Pith.

Methodology · arXiv:2609.07251 v1 · submitted 7 Sep 2026 · 16 references

Adaptive Conformal Inference Under Delayed Feedback: Coverage Guarantees and a Delay-to-Memory Diagnostic

Lama El Halabi, Adam Brandt

The paper says

“Adaptive Conformal Inference (ACI) extends conformal prediction to non-exchangeable settings by adjusting the nominal miscoverage level online in response to recent coverage errors.”

more of the abstract

When forecasts are issued with horizon τ, however, the outcome needed to evaluate a prediction is observed only τ steps later, so these adaptive updates must rely on delayed feedback. We study this setting and ask how the effect of delay depends on the persistence of the residual process. First, we show that the τ-delayed ACI recursion can be decomposed into τ interleaved ACI-like sequences. This representation yields a finite-sample bound on long-run empirical coverage with explicit dependence on τ. We also derive an approximate marginal coverage bound that relates coverage deviation to changes in the underlying environment across the forecast horizon and to the adaptation rate γ. We then introduce the delay-to-memory ratio r = τ/L, where L is the time scale over which the temporal signal driving non-exchangeability decays. Simulation results show that the usefulness of this ratio depends on the form of temporal dependence: it strongly organizes performance under AR(1) dependence, is less predictive of overall performance under GARCH(1,1) and Markov switching, but more clearly characterizes when scale normalization remains useful in those settings. Abrupt mean- and variance-shift experiments further show that the preferred adaptation rate depends on the residual dynamics. Overall, the results show that the effect of forecast delay is best understood relative to the time scale over which past residual information remains relevant.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. Solid paper that fills an open direction in adaptive conformal inference; the phase decomposition and delay-to-memory ratio are genuine contributions, and the simulations are thorough. The marginal bound has strong assumptions, but they are stated clearly. Accept. This line is wrong

1 major objection · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 16 references say.

The paper says, at Section 4.4; Eq. (12); Appendix A.3

The review did not quote this passage.

Pith says: Objection 1 of 1 · major

The marginal bound (12) is derived under stationarity of the closed-loop process (α_t, A_t, err_t), which is an assumption on the full feedback system rather than only on the hidden Markov environment. The assumption is stated and the bound is honestly labeled approximate, but the interpretation of the first term needs one clarifying sentence: under stationarity, E|α*_{A_{t+τ}} − α*_{A_t}| is …

more of objection 1

… the expected change of the optimal miscoverage level over a random τ-step window of a stationary chain, i.e., a steady-state fluctuation, not the transient distributional shift studied in the mean- and variance-shift experiments of Section 7.4. Because abrupt shifts are precisely the settings where a practitioner would want a marginal guarantee, I recommend stating explicitly in Section 4.4 that (12) characterizes steady-state miscoverage deviation and does not apply to transient shifts.

6 minor objections

The paper says, at Section 4.4 (assumptions); Appendix A.3 around Eq. (36)

Pith says: minor objection 1 of 6

The proof of (12) uses the sign-alignment identity sign(M(α_t|A_t) − α) = sign(α_t − α*_{A_t}), which requires M(·|a) to be nondecreasing in α for each a. This monotonicity is asserted inside the proof but omitted from the assumption list in Section 4.4; it follows from the nestedness of the two-sided intervals in α given the smoothed quantile properties stated in Section 4, so please add it to the stated assumptions or note explicitly that it is implied by the construction.

The paper says, at Section 6.4, Eqs. (22)–(23)

Pith says: minor objection 2 of 6

The claim that squared residuals inherit geometric autocorrelation decay at rate α+β presumes finite fourth moments; for Gaussian innovations the condition is 3α² + 2αβ + β² < 1. The most persistent configuration (α+β = 0.99, α = 0.1, β = 0.89) gives 3α² + 2αβ + β² ≈ 1.0001, so for this configuration the e-folding interpretation of L_vol rests on a borderline or nonexistent fourth moment. Please add a caveat or choose parameter values that satisfy the moment condition.

show 4 more

The paper says, at Section 6, Eq. (17); Tables 6 and 7

Pith says: minor objection 3 of 6

The curve-collapse reductions are reported as point values (79%, 46%, 9–27%, 83–84%) without uncertainty quantification, and with five persistence configurations spread over twelve bins the within-bin scatter is estimated from very few points per bin. Reporting seed-level dispersion or a permutation-style null (for example, replacing L by a misspecified or randomly permuted memory scale) would strengthen the claim that the collapse is not an artifact of the binning procedure.

