Marginal Distribution Shift Asymmetry

• Marginal distribution shift asymmetry is a framework describing directional changes in marginal probabilities between source and target distributions, with applications in regression, classification, and copula analysis.
• The literature proposes multiple methods—such as cosine similarity with Fisher-Rao distance and copula-based measures—to capture asymmetric shifts that standard divergence metrics may miss.
• Recent approaches adjust conformal prediction, target-aware regression, and skewed pivot-blend modeling to address undercoverage, skewness, and reweighting challenges in shifted distributions.

Marginal distribution shift asymmetry denotes directional, nonexchangeable departures between paired or source–target distributions when the departure is expressed through marginal probabilities, conditional marginals tied to subgroups, lower-versus-upper tail mass, or asymmetric movement between matched categorical states. Recent work treats this phenomenon through several mathematically distinct formalisms: symmetry departure in square contingency tables, target shift in regression, conditional probability shift in classification, coverage degradation under distribution shift in conformal prediction, copula-based tail asymmetry, target-aware regression with known target marginals, and flexible skewed density modeling (Urasaki et al., 2024, Yi et al., 13 Oct 2025, Teisseyre et al., 4 Mar 2025, Siahkali et al., 16 Feb 2026, Kato et al., 2020, Hou et al., 22 Jun 2026, She et al., 2024). This suggests that the topic is not a single model class, but a unifying perspective on how asymmetric marginal change is quantified, inferred, and exploited.

1. Shift taxonomies and the formal meaning of asymmetry

A central distinction in the literature is between shifts that modify marginals directly and shifts that modify conditional marginals. In regression under target shift, the source and target differ only in the marginal distribution of the response variable $Y$, while the conditional distribution of covariates given outcome remains invariant: $$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$ The same paper states that asymmetry is permitted because $P(y)$ and $Q(y)$ can have different supports, shapes, and skewness. In classification, the Conditional Probability Shift Model addresses the case in which the conditional probability of the label given part of the features shifts, $p(y\mid z)\neq q(y\mid z)$, while the conditional distribution of the remaining features is preserved, $p(x\mid y,z)=q(x\mid y,z)$. The paper explicitly states that covariate shift and label shift “do not encompass all forms of distributional shift,” and that CPSM covers cases where $p(y)=q(y)$ and $p(z)=q(z)$ but $q(y\mid z)\neq p(y\mid z)$ (Teisseyre et al., 4 Mar 2025, Yi et al., 13 Oct 2025).

Shift model	Source joint	Target joint
Covariate Shift (CS)	$p(x,z)p(y\mid x,z)$	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$0
Label Shift (LS)	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$1	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$2
Sparse Joint Shift (SJS)	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$3	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$4
CPS (case 1)	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$5	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$6
CPS + Marginal Shift	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$7	$$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$8

The CPS literature also gives a posterior correction formula: $$p(y) \neq q(y), \qquad p(\mathbf{x}\mid y) = q(\mathbf{x} \mid y).$$9 This establishes that asymmetry need not be a simple prior shift. A common misconception is that class-prior change is the only practically relevant marginal asymmetry; the cited work shows that subgroup-conditional change can occur even when global priors and feature marginals match. This suggests that “marginal distribution shift asymmetry” includes structurally localized changes, not only whole-population reweighting.

2. Symmetry departure in square contingency tables

In categorical analysis, marginal distribution shift asymmetry is formalized through square contingency tables that summarize survey results between two time points or cohorts represented by the same categorical variables. The symmetry model is

$P(y)$0

and the paper identifies this as the relevant notion of marginal symmetry for transition asymmetries and distribution shifts. The proposed asymmetry measure is based on cosine similarity and the Fisher-Rao distance, so that asymmetry is interpreted as a geodesic distance between two distributions (Urasaki et al., 2024).

The resulting index,

$P(y)$1

aggregates cell-wise geodesic distances from the symmetry point and is standardized so that $P(y)$2. The paper states that $P(y)$3 iff the table is perfectly symmetric and $P(y)$4 iff the asymmetry is maximal, with one of $P(y)$5 equal to zero for each $P(y)$6. Because the Fisher-Rao distance is a true geodesic in the space of statistical models, the measure is presented as a distance-based quantification rather than a divergence-type pseudo-distance.

