API: calibrators¶

base ¶

BaseCalibrator: the common fit / predict_proba / interpret contract.

UnattainableTargetError ¶

Bases: ValueError

The requested calibrated target is unattainable.

Raised instead of silently clamping when an interval does not intersect the calibrator's output range (or was emptied by buffer_logit), when a point-inverse target lies outside the open interval (0, 1), or when a probability-space point-inverse result would round to 0.0/1.0 (raw logit beyond logit(1 - 1e-12)).

BaseCalibrator ¶

Bases: JsonIO, ABC

Common contract for all probcal calibrators.

Subclasses implement _fit (estimation on validated arrays), _predict (the fitted map on clipped scores), and interpret. Everything else — validation, sklearn-style parameter handling without an sklearn import, the 2-D probability helper — lives here.

ATTRIBUTE	DESCRIPTION
`is_monotone_`	Whether the fitted map is guaranteed non-decreasing. Class-level default `True`; non-monotone calibrators (e.g. ENIR) override it. TYPE: `bool`
`fitted_`	Set by :meth:`fit`. TYPE: `bool`

affine_logit_coeffs_ `property` ¶

affine_logit_coeffs_: tuple[float, float] | None

Coefficients (a, b) of logit g(s) = a * logit(s) + b, if affine.

None for calibrators that are not affine on the logit scale. Consumed by the attribution adjustment.

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank for selector tie-breaks; lower wins a tie.

Default 100.0 means "unknown — override in subclasses". Custom calibrators declare their place in the tie-break by overriding this property.

fit ¶

fit(s: object, y: object, sample_weight: object = None) -> Self

Fit the calibration map on scores and binary outcomes.

PARAMETER	DESCRIPTION
`s`	Raw scores/probabilities in `[0, 1]` (clipped to `[1e-12, 1 - 1e-12]`). Users holding raw logits convert with :func:`probcal.expit` first. TYPE: `array_like`
`y`	Binary outcomes in `{0, 1}`; both classes must be present. TYPE: `array_like`
`sample_weight`	Positive observation weights. TYPE: `array_like or None` DEFAULT: `None`

RETURNS	DESCRIPTION
`Self`	The fitted calibrator.

Source code in src/probcal/base.py

def fit(self, s: object, y: object, sample_weight: object = None) -> Self:
    """Fit the calibration map on scores and binary outcomes.

    Parameters
    ----------
    s : array_like
        Raw scores/probabilities in ``[0, 1]`` (clipped to
        ``[1e-12, 1 - 1e-12]``). Users holding raw logits convert with
        :func:`probcal.expit` first.
    y : array_like
        Binary outcomes in ``{0, 1}``; both classes must be present.
    sample_weight : array_like or None
        Positive observation weights.

    Returns
    -------
    Self
        The fitted calibrator.
    """
    s_arr = validate_scores(s)
    y_arr = validate_binary_y(y)
    if s_arr.shape[0] != y_arr.shape[0]:
        raise ValueError(
            f"s and y must have equal length, got {s_arr.shape[0]} and {y_arr.shape[0]}"
        )
    w_arr = validate_weights(sample_weight, s_arr.shape[0])
    self._fit(s_arr, y_arr, w_arr)
    self.fit_meta_ = {
        "n_obs": int(s_arr.shape[0]),
        "n_events": float(np.sum(w_arr * y_arr)),
        "weight_sum": float(w_arr.sum()),
        "fitted_at_utc": datetime.now(UTC).isoformat(timespec="seconds"),
        "data_fingerprint": data_fingerprint(s_arr, y_arr, w_arr),
    }
    self.fitted_ = True
    return self

predict_proba ¶

predict_proba(s: object) -> ndarray

Calibrated probabilities P(y = 1) for new scores.

PARAMETER	DESCRIPTION
`s`	Raw scores/probabilities in `[0, 1]`. TYPE: `array_like`

RETURNS	DESCRIPTION
`numpy.ndarray of shape (n,)`	Calibrated probabilities.

Source code in src/probcal/base.py

def predict_proba(self, s: object) -> np.ndarray:
    """Calibrated probabilities ``P(y = 1)`` for new scores.

    Parameters
    ----------
    s : array_like
        Raw scores/probabilities in ``[0, 1]``.