The paper says, at Throughout (presentation)

Pith says: minor objection 4 of 6

Several typos and formatting slips should be corrected: 'Int-space' in Section 4.1 should read 'in t-space'; Appendix A.2 Step 3 begins with 'F rom'; the Section 4 heading reads '4τ-Delayed' without a space; Section 3 contains 'We therefore definet0 = 0'; and the Table 1 caption runs 'global time indextθ(m)' without a space.

The paper says, at Section 6.5; Tables 4 and 5

Pith says: minor objection 5 of 6

Please state explicitly whether the adaptation durations in Tables 4 and 5 are averaged over the 10 seeds or shown for a representative seed, and confirm in the captions that right-censored durations (coverage not recovered by the segment boundary) are recorded at the boundary as described in Section 6.5.

The paper says, at Reproducibility

Pith says: minor objection 6 of 6

No code or data availability statement is included; given the simulation-heavy design, a brief code-release statement would materially improve reproducibility and would also let readers verify the curve-collapse percentages in Tables 6 and 7.

What would settle it

As a verification step worth running, implement Algorithm 2 exactly on a stationary two-state HMM with τ=2 and fixed γ, estimate the left-hand side of Equation 12 by Monte Carlo using the true conditional miscoverage at each (A_t, α_t), and compare it with the right-hand side over T=10^5 steps; a persistent violation would indicate that Equation 12 relies on an unstated conditional-coverage assumption.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 10 Sep 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 2 min 53 s · 30,493 tokens read, 19,779 written
  3. Editor wrote the verdict above 12 s · 31,172 tokens read, 1,077 written
  4. Referee minor revision: 1 major objection, 6 minor 3 min 16 s · 30,867 tokens read, 25,089 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This is a solid paper. The main contribution—the phase-based decomposition of the τ-delayed ACI recursion—is clean and yields a genuine new bound on long-run coverage with an explicit τ/T term. The delay-to-memory ratio is a nice diagnostic that organizes performance across persistence levels, especially for AR(1) dependence. The simulation suite is unusually thorough: six residual families, controlled experiments, and careful step-size selection. The authors do not oversell; they are honest about when the ratio works and when it does not.

What the paper does well: (1) The decomposition in Section 4.1 is simple once you see it, but it is new and it directly answers the open direction Gibbs and Candès flagged. (2) The long-run coverage bound (Equation 11) is distribution-free and tightens cleanly as τ shrinks. (3) The curve-collapse diagnostic is a clever way to measure how well a reparameterization explains performance. (4) The normalization experiments under GARCH and Markov switching are well motivated and isolate the delay-induced scale-mismatch problem.

Soft spots: The approximate marginal coverage bound (Equation 12) requires stationarity, Lipschitz miscoverage, and a unique optimal level. Those are restrictive, but the bound is labeled approximate and the assumptions are stated. The practical utility of the ratio depends on knowing L, which is defined from model parameters. The paper does not offer a method to estimate L from data, so practitioners with unknown persistence will have to approximate. The Markov-switching results show the ratio organizes normalization effects more than overall score, which is a nuance the paper handles honestly. I do not see a load-bearing flaw.

Who this is for: researchers working on online conformal prediction, especially multi-step forecasting or delayed feedback. Practitioners who run rolling conformal predictions with a forecast horizon will also benefit from the diagnostic intuition. The paper deserves a serious referee; I would accept it after a round of minor revision—mostly asking for a discussion of how one might estimate L in practice, and perhaps a note about extending to time-varying τ.

Yes, I would bring this to reading group. Would cite it if I were working in this area.

The strongest objection

I checked the phase decomposition (Section 4.1), the boundedness argument (Appendix A.1), the long-run coverage bound (Equation 11 / Appendix A.2), and the approximate marginal coverage proof (Equation 12 / Appendix A.3). The decomposition is valid: reindexing t = θ + (m−1)τ removes the explicit delay phase-wise, and the recombination in Appendix A.2 yields the stated τ/(γT) and τ/T terms with no missing factor or hidden dependence on the data-generating process. The marginal bound also checks: the cross-term bound uses only Lipschitzness plus sign alignment of (M−α) and (α−α*), the (I) and second (II) terms cancel exactly under stationarity, and the remaining term gives L_Lip(1+γ)/γ · E|α*_{A_{t+τ}} − α*_{A_t}| + L_Lipγ/2. The softest point is the one the reader identified: Equation 12 requires a stationary closed-loop process and a Lipschitz, unique-optimal-level model, which are restrictive for real time series; but these assumptions are stated, and the bound is explicitly labeled approximate. I do not see an internal inconsistency or a failed step that would overturn the central claim.

