Local Conformal Quantile Estimation

• Local conformal quantile estimation is a framework that combines covariate-dependent quantile regression with localized conformal calibration to yield adaptive prediction intervals.
• These methods employ local adjustments via clustering, neighborhood-based corrections, or model-native partitions to handle heteroscedasticity and spatial variability.
• They tackle the challenge of conditional validity by offering marginal or subgroup-specific guarantees, balancing global and local calibration trade-offs.

Local conformal quantile estimation denotes a family of predictive inference methods that combines quantile estimation with conformal calibration to construct prediction sets for a future response whose width, shape, or calibration rule varies with the covariates. In the canonical formulation, lower and upper conditional quantiles are estimated as functions of $x$, and a conformal step restores finite-sample marginal coverage under exchangeability. More localized formulations replace the single global calibration correction by cluster-specific, neighborhood-based, subgroup-specific, spatial, or model-native local quantiles. A fundamental boundary in this literature is between prediction for a future $Y$ and inference for a fixed localized quantile functional: the latter is exemplified by localized quantile confidence intervals under a reweighted law $dQ/dP(x,y)\propto K((x_0-x)/h)$, which target a fixed parameter $\theta_p=Q_Y^{-1}(p)$ rather than a predictive set for a new response (Romano et al., 2019, Jang et al., 2023).

1. Conceptual foundations

The standard regression setting is i.i.d. data $(X_1,Y_1),\dots,(X_n,Y_n)$, with the objective of constructing a set $\widehat C_n(X_{n+1})$ such that

$$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$

Within this framework, “local” does not have a single meaning. In some papers it means that the interval width varies with $x$ because the lower and upper quantile functions are themselves covariate-dependent; in others it means that calibration is performed within a neighborhood, subgroup, or partition cell associated with the query point; in still others it means validity with respect to a localized population around the test point rather than the global test distribution (Romano et al., 2019, Alaa et al., 2023).

A second conceptual distinction concerns the inferential target. Conformalized quantile methods target prediction sets for a random future outcome. By contrast, “Tight Distribution-Free Confidence Intervals for Local Quantile Regression” defines locality by reweighting the covariate distribution around $x_0$ while leaving $P_{Y\mid X}$ unchanged, and then studies confidence intervals for the localized marginal quantile

$Y$0

where $Y$1 is the reweighted law. That target is a fixed parameter under a localized distribution, not a predictive interval for a new $Y$2. This distinction is central because prediction intervals must absorb irreducible outcome noise, whereas confidence intervals for fixed localized quantiles can shrink to zero width with growing sample size (Jang et al., 2023).

A third distinction concerns where localization is introduced. Some methods localize the base estimator of conditional quantiles; some localize the conformal calibration step; some localize both. This yields several non-equivalent notions of “local conformal quantile estimation,” ranging from locally adaptive but globally calibrated intervals to subgroup-localized or neighborhood-localized coverage statements (Romano et al., 2019, Jiang et al., 2024).

2. Canonical formulation: conformalized quantile regression

The baseline construction is Conformalized Quantile Regression (CQR), introduced by Romano et al. It begins with two quantile regressors fitted on a training split,

$Y$3

typically with $Y$4 and $Y$5. On a separate calibration split, the nonconformity score is

$Y$6

and the conformal threshold is the order statistic

$Y$7

The final interval is

$Y$8

Under exchangeability, this interval satisfies finite-sample marginal coverage, and the achieved coverage lies between $Y$9 and $dQ/dP(x,y)\propto K((x_0-x)/h)$0 (Romano et al., 2019).

The local adaptivity of CQR is structural rather than inferential. Its width

$dQ/dP(x,y)\propto K((x_0-x)/h)$1

varies with $dQ/dP(x,y)\propto K((x_0-x)/h)$2 through the learned quantile spread, so it adapts to heteroscedasticity and skewness whenever the quantile regressors are accurate. However, the conformal correction $dQ/dP(x,y)\propto K((x_0-x)/h)$3 is still global, so the formal guarantee remains marginal rather than conditional or neighborhood-specific (Romano et al., 2019).