    Returns
    -------
    numpy.ndarray of shape (n,)
        Calibrated probabilities.
    """
    self._check_fitted()
    return self._predict(validate_scores(s))

__sklearn_is_fitted__ ¶

__sklearn_is_fitted__() -> bool

Fitted state for sklearn's check_is_fitted (sklearn >= 1.6).

RETURNS	DESCRIPTION
`bool`	`True` once :meth:`fit` has run.

Source code in src/probcal/base.py

def __sklearn_is_fitted__(self) -> bool:
    """Fitted state for sklearn's ``check_is_fitted`` (sklearn >= 1.6).

    Returns
    -------
    bool
        ``True`` once :meth:`fit` has run.
    """
    return bool(self.fitted_)

__sklearn_tags__ ¶

__sklearn_tags__() -> Tags

Estimator tags for sklearn >= 1.6, built from the public constructors.

sklearn is imported inside the body, never at module or class level: import probcal stays numpy-only and this hook costs nothing until sklearn itself calls it. Only the fields that are actually true of a calibrator are set — it takes 1-D scores, not a 2-D feature matrix, it is neither a classifier nor a regressor, and it must be fitted.

RETURNS	DESCRIPTION
`Tags`	The tag object sklearn's `get_tags` expects.

RAISES	DESCRIPTION
`ImportError`	If sklearn is not installed (only reachable by calling the hook by hand).

Source code in src/probcal/base.py

def __sklearn_tags__(self) -> "Tags":
    """Estimator tags for sklearn >= 1.6, built from the public constructors.

    sklearn is imported inside the body, never at module or class level:
    ``import probcal`` stays numpy-only and this hook costs nothing until
    sklearn itself calls it. Only the fields that are actually true of a
    calibrator are set — it takes 1-D scores, not a 2-D feature matrix,
    it is neither a classifier nor a regressor, and it must be fitted.

    Returns
    -------
    sklearn.utils.Tags
        The tag object sklearn's ``get_tags`` expects.

    Raises
    ------
    ImportError
        If sklearn is not installed (only reachable by calling the hook
        by hand).
    """
    from sklearn.utils import InputTags, Tags, TargetTags

    return Tags(
        estimator_type=None,
        target_tags=TargetTags(required=True),
        requires_fit=True,
        input_tags=InputTags(two_d_array=False),
    )

interpret `abstractmethod` ¶

interpret() -> Interpretation

Fitted parameters with a plain-language, domain-aware reading.

Source code in src/probcal/base.py

@abstractmethod
def interpret(self) -> Interpretation:
    """Fitted parameters with a plain-language, domain-aware reading."""

interval_inverse ¶

interval_inverse(lo: float, hi: float, *, space: str = 'probability', buffer_logit: float = 0.0) -> tuple[float, float]

Generalized-inverse preimage (raw_lo, raw_hi) of a calibrated interval.

For a non-decreasing fitted map g: raw_lo = inf{s : g(s) >= lo} and raw_hi = sup{s : g(s) <= hi}.

PARAMETER	DESCRIPTION
`lo`	Calibrated-probability bounds; `lo=0`/`hi=1` map to the full raw range (−inf/+inf on the logit scale). TYPE: `float`
`hi`	Calibrated-probability bounds; `lo=0`/`hi=1` map to the full raw range (−inf/+inf on the logit scale). TYPE: `float`
`space`	Scale of the returned raw bounds. `"logit"` is what a SIGMOID-link raw-margin consumer (e.g. a counterfactual engine's `Target.raw`) expects. TYPE: `(probability, logit)` DEFAULT: `"probability"`
`buffer_logit`	Shrink the calibrated interval by this margin in logit space before inverting — robustness against future recalibration drift (a central-tendency update of magnitude <= buffer cannot invalidate the result). TYPE: `float` DEFAULT: `0.0`

RETURNS	DESCRIPTION
`tuple of float`	`(raw_lo, raw_hi)` preimage bounds, on the scale requested by `space`.

RAISES	DESCRIPTION
`UnattainableTargetError`	If the (buffered) interval does not intersect the output range — never silently clamped.
`NotImplementedError`	For non-monotone calibrators (`is_monotone_ = False`), whose preimage may be a union of intervals.