Several methodological claims are specific. First, the paper reports that visualizing the relationship of quantification methods on a two-dimensional plane shows that the proposed measure provides “the geometrically simplest and most natural quantification.” Second, the visualized figure indicates that the method is less affected by very few cells with extreme asymmetry than Euclidean or divergence-type measures. Third, a simulation study under conditional symmetry models with parameter $P(y)$7 shows that the method can directly extract and quantify only the asymmetric structure of the model and can more sensitively detect departures from symmetry than divergence-type measures. Standard errors and confidence intervals for $P(y)$8 can be estimated via the delta method. These points matter because they distinguish global asymmetry from sparse, extreme-cell effects and make the measure suitable for comparisons across tables with different distributions or sparse entries.

3. Distribution shift asymmetry in conformal prediction

Conformal prediction provides distribution-free marginal coverage guarantees under exchangeability, but the guarantees can fail if the data distribution shifts. One paper analyzes pseudo-calibration under bounded label-conditional covariate shift, where $P(y)$9 while $Q(y)$0 and $Q(y)$1 may differ, and derives lower bounds on target-domain marginal coverage in terms of the classifier’s source-domain loss and a Wasserstein measure of the shift (Siahkali et al., 16 Feb 2026).

The shift asymmetry enters at several levels. Hard pseudo-calibration uses pseudo-labels from a source-trained classifier when calibrating on unlabeled target data. The paper states that this is pessimistic: the pseudo-score for $Q(y)$2 is always less than or equal to the score for the true label, so the conformal quantile can be smaller than the oracle threshold and the resulting prediction sets can be too small, causing undercoverage. To address this, the paper proposes relaxed pseudo-calibrated sets that inflate the conformal threshold by a slack parameter and gives a rule connected to excess hinge loss and ramp loss. It also proposes a source-tuned pseudo-calibration algorithm that interpolates between hard pseudo-labels and randomized labels as a function of classifier uncertainty, with the explicit goal of making calibration more conservative when uncertainty is high.

The same work states that when marginal label distributions shift, $Q(y)$3, additional coverage loss can occur, quantified via the total variation distance between $Q(y)$4 and $Q(y)$5. It further reports that coverage degradation is not symmetric in marginal shifts, especially on class-imbalanced or high-class datasets, and that the source-tuned method mitigates coverage degradation while maintaining nontrivial prediction set sizes. This is significant because it separates two effects often conflated in practice: degradation from conditional covariate drift and degradation induced by asymmetric class-marginal change.

4. Target shift, density adjustment, and distribution-free prediction sets for regression

For regression, target shift is treated as a marginal distribution shift in which $Q(y)$6 but $Q(y)$7. The target is often fully unlabeled, so the crucial object is the likelihood ratio

$Q(y)$8

which simplifies under target shift to $Q(y)$9. Because $p(y\mid z)\neq q(y\mid z)$0 is not directly observed, the paper identifies the likelihood ratio by matching source and target covariate distributions within a finite B-spline space, using a sieve approximation $p(y\mid z)\neq q(y\mid z)$1 and least-squares distribution matching based only on observed covariates (Yi et al., 13 Oct 2025).

The nonconformity score is then built from a weight-adjusted conditional density estimator. The source conditional density is estimated through a quantile process, and Bayes’ rule is used to construct

$p(y\mid z)\neq q(y\mid z)$2

Prediction sets are highest predictive density sets implemented through weighted conformal ranks. The paper states that this construction can accommodate complex error structures such as asymmetry and multimodality, and that prediction sets adapt to skewed and heteroscedastic distributions rather than imposing symmetric intervals.

Several consequences are explicit. The method has a coverage guarantee, up to estimation error, provided the target is absolutely continuous with respect to the source. Standard conformal methods using residuals or symmetric score functions can be invalid or unnecessarily wide when distributional shift is asymmetric. In simulations and in a real-data application to the MIMIC-III clinical database, the method produces tighter prediction sets and accurately reflects skewed and heteroscedastic distributions. This suggests that marginal shift asymmetry in regression is not only a problem of reweighting but also a problem of choosing nonconformity scores whose geometry matches the shifted target density.

5. Tail-probability asymmetry and copula representations

In bivariate settings, asymmetry may be defined as imbalance between lower and upper tail joint probabilities. A copula-based measure captures this through

$p(y\mid z)\neq q(y\mid z)$3

The paper also gives the copula representation

$p(y\mid z)\neq q(y\mid z)$4

where $p(y\mid z)\neq q(y\mid z)$5 is the copula and $p(y\mid z)\neq q(y\mid z)$6 the survival copula (Kato et al., 2020).

