Pith. sign in

REVIEW 4 major objections 4 minor 35 references

Simultaneous Coverage and Efficiency Guarantee in Online Conformal Prediction

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Online conformal prediction can control coverage failures and prediction-set size simultaneously against time-varying benchmarks, in adversarial and stochastic settings.

desk verdict Model I reduction is solid and worth citing; Model II's rate-optimality claim overreaches for near-static data. read the letter →

arxiv 2607.26577 v1 pith:CMKG64JX submitted 2026-07-29 cs.LG cs.DS

Rahul Vaze This is my paper
classification cs.LGcs.DS MSC 62L2062G0568W27
keywords adaptiveconformalinferenceonlinepredictioncoverageguaranteeprediction-setefficiencydynamicregretdistributionshiftpinballlossminimaxoptimality
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper targets two blind spots in adaptive conformal inference: existing coverage guarantees are signed, so prolonged under-coverage can be cancelled by later over-coverage, and existing theory says nothing about prediction-set size, so validity can be bought with needlessly wide intervals. The author claims a unified framework in which a single online update rule yields simultaneous bounds on absolute, non-cancelling coverage violation and on efficiency regret against a benchmark that moves with the data: for adversarial conformity scores at rate O(√(T(1+P_T))), for drifting stochastic scores at the minimax-optimal rate O(T^{2/3}V_T^{1/3}), and for covariate-dependent scores at O(T^{(5β+d)/(6β+d)} S_T^{2β/(6β+d)}). If correct, these are the first joint coverage-and-efficiency guarantees under distribution shift, and they quantify how much harder the problem becomes as feedback coarsens from full scores to one covered/not-covered bit and as the target becomes a function of covariates. The price is a global density lower bound on conformity scores, which the paper argues is a necessity rather than a proof artifact for the efficiency half of the guarantee.

What carries the argument

The load-bearing identity: the ACI update is projected online gradient descent on the pinball loss ψ_t(q) = (1−α)(rt−q)+ + α(q−rt)+, a convex loss charging thresholds that miss the realized score. A dynamic-regret bound for ψ_t against u*_t = rt transfers to any monotone L-Lipschitz set-size function Λ via the pointwise inequalities (rt−qt)+ ≤ ψ_t/(1−α) and |qt−rt| ≤ ψ_t/min(α,1−α); this decoupling of the convex surrogate from the possibly non-convex objective is what carries the adversarial result. The stochastic results rely on a second mechanism: a quantile-crossing lemma bounds the coverage error |Ft(qt)−(1−α)| by estimation noise plus window drift plus one 1/W jump, and a global density

What would settle it

Run the ACI update on the square-wave conformity sequence with the theorem-prescribed step size η = √((1+P_T)/T), sweeping T and P_T over several decades: the predicted log-log slope of R(T) is 0.5 in each variable, so a slope that stays materially above 0.5 rather than drifting down to it would show the O(√(T(1+P_T))) bound is not tight. Separately, feed the sliding-window tracker drifting score distributions whose density is zero on an interval just below the oracle quantile, and measure efficiency regret; the paper's Remark 10 predicts R(T) must then grow linearly in T for any algorithm, so

Watch

Extended reading notes

Core claim

Coverage and prediction-set efficiency are certified together, against a benchmark that moves with the data, by one regret certificate: the standard ACI update is exactly projected gradient descent on the pinball loss, a convex surrogate independent of the efficiency objective; since that loss dominates both the one-sided coverage shortfall (rt − qt)+ and the threshold distance |qt − rt|, the same regret bound gives Q(T), R(T) = O(√(T(1+P_T))) against the pointwise optimal comparator u*_t = rt for any monotone L-Lipschitz set-size function, with no distributional or convexity assumptions. Drifting stochastic scores are tracked at the minimax-optimal rate T^{2/3}V_T^{1/3} by a sliding-window

Load-bearing premise

The guarantee rests on a global floor for the conformity-score density: the CDF of scores must rise no slower than some f_min > 0 across the entire threshold range. If the density vanishes between the current threshold and the ideal one, the proofs cannot convert coverage error into threshold error, and the paper argues no algorithm could control efficiency then — so the efficiency half of the result collapses even if coverage holds.

