee.
[Table 2, Figure 3] Table 2 and Figure 3 (Real-data experiments): The reported coverage is close to the nominal level on average, yet the experiments do not stratify results by local density or by regions where the bandwidth yields fewer than, say, 50 observations. If coverage degrades in low-density strata, the claim of uniformly valid long-run coverage would be weakened.

minor comments (2)

The notation for the localization kernel and the online update rule is introduced without a compact summary table; adding one would improve readability.
The abstract states that OLCP-Hedge 'performs bandwidth selection as an online expert aggregation problem' but does not mention the specific loss function or the constraint set; these details appear only later and should be previewed.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments highlight important aspects of the theoretical conditions and empirical validation that we will strengthen in the revision. We address each major comment below.

read point-by-point responses

Referee: [§3] §3 (Coverage Guarantees): The long-run coverage claim for OLCP rests on sublinear regret of the online update applied to a local empirical quantile. Under covariate heterogeneity the effective local sample size is controlled by the bandwidth and the local density; the manuscript does not state density or mixing conditions that would guarantee the local sample size grows sufficiently fast for the quantile to concentrate. Without such conditions the regret argument does not automatically translate into long-run coverage at level α.

Authors: We agree that the long-run coverage argument requires the local empirical quantile to concentrate, which depends on the effective local sample size growing sufficiently fast. The current manuscript implicitly relies on the bandwidth choice to ensure this but does not state explicit conditions. We will add assumptions on the covariate density being bounded away from zero in the relevant support and on the mixing rate of the underlying process. Under these conditions we will show that the sublinear regret of the online update implies the desired long-run coverage at level α, and we will revise the statement and proof of the relevant theorem in Section 3. revision: yes
Referee: [§4.1] §4.1 (OLCP-Hedge Algorithm): The constrained online convex optimization formulation for bandwidth selection is presented, but the regret bound is stated only with respect to the expert loss; it is not shown that the selected bandwidth sequence preserves the coverage property of the underlying OLCP procedure when the local sample size is small. A concrete bound linking the Hedge regret to the deviation of the local quantile would be needed to support the joint guarantee.

Authors: The referee correctly notes that the existing regret bound is stated relative to the best fixed expert and does not yet explicitly connect to coverage preservation when local samples are limited. We will add a supporting lemma that uses the sublinear regret of the constrained Hedge algorithm to bound the probability that the selected bandwidth yields an insufficient local sample size. This will establish that the coverage property of the base OLCP procedure is preserved with high probability, thereby completing the joint guarantee for OLCP-Hedge. The new analysis will appear in Section 4. revision: yes
Referee: [Table 2, Figure 3] Table 2 and Figure 3 (Real-data experiments): The reported coverage is close to the nominal level on average, yet the experiments do not stratify results by local density or by regions where the bandwidth yields fewer than, say, 50 observations. If coverage degrades in low-density strata, the claim of uniformly valid long-run coverage would be weakened.

Authors: We acknowledge that average coverage does not fully address potential variation across density regimes. We will augment the experimental section with a new stratification of both coverage and interval width by estimated local density and by bins of realized local sample size. This will include separate reporting for low-density regions (e.g., fewer than 50 local observations). Any observed degradation will be discussed explicitly as a limitation of the method in heterogeneous settings. revision: yes

Circularity Check

0 steps flagged

No significant circularity; coverage guarantees remain independent of method definitions.

full rationale

The provided abstract and context contain no quoted equations or self-citations that reduce the claimed long-run coverage to a fitted parameter, self-defined quantity, or prior author result by construction. OLCP and OLCP-Hedge are defined via localization plus online convex optimization, with coverage asserted as a separate guarantee (likely from regret analysis) and supported by simulations. This matches the reader's assessment of no reduction to inputs. No load-bearing self-citation chains or ansatz smuggling appear in the given text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. Coverage guarantees presumably rest on standard online learning assumptions not detailed here.

pith-pipeline@v0.9.0 · 5432 in / 1004 out tokens · 35236 ms · 2026-05-11T01:46:13.128682+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 2 internal anchors

[1]

Conformal prediction: A gentle introduction

Anastasios N Angelopoulos, Stephen Bates, et al. Conformal prediction: A gentle introduction. Foundations and trends® in machine learning, 16(4):494–591, 2023. 9

work page 2023
[2]

Conformal prediction beyond exchangeability.The Annals of Statistics, 51(2):816–845, 2023