Source code in src/probcal/base.py

def interval_inverse(
    self,
    lo: float,
    hi: float,
    *,
    space: str = "probability",
    buffer_logit: float = 0.0,
) -> tuple[float, float]:
    """Generalized-inverse preimage ``(raw_lo, raw_hi)`` of a calibrated interval.

    For a non-decreasing fitted map ``g``:
    ``raw_lo = inf{s : g(s) >= lo}`` and ``raw_hi = sup{s : g(s) <= hi}``.

    Parameters
    ----------
    lo, hi : float
        Calibrated-probability bounds; ``lo=0``/``hi=1`` map to the full
        raw range (−inf/+inf on the logit scale).
    space : {"probability", "logit"}
        Scale of the returned raw bounds. ``"logit"`` is what a
        SIGMOID-link raw-margin consumer (e.g. a counterfactual engine's
        ``Target.raw``) expects.
    buffer_logit : float
        Shrink the calibrated interval by this margin in logit space
        *before* inverting — robustness against future recalibration
        drift (a central-tendency update of magnitude <= buffer cannot
        invalidate the result).

    Returns
    -------
    tuple of float
        ``(raw_lo, raw_hi)`` preimage bounds, on the scale requested by
        ``space``.

    Raises
    ------
    UnattainableTargetError
        If the (buffered) interval does not intersect the output range —
        never silently clamped.
    NotImplementedError
        For non-monotone calibrators (``is_monotone_ = False``), whose
        preimage may be a union of intervals.
    """
    self._check_fitted()
    self._check_monotone()
    validate_interval(lo, hi)
    validate_space(space)
    lo_b, hi_b = shrink_interval(lo, hi, buffer_logit)
    gmin, gmax = self._output_range()
    if lo_b > gmax or hi_b < gmin:
        raise UnattainableTargetError(
            f"calibrated target [{lo_b:.6g}, {hi_b:.6g}] does not intersect the "
            f"calibrator's output range [{gmin:.6g}, {gmax:.6g}]"
        )
    raw_lo = 0.0 if lo_b <= gmin else float(self._inverse_left(lo_b))
    raw_hi = 1.0 if hi_b >= gmax else float(self._inverse_right(hi_b))
    if space == "logit":
        lo_out = -np.inf if raw_lo <= 0.0 else logit1(raw_lo)
        hi_out = np.inf if raw_hi >= 1.0 else logit1(raw_hi)
        return lo_out, hi_out
    return raw_lo, raw_hi

point_inverse ¶

point_inverse(p: object, *, space: str = 'probability') -> ndarray

Raw scores whose calibrated probabilities equal p (exact preimage).

Defined only for strictly monotone calibrators with an exact inverse: affine-logit maps (logit g(s) = a * logit(s) + b) invert in closed form here, covering Platt scaling, temperature scaling, and the tied Beta variants ("a", "ab"); BetaCalibrator overrides this method with its own exact construction for the full "abm" variant. Others raise NotImplementedError and should use :meth:interval_inverse instead.

PARAMETER	DESCRIPTION
`p`	Calibrated probabilities strictly inside `(0, 1)`; boundary and out-of-range values raise `UnattainableTargetError` (all-or-nothing, no silent clamp). TYPE: `array_like`
`space`	Scale of the returned raw values. `"logit"` returns the raw logit directly; `"probability"` (default) returns the raw score. TYPE: `(probability, logit)` DEFAULT: `"probability"`

RETURNS	DESCRIPTION
`ndarray`	Raw scores (or logits, if `space="logit"`) whose calibrated probability equals `p`.

RAISES	DESCRIPTION
`RuntimeError`	If not yet fitted.
`ValueError`	If `space` is not `"probability"` or `"logit"`.
`NotImplementedError`	If the calibrator is not monotone (`is_monotone_ = False`), or has no affine-logit closed form (`affine_logit_coeffs_` is `None`).
`UnattainableTargetError`	If any element of `p` lies outside the open interval `(0, 1)`; or if `space="probability"` and the raw logit of any result exceeds `logit(1 - 1e-12)` in magnitude — the probability representation would round to 0.0/1.0 and silently fail to round-trip; `space="logit"` is exact there.