Editorial extensions

If this is right

  • The unchanged standard ACI update, using only one covered/not-covered bit per round, gains a simultaneous absolute-coverage and efficiency guarantee against the moving benchmark u*_t = rt, with no distributional or convexity assumptions.
  • In the drifting stochastic model the sliding-window quantile tracker is rate-optimal: no algorithm can beat T^{2/3}V_T^{1/3}, so the coverage-versus-efficiency tradeoff in that regime is now quantified exactly.
  • For covariate-dependent scores, a partition-plus-ACI scheme achieves conditional coverage and efficiency simultaneously, with a rate that degrades smoothly as covariate dimension grows and oracle smoothness shrinks, making covariate-aware thresholding a structural necessity rather than an option.
  • Measuring coverage by non-cancelling violation rather than signed error changes the evaluation standard: methods can no longer hide prolonged under-coverage behind later over-coverage, and the combined cost Q(T)+R(T) exposes trivial wide-set solutions.
  • The minimax-optimal rate for drifting scores requires knowing the variation budget V_T; with an adaptive window the guarantee degrades to about T^{5/6}(1+V_T)^{1/6}, quantifying the price of not knowing how fast the environment changes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the pinball-regret transfer is a recipe, not just an analysis — any online learner whose dynamic regret on the pinball surrogate is known (multi-scale or strongly adaptive learners, proportional-integral controllers) should inherit a joint coverage-and-efficiency guarantee once the same |q − rt| transfer step is applied, since the ACI identification is only one instance of the
  • Editorial inference: the density-floor requirement is the main practical bottleneck. The paper shows a local floor makes the localization proof circular and vanishing density defeats efficiency entirely; an obvious testable extension, which the paper only sketches, is an algorithm with explicit inertia — restarting or re-anchoring when a threshold drifts into a low-density corridor — to lift the g
  • Editorial inference: the paper's own log-log rate check for ACI finds realized slopes near 0.6 against the predicted 0.5, which the author attributes to lower-order curvature. My reading is that the O(√(T(1+P_T))) bound may be loose rather than tight, and the square-wave input family is a clean test bed to settle whether the slope drifts to 0.5 over wider ranges.
  • Editorial note on the write-up: the proof of the conditional tracking bound (Section 11, step 5) invokes an 'Assumption C6' — local Lipschitzness of the efficiency cost in threshold and covariate — that never appears among the stated assumptions C1–C5; the argument needs it spelled out, though it is the same Lipschitz property used throughout. Separately, if the paper's conjecture in Remark 11 is
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a unified online conformal prediction framework that controls both absolute, non-cancelling coverage violation and prediction-set efficiency against dynamic oracle thresholds. Three settings are analyzed: (I) fully adversarial conformity scores, where the standard ACI update is identified with projected online gradient descent on a pinball loss, yielding Q(T),R(T)=O(sqrt(T(1+P_T))) for any monotone L-Lipschitz efficiency objective; (II) nonstationary stochastic scalar scores, where a sliding-window empirical quantile algorithm is claimed to attain O(T^{2/3}(1+V_T)^{1/3}) with a matching minimax lower bound; and (III) covariate-dependent scores, where a Hölder-partitioned ACI algorithm is claimed to track a function-valued oracle threshold at rate O(T^{(5β+d)/(6β+d)}S_T^{2β/(6β+d)}). The paper discusses limitations of density assumptions in Remarks 10 and 13, and acknowledges the need for adaptive step-size/window selection.

