Set the target

Shared objects

Snippets on this page continue from the objects the quickstart builds with credit_demo(): exp, x, target, X_bg, the solved res and batch, and cal, a fitted monotone calibrator (see the FAQ).

A target is an interval on the model's output: the counterfactual is feasible when the model's raw score lands inside it. Everything else — probability cutoffs, rating ladders, calibrated policies — is a way of constructing that interval. The full semantics are in Targets; this page is the workflow.

Probability and raw

from treecf import Target

Target.probability(op="<=", value=0.04)   # under a 4% PD cutoff (via logit)
Target.probability(range=(0.0, 0.05))     # inside a probability band
Target.raw(op=">=", value=1.5)            # raw margin / regression units
Target.raw(range=(-1.2, 0.5))

Target.probability inverts the model's own sigmoid, so it only exists for models with a sigmoid link; a RandomForestClassifier (identity link over averaged probabilities) takes Target.raw with probability-scale numbers instead, and asking for Target.probability there raises TargetError rather than silently targeting the wrong scale.

Rating ladders: Target.bands

One call, one counterfactual (or certified infeasibility) per grade:

from treecf import Target

ladder = exp.explain(x, target=Target.bands({
    "A": (0.00, 0.01),
    "B": (0.01, 0.03),
    "C": (0.03, 0.07),
}), seed=0)
sorted(ladder)   # ["A", "B", "C"], each a Counterfactual or Infeasible

Calibrated policies

When the deployed decision applies a post-hoc calibrator to the model's probability, the policy lives on the calibrated scale, and Target.probability is the wrong tool — it inverts the model's sigmoid, not the calibrator. Target.calibrated takes any object with the duck-typed calibrator protocol (is_monotone_, interval_inverse; every probcal calibrator conforms):

import treecf

res = exp.explain(x, target=treecf.Target.calibrated(cal, op="<=", value=0.02), seed=0)
res.score_calibrated   # the calibrated read-out at the counterfactual

buffer_logit= shrinks the interval before inversion so a bounded future recalibration cannot invalidate the plan; Target.bands(..., space="calibrated", calibrator=cal) puts a whole masterscale on the calibrated scale. Calibration covers the trap, the protocol, and provenance in certificates.

Next

With the target set, declare what may change and by how much: constrain the search. Or step back to bring your model.