Covariate-Shift Weighted Model Overview

• Covariate-shift weighted models are techniques that correct for differences between training and test covariate distributions using density ratio reweighting.
• They integrate methods like stratification, clipping, and control-variate correction to improve empirical risk minimization and calibration under distribution shift.
• The approach balances bias and variance while offering theoretical guarantees and practical trade-offs across prediction, feature selection, and robust inference.

Searching arXiv for the core paper and closely related covariate-shift weighting work. A covariate-shift weighted model is a predictive or inferential model that explicitly corrects for a discrepancy between training and test covariate distributions by reweighting observations, risks, scores, or calibration statistics with a density ratio such as $w(x)=P_{\mathrm{te}}(X=x)/P_{\mathrm{tr}}(X=x)$ or equivalent propensity-based forms, while maintaining the covariate-shift assumption that the conditional law of the response given the covariates is unchanged across domains (Xu et al., 2021). In this setting, the central objective is to approximate target-domain risk or target-domain validity using labeled source data and, often, unlabeled target covariates. Across the literature, covariate-shift weighted models appear in empirical risk minimization, feature selection, kernel and spectral methods, federated adaptation, conformal prediction, selective inference, and semi-supervised or tree-based learning, with the common rationale that weighted source averages can recover target-domain quantities under appropriate support and regularity conditions (Tibshirani et al., 2019).

1. Formal setting and core weighting principle

The canonical setup distinguishes a source or training distribution from a target or test distribution, with unchanged conditionals and shifted covariate marginals. This is written in several equivalent forms, including $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$ with $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$ (Xu et al., 2021), $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$ with $p_{\rm tr}(x)\neq p_{\rm te}(x)$ (Segovia-Martín et al., 2023), and $P'_{Y|X}=P_{Y|X}$ with $P'_X\neq P_X$ (Zecchin et al., 20 Jan 2025). Under this assumption, target risk can be rewritten as a weighted source expectation. For example, $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$ becomes an expectation under the source law multiplied by the importance weight $w(x)=q(x)/p(x)$ (Lam et al., 2019).

This weighting principle underlies both estimation and evaluation. In standard importance-weighted empirical risk minimization, the empirical criterion takes the form $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$ (Xu et al., 2023). In related formulations, the weight is expressed through a propensity score. In stratified learning, the propensity score is $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$0, with $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$1 (Autenrieth et al., 2021). In calibrated prediction, a domain discriminator $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$2 yields $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$3 (Park et al., 2020).

A central technical issue is overlap or support. Many results assume bounded weights, such as $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$4 for all $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$5 (Zecchin et al., 20 Jan 2025), $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$6 (Park et al., 2021), or uniformly bounded likelihood ratio $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$7 (Ma et al., 2022). Other works explicitly target unbounded or heavy-tailed ratios by truncation, clipping, or moment conditions (Fan et al., 17 Apr 2025). This suggests that the distinction between bounded-ratio and unbounded-ratio regimes is foundational for the design of covariate-shift weighted models.

2. Weighted empirical risk, joint optimization, and stable-variable recovery

The most direct covariate-shift weighted model is weighted empirical risk minimization. In the two-step paradigm, one first estimates the density ratio and then minimizes a weighted loss, for instance $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$8 with $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$9 (Zhang et al., 2020). A one-step alternative jointly learns the predictive model and the associated weights by minimizing an upper bound on the test risk. In that formulation, for bounded loss $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$0 and measurable $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$1,

$$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$2

and the empirical counterpart jointly minimizes over $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$3 and nonnegative $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$4 (Zhang et al., 2020). The same work gives an alternating-minimization scheme in which the $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$5-step is quadratic and the $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$6-step is a weighted empirical-risk problem (Zhang et al., 2020).

A distinct line of work interprets weighting as a feature-selection mechanism. The independence-driven importance weighting framework defines a stable variable set $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$7 by the condition $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$8 under $$\mathrm{supp}(P_{\mathrm{te}}(X))=\mathrm{supp}(P_{\mathrm{tr}}(X))$$9, and then defines the minimal stable variable set

$p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$0

(Xu et al., 2021). Under ideal conditions, the weighting stage learns sample weights so that the reweighted empirical feature distribution becomes approximately independent, after which weighted least squares is fitted. The resulting coefficient magnitudes $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$1 serve as feature scores (Xu et al., 2021).

The theoretical claims are strong under the idealized independence class. If $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$2, then for any weighting $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$3 that renders the features independent, the corresponding weighted least-squares coefficient satisfies $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$4; conversely, for each $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$5, there exists some independence-enforcing weighting for which $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$6 (Xu et al., 2021). Finite-sample behavior is controlled by $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$7 plus weight-estimation terms, and a more explicit bound of the form $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$8 with $p_{\rm tr}(y\mid x)=p_{\rm te}(y\mid x)$9 (Xu et al., 2021). A plausible implication is that, in this framework, weighting does not merely correct a risk estimate; it changes the effective representation of the problem by isolating variables whose predictive role is stable under covariate shift.