A later comparison of Romano et al.’s CQR with the normalized method of Kivaranovic et al. showed that both approaches are asymptotically efficient under additional assumptions, but that the Romano et al. construction typically yields tighter finite-sample intervals. The same study also discussed CQR-m, which normalizes deviations by estimated lower and upper half-widths around the median, and CQR-r, which normalizes by total interval width. Those variants are more explicitly local in their score construction, but can be less stable when denominators are small (Sesia et al., 2019).

3. Localization mechanisms beyond global CQR

One line of work replaces the single global conformal correction with region-specific corrections. “Improved conformalized quantile regression” clusters the explanatory variables in a feature space reweighted by permutation importance, chooses the number of clusters by an explained-variance criterion, and computes a separate conformal threshold $dQ/dP(x,y)\propto K((x_0-x)/h)$4 in each cluster. The resulting interval

$dQ/dP(x,y)\propto K((x_0-x)/h)$5

is piecewise local in the cluster partition rather than globally calibrated, and is presented as a simple partition-based local calibration strategy for tabular data (Sousa et al., 2022).

A more continuous local-global interpolation appears in Density-Calibrated Conformal Quantile Regression. There, the global CQR score

$dQ/dP(x,y)\propto K((x_0-x)/h)$6

is supplemented by a $dQ/dP(x,y)\propto K((x_0-x)/h)$7-nearest-neighbor local quantile $dQ/dP(x,y)\propto K((x_0-x)/h)$8. A density proxy

$dQ/dP(x,y)\propto K((x_0-x)/h)$9

determines weights

$\theta_p=Q_Y^{-1}(p)$0

and the conformal radius becomes

$\theta_p=Q_Y^{-1}(p)$1

The method is explicitly designed so that dense regions rely more heavily on local calibration and sparse regions fall back toward the global correction; the paper claims marginal coverage at level $\theta_p=Q_Y^{-1}(p)$2, where $\theta_p=Q_Y^{-1}(p)$3 is attributed to numerical optimization tolerance (Lu, 2024).

A different localization strategy is developed in Conformalized Unconditional Quantile Regression. Instead of directly estimating $\theta_p=Q_Y^{-1}(p)$4, it uses recentered influence functions from unconditional quantile regression to build nested candidate intervals for residuals and then conformalizes them within relevance subgroups $\theta_p=Q_Y^{-1}(p)$5. The resulting guarantee is subgroup-localized rather than fully conditional: $\theta_p=Q_Y^{-1}(p)$6 with a PAC-style subgroup theorem that incurs a slack term of order $\theta_p=Q_Y^{-1}(p)$7. In this formulation, “local” refers to validity relative to a relevance subgroup or localized marginal population around the query point, not to exact conditional coverage at $\theta_p=Q_Y^{-1}(p)$8 (Alaa et al., 2023).

For gradient-boosted trees, LoBoost introduces a model-native local conformal construction. Each input is represented by the sequence of leaves visited across the boosting ensemble, $\theta_p=Q_Y^{-1}(p)$9. At resolution $(X_1,Y_1),\dots,(X_n,Y_n)$0, the local group is

$(X_1,Y_1),\dots,(X_n,Y_n)$1

and the local conformal quantile is the empirical quantile of calibration residuals within that group. This yields intervals

$(X_1,Y_1),\dots,(X_n,Y_n)$2

with exact finite-sample group-conditional validity on each fixed boosting-induced region, together with asymptotic conditional coverage under additional regularity assumptions (Santos et al., 25 Feb 2026).

In spatial statistics, Localized Spatial Conformal Prediction estimates local residual quantiles from neighborhood residual structure. Given a base predictor $(X_1,Y_1),\dots,(X_n,Y_n)$3, residuals $(X_1,Y_1),\dots,(X_n,Y_n)$4, and neighborhood features $(X_1,Y_1),\dots,(X_n,Y_n)$5, it fits a local quantile regressor $(X_1,Y_1),\dots,(X_n,Y_n)$6 for residuals and chooses

$(X_1,Y_1),\dots,(X_n,Y_n)$7

The final interval is

$(X_1,Y_1),\dots,(X_n,Y_n)$8

This is explicitly a localized residual-quantile construction rather than a global residual-quantile correction (Jiang et al., 2024).

