Bayesian pseudo-posterior for finite mixtures of Clayton, Gumbel, Frank and Gaussian copulas: consistency, identifiability and applied insights

Table of Contents

1. Key Highlights
2. Introduction
3. Why tail asymmetry and mixtures matter
4. Finite mixtures of copulas: model and interpretation
5. Two-stage Bayesian pseudo-posterior: practical method
6. The logarithmic boundary envelope and why it guarantees consistency
7. Identifiability through finite linear independence: what it means and why it matters
8. Tail-dependence coefficients and conditional tail probabilities: consistent estimation
9. Simulation evidence: stability, efficiency and behavior near independence
10. Case studies: mirror-image tail asymmetries in physical fitness data
11. Practical recommendations for applied researchers
12. Limitations and open questions
13. How to compute tail-dependence measures and conditional probabilities from the pseudo-posterior
14. Practical examples and an implementation sketch
15. Policy and applied implications
16. FAQ

Key Highlights

• A two-stage Bayesian pseudo-posterior that evaluates copula likelihood on rank- or kernel-based pseudo-observations yields strong consistency for copula densities in a wide class of finite mixtures, including Clayton, Gumbel, Frank, Student and Gaussian components.
• Clayton, Gumbel, Frank and Gaussian copulas are jointly finitely linearly independent, which secures identifiability of mixture weights and components; simulations show the pseudo-posterior is stable in small samples and efficient near independence where tail structure emerges before weights stabilize.
• Applied analyses of two large physical-fitness datasets reveal opposing tail asymmetries: elite athletes show upper-tail coupling of sprint and jump performance, while older adults display lower-tail clustering of weak grip strength and low daily activity—patterns a single Gaussian copula misses.

Introduction

Modeling dependence between multiple variables is essential across disciplines: finance needs accurate joint tail risk measures, epidemiology must identify subpopulations at heightened risk, and sports science seeks combinations of traits that predict elite performance. Copulas provide a principled way to separate marginal behavior from dependence structure, but real-world dependence often deviates from the symmetric patterns captured by a single parametric copula. Finite mixtures of copulas offer flexibility to capture different dependence regimes across the distribution—distinct behaviors in the lower and upper tails, for example—but raise statistical and computational challenges: how can inference proceed when margins are unknown or when some components approximate independence?

A recent methodological development addresses these challenges through a two-stage Bayesian approach. Margins are replaced by ranks or kernel estimates to obtain pseudo-observations; a copula likelihood is then evaluated on these transformed data, producing a pseudo-posterior for the mixture parameters. The approach comes with rigorous guarantees: the pseudo-posterior concentrates on the true copula density under a mild technical condition (a logarithmic boundary envelope), a condition satisfied by finite mixtures of popular copulas in any fixed dimension. Additional theoretical work shows that Clayton, Gumbel, Frank and Gaussian copulas are finitely linearly independent, which implies identifiability of mixture weights and component parameters. Simulations and applications to large physical-fitness datasets illustrate the method’s practical strengths and its capacity to reveal asymmetric tail dependence that a Gaussian copula misses.

The following sections unpack these results, explain the methodology in accessible terms, connect them to practical considerations, and highlight implications for applied researchers.

Why tail asymmetry and mixtures matter

Dependence between variables is rarely homogeneous across their ranges. Two variables might be nearly independent for typical values but strongly co-move at extremes; alternatively, they may cluster at low values but not at high ones. Classic parametric copulas capture specific dependence patterns: Clayton emphasizes lower-tail dependence, Gumbel emphasizes upper-tail dependence, Frank is symmetric and tail-neutral, and Gaussian copulas impose a symmetric, tail-independent dependence that can miss heavy-tail clustering.

Consider three concrete situations:

• In systemic-risk assessment, asset returns may exhibit stronger co-movement during market crashes (lower tails) than during booms, so a model that captures lower-tail dependence is crucial.
• In medicine, two biomarkers might jointly indicate disease only when both are low, creating lower-tail clustering that a symmetric copula would underestimate.
• In sports science, speed and power may be coupled primarily among top performers, a phenomenon manifesting as upper-tail dependence.

Finite mixtures of copulas combine components with different tail behaviors, allowing models to represent such asymmetries. A two-component mixture with a Clayton and a Gumbel component, for instance, can capture simultaneous lower- and upper-tail clustering; mixing weights quantify the relative importance of each dependence regime. This flexibility improves fit and predictive performance, but learning mixture weights and component parameters reliably requires identifiability and robust inference procedures that handle unknown margins.