Source code in src/probcal/base.py

def point_inverse(self, p: object, *, space: str = "probability") -> np.ndarray:
    """Raw scores whose calibrated probabilities equal ``p`` (exact preimage).

    Defined only for strictly monotone calibrators with an exact
    inverse: affine-logit maps (``logit g(s) = a * logit(s) + b``)
    invert in closed form here, covering Platt scaling, temperature
    scaling, and the tied Beta variants (``"a"``, ``"ab"``);
    ``BetaCalibrator`` overrides this method with its own exact
    construction for the full ``"abm"`` variant. Others
    raise ``NotImplementedError`` and should use :meth:`interval_inverse`
    instead.

    Parameters
    ----------
    p : array_like
        Calibrated probabilities strictly inside ``(0, 1)``; boundary
        and out-of-range values raise ``UnattainableTargetError``
        (all-or-nothing, no silent clamp).
    space : {"probability", "logit"}, keyword-only
        Scale of the returned raw values. ``"logit"`` returns the raw
        logit directly; ``"probability"`` (default) returns the raw
        score.

    Returns
    -------
    numpy.ndarray
        Raw scores (or logits, if ``space="logit"``) whose calibrated
        probability equals ``p``.

    Raises
    ------
    RuntimeError
        If not yet fitted.
    ValueError
        If ``space`` is not ``"probability"`` or ``"logit"``.
    NotImplementedError
        If the calibrator is not monotone (``is_monotone_ = False``), or
        has no affine-logit closed form (``affine_logit_coeffs_`` is
        ``None``).
    UnattainableTargetError
        If any element of ``p`` lies outside the open interval
        ``(0, 1)``; or if ``space="probability"`` and the raw logit of
        any result exceeds ``logit(1 - 1e-12)`` in magnitude — the
        probability representation would round to 0.0/1.0 and silently
        fail to round-trip; ``space="logit"`` is exact there.
    """
    self._check_fitted()
    self._check_monotone()
    validate_space(space)
    arr = _validate_point_targets(p)
    z = self._point_inverse_logit(arr)
    _check_representable(z, space)
    return z if space == "logit" else expit(z)

get_params ¶

get_params(deep: bool = True) -> dict[str, object]

Constructor parameters as a dict (manual sklearn-compatible clone info).

Source code in src/probcal/base.py

def get_params(self, deep: bool = True) -> dict[str, object]:
    """Constructor parameters as a dict (manual sklearn-compatible clone info)."""
    sig = inspect.signature(type(self).__init__)
    return {
        name: getattr(self, name)
        for name in sig.parameters
        if name not in ("self", "args", "kwargs")
    }

set_params ¶

set_params(**params: object) -> Self

Set constructor parameters; unknown names raise ValueError.

Source code in src/probcal/base.py

def set_params(self, **params: object) -> Self:
    """Set constructor parameters; unknown names raise ``ValueError``."""
    valid = self.get_params()
    for key, value in params.items():
        if key not in valid:
            raise ValueError(
                f"unknown parameter {key!r} for {type(self).__name__}; valid: {sorted(valid)}"
            )
        setattr(self, key, value)
    return self

to_dict ¶

to_dict() -> dict[str, object]

Versioned JSON-native snapshot of the fitted object.

RETURNS	DESCRIPTION
`dict`	`{"probcal_schema", "probcal_version", "class", "params", "state", "fit_meta"}`. `fit_meta` records `n_obs`, `n_events`, `weight_sum`, `fitted_at_utc` (ISO 8601), the `data_fingerprint` (SHA-256 of the sorted `(s, y, w)` triple), and convergence flags where they exist.

RAISES	DESCRIPTION
`RuntimeError`	If not yet fitted.

Source code in src/probcal/base.py

def to_dict(self) -> dict[str, object]:
    """Versioned JSON-native snapshot of the fitted object.

    Returns
    -------
    dict
        ``{"probcal_schema", "probcal_version", "class", "params",
        "state", "fit_meta"}``. ``fit_meta`` records ``n_obs``,
        ``n_events``, ``weight_sum``, ``fitted_at_utc`` (ISO 8601), the
        ``data_fingerprint`` (SHA-256 of the sorted ``(s, y, w)``
        triple), and convergence flags where they exist.