Significance. If the stated results hold, the paper would be a valuable contribution: the Model I proof is a clean and non-circular reduction showing that ACI already controls a dynamic efficiency benchmark through the pinball-loss surrogate; Model II provides a coherent estimation-drift analysis under a density floor; and Model III extends the framework to function-valued oracles with a plausible cell-based contraction argument. The paper is generally honest about the strong density assumptions and about the need for knowledge of oracle budgets. However, the advertised rate-optimality of Model II is overstated, and the Model III proof rests on an unstated assumption. These issues are fixable but currently affect the central claims.

major comments (4)
  1. [Section 3.3, Theorem 7 and proof Steps 2–4] The minimax lower bound is not informative over the full stated range V_T≤c_2T. For V_T≲T^{-1/2}, the block length B=⌊1/(8CΔ²)⌋ with Δ=c_0(V_T/T)^{1/3} exceeds T, so the K=T/B block structure collapses and the proof's lower bound becomes vacuous. Consequently the 'matching minimax lower bound' does not establish rate-optimality in the near-static regime. Indeed at V_T=0 the true minimax risk for threshold estimation is Ω(√T), whereas the sliding-window algorithm with W†=O(T^{2/3}) only attains O(T^{2/3}); a full-history estimator would improve it. The rate-optimality claim in the abstract and Section 1.6 is therefore valid, at best, only in the drift-dominated regime V_T≳T^{-1/2}. The theorem should either be restricted to that regime or the lower bound strengthened to include the additive static-rate term.
  2. [Section 11.5, proof of Theorem 9] The final R(T) bound is justified by invoking 'Assumption C6', but Assumption 8 lists only C1–C5. No explicit Lipschitz or regularity condition on the efficiency functional Λ is stated for Model III. Thus the headline efficiency guarantee R(T)=O((L/f_s_min)T^{(5β+d)/(6β+d)}S_T^{2β/(6β+d)}) rests on a missing assumption. Please add C6 to Assumption 8, or derive the Lipschitz transfer from already-stated assumptions, and clarify that the symbol L is not being reused inconsistently with the Hölder constants in C1–C2.
  3. [Section 2, Theorem 3 and footnote 3] The theorem states that the plain ACI update (9) satisfies Q(T),R(T)=O(sqrt(T(1+P_T))), but the proof chooses η=O(sqrt((1+P_T)/T)), which requires knowing the comparator path length P_T in advance. The paper notes that without this knowledge one must run multiple step sizes in parallel with a Hedge meta-algorithm, which changes the update and introduces extra log T terms. The statement should distinguish the tuned single-ACI result from the fully adaptive variant, and the adaptive variant should include the log factors. As written, the abstract's unqualified claim about the standard ACI update is stronger than the algorithm actually delivers.
  4. [Section 8, Theorem 6 and Section 1.6] The simplified claim in Section 1.6 that the algorithm achieves O(T^{2/3}V_T^{1/3}) is misleading at V_T=0, where it vanishes. Theorem 6 correctly uses (1+V_T) terms, giving O(T^{2/3}) at V_T=0, but that rate is not minimax (the static lower bound is Ω(√T)). Please state all rates with (1+V_T) consistently and avoid presenting a rate that is zero at V_T=0 as a characterization of the near-static regime.
minor comments (4)
  1. [Section 10] The definitions of K and B should be made before they are used: B=⌊1/(8CΔ²)⌋ is introduced after K=T/B is already implied, and the floor effects are stated loosely. This matters because for small V_T the construction does not define a valid block sequence.
  2. [Section 11.5] There is a missing Assumption C6 and also a typographical issue: 'Holder's inequality' should be 'Hölder's inequality'. The notation L is overloaded (Hölder constant, Lipschitz constant for Λ) and should be disambiguated.
  3. [Section 5.1.1] The numerical replication uses DAX returns as a substitute for the [13] data, with a compute-saving re-estimation schedule. This is a reasonable approximation, but the text should state plainly that this is not the original benchmark data and that constants are not reused from the theory.
  4. [References] Reference [5] is incomplete ('arXiv:2307.xxxxx'). Please update or remove it. Reference [29] is a duplicate of [28] and can be merged.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main guarantees are derived from explicit assumptions and external, independently-checkable results, with no fitted parameter renamed as a prediction.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs. In Model I, Theorem 3 identifies the ACI update (9) with projected online gradient descent on the pinball loss (Eq. 10) and imports a standard dynamic-regret bound from [34]; the comparator u*_t = r_t is forced only by monotonicity of Lambda together with the coverage constraint u_t >= r_t, so the efficiency and coverage functionals are defined independently of the algorithm's iterates, not fitted to them. The proof then translates the pinball regret into Q(T) and R(T) via explicit inequalities involving the pinball subdifferential; no step equates the conclusion with an assumption. In Model II, the sliding-window quantile is compared with the oracle quantile F_t^{-1}(1-alpha), and the upper bound is obtained from concentration, drift, and a quantile-crossing lemma; the lower bound constructs a two-distribution hypothesis-testing instance and does not use the algorithm's update rule as an ingredient. Model III similarly tracks a function-valued oracle independent of the algorithm. The only passages that mention circularity are cautionary: Remark 10 explicitly states that replacing the global density floor by a merely local bound would make the localization argument circular, and the proof keeps the global floor f_min > 0, so the paper does not commit that circular step. Remark 11 concedes that the Q(T) lower bound is not tight in f_min; this is a stated limitation, not a circular reduction. There are no self-citations carrying the argument, no fitted constants presented as predictions, and no renaming of known results as new derivations. Numerical sections do not fit theoretical constants. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 10 assumptions · 0 invented entities

