op-attr-reg-predict - Pulse Documentation

name: op-attr-reg-predict
description: Per-row score from a bound model (a compose `fits` entry or a loaded `models` file) — no refit; whole model or a terms subset.
kind: operator
category: ATTR
operator: ATTR_REG_PREDICT
type: reference
applies_to: compose, predict
examples_tags: [regression, ols, compose]

Attributes emit row-level scalars; they do not produce Response.Components.

Use when

Adds to every row the value a model fitted once, elsewhere in the request, predicts for it from the row’s own predictor values.

Questions it answers:

How close are the model’s predictions on weeks it was not fitted to (a holdout)?
How much of each week’s sales does the TV term of the fitted model account for?

Use something else:

ATTR_REG_FITTED when the model should be fitted to these same rows inside the slot.

Params

Name	Type	Description
`fit`	string	Bound model name (required).
`terms`	[]string	Score only these terms, `(intercept)` opt-in; fit order. Omit: whole model.
`scale`	enum	GLM only: `response` (default, inverse link) or `link`.
`bound`	enum	`point` (default), `lower`, `upper`: ci bound × column. Needs `terms` = one predictor.

Inputs

Param	Accepted field types
model predictors	numeric slot columns (schema, features, earlier labels)
`label`	output name; default `ATTR_REG_PREDICT_<fit>`

Output

One f64 per record; null when a column the score reads is null. Training rows equal the fit’s own fitted values bit for bit.

Reading the output

value: The target value the named fitted model predicts for the row from the row’s own predictor values: intercept + sum(coefficient * predictor) at the fit’s full precision, in the target’s units — for a GLM on the response scale (the inverse link applied: a probability, a count) unless scale is link. With terms, only the listed terms are summed: that part of the prediction, a contribution in the target’s units (link units for a GLM).
- Caveat: A coefficient’s confidence bound gives the range of that one term, not a prediction interval for a row.

Gotchas

The fit’s predictor list is authoritative: build the same columns (same features) in the slot. An absent one-hot column is the reference level; scoring never infers a level.
A coefficient bound is not a prediction interval. A term with no bound scores null.
Row-local: streams, merges, materializes. Row weight ignored (the fit’s own applied).
Under selection a dropped feature scores 0.
Save, then load: save: "models/m.pulse-model.json" on the fit; a later request declares models: [{name: "m", path: "models/m.pulse-model.json"}] and scores fit: "m" — bit-equal to the in-compose fit, any cohort with the same predictor columns.

See

pulse_examples_search tags=[compose]
Skills: compose-fits, regression-modeling, op-attr-reg-fitted, expression-language (coef())

Keyboard shortcuts