Skip to main content

Crate skaters

Crate skaters 

Source
Expand description

skaters-core: Rust port of the skaters online distributional forecasting library. Parity-gated against the Python reference via parity/vectors.json (seven probes per step and horizon, atol = rtol = 1e-6).

All transcendental math routes through the pure-Rust libm crate, so the core is bit-deterministic across platforms (x86 / ARM / wasm32).

Re-exports§

pub use api::build_candidates;
pub use api::laplace;
pub use api::Forecaster;
pub use granger::Granger;
pub use granger::GrangerParams;
pub use granger::Mixture;
pub use cov::EmaCov;
pub use cov::LedoitWolfCov;
pub use cov::RunningCov;
pub use dist::Dist;
pub use periodicity::top_periods;
pub use periodicity::PeriodDetector;
pub use search::search;
pub use search::Search;
pub use spec::Spec;
pub use spec::TransformSpec;
pub use skater::bayesian_ensemble;
pub use skater::conjugate;
pub use skater::ema;
pub use skater::multiscale;
pub use skater::precision_weighted_ensemble;
pub use skater::sticky;
pub use skater::terminal_leaf_ensemble;
pub use skater::Sk;
pub use tails::gpdtails;
pub use tails::PDist;
pub use tails::SplicedDist;

Modules§

api
User-facing API: the candidate population and the laplace composition. Port of src/skaters/api.py (search/spec paths are out of scope, as in the R port).
cov
Online covariance estimation: running (Welford), EMA, and Ledoit-Wolf shrinkage toward identity. Port of src/skaters/cov/.
dist
Gaussian mixture distribution: the distributional prediction type. Port of src/skaters/dist.py. Operation order mirrors Python exactly, including the ulp-tolerant prune and CPython-compatible compensated sums.
granger
granger: laplace with trained parameters, mixed with laplace so it cannot do much worse. Port of src/skaters/granger.py (horizon 1), parity-gated against it.
leaf
Residual distribution estimators: the leaves of every prediction tree. Port of src/skaters/leaf.py (leaf, scale_mixture_leaf, crps_leaf, garch_leaf).
mathx
Portable math helpers. All transcendentals route through libm (pure Rust) so results are bit-identical across x86 / ARM / wasm32. sqrt is IEEE-exact and may use the std method.
parade
The prediction parade: online PIT/z calibration diagnostics in the state, plus the finite-input gate. Port of src/skaters/parade.py.
periodicity
Online periodicity detection via running autocorrelation. Port of src/skaters/periodicity.py.
runstats
Welford online mean/variance. Port of src/skaters/runstats.py.
search
Adaptive search over the transform tree. Port of src/skaters/search.py.
skater
The skater type: every forecaster is a struct with step(&mut self, y) -> Vec<Dist> whose fields are plain serde-serializable data (a product contract).
spec
Symbolic specification for skater pipelines. Port of src/skaters/spec.py.
tails
GPD tail splice: the conditional tail fit, in the predictive itself. Port of src/skaters/tails.py (Acklam phi_inv, censored-ML GPD fit, SplicedDist, and the gpdtails wrapper).
transform
Invertible online transforms for conjugation. Port of src/skaters/transform.py.