3. Variance control, robustness, and alternatives to global reweighting

A recurrent limitation of classical importance weighting is variance inflation. The literature repeatedly notes that large or unstable weights can degrade performance. Reweighting-based methods are described as suffering high variance when the distributional discrepancy is large and the weights are poorly estimated (Lam et al., 2019), and standard importance weighting can perform poorly under support mismatch or when ratios take large values (Segovia-Martín et al., 2023). This has led to several alternative weighted-model constructions.

One strategy is stratification on the propensity score. Stratified learning estimates $p_{\rm tr}(x)\neq p_{\rm te}(x)$0, partitions the pooled data into $p_{\rm tr}(x)\neq p_{\rm te}(x)$1 bins based on empirical quantiles of $p_{\rm tr}(x)\neq p_{\rm te}(x)$2, and fits a learner separately within each stratum (Autenrieth et al., 2021). The theoretical motivation is the balancing-score property $p_{\rm tr}(x)\neq p_{\rm te}(x)$3 and the exact-alignment proposition $p_{\rm tr}(x)\neq p_{\rm te}(x)$4 (Autenrieth et al., 2021). The method explicitly trades a little bias due to binning for reduced variance, and the summary states that the convergence rate of the stratified estimator avoids large $p_{\rm tr}(x)\neq p_{\rm te}(x)$5-type moments appearing in classical importance-weighted bounds (Autenrieth et al., 2021).

A second strategy is control-variate correction. In robust importance weighting, a regression estimate $p_{\rm tr}(x)\neq p_{\rm te}(x)$6 is combined with kernel mean matching weights $p_{\rm tr}(x)\neq p_{\rm te}(x)$7 to form

$p_{\rm tr}(x)\neq p_{\rm te}(x)$8

This reweights only residuals rather than the full loss (Lam et al., 2019). The summary states that the estimator can either strictly outperform or match the best-known existing rates for both KMM and NR, and under suitable smoothness assumptions achieves $p_{\rm tr}(x)\neq p_{\rm te}(x)$9 (Lam et al., 2019).

A third strategy is double weighting. The double-weighting formulation introduces training weights $P'_{Y|X}=P_{Y|X}$0 and test-side weights $P'_{Y|X}=P_{Y|X}$1 constrained by $P'_{Y|X}=P_{Y|X}$2 for all $P'_{Y|X}=P_{Y|X}$3 (Segovia-Martín et al., 2023). Exact reweighting and exact robust weighting appear as special cases, but the method caps both sets of weights through a trade-off parameter $P'_{Y|X}=P_{Y|X}$4 and constructs a minimax-risk classifier using a weighted feature-expectation uncertainty set (Segovia-Martín et al., 2023). The corresponding generalization bound scales with $P'_{Y|X}=P_{Y|X}$5 rather than $P'_{Y|X}=P_{Y|X}$6, which is described as an effective sample-size increase by factor $P'_{Y|X}=P_{Y|X}$7 (Segovia-Martín et al., 2023). The same source notes a limitation: predictions can be uninformative in regions where $P'_{Y|X}=P_{Y|X}$8 is driven near zero (Segovia-Martín et al., 2023).

A fourth strategy is doubly robust adaptation. The doubly robust estimator augments importance weighting with an auxiliary regression nuisance $P'_{Y|X}=P_{Y|X}$9 and constructs

$P'_X\neq P_X$0

With cross-fitting and orthogonalization, the first-order effect of nuisance-estimation error vanishes, and consistency holds if either the density-ratio estimator or the regression-function estimator is consistent in $P'_X\neq P_X$1 (Kato et al., 2023). This suggests a broader pattern: robust covariate-shift weighted models increasingly combine weighting with orthogonalization, augmentation, or structural constraints rather than relying on a single plug-in weight estimate.

4. Kernel, spectral, and Wasserstein-weighted models

Kernel and spectral methods provide some of the sharpest rate analyses for covariate-shift weighting. In RKHS-based nonparametric regression, one line of work shows that when the likelihood ratio is uniformly bounded, unweighted KRR with carefully chosen regularization is minimax rate-optimal up to a log factor (Ma et al., 2022). In that bounded-ratio regime, the paper states that KRR does not require full knowledge of likelihood ratios apart from an upper bound on them (Ma et al., 2022). By contrast, a naive constrained empirical-risk minimizer can be strictly sub-optimal under covariate shift, with lower-bound behavior $P'_X\neq P_X$2 in one construction, compared with the optimal order $P'_X\neq P_X$3 (Ma et al., 2022).

