Online Localized Conformal Prediction

• OLCP is a method that combines sequential online calibration with covariate-localized weighting to provide efficient prediction intervals in heterogeneous settings.
• It replaces a global score calibration rule with localized, weighted score distributions that adapt dynamically as covariate conditions change.
• OLCP incorporates online updates and expert aggregation (OLCP-Hedge) to balance prediction set width and miscoverage, ensuring long-run marginal coverage.

Online Localized Conformal Prediction (OLCP) denotes online conformal procedures that combine sequential adaptation with covariate-dependent localization, rather than relying on a single global calibration rule. In the formulation introduced under the name OLCP, the objective is to retain online long-run validity in non-exchangeable settings while improving efficiency under covariate heterogeneity by replacing global score calibration with localized, weighted score distributions around the current covariate and updating the nominal miscoverage level online (Lai et al., 6 May 2026). In current usage, the term also covers closely related structured variants in which localization is imposed over input regions, over peer units in panel data, or through input-dependent threshold functions, although the precise validity notion remains long-run marginal rather than exact conditional coverage (Kim et al., 2024).

1. Conceptual scope and lineage

OLCP emerged from the intersection of two previously separate developments. The first is online conformal prediction for sequential and time-series settings, where exchangeability fails and methods such as adaptive conformal inference replace finite-sample marginal validity by long-run calibration of empirical miscoverage. The second is localized conformal prediction, where calibration is made covariate-dependent through weights, local neighborhoods, or related mechanisms in order to reflect heterogeneity in the score distribution across the feature space (Lai et al., 6 May 2026).

Offline and batch precursors established the main localization motifs later reused by OLCP. "On training locally adaptive CP" proposes localized adaptivity through an $X$-dependent monotone transformation of the conformity score, preserving marginal validity while producing $X$-dependent interval sizes (Colombo, 2023). "Localized conformal model selection" develops a localized conformal model-selection framework that integrates local adaptivity with post-selection validity by using weighted residual quantiles and symmetry-preserving surrogate intervals (Wang et al., 22 Feb 2026). In image classification with vision-LLMs, localized split-conformal methods use a test-specific weighted empirical distribution of calibration scores, showing that raw cosine similarity is often insufficient and that a nonlinear similarity transformation can materially reduce mean set size while preserving marginal coverage (Fuchs et al., 30 Jun 2026).

The online side of the lineage supplied the sequential calibration machinery. Strongly adaptive online conformal methods showed that standard regret is insufficient in changing environments and introduced interval-local regret over all contiguous time windows, but their localization is temporal rather than covariate-based (Bhatnagar et al., 2023). Parameter-free global online conformal algorithms based on universal portfolios and mirror descent established that one can retain long-run online coverage under adversarial or intermittent-feedback regimes without feature-local calibration (Liu et al., 3 Feb 2026, Wang et al., 13 Mar 2025). OLCP, in the strict modern sense, is the step that combines such online adaptation with explicit covariate-local weighting (Lai et al., 6 May 2026).

2. Core mechanics of OLCP

In the canonical OLCP formulation, data arrive sequentially as $\{(X_t,Y_t)\}_{t=1}^T$, a score function $s(x,y)$ is fixed, and at time $t$ the learner observes $X_t$, outputs a prediction set $C_t(X_t)$, then observes $Y_t$. In regression experiments, the score is the absolute residual $S_t = |Y_t-\hat Y_t|$, equivalently $s(x,y)=|y-\hat f(x)|$ for a pretrained predictor $X$0 (Lai et al., 6 May 2026).

Instead of calibrating with a global empirical score distribution, OLCP uses a rolling window

$X$1

past realized scores $X$2, and a bandwidth-dependent localizer $X$3. For a query covariate $X$4, the normalized local weights are

$X$5

These weights define the localized empirical score distribution

$X$6

Prediction sets are then indexed by a nominal level $X$7: $X$8 OLCP predicts with $X$9, where $\{(X_t,Y_t)\}_{t=1}^T$0 is updated online (Lai et al., 6 May 2026).

The online update is inherited from adaptive conformal inference but applied to the localized family: $\{(X_t,Y_t)\}_{t=1}^T$1 This gives OLCP two simultaneous adaptation axes. The localized weighted quantile changes with $\{(X_t,Y_t)\}_{t=1}^T$2, so interval width can vary across covariate space; the scalar recursion on $\{(X_t,Y_t)\}_{t=1}^T$3 changes over time, so the method can react to nonstationarity even when the covariate-local score law drifts (Lai et al., 6 May 2026).