    Raises
    ------
    RuntimeError
        If not yet fitted.
    """
    self._check_fitted()
    fit_meta = dict(getattr(self, "fit_meta_", {}))
    for flag in ("converged_", "separation_fallback_"):
        if hasattr(self, flag):
            fit_meta[flag] = bool(getattr(self, flag))
    return envelope(
        self,
        params=encode_value(self._params_for_dict()),
        state=self._state(),
        fit_meta=encode_value(fit_meta),
    )

from_dict `classmethod` ¶

from_dict(d: dict) -> BaseCalibrator

Rebuild a fitted object from :meth:to_dict output.

Called on :class:BaseCalibrator itself, dispatches through the class registry to whatever class wrote d; called on a subclass, requires d["class"] to match that subclass.

PARAMETER	DESCRIPTION
`d`	Output of :meth:`to_dict` (parsed JSON). TYPE: `dict`

RETURNS	DESCRIPTION
`BaseCalibrator`	A fitted instance.

RAISES	DESCRIPTION
`ValueError`	If the schema version is unknown (naming the writing version), the class is not registered, or `d["class"]` does not match the subclass this was called on.

Source code in src/probcal/base.py

@classmethod
def from_dict(cls, d: dict) -> "BaseCalibrator":
    """Rebuild a fitted object from :meth:`to_dict` output.

    Called on :class:`BaseCalibrator` itself, dispatches through the
    class registry to whatever class wrote ``d``; called on a subclass,
    requires ``d["class"]`` to match that subclass.

    Parameters
    ----------
    d : dict
        Output of :meth:`to_dict` (parsed JSON).

    Returns
    -------
    BaseCalibrator
        A fitted instance.

    Raises
    ------
    ValueError
        If the schema version is unknown (naming the writing version),
        the class is not registered, or ``d["class"]`` does not match
        the subclass this was called on.
    """
    if cls is BaseCalibrator:
        from ._registry import load

        return load(d)  # type: ignore[return-value]
    check_payload(cls, d)
    params = cls._params_from_dict(
        decode_value(d.get("params", {}), arrays=False)  # type: ignore[arg-type]
    )
    obj = cls(**params)
    obj._set_state(dict(d.get("state", {})))
    obj.fit_meta_ = dict(decode_value(d.get("fit_meta", {}), arrays=False))  # type: ignore[call-overload]
    obj.fitted_ = True
    return obj

shrink_interval ¶

shrink_interval(lo: float, hi: float, buffer_logit: float) -> tuple[float, float]

Shrink [lo, hi] by buffer_logit on the logit scale (0/1 ends stay put).

Shared by every interval_inverse: robustness against future recalibration drift of magnitude <= buffer_logit.

RAISES	DESCRIPTION
`ValueError`	If `buffer_logit` is negative or not finite.
`UnattainableTargetError`	If the buffer empties the interval.

Source code in src/probcal/base.py

def shrink_interval(lo: float, hi: float, buffer_logit: float) -> tuple[float, float]:
    """Shrink ``[lo, hi]`` by ``buffer_logit`` on the logit scale (0/1 ends stay put).

    Shared by every ``interval_inverse``: robustness against future
    recalibration drift of magnitude ``<= buffer_logit``.

    Raises
    ------
    ValueError
        If ``buffer_logit`` is negative or not finite.
    UnattainableTargetError
        If the buffer empties the interval.
    """
    buffer_logit = float(buffer_logit)
    if not buffer_logit >= 0.0 or buffer_logit == float("inf"):
        raise ValueError(f"buffer_logit must be a finite value >= 0, got {buffer_logit!r}")
    lo_b, hi_b = float(lo), float(hi)
    if buffer_logit > 0.0:
        if lo > 0.0:
            lo_b = expit1(logit1(lo) + buffer_logit)
        if hi < 1.0:
            hi_b = expit1(logit1(hi) - buffer_logit)
        if lo_b > hi_b:
            raise UnattainableTargetError(
                f"buffer_logit={buffer_logit} empties the calibrated interval [{lo}, {hi}]"
            )
    return lo_b, hi_b

validate_interval ¶

validate_interval(lo: float, hi: float) -> None

0 <= lo <= hi <= 1.