When the ratio is possibly unbounded but has finite second moment, the same line introduces truncated reweighted KRR

$P'_X\neq P_X$4

and proves minimax rate-optimality up to logarithmic factors (Ma et al., 2022). This bounded-versus-truncated dichotomy reappears in a more general spectral-regularization framework.

In spectral algorithms under covariate shift, the weighted model is

$P'_X\neq P_X$5

with operator form $P'_X\neq P_X$6 and $P'_X\neq P_X$7 (Fan et al., 17 Apr 2025). Under a Rényi-type weight-moment condition, a source condition, and an effective-dimension bound, Theorem 1 gives the rate

$P'_X\neq P_X$8

for a prescribed choice of $P'_X\neq P_X$9 (Fan et al., 17 Apr 2025). With clipping $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$0 and a slightly overregularized choice of $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$1, Theorem 2 yields $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$2, approaching the minimax rate up to an arbitrary $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$3 (Fan et al., 17 Apr 2025). The same summary explicitly states that clipping resolves the suboptimality issue in unbounded density-ratio scenarios (Fan et al., 17 Apr 2025).

A different alternative avoids density-ratio estimation altogether by matching empirical measures in Wasserstein distance. The reweighting problem is

$$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$4

and it admits an explicit nearest-neighbor solution in which each target point sends its mass to its nearest source point (Reygner et al., 2020). The optimal weights are therefore $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$5 (Reygner et al., 2020). The summary emphasizes that no assumption of absolute continuity of $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$6 with respect to $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$7 is needed, only a support condition that every target region be seen by the source (Reygner et al., 2020). This method broadens the notion of a covariate-shift weighted model from density-ratio weighting to transport-based reweighting.

5. Weighted conformal, predictive-distribution, and selective-inference models

Covariate-shift weighted models are not limited to point prediction. A substantial literature extends conformal and selective inference by reweighting calibration or p-value constructions.

Weighted conformal prediction under covariate shift defines $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$8 and replaces the uniform empirical distribution of conformity scores with a weighted version (Tibshirani et al., 2019). In split conformal form, calibration residuals $$R_{\mathrm{t}}(f)=E_{(x,y)\sim P_{\mathrm{t}}}[\ell(f(x),y)]$$9 are combined with weights

$w(x)=q(x)/p(x)$0

and the prediction interval is obtained from a weighted quantile of $w(x)=q(x)/p(x)$1 (Tibshirani et al., 2019). If the true likelihood ratio is used, coverage is guaranteed: $w(x)=q(x)/p(x)$2 (Tibshirani et al., 2019). A closely related extension to Conformal Predictive Systems defines a weighted conformal transducer $w(x)=q(x)/p(x)$3 using the same normalized likelihood-ratio weights and aims at a predictive CDF calibrated for the target distribution (Jonkers et al., 2024). The same summary states that simulation experiments indicate that WCPS are probabilistically calibrated under covariate shift (Jonkers et al., 2024).

Weighted conformal risk control reweights calibration losses rather than residual quantiles. Given calibration scores $w(x)=q(x)/p(x)$4 and weights $w(x)=q(x)/p(x)$5, the weighted augmented empirical risk is

$w(x)=q(x)/p(x)$6

and the smallest $w(x)=q(x)/p(x)$7 satisfying $w(x)=q(x)/p(x)$8 determines the prediction set (Zecchin et al., 20 Jan 2025). The resulting guarantee is $w(x)=q(x)/p(x)$9 under the stated assumptions (Zecchin et al., 20 Jan 2025). The paper also provides an inefficiency bound showing that larger $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$0 loosens the bound and larger $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$1 or $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$2 can improve informativeness (Zecchin et al., 20 Jan 2025).

PAC prediction sets under covariate shift use rejection sampling based on exact or interval-valued importance weights. With exact weights bounded by $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$3, accepted source calibration points become IID draws from the target law, and a Clopper–Pearson upper bound determines the largest feasible threshold (Park et al., 2021). With interval weights, a worst-case choice of weights is used inside the same calibration logic, and the paper states that the output satisfies the PAC constraint with probability at least $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$4 (Park et al., 2021).

Weighted conformal p-values have also been used for model-free selective inference. In that setting, under covariate shift with $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$5, the weighted conformal p-values satisfy $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$6 for all $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$7 (Jin et al., 2023). Because these p-values may not obey the dependence property required by BH, the weighted conformalized selection procedure introduces a two-stage leave-one-out calibration and pruning scheme; Theorem 3.1 states finite-sample FDR control under correct weights (Jin et al., 2023). A plausible implication is that covariate-shift weighting has become a general device for transporting calibration validity, not only predictive risk, from source to target domains.

