samesame

Did the target shift? Did it get worse?

Feature-level monitoring can produce alerts that are difficult to interpret, and labels may arrive too late to support early action. samesame compares one interpretable score per observation — predicted risk, prediction error, confidence, or outlier score — between source (the reference) and target (the current deployment). A model score reduces each observation to one interpretable score so each row has one number to monitor.

It separates two questions:

• Any shift? Use ss.test_shift to ask whether the score distribution changed between source and target. This is a two-sided AUC test.
• Harmful shift? Use ss.test_harmful_shift(..., worse="higher"|"lower") to ask whether the target moved in a specified harmful direction. This is a one-sided test based on the weighted AUC.

Quick example

import numpy as np
import samesame as ss

rng = np.random.default_rng(12345)
source_scores = rng.normal(loc=0.0, scale=1.0, size=600)
target_scores = rng.normal(loc=0.6, scale=1.0, size=600)

shift = ss.test_shift(source=source_scores, target=target_scores, rng=rng)
harm = ss.test_harmful_shift(
    source=source_scores,
    target=target_scores,
    worse="higher",  # larger = more harm (e.g., risk)
    rng=rng,
)

print(f"Shift statistic: {shift.statistic:.3f}, p-value: {shift.pvalue:.4f}")
print(f"Harm  statistic: {harm.statistic:.3f}, p-value: {harm.pvalue:.4f}")

Interpret p-values alongside the statistic and score distributions: a small p-value is evidence against label exchangeability — not business impact, causality, or the probability the null is true. Evidence of a shift is not evidence of harm.

Toy scores vs. real scores: The example uses synthetic normal scores for brevity. For real features, build the score with a domain classifier and generate it out of sample - using cross_val_predict, oob_decision_function_, or held-out data - to avoid in-sample bias. See Get started.

Workflow

1. Build one score per observation, generating it out of sample if it comes from a fitted model.
2. Test for any shift with ss.test_shift.
3. Test for harmful shift with ss.test_harmful_shift(..., worse=...).
4. If poor feature overlap is a real concern, use ss.domain_weights to reweight the comparison — weighting changes the population the comparison describes and is not a default correction.

Explore the documentation:

• Get started - learn the workflow and run both tests in five minutes.
• Monitor a credit model - work through risk, confidence, and error monitoring with HELOC data.
• Weight for common support - learn when and how to reweight comparisons.

Read the full documentation.

Installation

python -m pip install samesame

Requires Python 3.12+, numpy, scipy, and scikit-learn.

For supported use cases and limitations, see the full documentation.