gamlss

Type-driven Rust crates for GAMLSS-style distributional regression.

Status: Actively developed. Public APIs and numerical behavior may change before 1.0.

Overview

GAMLSS models the full conditional response distribution rather than only its mean:

$$ Y_i \mid x_i \sim D(\theta_{i1}, \ldots, \theta_{iK}), \qquad g_k(\theta_{ik}) = X_{k,i}\beta_k + \sum_j f_{k,j}(x_i). $$

Each distribution parameter can have its own link, covariates, smooth terms, and penalties. The typed API expresses parameter domains and model structure at compile time while keeping model evaluation backend- and optimizer-agnostic.

The library supports scalar distributional models, optional multivariate families, finite mixtures, smooth predictors, target transforms, post-fit diagnostics, and a lightweight Bayesian posterior-potential layer. Distribution functions, quantiles, CRPS, and sampling are exposed through capability traits and vary by family. See the family capability matrix for exact coverage.

Fitting loops and optimizer integrations intentionally remain outside the core API.

Crates

Most users should depend on the gamlss facade. The workspace also publishes focused crates:

• gamlss-core and gamlss-family provide typed model abstractions, distributions, likelihoods, and scores.
• gamlss-spline and gamlss-special provide predictors, penalties, special functions, and numerical helpers.
• gamlss-transform and gamlss-diagnostics cover response preprocessing and post-fit diagnostics.
• gamlss-formula is an experimental builder layer for curated workflows; gamlss-bayes provides priors and posterior potentials without a sampler.

See project structure for API layers and crate boundaries.

Features

• formula (default) re-exports the experimental builder API.
• bayes enables gamlss::bayes and its common prelude types.
• rand enables sampling for supported families.
• multivariate enables multivariate families.

Use default-features = false for the facade without the formula layer, or depend on individual crates for tighter dependency control.

Development

Run the workspace checks with:

cargo fmt --all
cargo clippy --workspace --all-targets --all-features
cargo test --workspace --all-features

Run the main benchmarks with:

cargo bench -p gamlss-family --bench objective --all-features
cargo bench -p gamlss-spline --bench spline_hot_paths
cargo bench -p gamlss-special --bench numeric_kernels

Each crate's benches/README.md documents focused Criterion filters and coverage. Compare benchmark results only on the same hardware and toolchain.