The central theorems are obtained under substantial regularity and side-information assumptions: global density floors, Hölder smoothness, and oracle path-length or variation budgets. The paper is explicit about most of these costs, but the proof of Theorem 9 silently uses an unlisted Assumption C6. No new physical or conceptual entities are introduced beyond the algorithmic constructions.

free parameters (4)
  • ACI learning rate η (Model I) = Θ(sqrt((1+P_T)/T))
    Chosen analytically from the path length P_T, which is unknown a priori. The paper notes that without P_T, a Hedge construction over step sizes gives the same bound up to log factors.
  • Sliding-window size W† (Model II) = min(ceil((C_VC T / (2(1+V_T)))^{2/3}), T)
    Selected to balance estimation error and drift; depends on the unknown variation budget V_T. Theorem 12 removes the need to know V_T but degrades the rate to O(T^{5/6}).
  • HP-ACI cell width h* (Model III) = T^{-1/(6β+d)} S_T^{2/(6β+d)}
    Balances spatial discretization error against per-cell estimation error; requires knowledge of S_T and β.
  • HP-ACI step size η* (Model III) = Θ((S_T^2 / (T h^d))^{1/3})
    Must satisfy the stability bound C5 and depends on S_T, f_s_min, f_s_max. The diminishing-step variant avoids knowing these constants but loses optimality.
assumptions (10)
  • domain assumption Scores and thresholds lie in a bounded action set A=[0,Q_max]
    Used throughout for bounded subgradients, projections, and deterministic O(T) worst-case errors.
  • domain assumption Efficiency objective Λ is monotone nondecreasing and L-Lipschitz
    Stated in Remark 1. The proof of Theorem 9 refers to this as 'Assumption C6', but C6 is not listed in Assumption 8.
  • domain assumption Model II A1: scores r_t ~ F_t are independent across t
    Needed for the sliding-window concentration argument and for the lower-bound construction.
  • domain assumption Model II A2: global density lower bound f_min > 0 on A
    Used to convert F-space coverage error into threshold-space tracking error. The paper admits that without it R(T) cannot be controlled.
  • standard math Model II A3: uniform concentration of the empirical CDF (DKW-style)
    The paper shows it follows from DKW and independence; retained as an assumption only to make constants explicit.
  • domain assumption Model III C1: oracle threshold is β-Hölder in the covariate
    Needed to control spatial discretization error and to set the partition width.
  • domain assumption Model III C2: spatial continuity of conditional score CDF
    Bridges the score distribution at a cell center and at the actual covariate.
  • domain assumption Model III C3: oracle path-length budget S_T is finite
    Bounds the total temporal drift of the oracle threshold function; the rate is expressed in terms of S_T.
  • domain assumption Model III C4: independence and global density bounds on f_t(x) and f_t(q|x)
    Global density lower bound is needed for iterate confinement; without it the recursion can lose its restoring force.
  • domain assumption Model III C5: step-size bound η ≤ f_s_min / (2(f_s_max)^2)
    Ensures the contraction coefficient stays in [0,1) in the Lyapunov analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simultaneous Coverage and Efficiency Guarantee in Online Conformal Prediction." pith.science (2026). https://pith.science/paper/CMKG64JX

