Pulse Documentation

name: op-reg-mod-selection
description: Spec-level subset-selection modifier (forward / backward / stepwise) that composes with an OLS or GLM fit; drives a greedy search over predictors against an information criterion. Non-empty value forces the buffered path.
kind: operator
category: REG
operator: REG_SELECTION
type: reference
applies_to: process, compose, predict
examples_tags: [regression, selection, stepwise, buffered-pipeline]

Regression operators emit coefficient + diagnostics; no Response.Components. This modifier shrinks the active predictor set in Response.Regressions[i].SelectedFeatures and drops non-selected entries from Coefficients / StdErrors.

Params

Top-level selection on RegressionSpec plus its required criterion companion. A wrapper, not a fit: the engine refits the host model against candidate subsets and keeps the lowest-criterion winner.

• selection — enum, default "". "", forward (intercept-only, add best each step), backward (full, drop best each step), stepwise (bidirectional add/drop until no move improves).
• criterion — enum, default required when set. aic or bic.

Inputs

Inherits target / predictors from the host RegressionSpec. No additional field inputs.

Output

• SelectedFeatures []string — chosen predictors in fit order.
• Coefficients for selected predictors plus (intercept) only. Absent ≠ zero — non-selected predictors were dropped entirely.
• StdErrors / PValues for selected predictors only. Selection / Criterion echo back.

Gotchas

• Any non-empty selection forces buffered — RegressionSpec.Streamable() → false. Predict reports the downgrade.
• criterion missing when selection set → SERVICE_VALIDATION. bic penalizes complexity harder than aic.
• Worst-case O(p²) fits (stepwise); on wide tables prefer forward.
• Composes with resample — the resample wraps the selected-feature fit.
• Selection inflates type-I error on retained predictors; treat the model as exploratory.
• Never weighted: any row weight in force on the regression (slot, request or Options.DefaultWeight) is PULSE_WEIGHT_UNSUPPORTED with details.reason; "weight": null on the slot runs it unweighted.

See

• pulse_examples_search tags=[stepwise]
• Skills: regression-modeling, op-reg-ols, op-reg-glm, op-reg-mod-resample