Source code in src/probcal/base.py

def validate_interval(lo: float, hi: float) -> None:
    """``0 <= lo <= hi <= 1``."""
    if not 0.0 <= lo <= hi <= 1.0:
        raise ValueError(f"need 0 <= lo <= hi <= 1, got lo={lo}, hi={hi}")

clone_unfitted ¶

clone_unfitted(obj: object) -> object

A fresh, unfitted copy built from obj.get_params(deep=False).

Parameters that are themselves estimators (anything with get_params) — directly or inside a list, tuple, or dict, such as a :class:probcal.Chain's stages or a selector's candidates — are cloned recursively, so the copy never shares a sub-object with obj.

Source code in src/probcal/base.py

def clone_unfitted(obj: object) -> object:
    """A fresh, unfitted copy built from ``obj.get_params(deep=False)``.

    Parameters that are themselves estimators (anything with
    ``get_params``) — directly or inside a list, tuple, or dict, such as a
    :class:`probcal.Chain`'s stages or a selector's candidates — are cloned
    recursively, so the copy never shares a sub-object with ``obj``.
    """
    try:
        params = obj.get_params(deep=False)  # type: ignore[attr-defined]
    except TypeError:
        params = obj.get_params()  # type: ignore[attr-defined]
    return type(obj)(**{k: _clone_param(v) for k, v in params.items()})

parametric ¶

Parametric calibrators: Platt, temperature, and beta calibration.

Theory, derivations, and parameter interpretation: docs/concepts/methods-parametric.md.

References

Platt (1999); Lin, Lin & Weng (2007); Guo et al. (2017); Kull, Silva Filho & Flach (2017, AISTATS and EJS) — full records in the documentation.

PlattCalibrator ¶

Bases: BaseCalibrator

Logistic recalibration on the logit scale (Platt scaling).

Fits logit g(s) = a * logit(s) + b by IRLS with Lin–Lin–Weng smoothed targets (N+ + 1)/(N+ + 2) and 1/(N- + 2) for stability on small samples, where N+/N- are the weighted class masses (row counts under unit weights), so that integer weights match row duplication. The identity map is (a, b) = (1, 0).

ATTRIBUTE	DESCRIPTION
`a_`	Fitted slope — spread correction: `a < 1` shrinks overconfident scores toward the base rate, `a > 1` sharpens underconfident ones. TYPE: `float`
`b_`	Fitted intercept — calibration-in-the-large shift in log-odds. TYPE: `float`
`converged_`	Whether IRLS converged; if `False` a warning was raised at fit time and `interpret()` records it. TYPE: `bool`

References

Platt (1999); Lin, Lin & Weng (2007). The logistic family fitted on raw SVM outputs (Platt's original setting) does not contain the identity; on logits it does — see the parametric-methods chapter.

affine_logit_coeffs_ `property` ¶

affine_logit_coeffs_: tuple[float, float] | None

(a, b): Platt scaling is affine on the logit scale.

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 2.0: a two-parameter map, simpler than the nonparametric methods.

interpret ¶

interpret() -> Interpretation

Read the fitted slope and intercept against the identity (1, 0).

If IRLS did not converge at fit time (a UserWarning was raised), the messages include a note not to trust the coefficients.

Source code in src/probcal/parametric.py

def interpret(self) -> Interpretation:
    """Read the fitted slope and intercept against the identity ``(1, 0)``.

    If IRLS did not converge at fit time (a ``UserWarning`` was raised),
    the messages include a note not to trust the coefficients.
    """
    self._check_fitted()
    if self.a_ < 1.0:
        slope_msg = (
            f"slope a = {self.a_:.3f} < 1: scores were overconfident (too spread out); "
            "predictions are shrunk toward the base rate"
        )
    else:
        slope_msg = (
            f"slope a = {self.a_:.3f} >= 1: scores were underconfident (too flat); "
            "predictions are sharpened"
        )
    int_msg = (
        f"intercept b = {self.b_:.3f}: base-rate (calibration-in-the-large) shift of "
        f"{self.b_:+.3f} log-odds, odds factor {np.exp(self.b_):.3f}"
    )
    messages = [slope_msg, int_msg, "identity map corresponds to (a, b) = (1, 0)"]
    if not self.converged_:
        messages.append("IRLS did not converge; coefficients may be unreliable")
    return Interpretation(
        method=type(self).__name__,
        param_names=("a", "b"),
        param_values=(self.a_, self.b_),
        messages=tuple(messages),
    )

TemperatureCalibrator ¶

Bases: BaseCalibrator

Temperature scaling: g(s) = sigma(logit(s) / T).