The paper also gives a pinball-loss interpretation. Defining

$\{(X_t,Y_t)\}_{t=1}^T$4

the update is exactly projected online gradient descent on a sequence of pinball losses $\{(X_t,Y_t)\}_{t=1}^T$5. This places OLCP within the online convex optimization view of online conformal prediction while making the underlying prediction family explicitly local rather than global (Lai et al., 6 May 2026).

3. Bandwidth aggregation and OLCP-Hedge

Bandwidth choice is the main hyperparameter sensitivity in localized conformal methods. Small $\{(X_t,Y_t)\}_{t=1}^T$6 yields stronger localization and potentially sharper adaptation to heterogeneity, but with higher variance because fewer effective calibration points contribute. Large $\{(X_t,Y_t)\}_{t=1}^T$7 stabilizes the quantile estimate but can collapse the method back toward global calibration. OLCP-Hedge addresses this by turning bandwidth selection into an online expert aggregation problem with an explicit coverage constraint (Lai et al., 6 May 2026).

Suppose $\{(X_t,Y_t)\}_{t=1}^T$8 OLCP experts are defined by bandwidths $\{(X_t,Y_t)\}_{t=1}^T$9. Expert $s(x,y)$0 outputs a prediction set $s(x,y)$1, has size

$s(x,y)$2

and error

$s(x,y)$3

The meta-learner maintains a distribution $s(x,y)$4, samples $s(x,y)$5, and deploys $s(x,y)$6. For $s(x,y)$7, OLCP-Hedge defines

$s(x,y)$8

so $s(x,y)$9 is expected set size and $t$0 is expected excess miscoverage (Lai et al., 6 May 2026).

The constrained objective is to minimize cumulative expected size while keeping cumulative positive excess miscoverage sublinear. To enforce this, OLCP-Hedge maintains a virtual queue

$t$1

and uses the surrogate loss

$t$2

AdaHedge is used as the simplex optimizer. Operationally, the queue grows when excess miscoverage accumulates, thereby increasing the penalty on aggressive experts with small sets but poor coverage; when the queue is small, the optimizer can favor narrower experts more strongly (Lai et al., 6 May 2026).

OLCP-Hedge therefore does not treat bandwidth selection as unconstrained model selection. It is a constrained online convex optimization scheme in which coverage is encoded as a running resource constraint and size is the primary objective. This makes the method structurally different from ordinary expert averaging or from post hoc bandwidth tuning on a held-out set (Lai et al., 6 May 2026).

4. Validity notions and theoretical guarantees

The validity target in OLCP is long-run marginal coverage, not exact finite-sample exchangeable validity and not covariate-conditional coverage. The paper states the goal as

$t$3

For OLCP itself, the main proposition is a boundary-corrected telescoping identity. Writing

$t$4

the recursion implies

$t$5

If $t$6, then $t$7; if $t$8, then $t$9 (Lai et al., 6 May 2026). The theoretical obstruction is therefore not localization itself but the boundary behavior induced by projection of $X_t$0 into $X_t$1.

For OLCP-Hedge, the guarantees are expectation-based and comparator-relative. Under bounded expert sizes and a uniformly feasible comparator $X_t$2 satisfying $X_t$3 for all $X_t$4, the paper proves expected size regret

$X_t$5

and cumulative expected excess miscoverage

$X_t$6

These are sublinear, so average excess miscoverage vanishes while expected size competes with the best fixed feasible mixture of bandwidth experts in hindsight (Lai et al., 6 May 2026).

Related OLCP-style methods adopt different validity notions. In Bayesian optimization, LOCBO uses an input-dependent threshold function $X_t$7 in an RKHS and proves long-run deterministic coverage for noisy observations plus a long-run probabilistic guarantee for latent objective values under minimal noise assumptions (Kim et al., 2024). In non-exchangeable panel data, W-TQA proves a stepwise conditional miscoverage bound of the form

$X_t$8

where $X_t$9 is a localization-dependent mismatch term, together with long-run guarantees on the revealed-feedback subsequence and an all-round time-averaged guarantee under MCAR feedback (Tu et al., 18 May 2026). These results clarify that OLCP literature presently centers on online long-run or stepwise approximate guarantees, rather than exact local conditional validity.

5. Terminology, boundaries, and common confusions

The term OLCP is narrower than “online conformal prediction with any local flavor.” In the strict sense established by "Online Localized Conformal Prediction," OLCP requires both online adaptation and explicit covariate-dependent localization through local weights or an equivalent input-dependent calibration device (Lai et al., 6 May 2026). Several neighboring literatures are closely related but technically distinct.