Finite mixtures of copulas: model and interpretation

A copula is a multivariate distribution on the unit hypercube [0,1]^d with uniform marginals. For continuous margins, Sklar’s theorem guarantees representation of the joint distribution as the copula applied to the marginal cumulative distribution functions. In practice, margins are rarely known and must be estimated or removed by transformation.

A finite mixture of copulas posits that the dependence structure arises from a convex combination of K copula densities: C(u) = sum_{k=1}^K w_k C_k(u; θ_k), where u ∈ [0,1]^d are pseudo-observations, C_k are component copula densities parameterized by θ_k, and w_k ≥ 0 sum to 1. Each component can encode distinct tail behavior and dependence geometry. Interpretation is straightforward: the mixture represents a population-level dependence that arises from a latent categorical mechanism or simply as a flexible approximation to a complex, possibly multimodal dependence surface.

Key quantities of interest include:

• Mixture weights w_k, which reflect relative prevalence of different dependence regimes.
• Component parameters θ_k, which determine tail dependence and local curvature.
• Derived measures such as tail-dependence coefficients λ_L and λ_U (for lower and upper tails) and conditional tail probabilities, which are central to risk assessment and subgroup detection.

Identifiability is critical: distinct parameter settings should lead to distinct mixture densities so that the data can, in principle, inform weights and components. Without identifiability, multiple configurations could explain the same observed dependence, undermining inference.

Two-stage Bayesian pseudo-posterior: practical method

The two-stage approach separates margin handling from copula inference. Stage 1 transforms raw observations into pseudo-observations on [0,1] for each margin; Stage 2 treats these pseudo-observations as data for a copula model and computes a pseudo-likelihood, producing a pseudo-posterior over mixture parameters.

Stage 1 — constructing pseudo-observations:

• Rank-based approach: For continuous data with no ties, transform each margin by empirical ranks: u_i = rank(x_i)/(n+1). Ranks produce uniform[0,1] pseudo-observations and are robust to marginal misspecification. They are especially appealing when margins are complicated or when interest is focused on dependence.
• Kernel-based smoothing: When margins need more refined estimation—e.g., to reduce finite-sample discreteness—kernel density estimators can be used to create smoothed marginal CDFs and corresponding pseudo-observations. This approach suits moderate-to-large samples and when margins exhibit smooth structure.

Stage 2 — copula pseudo-likelihood and Bayesian computation:

• The copula density is evaluated at u_i and the mixture parameters θ and weights w. Because the margins were estimated rather than treated as unknown parameters, the resulting posterior is a pseudo-posterior: it conditions on data-derived transforms rather than on the full likelihood.
• Standard Bayesian machinery applies: choose priors on weights and component parameters, run MCMC targeting the pseudo-posterior, and compute posterior summaries for weights, parameters, tail coefficients and predictive densities.

Computational choices and sampler design:

• A marginal Metropolis sampler that updates component parameters and weights directly on the pseudo-posterior offers good performance; the authors find it up to twice as efficient as a data-augmentation sampler that introduces latent component indicators.
• Multistart maximum pseudo-likelihood (MPL) can provide a frequentist benchmark but may suffer from multiple local optima; the pseudo-posterior brings the benefits of Bayesian regularization and uncertainty quantification.

Benefits of the two-stage pseudo-posterior:

• Avoids modeling and sampling over complex marginal distributions simultaneously with the copula, reducing model complexity.
• Ranks allow use of the same dataset for both stages without violating consistency results under the envelope condition.
• When the true copula is in a finite mixture class satisfying the envelope, the pseudo-posterior concentrates on the true copula density.

The logarithmic boundary envelope and why it guarantees consistency

Consistency of Bayesian procedures depends on regularity conditions on the model and the data-generating mechanism. For the pseudo-posterior constructed on transformed data, the critical technical condition identified is that the log copula density admits a logarithmic boundary envelope. Explaining this concept requires attention to how copula densities behave near the corners of the unit hypercube.