T minimizes the calibration-set negative log-likelihood: the score in u = 1/T is monotone, so its root is found by bisection on u in [1e-6, 1e6]; without an interior minimum the nearer boundary is used and a UserWarning is raised.

ATTRIBUTE	DESCRIPTION
`T_`	Fitted temperature. `T > 1`: the model was overconfident and is softened; `T < 1`: underconfident and sharpened. Temperature cannot fix base-rate error — `s = 0.5` is a fixed point for every `T`; use Platt scaling or `LogitOffset` for level shifts. TYPE: `float`

References

Guo, Pleiss, Sun & Weinberger (2017).

affine_logit_coeffs_ `property` ¶

affine_logit_coeffs_: tuple[float, float] | None

(1/T, 0): temperature scaling is affine on the logit scale.

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 1.0: the simplest map, a single parameter.

interpret ¶

interpret() -> Interpretation

Read the fitted temperature against the identity T = 1.

Source code in src/probcal/parametric.py

def interpret(self) -> Interpretation:
    """Read the fitted temperature against the identity ``T = 1``."""
    self._check_fitted()
    if self.T_ > 1.0:
        msg = (
            f"T = {self.T_:.3f} > 1: the model was overconfident; logits are divided "
            "by T (softening toward 1/2)"
        )
    else:
        msg = (
            f"T = {self.T_:.3f} <= 1: the model was underconfident; logits are divided "
            "by T (sharpening away from 1/2)"
        )
    return Interpretation(
        method=type(self).__name__,
        param_names=("T",),
        param_values=(self.T_,),
        messages=(
            msg,
            "temperature cannot fix base-rate error: s = 0.5 maps to 0.5 for every T "
            "(use PlattCalibrator or LogitOffset for level shifts)",
        ),
    )

BetaCalibrator ¶

BetaCalibrator(variant: str = 'abm')

Bases: BaseCalibrator

Beta calibration: logit g(s) = a·ln s − b·ln(1 − s) + c.

Variants: "abm" fits (a, b, c); "ab" ties a = b (equivalent to Platt scaling on logits); "a" additionally fixes c = 0 (a single-parameter map, the temperature family in a different parameterization). The monotonicity constraint a, b >= 0 is enforced by the betacal refit strategy: a negative exponent drops its feature and refits. :meth:point_inverse is exact for every variant, including the non-affine "abm" (see :meth:_point_inverse_logit).

PARAMETER	DESCRIPTION
`variant`	`"abm"` (default) fits the full `(a, b, c)`; `"ab"` ties `a = b`; `"a"` additionally fixes `c = 0`. TYPE: `(abm, ab, a)` DEFAULT: `"abm"`

ATTRIBUTE	DESCRIPTION
`a_`	Sensitivity near `s -> 0` — governs the low-probability tail (critical for low-PD credit portfolios). TYPE: `float`
`b_`	Sensitivity near `s -> 1`. TYPE: `float`
`c_`	Base-rate shift in log-odds. TYPE: `float`
`constraint_active_`	Whether the `a, b >= 0` constraint forced a refit. TYPE: `bool`
`converged_`	Whether the fit whose coefficients survive converged (`True` for closed-form paths); if `False` a warning was raised at fit time and `interpret()` records it. TYPE: `bool`
`separation_fallback_`	Whether any IRLS call during fitting detected separation and fell back to the ridge-regularized fit; recorded by `interpret()`. TYPE: `bool`

References

Kull, Silva Filho & Flach (2017), AISTATS 54 and EJS 11(2). The identity is (a, b, c) = (1, 1, 0): beta calibration cannot un-calibrate an already calibrated model. a != b captures asymmetric tail distortion; temperature is the special case a = b = 1/T, c = 0.