The properties are straightforward. $p(y\mid z)\neq q(y\mid z)$7 if and only if lower and upper tail probabilities coincide for all $p(y\mid z)\neq q(y\mid z)$8, positive values indicate larger upper-tail mass, and negative values indicate larger lower-tail mass. The limit as $p(y\mid z)\neq q(y\mid z)$9 is $p(x\mid y,z)=q(x\mid y,z)$0. As $p(x\mid y,z)=q(x\mid y,z)$1, the limiting behavior depends on tail dependence coefficients, tail orders, or second derivatives of the copula, and the paper provides formulas for each case. This makes the measure local in the tail index $p(x\mid y,z)=q(x\mid y,z)$2, unlike global asymmetry summaries.

The paper also develops sample analogues for both copula data and ordinary real-valued data, establishes weak convergence to a Gaussian process in the copula-sample case, and gives simple methods for interval estimation and nonparametric testing. In an application to daily returns of S&P500 and Nikkei225, the estimated asymmetry is negative for small $p(x\mid y,z)=q(x\mid y,z)$3, with confidence intervals excluding $p(x\mid y,z)=q(x\mid y,z)$4, indicating that joint plunging events are more probable than joint soaring events. This is a concrete example of asymmetric marginal shift in tail risk rather than in class priors or response marginals.

6. Known target marginals and target-aware linear regression

Another line of work assumes that the target marginals of covariates and response are known or specified through population-level observations, boundary conditions, simulator configurations, or alignment-time constraints. In multivariate linear regression with stable conditional mean $p(x\mid y,z)=q(x\mid y,z)$5, this information can be used to construct target-aware estimators. The hybrid-loss estimator incorporates both target marginals through penalties that align the predicted target mean and target variance of $p(x\mid y,z)=q(x\mid y,z)$6; the constrained moment-matching estimator imposes the corresponding mean and variance constraints as hard constraints; and the two-stage estimator rescales ordinary least squares and recalibrates the intercept so that predictions match the target mean and variance (Hou et al., 22 Jun 2026).

The paper derives closed-form asymptotic mean squared errors for all three estimators and compares them with OLS. Its key factual conclusions are that the hybrid estimator is the benchmark target-aware estimator but requires coupled nonlinear optimization, that the constrained moment-matching estimator has essentially OLS-scale computation plus root-finding, and that the two-stage estimator nearly matches the hybrid benchmark in the high signal-to-noise regime at essentially no additional cost. Theoretical and Monte Carlo analyses show that performance depends on the alignment of mean and covariance shifts with $p(x\mid y,z)=q(x\mid y,z)$7, not merely on the amount of shift.

This has a direct bearing on asymmetry. The paper states that distribution shift asymmetry is geometry-sensitive: when mean or covariance shifts place more mass orthogonal to $p(x\mid y,z)=q(x\mid y,z)$8, moment-matching can outperform two-stage calibration; two-stage is insensitive to mean shift in those directions; and in low signal-to-noise regimes target constraints can increase error and even make moment-matching or calibration worse than OLS. A common misconception is that any access to target marginals necessarily improves estimation. The reported asymptotic mean squared errors and controlled experiments show that the benefit is conditional on signal strength and on the directional structure of the shift.

7. Skewed pivot-blend modeling and asymmetric outcome distributions

Skewness is treated explicitly in a separate modeling literature that argues the point at which unbalanced scales occur may be any quantile and cannot be reparametrized as an ordinary shift parameter in the presence of skewness. The skewed pivot-blend framework constructs a skewed density from any continuous density, including asymmetric and nonunimodal bases, by introducing distinct left and right scales around an adaptively estimated pivotal point $p(x\mid y,z)=q(x\mid y,z)$9. Its density is

$p(y)=q(y)$0

with continuity guaranteed at $p(y)=q(y)$1 (She et al., 2024).

This framework is relevant to marginal distribution shift asymmetry because it permits the pivotal point to be any quantile and estimates it jointly with the scales and other location parameters. The paper emphasizes that $p(y)=q(y)$2 is not a simple intercept when $p(y)=q(y)$3; it specifies where in the distribution the scale change occurs. This makes the model suitable when asymmetry appears away from the mean, median, or mode, including after data transformation. For semicontinuous outcomes, the paper introduces a skewed two-part model that identifies relevant variables across the entire population and mitigates the additional skewness induced by commonly used transformations.

The theoretical analysis is nonasymptotic and uses Bregman divergences and Orlicz norms. The paper states that uniform excess risk bounds increase as the asymmetry between $p(y)=q(y)$4 and $p(y)=q(y)$5 increases, quantifying how estimation becomes harder as the distribution becomes more skewed. Simulation studies and applications to abalone age and medical expenditure are reported to show improved recovery of scales, pivotal point, and coefficients, along with residual diagnostics based on backward pivot-blend transformations. This suggests that marginal shift asymmetry is not only a property to be detected after the fact; it can also be embedded directly into the generative form of the outcome model.