Copula densities can become unbounded near boundaries. For instance, Clayton and Gumbel copulas exhibit power-law singularities at lower and upper boundaries respectively. If the log density diverges too quickly near the boundary, approximating the true copula or ensuring posterior concentration becomes problematic. A logarithmic boundary envelope controls that divergence: it requires that the log density near the boundary grows at most on the order of the logarithm of the reciprocal of the distance to the boundary. Informally, it prevents excessively sharp spikes.

Why does this matter for rank-based pseudo-posteriors? The rank transformation creates pseudo-observations that are random and discrete in finite samples, and their behaviour near boundaries is delicate. The logarithmic envelope ensures that small estimation errors in the margins do not translate into uncontrolled errors in the copula likelihood, allowing the pseudo-posterior to concentrate as sample size increases.

Crucially, the envelope condition is satisfied for finite mixtures of widely used copulas: Gaussian, Student, Clayton, Gumbel and Frank. That means the theoretical guarantee is applicable to practical mixture models used by researchers. Under the envelope, the pseudo-posterior is strongly consistent for the copula density: with growing sample size, posterior mass concentrates in arbitrarily small neighborhoods of the true copula density.

Practical implication: applied researchers can use rank-based pseudo-posteriors with confidence when working with these common copulas and finite mixtures of them, knowing that asymptotic correctness holds under the envelope.

Identifiability through finite linear independence: what it means and why it matters

Identifiability of mixture models is subtle. If component densities are linearly dependent in a finite sense, different sets of weights and parameters can produce the same mixture density. The analysis establishes that Clayton, Gumbel, Frank and Gaussian copulas are jointly finitely linearly independent. In plain language, no nontrivial finite linear combination of these component densities can vanish almost everywhere unless all coefficients are zero.

Consequences:

• Mixture weights and component parameters are identifiable (under usual regularity), so consistent estimation is possible.
• Posterior concentration extends beyond the copula density to mixture weights and component parameters: as sample size grows, the pseudo-posterior concentrates on the true parameters, not just on their induced density.
• Practitioners can interpret estimated weights as meaningful relative frequencies or regime prevalences, not merely as overparameterized artifacts.

This finite linear independence is a mathematically nontrivial fact. Copula families derived from similar constructions might, in principle, exhibit dependencies that obstruct identifiability, but the result provides broad reassurance for these commonly used families.

Tail-dependence coefficients and conditional tail probabilities: consistent estimation

Two practically important summaries of dependence are tail-dependence coefficients and conditional tail probabilities:

• Lower-tail dependence λ_L is the limit of P(U2 ≤ t | U1 ≤ t) as t → 0; upper-tail dependence λ_U is the analogous limit as t → 1.
• Conditional tail probabilities at finite thresholds (e.g., P(U2 ≤ 0.05 | U1 ≤ 0.05)) give more granular views of extreme clustering.

Because the pseudo-posterior is consistent for the copula density under the envelope condition and because tail measures are continuous functionals of the copula in these families, posterior estimates of λ_L, λ_U and conditional tail probabilities are consistent. That matters for risk management, epidemiology and any application concerned with joint extremes.

A notable practical observation from simulation and empirical application is that tail coefficients often become well-identified earlier (in smaller samples) than the mixture weights, particularly near independence. When all components are close to the independence copula, their weights may be weakly identified, but the tail structure—if present—still stands out in the data and can be learned quickly. This hierarchical learning pattern is useful: it allows confident assessment of tail risks even when finer partitioning of dependence regimes remains uncertain.

Simulation evidence: stability, efficiency and behavior near independence

The pseudo-posterior was benchmarked against multi-start maximum pseudo-likelihood (MPL) and a data-augmentation MCMC on several fronts.

Large samples:

• The pseudo-posterior and MPL converge to similar estimates as sample size increases, validating both approaches in the asymptotic regime.

Small samples and near independence:

• The pseudo-posterior is more stable than MPL in small samples. MPL can be sensitive to starting values and can converge to local optima; Bayesian posterior averaging mitigates some of this instability.
• Near independence—where component copulas are themselves near the product copula—the data provide weak information about weights. The pseudo-posterior still learns tail coefficients ahead of weights. A marginal Metropolis sampler that focuses on parameters without introducing latent allocations shows computational advantages: it can be up to twice as efficient as a data-augmentation algorithm that samples latent component indicators at each iteration.

Interpretation:

• The pseudo-likelihood framework is robust to margin estimation error when ranks are used and the envelope holds.
• Bayesian regularization and uncertainty quantification prove valuable in settings where data do not strongly discriminate between similar mixture configurations.