The reader's ACCEPT verdict is well supported. The finite-sample long-run bound follows cleanly from the phase-wise telescoping argument, and the τ-dependence is correctly exposed in both terms. The approximate marginal bound is derived under explicitly stated stationarity, Lipschitz, and uniqueness assumptions; the algebra in Appendix A.3 is consistent, including the cancellation of the (I) term and the second part of (II) under stationarity. The simulation study is extensive, uses multiple residual families, and presents honest caveats about where the delay-to-memory ratio does and does not collapse performance. The weakest component remains the marginal bound's modeling assumptions, but this is a limitation of scope rather than a defect in the argument. No change to the verdict is warranted.

The referee's report

If the results hold — and my reading of the proofs and numerics found no error — the paper fills a genuine gap, since Gibbs and Candès (2021) left delayed feedback as an open direction. The phase-based decomposition and the recombination bound in Eq. (11) give a clean, correct statement of how delay degrades long-run coverage guarantees; the bound is distribution-free with explicit constants, and the a.s. convergence claim is actually implied pathwise by the deterministic inequality. Honest reporting is a real strength: the weak Markov-switching collapse (9–27%) and the AR(1) local-coverage non-collapse (Appendix C.1) are reported as limits of the diagnostic rather than concealed, Eq. (12) is explicitly labeled approximate with stated assumptions, and the oracle normalization constructions clearly separate deployable and non-deployable information. Concrete falsifiable predictions include the r ≈ 1 transition for AR(1) interval scores and selected step sizes (Figures 2–3) and the mismatch-frequency formula Pr(M_{t,τ}=1) = (1/2)(1 − e^{−r}) in Eq. (30). The main limitations are the steady-state scope of Eq.

Pith's reading of the paper (a machine's summary, not the authors')

This paper studies how forecast delay affects Adaptive Conformal Inference (ACI), a method that adjusts coverage levels online. It shows that the $ au$-delayed ACI update can be decomposed into $ au$ interleaved standard-ACI sequences, which yields a finite-sample coverage bound with explicit dependence on $ au$. The paper also introduces a delay-to-memory ratio $r = \tau/L$, where $L$ is the time scale over which the temporal signal driving non-exchangeability decays. Simulations across AR(1), GARCH, Markov switching, and abrupt-shift processes demonstrate that $r$ organizes performance—especially interval score and the benefit of scale normalization—more effectively than the raw delay $ au$ alone.

What carries the argument. The key machinery is the phase-based representation (Section 4.1), which maps the global time index $t$ to a phase $ heta \in \{1,\ldots,\tau\}$ and an update index $m$, turning the $ au$-delayed recursion into $ au$ independent standard-ACI recursions. This allows the application of existing boundedness and long-run coverage results to each phase, leading to a finite-sample bound that scales with $\tau/T$. The second key piece is the delay-to-memory ratio $r = \tau/L$, where $L$ is defined from the decay rate of the relevant dependence structure (auto-correlation for AR(1), volatility persistence $\alpha+\beta$ for GARCH, latent-state eigenvalue $\lambda_2$ for Markov switching). This ratio serves as a diagnostic for whether past residual information remains relevant at the forecast target.

Core claim. The central claim is that the effect of forecast delay on adaptive conformal inference is best understood relative to the persistence of the residual process. The paper proves that the $ au$-delayed ACI recursion can be rewritten as $ au$ interleaved ACI-like phases, each evolving independently. From this representation it derives a distribution-free finite-sample bound on long-run empirical coverage (Equation 11) and, under a stationary hidden Markov model, an approximate marginal coverage bound (Equation 12) that decomposes coverage deviation into distributional shift over the forecast horizon and an adaptation penalty. The paper then defines the delay-to-memory ratio $r = \tau/L$ for three persistent process families and shows numerically that this ratio substantially collapses interval-score curves across different persistence levels for AR(1) residuals (79% scatter reduction) and strongly organizes the benefit of scale normalization for GARCH (50%) and Markov switching (83–84%). The results establish that the relevant memory scale depends on which feature of the residual distribution carries dependence—the level, the volatility, or the latent regime—and that the ratio is more informative than raw delay when the quantity of interest involves that feature.

