{"id":"77067416-fb2f-4b37-ab5c-6a9308d438a5","arxiv_id":"2412.19318","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Adaptive conformal inference can be made parameter-free with coin-betting updates (KT/ONS) while retaining a long-term coverage guarantee under bounded scores.","lead":"The paper proposes a parameter-free adaptive conformal prediction method based on coin-betting online learning, replacing learning-rate-tuned gradient descent. It proves asymptotic coverage under bounded scores and tests the method on simulated and real data streams.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix B Step 2 proves boundedness of the KT iterates only by assuming |s_t|≤D, which is the statement being proved; Theorem 2.1's coverage proof is therefore incomplete.","rationale":"Reader's weakest assumption is precisely the circularity in Step 2. I checked the proof: Eq. (8) is algebraically correct, but the bound after it uses |s_t|≤D to control both |s_t| and |s_t||Σg_i|; no prior result supplies this. Step 1's qualitative monotonicity cannot bound the size of an excursion because W_t can be large. Thus the contradiction in Step 3 has no support. The theorem is likely repairable: the feedback c_t is such that s_t>D forces c_t=−α and s_t<0 forces c_t=1−α, which should preclude unbounded growth, but the paper does not supply the argument. Credit is due for a clean algorithm and broad experiments; the empirical comparisons and the regret bound are independent of the fix. The appropriate verdict remains CONDITIONAL: revise the proof (or provide a formal verification of the boundedness lemma) before the long-term coverage claim is accepted.","tokens_in":19330,"tokens_out":16646,"duration_ms":147670,"concrete_test":"Re-derive the boundedness step without the circular assumption. Concretely, prove or disprove the uniform bound using only the explicit identities λ_{t+1}=(Σ_{i≤t}c_i)/(t+1), W_t=1+Σ_{i≤t}s_i c_i, c_t=1{S_t>s_t}−α, and the feedback rules (if s_t>D then c_t=−α; if s_t<0 then c_t=1−α). If the line following Eq. (8) is replaced by an induction that does not assume |s_t|≤D on the right-hand side, does |s_t|≤3D+1 follow? In parallel, run an adversarial search over score sequences S_t∈{0,D} for α∈{0.05,0.1,0.4} and horizon 10^5, recording max|s_t|; if the maximum exceeds 3D+1, Step 2's claim is false, and if it never does, the missing argument is still required.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step in Theorem 2.1 is the boundedness of the iterates |s_t|≤3D+1, asserted at the end of Step 2 of Appendix B. The proof uses Eq. (8) to bound |s_{t+1}−s_t|, and the displayed estimate immediately after (8) bounds |s_t| by D and |s_t Σ_{i≤t} g_i| by Dt. Both inequalities substitute the very conclusion being proved (|s_t|≤D) into the right-hand side. Eq. (7) controls only the weighted sum Σ g_i s_i, not |s_t| itself. Step 1 shows only that a radius above D is followed by a decrease and that a negative radius is followed by a positive one; it does not bound the size of an excursion. The contradiction in Step 3 requires the uniform bound to rule out the unbounded iterates that the KT wealth lower bound would otherwise produce when (1) fails. Since that uniform bound is not established, the long-term coverage guarantee is unproved as written. This is a technical gap internal to the proof, not the data assumption S_t≤D; the theorem may still be true, but the central claim currently rests on an unjustified assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a parameter-free adaptive conformal inference method based on coin-betting online convex optimization, specifically a Krichevsky-Trofimov (KT) betting strategy (with an Online Newton Step variant in the appendix). The principal theoretical claim (Theorem 2.1) is that, for bounded nonconformity scores with S_t ∈ [0, D], the KT-based algorithm achieves the long-term miscoverage guarantee (1) without any tuning. The paper also provides regret bounds, discusses the necessity of bounded scores, and presents simulations on changepoint, electricity demand, and stock price data comparing the betting approach against tuned OGD/SF-OGD baselines.","tokens_in":19539,"tokens_out":11288,"duration_ms":89208,"significance":"The contribution is potentially valuable: if the coverage guarantee is established, the method is a parameter-free adaptive conformal predictor that avoids learning-rate tuning, a recognized weakness of existing OGD-based methods. The paper gives an explicit, simple algorithm and a regret statement, and the empirical study is broad, covering simulated changepoint data, electricity demand, and multi-horizon stock price forecasting. The KT variant has no fitted constants, which is a strength. However, the central theorem's proof contains a gap in the boundedness step, so the main coverage guarantee is not yet proven as written; this limits the