These findings guide both method choice and computational strategy. For moderate-to-large samples, MPL and pseudo-posterior agree, but for small samples or near-independence regimes, the Bayesian approach offers practical advantages.

Case studies: mirror-image tail asymmetries in physical fitness data

Two large, real-world datasets illustrate the practical interpretability and predictive gains of finite copula mixtures inferred via the pseudo-posterior.

Case 1 — University students (n = 8,772): sprint and jump performance

• Data: Two performance metrics—sprint time and jump height—measured on a large cohort of university students.
• Finding: Dependence concentrates in the upper tail. Individuals who score highly on one metric tend to score highly on the other; this coupling is strongest among top performers.
• Interpretation: The joint occurrence of high sprint speed and high jump height reflects shared physiological attributes (explosive power, fast-twitch muscle fibers) and training specialization that manifest most strongly among elite performers. A mixture including a Gumbel-like component captures this upper-tail clustering.
• Predictive implication: Talent identification programs can use a model that allows upper-tail dependence to better detect multi-skill elites than a Gaussian copula which smooths over tail asymmetry.

Case 2 — National health survey adults (n = 5,336): grip strength and daily activity

• Data: Grip strength and daily activity recorded in a representative national health survey with adult participants across the age spectrum.
• Finding: Dependence concentrates in the lower tail and strengthens with age. Low grip strength co-occurs with low daily activity increasingly among older adults; high grip strength and high activity do not show similar coupling.
• Interpretation: Lower-tail clustering signals a vulnerable subgroup for whom both muscular weakness and sedentariness co-occur—identifying such clustering has clear public-health relevance for targeted interventions to prevent functional decline.
• Predictive implication: Models that allow lower-tail dependence (Clayton-like components) better capture risks of combined frailty than Gaussian models. A Gaussian copula notably fails to capture these asymmetric patterns in held-out data, underestimating the joint risk of low strength and low activity.

Both case studies highlight asymmetric dependence that a single symmetric Gaussian copula misses. Mixture copulas learned via the pseudo-posterior are able to surface these patterns and yield better predictive performance on held-out data.

Practical recommendations for applied researchers

The methodological and empirical results suggest clear guidance for practitioners who need flexible dependence models:

1. When to prefer finite mixtures:
   • Use mixtures when exploratory analysis or domain knowledge suggests asymmetric tail behavior or multiple distinct dependence regimes.
   • Mixtures are appropriate when a single parametric copula fails diagnostic checks or predictive accuracy on extremes.
2. Margin treatment:
   • Use rank-based pseudo-observations when margins are complicated or when robustness is paramount. Ranks are straightforward and compatible with the pseudo-posterior consistency result under the envelope.
   • Consider kernel-smoothed margins when sample size is moderate-to-large and smoothing is beneficial for finite-sample accuracy.
3. Component selection:
   • Include copulas that reflect hypothesized tail behavior: Clayton for lower-tail, Gumbel for upper-tail, Frank for symmetric moderate dependence, Student for heavy-tailed symmetric dependence, and Gaussian as a baseline.
   • Finite mixture sizes should balance flexibility and identifiability; overfitting with too many components can increase computational burden and interpretational complexity. Model selection criteria (e.g., marginal pseudo-likelihood comparisons, predictive log-scores) help choose K.
4. Priors and regularization:
   • Use weakly informative priors on weights (e.g., Dirichlet with moderate concentration) to avoid pathological weight estimates in small samples.
   • Priors on copula parameters can reflect plausible ranges of dependence; regularization stabilizes estimation in near-independence settings.
5. Sampling strategies:
   • Marginal Metropolis samplers that operate directly on mixture parameters are computationally efficient, especially when latent allocations mix slowly.
   • Data augmentation with allocation variables can be useful when component posterior separation is strong, but monitor mixing carefully.
6. Diagnostics:
   • Compare pseudo-posterior predictive distributions with observed joint behaviors, focusing on tail quantiles and conditional tail probabilities.
   • Conduct posterior predictive checks targeted at extremes (e.g., conditional exceedance frequencies) to ensure tail behaviors are captured.
   • Compare to Gaussian copula baselines; misfit in tails often indicates need for asymmetric components.
7. Reporting:
   • Report posterior credible intervals for tail-dependence coefficients and conditional tail probabilities, not only point estimates.
   • When weights are weakly identified, highlight uncertainty and focus interpretation on derived measures (e.g., tail coefficients) that may be better resolved.