@misc{pith2026260726577,
  author       = {Pith},
  title        = {Pith review of: Simultaneous Coverage and Efficiency Guarantee in Online Conformal Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMKG64JX}},
  note         = {Machine review of arXiv:2607.26577}
}
read the original abstract

Adaptive conformal inference (ACI) of Gibbs and Cand{\`e}s and its variants are the standard approach to online conformal prediction under distribution shift, but they suffer from three fundamental limitations. First, their guarantees control only the \emph{signed} long-run coverage error: persistent miscoverage in one direction can be masked by compensating errors later, so a method can satisfy the theoretical guarantee while being badly wrong for extended periods. Second, existing guarantees say nothing about prediction-set size, so validity can be achieved trivially at the cost of unduly wide prediction sets. Third, the efficiency guarantees that do exist compare against a \emph{fixed} predictor chosen in hindsight, a benchmark that becomes increasingly less meaningful once the data-generating distribution shifts, since the very notion of an optimal threshold then changes over time. We consider a unified online learning framework that simultaneously controls absolute, non-cancelling coverage violation and prediction-set efficiency against a dynamically evolving benchmark for three important models. In the fully adversarial setting, exploiting the fact that the standard ACI update is exactly projected online gradient descent on the pinball loss, we derive simultaneous coverage and efficiency guarantees for arbitrary monotone Lipschitz efficiency objectives, with no distributional or {\it convexity} assumptions. In the stochastic setting with full-score feedback, we propose a sliding-window quantile tracker and establish a matching minimax lower bound showing our algorithm is rate-optimal. In the covariate-dependent stochastic setting, we develop a partitioned ACI algorithm that tracks a function-valued oracle threshold, and derive simultaneous coverage and efficiency guarantees.

Figures

Figures reproduced from arXiv: 2607.26577 by the authors.

Figure 1
Figure 1. Raw coverage and efficiency traces on real DAX/GARCH scores. [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Efficiency measured against the dynamic (centered-window) oracle [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Combined coverage-violation-plus-efficiency-regret cost, [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Coverage violation Q(T) as a function of oscillation period p, fixed W = 150 and γ = 0.20. 101 102 103 0 0.2 0.4 0.6 0.8 1 ·104 oscillation period p (log scale) cumulative efficiency regret R(T) ACI (Model I) Sliding-window (Model II) p = W = 150 [PITH_FULL_IMAGE:figu…
Figure 5
Figure 5. Figure 5: Efficiency regret R(T) (tracking error against the exact oracle u ∗ t = rt) as a function of oscillation period p. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 6
Figure 6. Figure 6: Threshold trajectories for p = 100 < W = 150, illustrating the mechanism behind Figures 4–5. p Q(T) ACI Q(T) window R(T) ACI R(T) window combined winner 10 121.0 10.5 9985.8 9960.5 window 15 107.6 13.0 10091.9 9963.0 window 20 142.4 16.0 9764.3 9916.0 ACI 50 203.8 32.5…
Figure 7
Figure 7. Figure 7: Log-log rate check for ACI (Model I). Left: fixed [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Left: conditional coverage by covariate bin, marginalized over the whole horizon. Middle: thresh [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Local (windowed) conditional coverage deviation from target, jointly over covariate [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 4 linked inside Pith

  1. [5]