Rina Foygel Barber, Emmanuel J Candes, Aaditya Ramdas, and Ryan J Tibshirani. Conformal prediction beyond exchangeability.The Annals of Statistics, 51(2):816–845, 2023

work page 2023
[3]

Cboe V olatility Index (vix)

Cboe Global Markets. Cboe V olatility Index (vix). https://www.cboe.com/tradable_ products/vix/, 2026

work page 2026
[4]

CDC FluView: Influenza-like illness surveillance

Centers for Disease Control and Prevention. CDC FluView: Influenza-like illness surveillance. https://www.cdc.gov/fluview/, 2024

work page 2024
[5]

Follow the leader if you can, hedge if you must.The Journal of Machine Learning Research, 15(1):1281–1316, 2014

Steven De Rooij, Tim Van Erven, Peter D Grünwald, and Wouter M Koolen. Follow the leader if you can, hedge if you must.The Journal of Machine Learning Research, 15(1):1281–1316, 2014

work page 2014
[6]

A second-order bound with excess losses

Pierre Gaillard, Gilles Stoltz, and Tim Van Erven. A second-order bound with excess losses. In Conference on Learning Theory, pages 176–196. PMLR, 2014

work page 2014
[7]

Adaptive conformal inference under distribution shift

Isaac Gibbs and Emmanuel Candes. Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems, 34:1660–1672, 2021

work page 2021
[8]

Conformal inference for online prediction with arbitrary distribution shifts.Journal of Machine Learning Research, 25(162):1–36, 2024

Isaac Gibbs and Emmanuel J Candès. Conformal inference for online prediction with arbitrary distribution shifts.Journal of Machine Learning Research, 25(162):1–36, 2024

work page 2024
[9]

Localized conformal prediction: A generalized inference framework for confor- mal prediction.Biometrika, 110(1):33–50, 2023

Leying Guan. Localized conformal prediction: A generalized inference framework for confor- mal prediction.Biometrika, 110(1):33–50, 2023

work page 2023
[10]

Online convex optimization with hard constraints: Towards the best of two worlds and beyond.Advances in Neural Information Processing Systems, 35:36426–36439, 2022

Hengquan Guo, Xin Liu, Honghao Wei, and Lei Ying. Online convex optimization with hard constraints: Towards the best of two worlds and beyond.Advances in Neural Information Processing Systems, 35:36426–36439, 2022

work page 2022
[11]

Splice-2 comparative evaluation: Electricity pricing

Michael Harries. Splice-2 comparative evaluation: Electricity pricing. Technical report, The University of New South Wales, 1999

work page 1999
[12]

Darts: User-friendly modern machine learning for time series.Journal of Machine Learning Research, 23(124):1–6, 2022

Julien Herzen, Francesco Lässig, Samuele Giuliano Piazzetta, Thomas Neuer, Léo Tafti, Guil- laume Raille, Tomas Van Pottelbergh, Marek Pasieka, Andrzej Skrodzki, Nicolas Huguenin, Maxime Dumonal, Jan Ko´scisz, Dennis Bader, Frédérick Gusset, Mounir Benheddi, Camila Williamson, Michal Kosinski, Matej Petrik, and Gaël Grosch. Darts: User-friendly modern mac...

work page 2022
[13]

Conformal prediction with local weights: randomiza- tion enables robust guarantees.Journal of the Royal Statistical Society Series B: Statistical Methodology, 87(2):549–578, 2025

Rohan Hore and Rina Foygel Barber. Conformal prediction with local weights: randomiza- tion enables robust guarantees.Journal of the Royal Statistical Society Series B: Statistical Methodology, 87(2):549–578, 2025

work page 2025
[14]

Second-order quantile methods for experts and combina- torial games

Wouter M Koolen and Tim Van Erven. Second-order quantile methods for experts and combina- torial games. InConference on Learning Theory, pages 1155–1175. PMLR, 2015

work page 2015
[15]

Colin Lea, Rene Vidal, Austin Reiter, and Gregory D. Hager. Temporal convolutional networks: A unified approach to action segmentation, 2016

work page 2016
[16]

Distribution-free predictive inference for regression.Journal of the American Statistical Associ- ation, 113(523):1094–1111, 2018

Jing Lei, Max G’Sell, Alessandro Rinaldo, Ryan J Tibshirani, and Larry Wasserman. Distribution-free predictive inference for regression.Journal of the American Statistical Associ- ation, 113(523):1094–1111, 2018

work page 2018
[17]