Limitations and open questions

The pseudo-posterior approach and accompanying theory make substantial progress, but several limitations and directions for further work remain.

Higher-dimensional scaling:

• Finite mixtures of copulas in higher dimensions increase parameter counts and computational complexity. The envelope condition and linear independence results hold in fixed dimension, but practical performance in high dimensions requires careful component selection and scalable samplers.
• Vine copulas and factor copula models offer alternatives for high-dimensional dependence with structured sparsity; integrating mixture components into vines poses both modeling and identifiability challenges.

Model selection and number of components:

• Choosing K remains a practical challenge. Cross-validated predictive scores on pseudo-observations are useful, but principled Bayesian model selection under the pseudo-likelihood framework requires further study.

Missing data and censoring:

• Real datasets often include missing values or censoring. Extending the pseudo-posterior to handle missingness—potentially by combining rank-based treatment with imputation or latent-variable strategies—needs development and theoretical justification.

Time series and dynamic dependence:

• Dependence that evolves over time calls for dynamic copula mixtures or regime-switching copula models. Extending the pseudo-posterior approach to temporal settings requires careful use of temporal ranks and consideration of dependence across time.

Nonparametric component families:

• Finite mixtures of parametric copulas provide flexibility, but nonparametric copula components could further capture complex local dependence. Combining pseudo-posterior consistency results with nonparametric component classes is an open theoretical and computational problem.

Robustness to marginal estimation:

• While ranks offer robustness, kernel-based margins and small-sample margins can introduce bias. Quantifying the finite-sample impact of margin estimation choices and developing bias-correction strategies remain valuable.

How to compute tail-dependence measures and conditional probabilities from the pseudo-posterior

Applied users will want concrete steps for computing tail measures from the posterior samples.

Procedure:

1. For each posterior draw of mixture parameters {w_k, θ_k}:
   • Compute the copula C(u; w, θ) and its density as needed.
   • Evaluate finite-threshold conditional probabilities, e.g., P(U2 ≤ t | U1 ≤ t) at chosen t values, by integrating the joint copula mass over the relevant rectangle and dividing by t.
   • Approximate limiting tail-dependence coefficients by evaluating at a sequence of small t (for lower tail) or large t near 1 (for upper tail) and extrapolating or taking the limit numerically if stable.
2. Aggregate across posterior draws:
   • Report posterior means, medians and credible intervals for λ_L, λ_U and conditional tail probabilities at relevant thresholds.
   • Use posterior predictive simulations to obtain uncertainty in tail-event predictions.

Numerical considerations:

• When component copulas have singular behavior near boundaries, numerical integration requires dense grids or importance sampling targeted to corners.
• For bivariate analyses, closed-form expressions exist for tail coefficients for many families (e.g., Clayton and Gumbel), and these can be evaluated directly per posterior draw.
• For higher dimensions or non-closed-form situations, Monte Carlo integration using copula draws per posterior sample is straightforward.

Practical examples and an implementation sketch

A minimal outline for implementing the pseudo-posterior in an applied workflow:

1. Data preparation:
   • Clean data, handle ties (e.g., jitter small ties if ranks needed).
   • Inspect marginal distributions and dependence scatterplots focusing on tails.
2. Pseudo-observations:
   • Compute ranks u_{ij} = rank(x_{ij})/(n+1) for each margin j, or estimate marginals with kernel CDFs for smoother transforms.
3. Model specification:
   • Choose K and component families (e.g., Clayton + Gumbel + Gaussian).
   • Specify priors: Dirichlet(a) for weights (a=1 for uniform, >1 to regularize toward equal weights), weakly informative priors on θ_k.
4. Posterior sampling:
   • Implement marginal Metropolis-Hastings updates for θ_k and weights w; propose on transformed space for bounded parameters (e.g., logit for weights).
   • Monitor mixing, effective sample size, and convergence diagnostics.
   • Optionally run multiple chains with dispersed starts.
5. Summaries and checks:
   • Compute posterior summaries for weights, parameters, tail coefficients, conditional tail probabilities.
   • Perform posterior predictive checks comparing observed joint tail frequencies to predictive intervals.
   • Evaluate out-of-sample predictive scores where possible (log-score, tail-weighted scores).
6. Interpretation and action:
   • Interpret asymmetric dependence and quantify its practical implications (risk estimates, targeted intervention selection, talent identification).
   • Report uncertainty and sensitivity to K and marginal transforms.