6. Distributed, semi-supervised, tree-based, and application-specific constructions

Covariate-shift weighted models also arise in settings where training data are partitioned, partially unlabeled, or structurally heterogeneous. In federated covariate shift adaptation, each source $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$8 uses a local density ratio $$\widehat R_{\mathrm{IW}}(h)=\frac{1}{n}\sum_{i=1}^n w(x_i)L(h(x_i),y_i)$$9 and constructs a variance-reduced target-risk estimate with control variates (Xu et al., 2023). These local estimators are combined through a federated linear combination with weights

$P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$00

yielding the FedDAE target-risk estimator (Xu et al., 2023). The summary states asymptotic unbiasedness, asymptotic variance minimality among unbiased federated linear combinations, and the ordering $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$01 (Xu et al., 2023). The final target predictor is a convex combination of source models, $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$02 with $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$03 (Xu et al., 2023).

Semi-supervised constructions exploit unlabeled target data more directly. In semi-generative modelling, the causal graph $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$04 induces a factorization $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$05 (Kügelgen et al., 2018). Under covariate shift in $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$06, the weighted discriminative term $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$07 is combined with generative terms over both labeled source data and unlabeled target pairs $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$08 (Kügelgen et al., 2018). The summary explicitly frames this as combining adaptation with semi-supervised learning (Kügelgen et al., 2018).

Tree-based weighted models transplant the same principle into recursive partitioning. In the weighted CART framework, source observations receive weights estimated from a classifier for domain membership, typically via $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$09 with truncation and normalization (Cai et al., 2024). Every split criterion is rewritten in weighted form. For regression, the split score is

$P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$10

and pruning minimizes a weighted cost-complexity functional (Cai et al., 2024). The same framework extends to GLM-trees, bagging, random forests, and gradient boosting via weighted bootstrap, weighted splitting, or weighted gradient steps (Cai et al., 2024).

The application range is broad. Weighted conformal risk control is validated on fingerprinting-based localization with rural Sigfox RSSI fingerprints from $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$11 base stations (Zecchin et al., 20 Jan 2025). Stratified learning reports the best reported AUC $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$12 on the updated "Supernovae photometric classification challenge" and improvement on conditional density estimation of galaxy redshift from SDSS data (Autenrieth et al., 2021). Weighted survival conformal prediction combines the covariate-shift ratio with inverse-probability-of-censoring factors to form $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$13, then computes weighted p-values for right-censored survival intervals (Shin et al., 3 Dec 2025). In calibrated prediction, adversarial feature alignment is used alongside importance weighting and temperature scaling to improve calibration under dataset shift (Park et al., 2020).

7. Theoretical themes, misconceptions, and design trade-offs

Several common misconceptions are contradicted by the cited literature. One is that covariate-shift correction is synonymous with plugging a density-ratio estimate into weighted ERM. The literature contains propensity-stratified models (Autenrieth et al., 2021), control-variate estimators (Lam et al., 2019), joint one-step objectives (Zhang et al., 2020), double-weighting minimax formulations (Segovia-Martín et al., 2023), transport-based nearest-neighbor reweighting (Reygner et al., 2020), and weighted conformal or selective-inference procedures (Tibshirani et al., 2019, Jin et al., 2023). A second misconception is that more accurate weighting always requires explicit pointwise density-ratio estimation. In bounded-ratio RKHS regression, unweighted KRR can already be minimax optimal when regularized appropriately (Ma et al., 2022). A third misconception is that validity under covariate shift is purely asymptotic; weighted conformal and PAC constructions provide finite-sample guarantees when the required weights or intervals are available (Tibshirani et al., 2019, Park et al., 2021).

A recurring trade-off is bias versus variance. Global reweighting can be unbiased but unstable when weights are large, motivating stratification, clipping, truncation, capped double weights, or domain-invariant representations (Autenrieth et al., 2021, Fan et al., 17 Apr 2025, Segovia-Martín et al., 2023, Park et al., 2020). Another trade-off is robustness versus informativeness. Weighted conformal methods can preserve risk or coverage guarantees under shift, but stronger shift, larger $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$14, or wider weight intervals generally produce larger or less informative prediction sets (Zecchin et al., 20 Jan 2025, Park et al., 2021). In minimax double-weighting, local predictions may become uninformative where the test-side weight $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$15 is near zero (Segovia-Martín et al., 2023).

Across these developments, three design principles recur. First, the target quantity must be written in a form that source data can estimate through weighting, conditioning, or transport. Second, weight instability must be controlled, either structurally or statistically. Third, the object being transported need not be a training loss; it may be a feature set, a calibration distribution, a p-value law, a federated validation score, or a survival prediction interval. This suggests that the modern covariate-shift weighted model is best understood not as a single algorithmic template, but as a general statistical design pattern for targeting target-domain behavior from source-domain data under $P_{\mathrm{te}}(Y\mid X)=P_{\mathrm{tr}}(Y\mid X)$16 (Xu et al., 2021).