Fitting Architecture

The fitting layer should stay simple for ordinary users while leaving room for more algorithms over time.

Current keyword API

This must remain enough:

fit(RVineCopula, U)

Additional keywords control candidate families, criteria, rotations, structure, truncation, and thresholds:

fit(
    RVineCopula,
    U;
    family_set=:default,
    selection_criterion=:bic,
    tree_criterion=:tau,
    tree_algorithm=:kruskal,
    allow_rotations=true,
)

Simple scalar configuration should remain keyword-based. Do not introduce types only to replace a boolean, a threshold, or a symbol that is not an extension point.

Strategy objects, eventually

As the package grows, some concepts may deserve dispatchable strategies:

• tree criteria;

• maximum-spanning-tree algorithms;

• structure learners;

• pair selectors;

• truncation selectors;

• threshold selectors.

The reason to introduce a strategy object is extensibility: external users should be able to add a new algorithm without editing one large central if/elseif chain.

Do not over-type simple options

A strategy type should represent behavior. It should not be a wrapper around a scalar parameter unless dispatch genuinely simplifies the implementation.

Distribution versus fitted result

A vine copula object is the probabilistic model. It should contain the structure and pair-copulas needed to evaluate, simulate, and transform observations.

Fitting metadata belongs in CopulaModel:

• fit method;

• convergence flag;

• iterations;

• log-likelihood;

• AIC/BIC and future mBICV;

• selected truncation;

• edge-level diagnostics;

• selection traces.

This separation keeps model evaluation lightweight and avoids turning distribution objects into mutable analysis logs.

Near-term extension priorities

The next fitting-related additions should be staged:

1. document current behavior and edge metadata;

2. add automatic truncation selection;

3. add mBICV scoring;

4. expose richer diagnostics through fitted-result metadata;

5. only then consider public strategy objects.

The strategy architecture should follow real pressure from new algorithms rather than arriving as a speculative object model.