A first distinction is between feature-space localization and temporal localization. "Improved Online Conformal Prediction via Strongly Adaptive Online Learning" introduces strongly adaptive regret over all contiguous time intervals and is therefore a core paper for temporally localized online conformal prediction, but it does not provide context-conditional or covariate-local calibration (Bhatnagar et al., 2023). A second distinction is between feature-space localization and groupwise or discrete localization. "Parameter-Free and Group Conditional Online Conformal Prediction" studies group-conditional coverage under arbitrary nonstationarity; the paper can be read as developing a discrete, group-based version of localization, but its guarantees are attached to predefined soft groups rather than to a continuum of covariate neighborhoods (Bharti et al., 29 May 2026).

A third distinction concerns methods that are online but global. IM-OCP is a mirror-descent-based online conformal method with intermittent feedback, but it maintains a single global threshold and is therefore not localized in the usual covariate-conditional sense (Wang et al., 13 Mar 2025). UP-OCP is similarly global: it learns one online radius by a universal-portfolio update and targets long-run marginal coverage for arbitrary data streams, without feature-dependent thresholds (Liu et al., 3 Feb 2026). Methods enforcing monotonicity across multiple confidence levels are also complementary rather than localized: the corresponding online optimization machinery operates on a family of global thresholds $C_t(X_t)$0 and does not localize by covariates (Rivera et al., 12 May 2026).

A fourth distinction concerns nomenclature. "Sepsyn-OLCP" uses “OLCP” to mean online learning plus conformal prediction, but its conformal component is an EnbPI-style sequential interval method and the calibration offsets are pooled per arm rather than localized by patient similarity; the paper is broadly related through online/sequential conformal prediction but not localized conformal prediction proper (Zhou et al., 18 Mar 2025).

These distinctions matter because OLCP is often misread as a synonym for any online conformal method with context, prior information, group structure, or model selection. The present literature instead separates at least four notions: global online calibration, temporally localized online calibration, group-conditional online calibration, and covariate-localized online calibration. OLCP, in the narrow sense, refers to the fourth.

6. Variants, applications, and empirical patterns

OLCP has already diversified into several application-specific and structurally specialized forms. LOCBO embeds localized online conformal calibration directly into Bayesian optimization. It uses a GP-based conformity score,

$C_t(X_t)$1

and a function-valued threshold $C_t(X_t)$2, updated by online kernel gradient descent. The resulting intervals calibrate a corrected likelihood for the noisy observation, which is then denoised to obtain a calibrated posterior over the latent objective function. Under model misspecification, especially heteroscedastic noise, LOCBO consistently outperforms state-of-the-art Bayesian optimization baselines, and localized LOCBO with finite bandwidth outperforms its non-localized limit $C_t(X_t)$3 (Kim et al., 2024).

Panel-data OLCP takes a different structural form. W-TQA exploits the fact that when a forecast is needed for one target unit, contemporaneous outcomes from related units may already be observed and can serve as a calibration panel. Localization is imposed over units, using Gaussian-kernel weights on distances between running average feature profiles,

$C_t(X_t)$4

while a Gibbs–Candès-style update adapts $C_t(X_t)$5 whenever target feedback is revealed. Empirically, W-TQA improves worst-unit coverage through adaptive interval-width allocation rather than uniform inflation, and its two states are complementary: similarity weights matter most when feedback is sparse, while the adaptive level improves further as feedback accumulates (Tu et al., 18 May 2026).

Beyond direct OLCP methods, several adjacent strands indicate where the field is moving. Localized conformal model selection shows how feature-dependent local adaptation and data-dependent model choice can be combined without breaking marginal validity under exchangeability, suggesting a template for future online safe-model-selection mechanisms (Wang et al., 22 Feb 2026). Batch localized image classification with transformed cosine similarity shows that the choice of localizer is critical in high-dimensional embedding spaces and that raw similarity can be too diffuse, a practical lesson likely to transfer to online localized classification (Fuchs et al., 30 Jun 2026). Group-conditional online methods indicate that soft partitions or basis-defined regions may provide a tractable intermediate step between global online calibration and fully continuous local coverage (Bharti et al., 29 May 2026).

The main empirical pattern across the literature is consistent. When heterogeneity is primarily temporal, globally calibrated but strongly adaptive online methods already help (Bhatnagar et al., 2023). When heterogeneity is tied to covariates, peer groups, or input regions, explicitly localized online calibration yields narrower prediction sets or improved tail-group reliability relative to global baselines (Lai et al., 6 May 2026, Kim et al., 2024, Tu et al., 18 May 2026). This suggests that the central value of OLCP is not validity in isolation, but the ability to preserve long-run online calibration while reallocating uncertainty to the regions of the input space where it is actually needed.