Changelog

Source: NEWS.md

MetaHunt 0.1.0

Initial public release.

End-to-end fitter

• metahunt() chains denoised functional SPA basis hunting, constrained simplex projection, and Dirichlet weight modelling in a single call. Method dispatch for predict(), summary(), and plot() on the returned "metahunt" object.

Conformal prediction

• split_conformal() and cross_conformal() return distribution-free prediction intervals around the target function (pointwise on the grid or, with a wrapper, around a scalar summary).
• conformal_from_fit() adds intervals to an already-fit pipeline using a held-out calibration set.
• coverage(), summary(), and plot() methods for the "metahunt_conformal" class.

Rank and tuning selection

• reconstruction_error_curve() (unsupervised elbow) and cv_error_curve() (supervised CV) for picking K.
• select_denoising_params() cross-validates the (N, Delta) knobs of dfspa().

Pipeline building blocks

• dfspa() denoised functional Successive Projection Algorithm (Algorithm 1 of the paper).
• project_to_simplex() constrained simplex projection of each study’s function onto the recovered bases (quadratic program via quadprog).
• fit_weight_model() and predict.metahunt_weight_model() for Dirichlet regression of simplex weights on study-level covariates, with coef.metahunt_weight_model() for inspecting coefficients.
• predict_target() and apply_wrapper() for composing predictions and scalar summaries by hand.

Data preparation

• build_grid() constructs a shared evaluation grid from any reference patient-level dataset.
• f_hat_from_models() evaluates a list of fitted models on the shared grid with class-aware dispatch for ranger, grf (causal_forest, regression_forest), and a default branch that covers lm/glm/randomForest. Custom S4 classes can supply their own predict_fn.

Baselines

• minmax_regret() implements the covariate-free worst-case-regret aggregator of Zhang, Huang, and Imai (2024, arXiv:2412.11136).

Documentation

• Tutorials: metahunt-intro, data-prep, grid-weights, wrapper-scalar, plus get-started.
• Companion paper: Shi, Imai, and Zhang (2024, arXiv:2604.23847).