The premise it rests on. For the approximate marginal coverage bound (Equation 12), the paper assumes the data come from a stationary hidden Markov model with a Lipschitz marginal miscoverage function and a unique optimal miscoverage level per environment; these assumptions are strong for many real-world time series and if they fail, the bound may not hold.

If the paper is right (Pith's inferences, not the authors')

The finite-sample coverage bound (Equation 11) shows that long-run empirical coverage converges to the target at rate $\tau/T$, meaning longer forecast horizons require more time steps to achieve the same coverage guarantee; the phase-based representation makes this dependence explicit.

The approximate marginal coverage bound (Equation 12) predicts that the optimal step size $\gamma_\text{opt}$ scales with the expected change in the optimal coverage level over $\tau$ steps, providing a principled rule for selecting an adaptation rate when the distribution shift magnitude can be estimated.

For AR(1) residuals, the delay-to-memory ratio $r_\varepsilon$ produces a strong curve collapse in interval score near $r \approx 1$, suggesting that practitioners can treat $r$ as a threshold: when $r \ll 1$, delay is not yet harmful; when $r \gg 1$, performance approaches the i.i.d. baseline and aggressive adaptation becomes wasteful.

Under GARCH and Markov switching, normalization of residuals by conditional scale (or state-dependent scale) improves interval score and local coverage when $r < 1$, but becomes detrimental when $r > 1$ due to scale mismatch between the construction time and the target time; the target-scale oracle shows that better forecasts of future scale could recover this loss.

For abrupt mean shifts, shortening the calibration window generally improves adaptation speed (lower adaptation duration in 24 of 28 configurations); for variance shifts the opposite holds, because a shorter window selects a smaller step size that slows recovery.

The delay-to-memory ratio could be estimated online from the residual autocorrelation or squared-residual autocorrelation, enabling dynamic step-size selection that responds to changes in persistence over time. This would pair naturally with expert-aggregation or fully adaptive ACI methods, as the paper suggests.

The phase representation implies that one could maintain $\tau$ separate adaptive sequences with different step sizes, one per forecast horizon, as in Szabadváry (2024); the delay-to-memory ratio provides a principled way to tie those step sizes to the residual persistence, rather than tuning them independently.

In environments where the residual process is non-stationary or the memory length changes abruptly (e.g., after a regime shift), the diagnostic power of $r$ will decrease; an online estimator of $L$ with changepoint detection could extend the ratio's usefulness to such settings.

The paper's coverage bounds focus on marginal and long-run coverage; a complementary analysis of conditional (worst-case) coverage versus $r$ could reveal whether the ratio also predicts the risk of severe local undercoverage, which the local coverage error results show does not collapse across persistence levels.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Adaptive Conformal Inference (ACI) extends conformal prediction to non-exchangeable settings by adjusting the nominal miscoverage level online in response to recent coverage errors. When forecasts are issued with horizon τ, however, the outcome needed to evaluate a prediction is observed only τ steps later, so these adaptive updates must rely on delayed feedback. We study this setting and ask how the effect of delay depends on the persistence of the residual process. First, we show that the τ-delayed ACI recursion can be decomposed into τ interleaved ACI-like sequences. This representation yields a finite-sample bound on long-run empirical coverage with explicit dependence on τ. We also derive an approximate marginal coverage bound that relates coverage deviation to changes in the underlying environment across the forecast horizon and to the adaptation rate γ. We then introduce the delay-to-memory ratio r = τ/L, where L is the time scale over which the temporal signal driving non-exchangeability decays. Simulation results show that the usefulness of this ratio depends on the form of temporal dependence: it strongly organizes performance under AR(1) dependence, is less predictive of overall performance under GARCH(1,1) and Markov switching, but more clearly characterizes when scale normalization remains useful in those settings. Abrupt mean- and variance-shift experiments further show that the preferred adaptation rate depends on the residual dynamics. Overall, the results show that the effect of forecast delay is best understood relative to the time scale over which past residual information remains relevant.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 1.00