4. Guarantees, impossibility, and the local-validity frontier

A central theoretical fact across this literature is that exact distribution-free conditional guarantees are generally unattainable. CQR itself emphasizes that its theorem is for marginal coverage,

$(X_1,Y_1),\dots,(X_n,Y_n)$9

not for $\widehat C_n(X_{n+1})$0. The motivation aligns with impossibility results for exact distribution-free conditional coverage in conformal prediction and with impossibility results for exact conditional quantile inference under nonatomic covariates (Romano et al., 2019, Jang et al., 2023).

The localized quantile-inference paper makes this point especially sharply. It reviews impossibility results for $\widehat C_n(X_{n+1})$1, conditional medians, and conditional quantiles, and avoids them by changing the estimand from the exact conditional quantile $\widehat C_n(X_{n+1})$2 to a localized marginal quantile under a reweighted law. This is a different response to the same impossibility frontier: rather than strengthening the guarantee, it changes the target to something distribution-free or asymptotically distribution-free inference can cover (Jang et al., 2023).

Within predictive conformal methods, the main guarantees therefore occupy an intermediate spectrum. CHR provides exact finite-sample marginal coverage and, under assumptions including consistency of the estimated conditional distribution, asymptotic conditional coverage and asymptotic oracle-optimal interval length. Its intervals are derived from shortest intervals under an estimated conditional histogram, so it is locally adaptive through the estimated conditional law even though the calibration quantile is global (Sesia et al., 2021).

Other methods make the localization explicit in the guarantee. CUQR offers exact marginal conformal validity and a subgroup-level PAC guarantee. LoBoost gives exact finite-sample group-conditional coverage on fixed boosting-induced cells and an asymptotic conditional-coverage result when the cells shrink appropriately. LSCP replaces exchangeability with stationarity and spatial mixing, derives finite-sample bounds on the conditional coverage gap, and establishes asymptotic conditional coverage in spatial settings (Alaa et al., 2023, Santos et al., 25 Feb 2026, Jiang et al., 2024).

These results delineate a recurring pattern. Exact finite-sample distribution-free validity is easiest to obtain marginally. Finer notions of validity—subgroup-localized, group-conditional, approximate local, or asymptotic conditional—are attainable, but only by weakening the target, changing the estimand, adding structural assumptions, or accepting finite-sample slack terms. That distinction separates local adaptivity of interval width from local validity of the coverage guarantee.

5. Base quantile estimators and adjacent non-conformal theory

Local conformal quantile methods inherit much of their behavior from the underlying quantile estimators. “Uniform bias study and Bahadur representation for local polynomial estimators of the conditional quantile function” analyzes the multivariate local polynomial estimator

$\widehat C_n(X_{n+1})$3

with $\widehat C_n(X_{n+1})$4. It proves uniform bias bounds of order $\widehat C_n(X_{n+1})$5, a uniform weak Bahadur representation, robustness to random bandwidths, and uniformity jointly in quantile level, covariate, and bandwidth. Those are exactly the types of simultaneous controls needed when lower and upper quantiles must be estimated over many calibration points (Guerre et al., 2011).

“Local Quantile Regression” studies pointwise adaptive local quantile smoothing over an ordered covariate by choosing bandwidths through local likelihood-ratio consistency tests. The candidate estimator at each location is the largest bandwidth whose fit remains compatible with all smaller-scale estimators, and the resulting adaptive procedure is shown to perform as well as an oracle that would minimize local estimation risk. This provides a locally adaptive base quantile engine for one-dimensional or ordered designs, especially when quantile curves exhibit spatially varying smoothness or changepoints (Spokoiny et al., 2012).