Open-source software:

• Many statistical languages support copula densities and sampling (R packages such as copula, VineCopula, and Bayesian tools). Custom implementation of the pseudo-posterior is straightforward by combining marginal transforms with MCMC libraries.

Policy and applied implications

The capacity to detect asymmetric tail dependence has immediate applications:

• Public health: pinpointing combined low-functioning subgroups (e.g., low activity and weak grip strength) enables targeted prevention programs for frailty.
• Sports science: identifying upper-tail coupling informs talent-spotting and specialized training design.
• Finance and insurance: accurate modeling of joint extremes improves portfolio stress testing and pricing of catastrophe-linked contracts.

Because the pseudo-posterior approach delivers consistent tail measures and quantifies uncertainty, decision-makers can base targeted interventions and resource allocation on statistically principled estimates rather than ad hoc thresholds.

FAQ

Q: What exactly is a copula and why use it? A: A copula is a multivariate distribution function with uniform marginals that captures the dependence structure among variables independently of their marginal distributions. Use copulas when dependence is of primary interest or when margins differ across variables, allowing flexible joint-modeling by combining marginal models with a copula.

Q: Why replace margins with ranks or kernel estimates rather than model them jointly? A: Modeling margins jointly with the copula increases model complexity and demands stronger assumptions. Rank-based transforms produce uniform pseudo-observations that isolate dependence, are robust, and simplify inference. Kernel estimates can give smoother finite-sample behavior. The two-stage pseudo-posterior leverages this simplification while maintaining consistency under the envelope condition.

Q: What is strong consistency of the pseudo-posterior? A: Strong consistency means the pseudo-posterior concentrates on arbitrarily small neighborhoods of the true copula density with probability one as the sample size grows. Practically, this ensures that posterior estimates of the copula converge to the true dependence structure.

Q: How do tail-dependence coefficients differ from correlation? A: Correlation measures average linear association and can miss asymmetric or tail-localized dependence. Tail-dependence coefficients quantify the probability of joint extremes (both very high or both very low), which is essential for risk assessment and identifying joint vulnerabilities.

Q: What does identifiability of mixture components mean for interpretation? A: Identifiability ensures that the estimated weights and component parameters correspond uniquely to the observed mixture density. This allows practitioners to interpret weights as regime prevalences and to attribute tail behaviors to specific copula components.

Q: Are these results limited to bivariate cases? A: The theoretical envelope and linear independence results hold for finite mixtures in any fixed dimension, but practical challenges increase with dimension. Computational scalability and model selection become more demanding in higher dimensions.

Q: How should I choose the number K of components? A: Balance flexibility and parsimony. Use predictive scoring on held-out pseudo-observations, compare marginal pseudo-likelihoods, and consider interpretability. Start with small K motivated by hypothesized tail behaviors and expand only if diagnostics indicate misfit.

Q: Does the method work when some components are close to independence? A: Yes. The pseudo-posterior is stable and often learns tail coefficients before weights when components approximate independence. Bayesian regularization helps prevent overfitting in such near-independence regimes.

Q: Can the approach handle missing data? A: Missingness requires additional modeling. Simple imputation or model-based approaches integrating missing-data mechanisms into the pseudo-posterior workflow can work but need careful validation and, in some cases, new theoretical support.

Q: Where can I find software to implement this? A: Standard statistical packages provide copula functions, and MCMC frameworks (Stan, JAGS, R tools) can be adapted. Several R packages (e.g., copula, VineCopula) provide component densities and tail measures; custom MCMC targeting the pseudo-posterior is typically required for mixtures.

The two-stage Bayesian pseudo-posterior for finite mixtures of copulas combines theoretical guarantees with practical flexibility. It enables consistent estimation of copula densities and tail measures for broadly used copula families and yields interpretable mixture weights under an identifiability result. Empirical applications demonstrate the method's ability to surface asymmetric dependence patterns that matter for prediction and policy. Applied researchers aiming to model complex dependence in tails will find this framework a principled option that balances robustness, interpretability and computational tractability.