“Overall, the results show that the effect of forecast delay is best understood relative to the time scale over which past residual information remains relevant.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.77 · human participants 0.02 · controlled comparison 0.06 · randomized assignment 0.02 · states quantitative result 0.70 · compares with prior results 0.14 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Adaptive Conformal Inference (ACI) extends conformal prediction to non-exchangeable settings by adjusting the nominal miscoverage level online in response to recent coverage errors. background 1.00 0.08 0.04 asserted 1.00
When forecasts are issued with horizon $\tau$, however, the outcome needed to evaluate a prediction is observed only $\tau$ steps later, so these adaptive updates must rely on delayed feedback. background 0.94 0.08 0.03 asserted 1.00
We study this setting and ask how the effect of delay depends on the persistence of the residual process. contribution 0.64 0.26 0.04 unclear 0.93
First, we show that the $\tau$-delayed ACI recursion can be decomposed into $\tau$ interleaved ACI-like sequences. method 0.34 0.94 0.05 asserted 1.00
This representation yields a finite-sample bound on long-run empirical coverage with explicit dependence on $\tau$. result 0.91 0.96 0.09 asserted 1.00
We also derive an approximate marginal coverage bound that relates coverage deviation to changes in the underlying environment across the forecast horizon and to the adaptation rate $\gamma$. result 0.71 0.96 0.06 asserted 1.00
We then introduce the delay-to-memory ratio $r=\tau/L$, where $L$ is the time scale over which the temporal signal driving non-exchangeability decays. contribution 0.61 0.84 0.06 asserted 0.90
Simulation results show that the usefulness of this ratio depends on the form of temporal dependence: it strongly organizes performance under AR(1) dependence, is less predictive of overall performance under GARCH(1,1) and Markov switching, but more clearly characterizes when scale normalization remains useful in those settings. result 0.98 0.95 0.11 asserted 1.00
Abrupt mean- and variance-shift experiments further show that the preferred adaptation rate depends on the residual dynamics. result 0.97 0.95 0.05 asserted 1.00
Overall, the results show that the effect of forecast delay is best understood relative to the time scale over which past residual information remains relevant. interpretation 0.99 0.92 0.87 asserted 0.99

What kind of paper the classifier reports this is: theoretical result 0.99 · evidence from proof or derivation 0.74

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 7,657 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (21)
Figure 1 of the paper
Figure 1. Interval score heatmap for Gaussian (i.i.d.) residuals across the
Figure 2 of the paper
Figure 2. Achieved interval score for AR(1) residuals at the selected step size
Figure 3 of the paper
Figure 3. Approximately score-minimizing step-size
Figure 4 of the paper
Figure 4. Interval score across the γ × τ grid for AR(1) residuals at three persistence levels (φ = 0.6, 0.85, 0.95). A dot marks mild undercoverage (0.86 ≤ coverage < 0.89) and a cross marks (coverage < 0.86); unmarked cells meet coverage. Local coverage error exhibits a different pattern from interval score. Although it generally in￾creases as the delay grows relative to the memory length, the curves do not collapse across persistence levels, indicating that rε alone does not explain differences in the magnitude of local calibration er￾ror. The full results are reported in Appendix C.1 (
Figure 5 of the paper
Figure 5. Achieved interval score at the selected step-size
Figure 6 of the paper
Figure 6. Ratio of the interval score obtained with normalized residuals to that obtained with
Figure 7 of the paper
Figure 7. Rolling coverage (top) and interval width (bottom) along the combined mean-shift trajec
Figure 8 of the paper
Figure 8. Rolling coverage (top) and interval width (bottom) along the combined variance-shift
Figure 9 of the paper
Figure 9. Achieved interval score at the selected step size
Figure 10 of the paper
Figure 10. Achieved interval score (top) and mismatch-minus-match conditional interval score (bot
Figure 11 of the paper
Figure 11. Normalized-to-unnormalized interval-score ratio for the two oracle normalization con
Figure 12 of the paper
Figure 12. Worst-case local coverage error at the approximately score-minimizing step size
Figure 13 of the paper
Figure 13. Worst-case local coverage error at the approximately score-minimizing step-size
Figure 14 of the paper
Figure 14. Rolling coverage (top) and interval width (bottom) along the combined mean-shift tra
Figure 15 of the paper
Figure 15. Rolling coverage (top) and interval width (bottom) along the combined variance-shift
Figure 16 of the paper
Figure 16. Rolling coverage (top) and interval width (bottom) along the combined mean-shift tra
Figure 17 of the paper
Figure 17. Rolling coverage (top) and interval width (bottom) along the combined variance-shift
Figure 18 of the paper
Figure 18. Selected step size γ ⋆ as a function of feedback delay for the mean-shift family (left) and variance-shift family (right). Colored curves show the step sizes selected for the isolated-shift experiments using R = 250. The solid and dashed black curves show the selections for the combined￾shift experiments using R = 250 and R = 100, respectively. Step sizes are selected separately for each configuration using the coverage-aware interval-score criterion. 52
Figure 19 of the paper
Figure 19. Local coverage error at γ ⋆ as a function of feedback delay. The top row presents the mean-shift family, and the bottom row presents the variance-shift family; the left and right columns use local evaluation windows of 100 and 250 observations, respectively. Colored curves show the isolated-shift experiments using R = 250. The solid and dashed black curves show the combined￾shift experiments using R = 250 and R = 100, respectively. Lower values indicate better local calibration. 53
Figure 20 of the paper
Figure 20. Achieved interval score (top) and mismatch-minus-match conditional interval score (bot
Figure 21 of the paper
Figure 21. Normalized-to-unnormalized interval-score ratio for the current-state and target-state