    Conformal prediction beyond exchangeability.arXiv:2307.xxxxx, 2023

    R Barber et al. Conformal prediction beyond exchangeability.arXiv:2307.xxxxx, 2023

    2023

  2. [12]

    V olume-regularized conformal prediction.arXiv:25xx.xxxxx, 2025

    R Gao et al. V olume-regularized conformal prediction.arXiv:25xx.xxxxx, 2025

    2025

  3. [1]

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification.arXiv preprint arXiv:2107.07511, 2023

    Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification.arXiv preprint arXiv:2107.07511, 2023

    arXiv 2023

  4. [2]

    Raps: Conformal pre- diction with ranked prediction sets

    Anastasios N Angelopoulos, Stephen Bates, Michael Jordan, and Jitendra Malik. Raps: Conformal pre- diction with ranked prediction sets. InInternational Conference on Artificial Intelligence and Statistics, 2021

    2021

  5. [3]

    Angelopoulos, Emmanuel J

    Anastasios N. Angelopoulos, Emmanuel J. Cand `es, and Ryan J. Tibshirani. Conformal PID control for time series prediction.arXiv preprint arXiv:2307.16895, 2023

    arXiv 2023

  6. [4]

    Online conformal predic- tion via online optimization

    Felipe Areces, Christopher Mohri, Tatsunori Hashimoto, and John Duchi. Online conformal predic- tion via online optimization. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of the 42nd Interna- tional Conference on Machine Learning, volume 267 ofProceedings...

    2025

  7. [6]

    Non-stationary stochastic optimization.Operations Research, 63(5):1227–1244, 2015

    Omar Besbes, Yonatan Gur, and Assaf Zeevi. Non-stationary stochastic optimization.Operations Research, 63(5):1227–1244, 2015

    2015

  8. [7]

    Online conformal prediction with improved regret

    Akshay Bhatnagar et al. Online conformal prediction with improved regret. InInternational Confer- ence on Machine Learning, 2024

    2024

Show all 35 references
  1. [8]

    Improved online conformal prediction via strongly adaptive online learning

    Akshay Bhatnagar, Huan Wang, Caiming Xiong, and Yu Bai. Improved online conformal prediction via strongly adaptive online learning. InInternational Conference on Machine Learning, 2023

    2023

  2. [9]

    From external to internal regret.Journal of Machine Learning Research, 8:1307–1324, 2007

    Avrim Blum and Yishay Mansour. From external to internal regret.Journal of Machine Learning Research, 8:1307–1324, 2007

    2007

  3. [10]

    Online conformal prediction with local adaptation

    S Feldman et al. Online conformal prediction with local adaptation. InICML, 2023. 30

    2023

  4. [11]

    Schapire

    Yoav Freund and Robert E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. InJournal of Computer and System Sciences, volume 55, pages 119–139, 1997

    1997

  5. [13]

    Adaptive conformal inference under distribution shift

    Isaac Gibbs and Emmanuel Cand `es. Adaptive conformal inference under distribution shift. InAd- vances in Neural Information Processing Systems, 2021

    2021

  6. [14]

    Cherian, and Emmanuel J

    Isaac Gibbs, John J. Cherian, and Emmanuel J. Cand `es. Conformal prediction with conditional guar- antees.arXiv preprint arXiv:2305.12616, 2023

    2023 arXiv

  7. [15]

    Flexible conformal prediction for regression.JASA, 2020

    Rafael Izbicki et al. Flexible conformal prediction for regression.JASA, 2020

    2020

  8. [16]

    Conformal prediction with conditional coverage.Journal of Machine Learning Research, 2022

    Rafael Izbicki et al. Conformal prediction with conditional coverage.Journal of Machine Learning Research, 2022

    2022

  9. [17]

    Conformal prediction for high-dimensional settings

    A Kiyani et al. Conformal prediction for high-dimensional settings. 2024

    2024

  10. [18]

    Distribution-free predictive inference for regression.JASA, 2013

    Jing Lei et al. Distribution-free predictive inference for regression.JASA, 2013

    2013

  11. [19]

