API guide¶

The public namespace is shape_adaptive_robust_regression. The table below is a curated entry point; the estimator source contains the exact signatures and implementation docstrings.

Primary estimator¶

ShapeAdaptiveEN(alpha=0.5, c_grid=(1.0, 1.5, 2.0, 3.0, 5.0),
                n_iter=2, n_lam=15, k=5, seed=0,
                pilot=None, kkt_tol=None)

Starts from a pilot, estimates a bounded score from standardized pilot residuals, and fits an elastic net to the resulting pseudo-responses. alpha controls elastic-net mixing; c_grid contains score-shape candidates; n_iter is the number of updates; n_lam is the penalty-path size and k the CV fold count. seed controls the reproducible fold construction. pilot optionally supplies the initial fit. kkt_tol adds a final-target numerical bound when set. Defaults are documented, not changed by this site.

Shape-penalized estimator¶

ShapePenalizedEN(lam=0.1, alpha=0.5, lam1=0.0, lam2=0.0,
                 robust=False, c_huber=1.345, c_clip=2.5,
                 scale=None, max_sweeps=1000, tol=1e-9,
                 beta_init=None, eps=1e-6, kkt_tol=None)

lam and alpha set elastic-net regularization. lam1 and lam2 weight residual skewness and excess-kurtosis penalties. robust=False uses least squares and raw centered residual moments; robust=True uses Huber loss and clipped standardized residual moments. c_huber, c_clip and scale govern that robust configuration. Solver controls are max_sweeps, tol, beta_init, eps and kkt_tol.

Both main classes expose fit(X, y), predict(X), coef_ and intercept_. Keep training and evaluation observations separate.

Comparison and tuning interfaces¶

Object	Role
`QuantileLasso(lam=0.01, taus=(0.5,))`	LAD Lasso at the median or composite quantile regression Lasso
`fit_en_cv`	Elastic net, or Huber elastic net, selected by internal CV
`fit_sken_cv`	CV selection of residual-shape weights at the supplied penalty level
`fit_quantile_lasso_cv`	CV selection for the quantile-Lasso comparator
`diagnostic_summary(model)`	Compact refit/CV KKT and iteration-cap summaries

Consult the source docstrings for full helper signatures and returned diagnostic fields. LP-based quantile fits fail explicitly on optimizer failure and do not expose the coordinate-descent KKT diagnostic.