Conformal prediction after data-dependent model selection

Ruiting Liang, Wanrong Zhu, and Rina Foygel Barber. Conformal prediction after efficiency- oriented model selection.arXiv preprint arXiv:2408.07066, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[18]

Online conformal prediction via universal portfolio algorithms.arXiv preprint arXiv:2602.03168, 2026

Tuo Liu, Edgar Dobriban, and Francesco Orabona. Online conformal prediction via universal portfolio algorithms.arXiv preprint arXiv:2602.03168, 2026

work page arXiv 2026
[19]

A Modern Introduction to Online Learning

Francesco Orabona. A modern introduction to online learning.arXiv preprint arXiv:1912.13213, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1912
[20]

Coin betting and parameter-free online learning.Advances in Neural Information Processing Systems, 29, 2016

Francesco Orabona and Dávid Pál. Coin betting and parameter-free online learning.Advances in Neural Information Processing Systems, 29, 2016. 10

work page 2016
[21]

Conformalized quantile regression

Yaniv Romano, Evan Patterson, and Emmanuel Candes. Conformalized quantile regression. Advances in neural information processing systems, 32, 2019

work page 2019
[22]

Scott.Multivariate Density Estimation: Theory, Practice, and Visualization

David W. Scott.Multivariate Density Estimation: Theory, Practice, and Visualization. Wiley, 1992

work page 1992
[23]

Silverman.Density Estimation for Statistics and Data Analysis

Bernard W. Silverman.Density Estimation for Statistics and Data Analysis. Chapman and Hall, 1986

work page 1986
[24]

Optimal algorithms for online convex optimization with adversarial constraints.Advances in Neural Information Processing Systems, 37:41274–41302, 2024

Abhishek Sinha and Rahul Vaze. Optimal algorithms for online convex optimization with adversarial constraints.Advances in Neural Information Processing Systems, 37:41274–41302, 2024

work page 2024
[25]

Stooq.https://stooq.com/, 2026

Stooq. Stooq.https://stooq.com/, 2026

work page 2026
[26]

Conformal prediction for time-series forecasting with change points

Sophia Sun and Rose Yu. Conformal prediction for time-series forecasting with change points. arXiv preprint arXiv:2509.02844, 2025

work page arXiv 2025
[27]

Conformal prediction under covariate shift.Advances in neural information processing systems, 32, 2019

Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal prediction under covariate shift.Advances in neural information processing systems, 32, 2019

work page 2019
[28]

Springer, 2005

Vladimir V ovk, Alexander Gammerman, and Glenn Shafer.Algorithmic learning in a random world. Springer, 2005

work page 2005
[29]

Conformal prediction interval for dynamic time-series

Chen Xu and Yao Xie. Conformal prediction interval for dynamic time-series. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 11559–11569. PMLR, 18–24 Jul 2021

work page 2021
[30]

Sequential predictive conformal inference for time series

Chen Xu and Yao Xie. Sequential predictive conformal inference for time series. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,Proceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 38707–38727. PMLR, 23–29 Jul 2023

work page 2023
[31]

Selection and aggregation of conformal prediction sets.Journal of the American Statistical Association, 120(549):435–447, 2025

Yachong Yang and Arun Kumar Kuchibhotla. Selection and aggregation of conformal prediction sets.Journal of the American Statistical Association, 120(549):435–447, 2025

work page 2025
[32]

Adaptive conformal predictions for time series

Margaux Zaffran, Olivier Féron, Yannig Goude, Julie Josse, and Aymeric Dieuleveut. Adaptive conformal predictions for time series. InInternational conference on machine learning, pages 25834–25866. PMLR, 2022. 11 A Proof of Proposition 3.1 Proof.By definition, zt =α t +γ(α−err t), L t = (−zt)+, U t = (zt −1) +. Since projection onto[0,1]satisfies Π[0,1](z...

work page 2022
[33]

The series is split chronologically into 70% training, 10% validation, and 20% testing, giving 913, 130, and 262 observations, respectively

Missing values are filled by interpolation followed by forward/backward filling. The series is split chronologically into 70% training, 10% validation, and 20% testing, giving 913, 130, and 262 observations, respectively. The response is standardized using the training mean and standard deviation, and intervals are constructed on this standardized scale. ...

work page 2008