Figures are the authors', reproduced from the paper.

References (16)
  1. [1] Angelopoulos, A. N., Barber, R. F., and Bates, S. (2024). Online conformal prediction with decaying step sizes.arXiv preprint arXiv:2402.01139. 54 on PithPith's reviewarXiv
  2. [2] Angelopoulos, A. N., Candès, E. J., and Tibshirani, R. J. (2023). Conformal pid control for time series prediction.arXiv preprint arXiv:2307.16895 on PithPith's reviewarXiv
  3. [3] Barber, R. F., Candès, E. J., Ramdas, A., and Tibshirani, R. J. (2023). Conformal prediction beyond exchangeability.The Annals of Statistics, 51(2):816–845 on Pith
  4. [4] Bhatnagar, A., Wang, H., Xiong, C., and Bai, Y. (2023). Improved online conformal prediction via strongly adaptive online learning. InInternational Conference on Machine Learning, pages 2337–2363. PMLR on Pith
  5. [5] Cauchois, M., Gupta, S., and Duchi, J. C. (2021). Knowing what you know: valid and validated confidence sets in multiclass and multilabel prediction.Journal of machine learning research, 22(81):1–42 on Pith
  6. [6] Gibbs, I. and Candès, E. (2021). Adaptive conformal inference under distribution shift.Advances in Neural Information Processing Systems, 34:1660–1672 on Pith
  7. [7] Gibbs, I. and Candès, E. J. (2024). Conformal inference for online prediction with arbitrary distri- bution shifts.Journal of Machine Learning Research, 25(162):1–36 on Pith
  8. [8] Hamilton, J. D. (1994).Time Series Analysis. Princeton University Press. Ch. 22 covers Markov- switching models and two-state chain properties on Pith
  9. [9] Li, R., Menacho, D., and Rodríguez, A. (2026). Optimization-based online conformal prediction for multi-step forecasting.arXiv preprint arXiv:2508.13362 on PithPith's reviewarXiv
  10. [10] Shafer, G. and Vovk, V. (2008). A tutorial on conformal prediction.Journal of Machine Learning Research, 9(12):371–421 on Pith
  11. [11] Skalse, J., Pona, E., Simeone, O., and Paoletti, N. (2026). Online conformal prediction beyond feedback.arXiv preprint arXiv:2608.07139. Szabadváry, J. H. (2024). Adaptive conformal inference for multi-step ahead time-series forecasting online.arXiv preprint arXiv:2409.14792 on PithPith's reviewarXiv
  12. [12] Wang, B., Zecchin, M., and Simeone, O. (2025). Mirror online conformal prediction with intermittent feedback.arXiv preprint arXiv:2503.10345 on PithPith's reviewarXiv
  13. [13] Wang, X. and Hyndman, R. J. (2026). Online conformal inference for multi-step time series fore- casting.arXiv preprint arXiv:2410.13115 on PithPith's reviewarXiv
  14. [14] Yang, Z., Candès, E., and Lei, L. (2024). Bellman conformal inference: Calibrating prediction intervals for time series.arXiv preprint arXiv:2402.05203 on PithPith's reviewarXiv
  15. [15] Zaffran, M., Féron, O., Goude, Y., Josse, J., and Dieuleveut, A. (2022). Adaptive conformal pre- dictions for time series. InInternational Conference on Machine Learning, pages 25834–25866. PMLR. 55 on Pith
  16. [16] Zucchini, W. and MacDonald, I. L. (2009).Hidden Markov Models for Time Series: An Introduction Using R. Chapman and Hall/CRC. 56 on Pith
Cited by 1 Pith paper

Reviewed papers on Pith that cite this one. The full citation record

  1. Capacity-Constrained Delayed Feedback for Online Conformal Risk Control cs.LG · Nov 2026

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment