Online Conformal Prediction

• Online Conformal Prediction is a sequential method that constructs calibrated prediction sets using feedback, ensuring long-run control over miscoverage rates.
• It measures efficiency through metrics like interval width, volume, and regret, comparing online performance to the best fixed hindsight intervals.
• The approach integrates techniques such as quantile loss optimization, mirror descent, and Bayesian updates to deliver robust, parameter-free updates even against adversarial data.

Online conformal prediction (OCP) is the sequential analogue of conformal prediction in which prediction sets are produced one round at a time and updated after feedback, typically without assuming exchangeability and often against arbitrary or adversarial data streams. Across the recent literature, the object output at time $t$ may be an interval $I_t\subseteq[0,1]$, a radius $\tau_t$ defining $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$, or a threshold $q_t$ or $r_t(\alpha)$ that induces a set $\{y:s_t(X_t,y)\le q_t\}$; the common aim is long-run control of the miscoverage rate at level $\alpha$ together with efficiency, usually measured by interval width, set size, volume, or regret to a fixed comparator (Srinivas, 3 Jul 2025, Liu et al., 3 Feb 2026, Zhang et al., 2024).

1. Core protocol and quantitative objectives

A standard OCP protocol fixes a horizon $T$ and a target miscoverage rate $\alpha\in[0,1]$. In the interval-only formulation, at each round $I_t\subseteq[0,1]$0 the learner outputs an interval $I_t\subseteq[0,1]$1, then Nature reveals $I_t\subseteq[0,1]$2. Coverage is tracked by

$I_t\subseteq[0,1]$3

and efficiency by

$I_t\subseteq[0,1]$4

The fixed-interval comparator is

$I_t\subseteq[0,1]$5

so OCP becomes an online analogue to the problem of constructing efficient confidence intervals (Srinivas, 3 Jul 2025).

A second formulation, used widely in regression-style OCP, starts from a predictor $I_t\subseteq[0,1]$6, a nonnegative radius $I_t\subseteq[0,1]$7, and the interval

$I_t\subseteq[0,1]$8

With residual score $I_t\subseteq[0,1]$9, coverage is the event $\tau_t$0, equivalently $\tau_t$1. Marginal miscoverage is encoded through

$\tau_t$2

and many results target $\tau_t$3 or explicit finite-time bounds (Bharti et al., 29 May 2026).

Several papers reduce marginal OCP to sequential quantile prediction. In that view, one observes a stream of realized scores $\tau_t$4 and, before seeing $\tau_t$5, predicts thresholds $\tau_t$6 that act as online $\tau_t$7-quantile estimates. The associated pinball or quantile loss appears in several equivalent forms, including

$\tau_t$8

and

$\tau_t$9

with subgradients that are directly tied to coverage frequency errors (Zhang et al., 2024, Liu et al., 3 Feb 2026).

2. Efficiency guarantees and the adversarial–exchangeable separation

A distinctive recent development is the explicit treatment of efficiency as a first-class online objective rather than an auxiliary empirical feature. In the efficiency-focused interval formulation, OCP asks simultaneously for coverage near $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$0 and average interval length competitive with the best fixed interval in hindsight. The paper "Online Conformal Prediction with Efficiency Guarantees" formalizes a $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$1-approximate efficiency requirement,

$C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$2

and studies the resulting trade-off between mistakes and volume under exchangeable versus arbitrary sequences (Srinivas, 3 Jul 2025).

For exchangeable sequences, the main positive result is that one can essentially match the best fixed hindsight interval with no volume blow-up. Running the one-parameter meta-algorithm with $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$3 and

$C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$4

yields, with probability at least $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$5 over the random order of $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$6,

$C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$7

and

$C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$8

The result is stated as recovery of coverage $C_t(X_t)=[\hat Y_t-\tau_t,\hat Y_t+\tau_t]$9 while matching the same length as the best hindsight interval up to $q_t$0.

For arbitrary sequences, the paper proves a qualitatively different picture. Any algorithm that guarantees on every sequence

$q_t$1

must, on some sequence, incur at least

$q_t$2

mistakes. Thus driving $q_t$3 close to $q_t$4 forces an unbounded blow-up in mistakes, while keeping mistakes near $q_t$5 requires a larger $q_t$6.

The same work gives a deterministic meta-algorithm that is robust to an adaptive adversary and recovers all Pareto-optimal trade-offs between the volume approximation $q_t$7 and mistake overhead. Its update rule maintains a candidate interval $q_t$8, resets when empirical coverage of past points falls below $q_t$9, computes a minimum-volume interval $r_t(\alpha)$0 covering the required fraction of past points, and replaces $r_t(\alpha)$1 by a centered expansion of $r_t(\alpha)$2. With $r_t(\alpha)$3, it guarantees

$r_t(\alpha)$4

and

$r_t(\alpha)$5

The paper’s final structural claim is a separation theorem: no single algorithm can simultaneously be Pareto-optimal for arbitrary sequences and optimal for exchangeable sequences. More concretely, any method that guarantees the adversarial max-volume bound on all sequences must, on some i.i.d. sequence, make at least

$r_t(\alpha)$6

mistakes in expectation. This gap is presented as being in contrast to the classical online learning problem.

3. Quantile-loss optimization, parameter-free updates, and prior-informed OCP

A large part of the current OCP literature is organized around online optimization of the pinball loss. One influential synthesis identifies linearized regret,

$r_t(\alpha)$7

as the key quantity controlling coverage. If an online algorithm guarantees $r_t(\alpha)$8 for all comparators $r_t(\alpha)$9, then a Fenchel-conjugate argument yields finite-time control of the cumulative coverage error. On that basis, "Online Conformal Prediction via Universal Portfolio Algorithms" proposes UP-OCP, a parameter-free method that reduces interval-valued OCP to a two-asset portfolio selection problem. Under the polynomial-growth condition $\{y:s_t(X_t,y)\le q_t\}$0, it proves a finite-time miscoverage bound of order $\{y:s_t(X_t,y)\le q_t\}$1, with the rate improving to $\{y:s_t(X_t,y)\le q_t\}$2 when $\{y:s_t(X_t,y)\le q_t\}$3 or $\{y:s_t(X_t,y)\le q_t\}$4, and implements the update in $\{y:s_t(X_t,y)\le q_t\}$5 time per step with no manual learning-rate tuning (Liu et al., 3 Feb 2026).

A separate line of work uses regularized empirical distributions rather than first-order linearizations. "The Benefit of Being Bayesian in Online Conformal Prediction" starts from the observation that gradient-based baselines run independently across many confidence levels can violate monotonicity, so that $\{y:s_t(X_t,y)\le q_t\}$6 may occur even when $\{y:s_t(X_t,y)\le q_t\}$7. Its Bayesian OCP algorithm forms a regularized belief

$\{y:s_t(X_t,y)\le q_t\}$8

and outputs $\{y:s_t(X_t,y)\le q_t\}$9. Because $\alpha$0 depends only on the multiset of past scores, the outputs are permutation-invariant, and monotonicity in $\alpha$1 holds by construction. The same paper proves that the method is a non-linearized FTRL algorithm on quantile loss, achieves $\alpha$2 adversarial regret for each $\alpha$3, supports multiple arbitrary confidence-level queries on the fly, and, under i.i.d. scores, attains actual coverage within $\alpha$4 of the target (Zhang et al., 2024).

Prior information and limited feedback are also incorporated through mirror methods. "Mirror Online Conformal Prediction with Intermittent Feedback" defines a strongly convex regularizer

$\alpha$5

uses the mirror map $\alpha$6, and performs online mirror descent with an importance-weighted subgradient $\alpha$7. The resulting IM-OCP algorithm retains deterministic long-term coverage,

$\alpha$8

so $\alpha$9 gives $T$0 coverage error, while also achieving $T$1 regret and $T$2 memory (Wang et al., 13 Mar 2025).

4. Group-conditional guarantees and nested risk spectra

Marginal validity is not the only coverage target studied in OCP. "Parameter-Free and Group Conditional Online Conformal Prediction" introduces soft group memberships $T$3, group counts $T$4, and group-conditional miscoverage

$T$5

Its algorithm POGO parameterizes the radius by

$T$6

and learns $T$7 through a separate two-stock Universal Portfolio for each group. Under the growth condition $T$8, the paper proves the finite-time bound

$T$9

with the stated consequence

$\alpha\in[0,1]$0

The same work describes POGO as parameter-free, group-conditional, and finite-time distribution-free in a fully online, non-exchangeable setting, and reports worst-group coverage within $\alpha\in[0,1]$1–$\alpha\in[0,1]$2 of target on MIMIC-IV and stock-forecasting benchmarks (Bharti et al., 29 May 2026).

A different extension concerns simultaneous calibration across many risk levels. "Online Conformal Prediction: Enforcing monotonicity via Online Optimization" studies thresholds $\alpha\in[0,1]$3 on a grid $\alpha\in[0,1]$4 under the nesting constraint

$\alpha\in[0,1]$5

so that $\alpha\in[0,1]$6. The paper gives two online mirror-descent instantiations: an exponentiated-gradient method over nonnegative gap variables and a projected-gradient method using isotonic projection onto

$\alpha\in[0,1]$7

Under a density lower bound $\alpha\in[0,1]$8, both methods obtain average squared quantile-tracking bounds, and under an additional Lipschitz-CDF assumption, empirical miscoverage at each level converges to $\alpha\in[0,1]$9 at $I_t\subseteq[0,1]$00. In the reported synthetic and inflation-forecasting experiments, the proposed methods have zero nestedness violations, and the exponentiated-gradient method yields smoother and narrower fan charts than independent single-$I_t\subseteq[0,1]$01 tracking (Rivera et al., 12 May 2026).

A related time-series development replaces the binary signal $I_t\subseteq[0,1]$02 by bounded relevance functions $I_t\subseteq[0,1]$03, where $I_t\subseteq[0,1]$04, and inserts these into PID and ECI threshold updates. Under boundedness and saturation assumptions, the proposed PID-Prop, PID-Full, PID-Half, and ECI-R updates retain long-run coverage, while experiments on Amazon, Google, Microsoft, and Delhi temperature report tighter intervals in some settings and wider intervals when needed to recover nominal coverage (Dupuy et al., 3 Oct 2025).

5. Partial feedback, noisy labels, and corrupted observations

Several OCP variants address settings in which the learner does not receive clean full feedback after every prediction. In "Online Conformal Prediction with Adversarial Semi-bandit Feedback via Regret Minimization," each thresholded conformal set is treated as a bandit arm, the true label is revealed only when it lies inside the constructed set, and the learner observes only a semi-bandit feedback otherwise. After discretizing thresholds into a grid $I_t\subseteq[0,1]$05, the paper designs the OCP-Unlock+ algorithm by adapting $I_t\subseteq[0,1]$06 to exploit full unlocking when $I_t\subseteq[0,1]$07 and monotonic partial unlocking when $I_t\subseteq[0,1]$08. Its high-probability regret bound implies

$I_t\subseteq[0,1]$09

so the long-run miscoverage approaches $I_t\subseteq[0,1]$10 even against an adaptive adversary (Yang et al., 20 Apr 2026).

Intermittent validation also appears in application-specific OCP. "Online Conformal Probabilistic Numerics via Adaptive Edge-Cloud Offloading" calibrates highest-probability-density sets from a probabilistic linear solver when cloud feedback is only sporadically available. With score

$I_t\subseteq[0,1]$11

prediction set $I_t\subseteq[0,1]$12, validation probability

$I_t\subseteq[0,1]$13

and intermittent update

$I_t\subseteq[0,1]$14

the paper proves

$I_t\subseteq[0,1]$15

under $I_t\subseteq[0,1]$16 and bounded scores (Hou et al., 18 Mar 2025).

Robustness to noisy supervision has become a separate topic. "Robust Online Conformal Prediction under Uniform Label Noise" shows that uniform label noise creates a persistent gap between the actual miscoverage rate and the desired rate $I_t\subseteq[0,1]$17 for standard updates, and introduces a robust pinball loss that is an unbiased estimator of the clean loss: $I_t\subseteq[0,1]$18 Its threshold updates eliminate the coverage gap for both constant and dynamic learning rates, with $I_t\subseteq[0,1]$19 convergence of both empirical and expected coverage errors (Xi et al., 30 Jan 2025).

A still harsher model replaces clean error indicators by arbitrarily corrupted binary feedback. "Online Conformal Prediction with Corrupted Feedback" defines corruption flags $I_t\subseteq[0,1]$20 and studies two robust schemes. F-ROCP exploits the fact that when $I_t\subseteq[0,1]$21, the true gradient is known from the threshold itself. AC-ROCP adds an active compensation term based on predicted corruption probabilities and probing rounds. The main guarantees bound the miscoverage error by a standard $I_t\subseteq[0,1]$22-type term plus an explicit bias term involving in-range corruption, and the empirical study on CIFAR-100 and AVA regression reports markedly improved calibration and significantly smaller prediction sets than baseline OCP under corrupted feedback (Wang et al., 19 May 2026).

6. Ensemble methods, drift adaptation, and safety-critical deployments

Because efficiency depends strongly on the underlying predictive model, several papers treat OCP as a model-selection or model-aggregation problem. "Conformal online model aggregation" considers $I_t\subseteq[0,1]$23 conformal experts $I_t\subseteq[0,1]$24 and forms a randomized weighted majority-vote set

$I_t\subseteq[0,1]$25

If each expert has marginal or conditional coverage at level $I_t\subseteq[0,1]$26, then the aggregated set satisfies

$I_t\subseteq[0,1]$27

and its total size obeys

$I_t\subseteq[0,1]$28

with only $I_t\subseteq[0,1]$29 additional arithmetic per round (Gasparin et al., 2024). Later multi-model methods replace full aggregation by adaptive subset selection. The graph-based GMOCP construction samples a subset $I_t\subseteq[0,1]$30 of candidate models through a bipartite graph, updates model-specific miscoverage parameters via scale-free OGD, and proves

$I_t\subseteq[0,1]$31

with per-step complexity $I_t\subseteq[0,1]$32, while the related graph-structured feedback and EGMOCP variants add prediction-set-size feedback and establish valid coverage together with sublinear regret (Hajihashemi et al., 4 Jan 2026, Hajihashemi et al., 26 Jun 2025).

Nonstationary calibration can also be handled by explicit drift detection or retrospective recalibration. "Optimal training-conditional regret for online conformal prediction" studies independent data under change-point and smooth-drift models, uses drift detection inside split-conformal and full-conformal style algorithms, and proves minimax-optimal training-conditional regret rates: $I_t\subseteq[0,1]$33 for pretrained scores, and analogous rates with $I_t\subseteq[0,1]$34 and a stability factor $I_t\subseteq[0,1]$35 for online-trained scores (Liang et al., 18 Feb 2026). "Online Conformal Inference with Retrospective Adjustment for Faster Adaptation to Distribution Shift" instead recomputes leave-one-out residuals and predictions through efficient update formulas inside a sliding-window Jackknife+ scheme; it preserves long-term coverage when combined with ACI-type updates and reduces naive $I_t\subseteq[0,1]$36 computation to $I_t\subseteq[0,1]$37 per step (Jun et al., 6 Nov 2025).

Safety-critical control has motivated a further generalization from one-step labels to delayed multi-step scores. "Staggered Integral Online Conformal Prediction for Safe Dynamics Adaptation with Multi-Step Coverage Guarantees" defines an integral nonconformity score

$I_t\subseteq[0,1]$38

runs $I_t\subseteq[0,1]$39 staggered OCP threads to accommodate the delay, and proves both long-run multi-step coverage and long-run safety when coupled with robust tube MPC: $I_t\subseteq[0,1]$40 (Cherenson et al., 7 Apr 2026).

These developments suggest that OCP is no longer a single threshold-updating heuristic but a family of online calibration problems spanning efficiency-competitive interval prediction, multigroup fairness, nested risk control, partial-feedback learning, model aggregation, drift adaptation, and safety-critical decision systems.