ESCP: Ensemble Spatial Conformal Prediction

• Ensemble Spatial Conformal Prediction is a framework that builds rigorous prediction sets for ensemble-based forecasts and spatial regression with finite-sample coverage guarantees.
• It integrates conformal prediction theory with ensemble sampling and localized model selection to handle spatial and temporal heterogeneity in data.
• Empirical results demonstrate that ESCP produces sharper intervals—often 25–35% shorter—while maintaining near-perfect coverage in complex forecasting tasks.

Ensemble Spatial Conformal Prediction (ESCP) is a modern statistical framework designed to construct finite-sample valid uncertainty quantification around ensemble-based point or trajectory forecasts. ESCP accommodates spatial and temporal heterogeneity, model-selection uncertainty, and multimodal predictive distributions by leveraging conformal prediction (CP) theory while integrating ensemble and spatial adaptation. The framework has been instantiated both for trajectory ensembles (e.g., CP-Traj) and in regression settings with localized model selection. ESCP offers practical and sharp prediction sets with rigorous marginal coverage guarantees across a broad array of spatiotemporal forecasting tasks (Li et al., 18 Aug 2025, Wang et al., 22 Feb 2026).

1. Problem Setup and Ensemble Sampling

ESCP unifies the construction of prediction intervals over ensembles of probabilistic forecasts. In the sequential/trajectory setting, observations consist of a streaming time series of features $x_t\in\mathbb{R}^N$ and scalar responses $y_t\in\mathbb{R}$. At each time $t$, a probabilistic forecaster $f$ emits $M$ sampled trajectories of length $H$: $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$ where $$\hat{\mathcal{Y}}_t^h=\{\hat y_t^{h,(1)},...,\hat y_t^{h,(M)}\}$$ denotes the ensemble at forecast horizon $h$. The central objective is to construct, for each time $t$ and horizon $y_t\in\mathbb{R}$0, a spatial prediction set $y_t\in\mathbb{R}$1 such that the empirical miscoverage rate converges to the nominal $y_t\in\mathbb{R}$2 and the set remains minimal (Li et al., 18 Aug 2025).

In the regression context, the calibration dataset $y_t\in\mathbb{R}$3 and a finite set of $y_t\in\mathbb{R}$4 regression models $y_t\in\mathbb{R}$5 are available. ESCP aims to adaptively select the most appropriate local model (or combination) based on spatially weighted residual distributions, providing intervals adaptive to both spatial heterogeneity and model complexity (Wang et al., 22 Feb 2026).

2. Nonconformity Scores, Surrogates, and Local Adaptation

At the heart of ESCP is the use of conformal prediction nonconformity measures tailored to the setting:

• Trajectory Ensemble Mode: The “probabilistic conformal prediction” (PCP) score at each time and horizon is $y_t\in\mathbb{R}$6. The CP set is constructed as

$y_t\in\mathbb{R}$7

where $y_t\in\mathbb{R}$8 is the $y_t\in\mathbb{R}$9 empirical quantile over past scores.

• Localized Model Selection Mode: Local conformal intervals use kernel-weighted empirical quantiles of model residuals:

$t$0

where $t$1 achieves a $t$2 weighted coverage under localizer $t$3. Surrogate intervals—constructed via hypothetical best-case and worst-case augmentations—enable safe selection among candidate models without conditioning on unobserved outcomes (Wang et al., 22 Feb 2026).

3. Online Update and Horizon-wide Optimization

For trajectory ensembles, ESCP maintains dynamic, per-horizon miscoverage thresholds $t$4 updated via a stochastic-gradient-type rule: $t$5 where $t$6 flags out-of-set observations and $t$7, $t$8 are step and relaxation parameters, respectively. Simultaneously, a joint horizon-wide optimization selects $t$9 to minimize a composite objective (e.g., coverage miss plus average interval width) constrained to $f$0 (Li et al., 18 Aug 2025).

In the localized model selection variant, empirical coverage tests over surrogate sets determine data-driven “safe model” index sets $f$1 and select a bracket $f$2 of admissible miscoverage levels for prediction at a new test point (Wang et al., 22 Feb 2026).

