Conformal API¶

Conformal prediction (CP) in torchregress lives in torchregress.losses.conformal. This page lists every conformal loss, calibrator, and predictor, plus the test-time adapters in torchregress.test_time.ot_conformal. For background and the unified coverage theorem, see Conformal prediction.

Method matrix¶

Symbol	Strategy	When to use
`SplitConformal`	Residual-based	Baseline; homoscedastic noise; fast
`CQR`	Conformalized Quantile Regression	Heteroscedastic, skewed noise; adaptive width
`UACQR`	Width-normalized CQR	Same as CQR; penalized by width
`DensityConformal`	Density-weighted residuals	Imbalanced or long-tail data
`MonteCarloConformal`	MC-sample normalized	Bayesian models, ensembles
`LocalConformal`	Local feature-space coverage	Coverage guaranteed in feature subgroups
`LocalConformalMAD`	Local + MAD scaling	Robust to outliers in calibration scores
`CTI`	Density level-set	Multimodal, complex distributions
`DistributionalConformal`	Distributional coverage	When full predictive CDF is available
`LevelSetConformalPredictor`	Level-set predictor	Construct smallest intervals from CDF
`MultiTargetConformal`	Multi-target coverage	Joint coverage on `[B, D]` targets
`MultivariateScoreConformal`	Scalar multivariate score (Mahalanobis or Gaussian NLL)	Joint ellipsoidal region for `[B, D]` targets from a predictive `(mu, Sigma)`; weighted under shift
`NonExchangeableConformalRegressor`	Weighted split conformal (NexCP, Barber et al. 2023)	Covariate shift or drift with importance weights `w(x)`; reports two-sided coverage bounds
`PrevalenceAdjustedCP`	Group-prevalence-adjusted	Subgroup prior shift
`R2CConformal`	Regression-as-classification	Multimodal / binned targets
`SLSConformal`	Super-level set conformal	Density-based / SLS regression
`CVPlus`	Out-of-fold residuals	Cross-validation ensembles. Leave-one-out (`K = n`) is the jackknife+ case.
`EnsembleBatchCP`	Out-of-bag residuals	Bootstrap-based ensemble predictors (EnbPI)
`ConformalLoss(method=…)`	Unified CQR/UACQR/split training wrapper	One-call training of supported methods

Calibrators (post-hoc)¶

Symbol	Description
`ConformalPredictor`	Base post-hoc calibrator: `calibrate(y_pred, target)` then `predict_interval(y_pred)`.
`MultiTargetConformal`	Multi-target version. Returns per-target coverage.

Test-time shift-aware conformal¶

In torchregress.test_time.ot_conformal:

Symbol	Description
`ScoreCDFReweighter`	Estimates sample weights `w(x) = p_target(x) / p_source(x)` via optimal transport.
`OptimalTransportCoverageGap`	Computes empirical coverage gap between source and target before/after reweighting.
`WeightedSplitConformalAdapter`	Weighted split conformal with per-test-point `w(x)`. Maintains coverage under covariate shift.
`weighted_split_classification_predictive_batch`	Classification-style adaptation; useful for ordinal / binned heads.

References: Tibshirani et al., "Conformal Prediction Under Covariate Shift" (NeurIPS 2019).

Quick example¶

from torchregress.losses import ConformalLoss, CQR

# CQR as a training wrapper (alpha = miscoverage; 1-alpha = target coverage)
loss_fn = ConformalLoss(method="cqr", alpha=0.1)
loss = loss_fn(y_pred, y_true)

# Or use CQR directly as a post-hoc calibrator
cp = CQR(alpha=0.1)
cp.calibrate(y_pred_cal, y_cal)
lower, upper = cp.predict_interval(y_pred_test)

Next steps¶

Conformal overview — unified coverage theorem, full method details
Conformal predictors — score formulas, parameter tables, best practices
Distributional conformal — CDF/density-based methods
Test-time shift — shift-aware conformal