“Fast and Locally Adaptive Bayesian Quantile Smoothing using Calibrated Variational Approximations” contributes a model-based alternative. It uses a Bayesian quantile trend filtering prior on discrete differences, a latent-variable representation of the asymmetric Laplace working likelihood, and a calibrated mean-field variational approximation with location-specific variance inflation

$\widehat C_n(X_{n+1})$6

The calibration step chooses $\widehat C_n(X_{n+1})$7 by bootstrap to target empirical coverage of marginal credible intervals at each location. This is not conformal prediction, but it is a directly relevant source of locally adaptive quantile fits and local scale surrogates for subsequent calibration (Onizuka et al., 2022).

Metric-space uncertainty quantification extends the same logic beyond scalar responses. In the heteroscedastic regime, “Conformal and kNN Predictive Uncertainty Quantification Algorithms in Metric Spaces” estimates the local conditional radius

$\widehat C_n(X_{n+1})$8

by a $\widehat C_n(X_{n+1})$9-nearest-neighbor empirical quantile of residual radii $$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$0. The paper explicitly notes that this local $$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$1-NN estimator is not conformally calibrated in the heteroscedastic setting, but it provides a blueprint for combining local radius estimation with a subsequent conformal correction (Lugosi et al., 21 Jul 2025).

6. Extensions, neighboring methods, and recurring limitations

Several neighboring methods broaden the notion of quantile-based conformal prediction without fitting the local-CQR template exactly. For downside risk forecasting, “Calibrated quantile prediction for Growth-at-Risk” uses the one-sided conformalized quantile estimator

$$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$2

which guarantees

$$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$3

under exchangeability. The correction is global rather than local, but the paper is important as a baseline for direct conformal calibration of one-sided quantile estimates (Bogani et al., 2024).

CHR moves in a different direction by estimating a conditional histogram or many conditional quantiles and then conformalizing the shortest estimated interval of mass $$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$4. Its conformity score is the smallest mass level at which the point enters the nested estimated intervals, and it achieves exact finite-sample marginal coverage together with asymptotic conditional coverage and asymptotically oracle-optimal length under assumptions. Here the interval is locally adaptive through the estimated conditional distribution rather than through a localized conformal quantile (Sesia et al., 2021).

Conformalized High-Density Quantile Regression shifts from interval quantiles to highest-density regions. It learns input-conditional probabilities over adaptive Voronoi cells in output space, uses dynamic prototype addition and deletion, and conformalizes cumulative density scores to obtain finite-sample marginal coverage for high-density predictive regions. This is relevant to local conformal quantile estimation because it replaces lower and upper quantile endpoints by locally adaptive, potentially non-convex output regions, but its calibration threshold remains global (Cengiz et al., 2024).

Semiparametric conformal prediction is another adjacent development. It estimates the joint CDF of vector-valued nonconformity scores by nonparametric vine copulas, computes a plug-in multivariate $$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$5 score quantile, and improves it with an influence-function-based one-step correction. The method is explicitly global in $$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$6: it models the unconditional joint score distribution rather than a covariate-local one. Its main relevance is methodological, since the same semiparametric quantile-estimation logic could in principle be transplanted to localized score distributions (Park et al., 2024).

Recurring limitations are consistent across the literature. Exact finite-sample conditional coverage is not achieved. Local calibration can become unstable when neighborhoods or partition cells are small. In high dimensions, nearest-neighbor, kernel, and local polynomial approaches face the usual curse of dimensionality. Efficiency remains highly dependent on the quality of the base quantile model, and normalized conformal scores can become unstable when estimated denominators are small. Quantile crossing, sparse calibration sets within local groups, and the gap between local adaptivity and local validity remain persistent technical issues (Sesia et al., 2019, Lu, 2024, Sousa et al., 2022).

In this sense, local conformal quantile estimation is less a single algorithm than a design space. Its unifying objective is to preserve the robustness of conformal calibration while replacing a single global quantile correction by an estimator that is sensitive to local heterogeneity in $$\mathbb{P}\{Y_{n+1}\in \widehat C_n(X_{n+1})\}\ge 1-\alpha.$$7, local neighborhoods in feature or spatial space, model-native partitions, or localized target populations. The literature differs primarily in where that localization is imposed, what object is estimated, and what notion of validity is ultimately claimed.