4. Construction of Prediction Sets and Theoretical Guarantees

The final ESCP prediction set depends on the context:

• Trajectory Ensemble: After computing the empirical quantile radius $f$3, the prediction set $f$4 at each horizon is the union of $f$5 balls of radius $f$6 about each trajectory sample, which may be discontinuous and adapt to multimodal and diverging forecasts.
• Model Selection Regression: For each admissible miscoverage level $f$7, the shortest among $f$8 local conformal intervals at $f$9 is found; the ESCP interval is the union of these intervals over $M$0.

Theoretical Coverage: For both the trajectory and regression formulations, ESCP guarantees marginal coverage at the user-specified level: $M$1

$M$2

5. Algorithmic Implementation and Pseudocode

Key computational steps are organized as follows:

Trajectory Ensemble ESCP (CP-Traj) (Li et al., 18 Aug 2025):

$$\hat{\mathcal{Y}}_t^h=\{\hat y_t^{h,(1)},...,\hat y_t^{h,(M)}\}$$1

Localized Model Selection ESCP (Wang et al., 22 Feb 2026):

1. Precompute best/worst case surrogate intervals for all calibration points, models, and miscoverage levels.
2. For each $M$3, construct safe index sets $M$4 for $M$5.
3. Compute empirical coverage statistics and define $M$6.
4. For $M$7, for each $M$8 in the bracket, evaluate $M$9.
5. Output the union of minimal intervals over admissible $H$0.

6. Empirical Performance

Empirical evaluations show:

• For trajectory tasks (cyclone tracks, lane forecasting, influenza, synthetic MarkovAR), ESCP (CP-Traj):
  • Achieves minimal calibration-score (mean absolute coverage error).
  • Produces interval widths 2–3× smaller than methods such as ACI, NEXCP, FACI, or SAOCP in high-dimensional settings.
  • Retains near-perfect 90% coverage across forecast steps while sharpening the size and adaptivity of intervals (Li et al., 18 Aug 2025).
• For spatial regression/model selection:
  • With $H$1, $H$2, and localizer bandwidth $H$3, ESCP yields average normalized interval lengths (ensemble) of 0.3470 compared to 0.4305 for the best single model.
  • In complex nonlinear examples, ensemble intervals are 25–35% shorter for large $H$4 and low noise, with empirical coverage always at or above $H$5.
  • Gains diminish as noise increases; adaptation to spatial signal and model structure is pronounced in low-noise or highly heterogeneous regimes (Wang et al., 22 Feb 2026).

Setting	Ensemble Interval	Best Single	Empirical Coverage
Piecewise sinusoid, $H$6, $H$7, $H$8	0.3470	0.4305	≥ 0.90
Nonlinear, $H$9, $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$0, $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$1	0.9000	1.3225	≥ 0.90

7. Computational Complexity and Practical Considerations

• Complexity: For trajectory ensembles, per-timestep cost is dominated by the number of ensemble samples, horizons, and quantile evaluations. For spatial model selection, $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$2 dominates calibration, with $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$3 per prediction. Parallelization across points, models, and $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$4 is straightforward.
• Implementation: Efficient weighted quantile algorithms and memory streaming are recommended; split-conformal approximations scale to large $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$5 or $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$6. Surrogates allow for other nonconformity losses including classification or quantile loss.
• Hyperparameter Tuning: Online step-size $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$7 and radius schedule $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$8 (e.g., $$\hat y_t^{(m)} = (\hat y_t^{1,(m)},...,\,\hat y_t^{H,(m)}) \sim \hat{p}\big(\hat y_t^1,...,\hat y_t^H\,|\,x_{1:t}\big), \quad m=1,...,M,$$9) should be validated empirically. Localizer bandwidth in regression is selected to match underlying function/noise variation, often by cross-validation.
• Robustness: If calibration drifts under nonstationarity, adaptive adjustment of $$\hat{\mathcal{Y}}_t^h=\{\hat y_t^{h,(1)},...,\hat y_t^{h,(M)}\}$$0 or trimming extreme surrogate values is suggested.

A plausible implication is that ESCP generalizes classical conformal prediction to modern ensemble forecasting and regression with spatial adaptivity, yielding sharp, valid, and computationally efficient uncertainty sets suitable for high-dimensional and complex domains (Li et al., 18 Aug 2025, Wang et al., 22 Feb 2026).