significance of the theoretical contribution and should be addressed before the results can be relied upon.","major_comments":[{"comment":"The proof of the uniform bound |s_t| ≤ 3D+1 is circular. The displayed estimate immediately after Eq. (8) bounds the right-hand side by (D + 1 + D(t−1) + 1 + Dt)/(t+1), where the term D for |s_t| and the term D(t−1) for |Σ_{i=1}^{t−1} g_i s_i| both assume that the iterates before time t satisfy |s_i| ≤ D. That is exactly the conclusion being proved. Step 1 only shows local directional behavior (a radius above D is followed by a decrease, and a nonnegative radius followed by a negative one is followed by a positive one); it does not bound the magnitude of an excursion above D or below 0, and the assertion that the decrease 'repeats until' the radius becomes ≤ D is not established. Consequently, the boundedness conclusion |s_t| ≤ 3D+1 is unsupported, and Step 3's contradiction depends on that bound to rule out the unbounded iterates that the KT wealth lower bound would produce when (1) fails. The coverage guarantee is therefore not proven as written.","section":"Appendix B, Step 2 (Eq. (8) and following display)"}],"minor_comments":[{"comment":"The text says code is provided as a supplement, but no URL or repository identifier appears in the manuscript; please add a working link or a detailed availability statement.","section":"Section 3 and Impact Statement"},{"comment":"The experiments report averages over 200–250 seeds or over a test period, but no standard errors, confidence intervals, or seed-level variability are given; several reported differences (e.g., KT versus OGD coverage around 0.9–1 percentage point) are small enough that variability information is needed to assess them.","section":"Figures 1–5 and Tables 1–7"},{"comment":"The caption contains a duplicated phrase 'versions of versions of online gradient descent'; also the footnote marker in Section 3 appears as 'Figure 6¹', which is confusing.","section":"Figure 5 caption"},{"comment":"The regret bound introduces a universal constant C without specifying its value or provenance; please state whether C is explicit and how it enters the proof of Theorem 2.1.","section":"Equation (6)"},{"comment":"Algorithm 2 is presented as part of the proposed approach and is evaluated in the experiments, but no coverage or regret guarantee is claimed for it; please add a sentence clarifying that the theoretical results apply only to the KT-based Algorithm 1.","section":"Appendix A, Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an arXiv preprint of a paper already published at ICML 2024 (PMLR 235); the present report is on the arXiv version. The missing boundedness argument in Theorem 2.1 is the main obstacle: the experimental study is suggestive but cannot substitute for the proof. If the authors can provide a valid bound on the iterates or an alternative argument, the contribution would be solid. No concerns about attribution or citation patterns; the only reproducibility gap is the missing code link."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely useful: it replaces the learning-rate dial that every OGD-based adaptive conformal predictor needs with a coin-betting update that has no tuning parameters. That is a real contribution to the conformal-inference subfield. The KT-based algorithm is simple, the regret bound is standard sub-linear, and the experiments are broad—changepoint regressions, electricity demand, multi-horizon stock forecasts—with honest comparisons. I believe the empirical claims: KT matches tuned OGD/SF-OGD in coverage, sometimes undercovers by a point or two, and produces narrower or similar intervals. The design has no fitted constants, and the betting machinery is squarely from Orabona–Pal and Cutkosky–Orabona, which is fine; the novelty is the bridge to adaptive conformal inference.\n\nThe problem is Theorem 2.1. Step 2 of Appendix B tries to prove the iterates stay in [-3D-1, 3D+1]. The derivation of Eq (7), the bound on the cumulative weighted subgradients, is fine. The trouble is Eq (8): after expressing s_{t+1}-s_t, the displayed bound uses |s_t| ≤ D on the right-hand side. That is precisely the kind of uniform control the step is meant to establish. Step 1 only shows qualitative behavior (if you are above D you get pushed down; a negative radius forces the next one positive); it does not cap an excursion. So the bound |s_t| ≤ 3D+1 is not proven, and the contradiction argument in Step 3, which relies on boundedness plus the KT wealth lower bound, does not close. The theorem may well be true, but as written the central coverage guarantee is unproved.\n\nMinor issues: no error bars in the plots, and the code link advertised in the Impact Statement does not appear in this version. The ONS variant has no coverage theorem, which is fine, but the abstract's 'we prove that our method controls long-term miscoverage' should be scoped to KT.\n\nIf this paper is under review, send it out. The idea is worth referee time, and the proof gap is identifiable and possibly fixable—maybe a direct induction or a potential-function argument. The reviewer should be asked to check whether Step 2 can be repaired, and to demand error bars and code. If it is an ICML paper already, the gap is worth an erratum or a fix in a revised version.\n\nMy take: conditional accept, with the main theorem repaired.","headline":"A genuinely useful parameter-free alternative to tuned OGD in adaptive conformal inference, but the long-term coverage theorem has a circular step in Appendix B and is not proven as written.","tokens_in":20093,"tokens_out":4149,"would_cite":false,"duration_ms":38474,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a coin-betting update for conformal prediction intervals can remove learning-rate tuning entirely while still achieving the nominal long-term miscoverage rate, and that the resulting parameter-free predictor is…","keywords":["adaptive conformal inference","conformal prediction","parameter-free online learning","coin betting","Krichevsky-Trofimov estimator","long-term coverage guarantee","distribution shift","pinball loss"],"falsifier":"Simulate Algorithm 1 on a long stream of bounded scores chosen adversarially from {0,D} while recording the running miscoverage; if any such stream yields a limiting miscoverage different from α, Theorem 2.1 is false. As a first check, verify whether |s_t| ever exceeds 3D+1 on such a stream.","tokens_in":19102,"feed_emoji":"🎯","tokens_out":7816,"duration_ms":67392,"temperature":0.7,"pith_summary":"This paper proposes a new way to build prediction intervals for data arriving one point at a time, with no assumptions about how the data are generated. Instead of tuning a learning rate as in online gradient descent, the method treats the interval radius as a bet placed on the outcomes of continuous coin flips and updates the radius by the Krichevsky-Trofimov betting strategy. The main theoretical claim, Theorem 2.1, is that for bounded nonconformity scores this parameter-free predictor drives the long-term miscoverage rate exactly to the nominal level α, with no user-set parameters. The experiments show it performs comparably to carefully tuned gradient-descent baselines, so a practitioner can automate prediction intervals without the most sensitive tuning decision.","feed_headline":"Coin betting tunes conformal intervals with no learning rate","feed_subtitle":"A parameter-free predictor provably hits the target miscoverage level and matches tuned gradient descent on real data streams.","key_machinery":"The central object is the coin-betting view of online convex optimization, specialized to pinball-loss quantile estimation. In Algorithm 1, the coin outcome is c_t = -g_t, where g_t is a subgradient of the pinball loss at the current radius; the gambler bets a fraction λ_{t+1} of current wealth W_t, so the next radius is s_{t+1} = λ_{t+1} W_t. The KT estimator provides the update λ_{t+1} = (t/(t+1))λ_t - g_t/(t+1). This machinery does two jobs: it yields a sub-linear regret bound with no tuned constants, and it converts any persistent deviation of the miscoverage rate from α into exponentially growing radii, which is the engine of Theorem 2.1.","core_discovery":"The central claim is that the price of learning-rate tuning can be eliminated entirely: a conformal predictor whose radii are updated by the Krichevsky-Trofimov betting rule (Algorithm 1 in the paper) satisfies the long-term coverage guarantee for any data stream whose nonconformity scores are bounded by an unknown constant D, with no user-set parameters. The proof builds a bridge between quantile estimation and betting: the coin outcomes are the negated subgradients of the pinball loss, the gambler's wealth is W_t = 1 - Σ g_i s_i, and the radius at time t+1 is s_{t+1} = λ_{t+1} W_t with λ updated by the KT rule. The argument shows that if the average miscoverage failed to converge to α, the wealth lower bound for the KT bettor would force the radii |s_t| to grow without bound, contradicting a claimed boundedness of the iterates; the paper concludes that coverage must converge. The claim is stated for α ∈ (0, 1/2) and relies on the scores lying in [0, D].","pith_inferences":["If the circular boundedness step in Appendix B is repaired or replaced, the same betting mechanism could plausibly extend to non-bounded or heavy-tailed scores through clipping or truncation, a direction the paper does not analyze.","The coverage guarantee is asymptotic; a finite-sample or non-asymptotic version, such as a probability bound on miscoverage deviation after T steps, is a natural next step that the paper leaves implicit.","Because the KT update carries no scale information at all, a direct comparison against alternative parameter-free or scale-free online optimizers on adversarially shifted streams would sharpen the claim that no tuning is needed."],"forward_implications":["If Theorem 2.1 stands, adaptive conformal inference can be run on arbitrary data streams with bounded scores and no learning-rate tuning while still converging to the target miscoverage frequency.","The same betting update can be run independently for each horizon in multi-step forecasting, so per-horizon intervals inherit the guarantee without extra tuning.","The method applies to any family of nested prediction sets parameterized by a single radius, including quantile-regression interval recalibration, not just symmetric absolute-error bands.","The sub-linear pinball-loss regret bound rules out trivial predictors that satisfy coverage by alternating between empty and full sets.","A practitioner loses at most a small finite-sample coverage gap relative to tuned gradient methods, while gaining invariance to the scale of residuals."],"supporting_citations":[{"why":"Supplies the coin-betting framework and the KT wealth lower bound used in the proof of Theorem 2.1.","marker":"Orabona & P´al (2016)"},{"why":"Provides the KT betting estimator whose wealth guarantee underlies the regret bound.","marker":"Krichevsky & Trofimov (1981)"},{"why":"Gives parameter-free reductions via coin betting in Banach spaces, used for the ONS variant and its regret analysis.","marker":"Cutkosky & Orabona (2018)"},{"why":"Introduced the OGD-based adaptive conformal inference baseline that the paper compares against and seeks to de-parameterize.","marker":"Gibbs & Cand`es (2021)"},{"why":"Provides the strongly adaptive SF-OGD conformal predictor used as a comparison baseline.","marker":"Bhatnagar et al. (2023)"},{"why":"Introduces scale-free online gradient descent, the basis of the SF-OGD baseline update.","marker":"Orabona & P´al (2018)"},{"why":"Supplies the Online Newton Step algorithm used as the second betting strategy.","marker":"Hazan et al., 2007"}],"fun_headline_variants":["Betting removes learning rate from conformal inference","No tuning: conformal prediction via betting","Parameter-free conformal inference by betting","Betting adapts conformal intervals sans learning rate","Adaptive conformal prediction without any tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee rests on an unproven claim that the prediction intervals stay within about three times the largest observed error, and the paper's proof of that claim assumes the intervals never exceed the largest observed error in the first place.","fun_headline_variants_meta":{"raw":{"variants":["Betting removes learning rate from conformal inference","No tuning: conformal prediction via betting","Parameter-free conformal inference by betting","Betting adapts conformal intervals sans learning rate","Adaptive conformal prediction without any tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000576,"raw_usage":{"total_tokens":2697,"prompt_tokens":902,"completion_tokens":1795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":1727}},"tokens_in":518,"tokens_out":1795,"duration_ms":13030,"temperature":1.0,"reasoning_tokens":1727,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:43:42.901443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate Algorithm 1 on a long stream of bounded scores chosen adversarially from {0,D} while recording the running miscoverage; if any such stream yields a limiting miscoverage different from α, Theorem 2.1 is false. As a first check, verify whether |s_t| ever exceeds 3D+1 on such a stream.","supporting_citations":[{"cited_title":"and Pál, D","cited_arxiv_id":null,"evidence_quote":"Supplies the coin-betting framework and the KT wealth lower bound used in the proof of Theorem 2.1."},{"cited_title":"and Trofimov, V","cited_arxiv_id":null,"evidence_quote":"Provides the KT betting estimator whose wealth guarantee underlies the regret bound."},{"cited_title":"and Orabona, F","cited_arxiv_id":null,"evidence_quote":"Gives parameter-free reductions via coin betting in Banach spaces, used for the ONS variant and its regret analysis."},{"cited_title":"and Cand \\`e s, E","cited_arxiv_id":null,"evidence_quote":"Introduced the OGD-based adaptive conformal inference baseline that the paper compares against and seeks to de-parameterize."},{"cited_title":"Improved online conformal prediction via strongly adaptive online learning","cited_arxiv_id":null,"evidence_quote":"Provides the strongly adaptive SF-OGD conformal predictor used as a comparison baseline."},{"cited_title":"and Pál, D","cited_arxiv_id":null,"evidence_quote":"Introduces scale-free online gradient descent, the basis of the SF-OGD baseline update."},{"cited_title":"Logarithmic regret algorithms for online convex optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the Online Newton Step algorithm used as the second betting strategy."}],"review_version":1}