    Distribution-free predictive inference for regression

    Jing Lei, Max G’Sell, Alessandro Rinaldo, Ryan J Tibshirani, and Larry Wasserman. Distribution-free predictive inference for regression. InJournal of the American Statistical Association, 2018

    2018

  12. [20]

    The tight constant in the DKW inequality.The Annals of Probability, 18(3):1269– 1283, 1990

    Pascal Massart. The tight constant in the DKW inequality.The Annals of Probability, 18(3):1269– 1283, 1990

    1990

  13. [21]

    Distribution-free conformal prediction for regression.arXiv:2102.06173, 2021

    D Podkopaev and A Ramdas. Distribution-free conformal prediction for regression.arXiv:2102.06173, 2021

    2021 arXiv

  14. [22]

    The relationship between no-regret learning and online conformal prediction

    Ramya Ramalingam, Shayan Kiyani, and Aaron Roth. The relationship between no-regret learning and online conformal prediction. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of t...

    2025

  15. [23]

    Conformalized quantile regression

    Yaniv Romano, Evan Patterson, and Emmanuel Cand `es. Conformalized quantile regression. InAd- vances in Neural Information Processing Systems, 2019

    2019

  16. [24]

    Least squares conformal prediction.JASA, 2019

    Mauricio Sadinle et al. Least squares conformal prediction.JASA, 2019

    2019

  17. [25]

    A tutorial on conformal prediction.Journal of Machine Learning Research, 9:371–421, 2008

    Glenn Shafer and Vladimir V ovk. A tutorial on conformal prediction.Journal of Machine Learning Research, 9:371–421, 2008

    2008

  18. [26]

    Srinivas

    V . Srinivas. Online conformal prediction with efficiency guarantees. arXiv:2507.02496, 2025

    2025

  19. [27]

    Conformal prediction for time series with modern statistical methods

    Kamile Stankeviciute et al. Conformal prediction for time series with modern statistical methods. In NeurIPS, 2021

    2021

  20. [28]

    Tsybakov.Introduction to Nonparametric Estimation

    Alexandre B. Tsybakov.Introduction to Nonparametric Estimation. Springer, 2008

    2008

  21. [29]

    Tsybakov.Introduction to Nonparametric Estimation

    Alexandre B. Tsybakov.Introduction to Nonparametric Estimation. Springer, New York, 2009

    2009

  22. [30]

    Vladimir V ovk, Alexander Gammerman, and Craig Saunders.Machine learning applications of algo- rithmic information theory. 1999. 31

    1999

  23. [31]

    Springer, 2005

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

    2005

  24. [32]

    Enbpi: Enhanced batch prediction intervals for time series

    Chen Xu and Yao Xie. Enbpi: Enhanced batch prediction intervals for time series. InInternational Conference on Learning Representations, 2021

    2021

  25. [33]

    Adaptive conformal predictions for time series.International Conference on Machine Learning, 2022

    Margaux Zaffran et al. Adaptive conformal predictions for time series.International Conference on Machine Learning, 2022

    2022

  26. [34]

    nearest-neighbour

    Lijun Zhang, Shiyin Lu, and Zhi-Hua Zhou. Adaptive online learning in dynamic environments. Advances in neural information processing systems, 31, 2018. 8 Proof of Theorem 6 The proof decomposes the total error into three sources: estimation noise in the empirical CDF, bias fr...

    2018

  27. [35]

    Step 3: Lower bounding the Bayes risk.Fix a blockk

    Taking expectation over(H k−1, U,Θk): P(ϕt ̸= Θk)≥ 3 8 . Step 3: Lower bounding the Bayes risk.Fix a blockk. Combining|q t −u ∗ Θk | ≥∆ 2 1{ϕt ̸= Θk}with P(ϕt ̸= Θk)≥ 3 8 from Step 2, and summing over theBrounds in blockk: X t∈blockk E|qt −u ∗ Θk | ≥∆ 2 X t∈blockk P(ϕt ̸= Θk)≥...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.