Glossary

The words the documentation uses with a fixed meaning. Credit-risk terms come first with their generic equivalent, because PD models are the archetype the package is built around; the package applies to any binary score.

PD (probability of default). The calibrated probability that an obligor defaults within the horizon. Generic equivalent: the positive-class probability of any binary classifier.

Obligor. One borrower, one row. Generic equivalent: an observation.

Event, event rate. A default (y = 1) and the share of rows that are events. Sample size in this documentation is counted in events, not rows: 500 rows at a 3% event rate carry fifteen events.

Score. The model's output before calibration: a probability in (0, 1), or a raw logit converted with probcal.expit. Every calibrator maps scores to calibrated probabilities.

Calibration set, held-out set. The rows a calibrator is fitted on, and the rows every metric is reported on. They must be disjoint from each other and from the model's training rows; see Data splitting.

Grade. A rating bucket: obligors whose calibrated PD falls in one band of the masterscale. Per-grade backtests compare each grade's realized default count with the PD assigned to it.

Masterscale. The fixed ladder of PD bands that defines the grades, shared across models and reporting periods. Masterscale is the object: it assigns grades (lo <= p < hi, top band closed), hands its bands to every translator, and serializes with a fingerprint. build_masterscale designs one from data; jeffreys_upper_bands builds a conservative band table from posteriors; calibrated_bands_to_raw translates bands into raw-score cutoffs.

Central tendency. The long-run average default rate a portfolio's PDs are required to aggregate to, set by policy rather than by the sample. Reaching it is an offset, not a refit.

Offset. A rigid shift on the logit scale, LogitOffset, applied after the calibration map and kept as its own auditable stage. audit_report() records the mean, the guardrails, a timestamp, and a fingerprint before and after.

Guardrails. calibration_guardrails: the logit-scale calibration slope (accepted in [0.9, 1.1]), the intercept (accepted within ±0.1), and Spiegelhalter's test (accepted at p > 0.05), reported together with an overall all_ok flag.

Traffic light. The per-grade verdict of binomial_grade_test and jeffreys_grade_test: green above p = 0.05, amber above 0.01, red at or below 0.01. The tests are one-sided: a small value means the grade's PD is likely understated.

LDP (low-default portfolio). A portfolio or grade with zero or very few observed defaults, where the naive default rate is uninformative.

Pluto–Tasche. The most-prudent-estimate method for LDPs: an upper confidence bound on each grade's PD that pools the grade with every worse (higher-PD) grade, so that a zero-default grade still receives a positive PD. metrics.pluto_tasche and pluto_tasche_from_arrays.

MoC (margin of conservatism). A deliberate upward shift of PDs to cover estimation uncertainty. monitor.moc_offset derives one from a monitor's confidence sequence for the current offset.

Scorecard points. The integer scale a deployed scorecard reports (for instance PDO = 20, 600 points at 50:1 odds). The optbinning integration keeps points unchanged and calibrates the PD behind them.

Calibrator. A fitted map from score to calibrated probability with the contract fit(s, y, sample_weight=None), predict_proba(s), interpret(). Thirteen are registered; the catalog lists them.

Reliability curve. Observed event rate against predicted probability, drawn on the probability or the logit axis. Binned, kernel, and CORP (isotonic) constructions exist; see Metrics and tests.

Inverse map. The raw score that produces a given calibrated probability. point_inverse is the exact preimage, defined only for strictly monotone maps with a closed form; interval_inverse is the generalized inverse of a calibrated interval, inf{s : g(s) >= lo} to sup{s : g(s) <= hi}, defined for every non-decreasing map, plateaus included. Unattainable targets raise UnattainableTargetError instead of clamping.

Cutoff. A raw-score threshold obtained by inverting a policy PD, so the decision rule can be expressed on the scale the deployed system uses.

Segment. A named subpopulation (a product line, a region) that receives its own shrunken offset on top of a shared base map in SegmentedCalibrator, or its own rows in grouped evaluation.

E-process. A non-negative process with initial value 1 whose expectation under the null stays at most 1 at every look; crossing 1/alpha is an alarm that keeps its type-I error at alpha under optional stopping. The monitor's null is that the deployed forecast is calibrated.

Anytime-valid. A guarantee that holds at every look at the data rather than at one pre-planned sample size. Fixed-sample p-values lose this under repeated checking; e-processes keep it.

Onset. The batch at which a monitored drift is estimated to have begun, as opposed to the batch at which the alarm fired.

Confidence sequence. A sequence of intervals that all contain the true value simultaneously with the stated probability, so it can be read after any batch. The monitor keeps one for the current offset.

Fingerprint. SHA-256 of an object's canonical serialized form, blind to library version and timestamp. Identical fits on identical data share it; it is the provenance id monitors, registries, and recourse engines record.

Golden file. A committed JSON artifact in tests/golden/ that a release must still load and reproduce bit for bit. It is how "every 0.x release reads schema 1" is enforced.

Chain. probcal.Chain: a calibrator followed by offsets, fitted and inverted as one object, so a recourse engine sees a single map.