Crate prospicio_models 

The model interface and its life cycle, shared by every engine (GLM, GAM, neural networks, boosting adapters, Bayesian models). See docs/design/models.md.

• Family and Link: exponential dispersion families and link functions, the likelihoods and losses engines share;
• Terms, Coding and Design: model terms, factor coding learned on training data, and the design matrix with offset and weights;
• Model and Fitted: fit, predict, and joint predictive distributions;
• metrics: deviance, Gini, lift, CRPS, coverage;
• resample: k-fold, grouped and time-ordered splits, cross-validation and grid search;
• artifact: the envelope of saved models (format, version, provenance) and its JSON helpers;
• simulate: predictive distributions from fitted means, for engines that give only a mean per row (the boosting adapters).

This crate has no heavy dependencies; engines live in their own crates.

Re-exports

pub use design::Coding;

pub use design::Column;

pub use design::Design;

pub use design::Frame;

pub use design::Term;

pub use design::Terms;

pub use family::Family;

pub use link::Link;

pub use model::Fitted;

pub use model::Model;

Modules

artifact

The envelope every model artifact shares (docs/design/models.md, “Model artifacts”): the format check and version rule, provenance, and the JSON helpers the engines’ readers and writers use. Non-finite numbers are written as the strings "NaN", "inf" and "-inf"; with the workspace’s float_roundtrip every other number round-trips bit for bit.

compare

One table of out-of-sample scores across models of any kind.

design

Design matrices: turning named columns into the numeric matrix a model fits, the same way on training data and on new data.

encode

Out-of-fold target encoding of a high-cardinality factor with credibility (Holmes and Casotto, CAS Monograph 16, chapter 6).

family

Exponential dispersion families: what a GLM, a neural network’s loss, a booster’s objective and a Bayesian likelihood all share.

link

Link functions: η = g(μ) between the mean and the linear predictor.

metrics

Scores for fitted models: deviance-based, ranking (Gini, lift), calibration (actual vs expected), and probabilistic (CRPS, coverage, log score, PIT).

model

The interface every model implements, whatever its engine.

monitor

Monitoring a stored model on new periods: actual against expected.

resample

Resampling: train/test splits, cross-validation, and grid and random search.

simulate

Predictive distributions from fitted means, for engines that give a mean per row and nothing else (gradient boosting through its Python or R package, an imported network): the family adds the process noise, and several mean vectors (bootstrap refits) add the parameter uncertainty.

stack

Model weights for blending from pointwise out-of-sample log predictive densities: stacking and pseudo-BMA (Yao, Vehtari, Simpson and Gelman, 2018, “Using stacking to average Bayesian predictive distributions”).