Source code in src/probcal/parametric.py

def __init__(self, variant: str = "abm") -> None:
    self.variant = variant

affine_logit_coeffs_ `property` ¶

affine_logit_coeffs_: tuple[float, float] | None

(a, c) for the tied variants; None for "abm".

With a = b the map reduces to logit g = a * logit(s) + c, which is affine on the logit scale; the full three-parameter map is not (see the shap-calibration chapter).

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank by variant: 1.5 ("a"), 2.5 ("ab"), 3.0 ("abm").

.get with a fallback because variant is validated only in _fit; the property must not raise pre-fit.

interpret ¶

interpret() -> Interpretation

Read the fitted exponents and intercept against the identity (1, 1, 0).

If IRLS did not converge at fit time (a UserWarning was raised), the messages include a note not to trust the coefficients.

Source code in src/probcal/parametric.py

def interpret(self) -> Interpretation:
    """Read the fitted exponents and intercept against the identity (1, 1, 0).

    If IRLS did not converge at fit time (a ``UserWarning`` was raised),
    the messages include a note not to trust the coefficients.
    """
    self._check_fitted()
    messages = [
        (
            f"a = {self.a_:.3f}: sensitivity near s -> 0; a < 1 raises the smallest "
            "probabilities (model was overconfident in the low tail), a > 1 deepens them"
        ),
        (f"b = {self.b_:.3f}: sensitivity near s -> 1; the mirrored reading for the high tail"),
        (
            f"c = {self.c_:.3f}: base-rate shift of {self.c_:+.3f} log-odds, odds factor "
            f"{np.exp(self.c_):.3f}"
        ),
        "identity map corresponds to (a, b, c) = (1, 1, 0)",
    ]
    if abs(self.a_ - self.b_) > 0.1:
        messages.append(
            f"a != b (gap {self.a_ - self.b_:+.3f}): asymmetric tail distortion that no "
            "symmetric (Platt/temperature) map could express"
        )
    if self.constraint_active_:
        messages.append(
            "monotonicity constraint a, b >= 0 was active: a negative exponent was "
            "dropped and the model refitted (betacal strategy)"
        )
    if not self.converged_:
        messages.append("IRLS did not converge; coefficients may be unreliable")
    if self.separation_fallback_:
        messages.append(
            "separation was detected during fitting; at least one fit fell back to "
            "the ridge-regularized solution (ridge=1e-6)"
        )
    return Interpretation(
        method=type(self).__name__,
        param_names=("a", "b", "c"),
        param_values=(self.a_, self.b_, self.c_),
        messages=tuple(messages),
    )

isotonic ¶

Isotonic calibrators: PAVA-based isotonic and centered isotonic regression (CIR).

Theory and worked example: docs/concepts/methods-nonparametric.md.

References

Barlow, Bartholomew, Bremner & Brunk (1972); Zadrozny & Elkan (2002); Oron & Flournoy (2017) — full records in the documentation.

IsotonicCalibrator ¶

IsotonicCalibrator(interpolation: str = 'none')

Bases: BaseCalibrator

Isotonic calibration: the PAVA step function.

Fits the least-squares non-decreasing map of outcomes on scores. The fitted map is a right-continuous step function with one level per pooled block; scores outside the calibration range clamp to the first/last level. interpolation="linear" instead joins block midpoints, removing the discontinuities.

PARAMETER	DESCRIPTION
`interpolation`	`"none"` (default) keeps the raw step function; `"linear"` joins the block midpoints, removing the tied-prediction plateaus. TYPE: `(none, linear)` DEFAULT: `"none"`

ATTRIBUTE	DESCRIPTION
`n_blocks_`	Number of pooled blocks — the effective complexity estimated from the data. TYPE: `int`
`block_mean_`	Event rate of each pooled block (the step levels). TYPE: `ndarray`
`block_first_s_, block_last_s_`	Score range covered by each block. TYPE: `ndarray`
`block_center_s_`	Weight-centered score coordinate of each block (used by CIR). TYPE: `ndarray`

References

Barlow et al. (1972) for PAVA; Zadrozny & Elkan (2002) for its use in classifier calibration.

Source code in src/probcal/isotonic.py

def __init__(self, interpolation: str = "none") -> None:
    self.interpolation = interpolation

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 50.0: nonparametric, data-driven block count.

interpret ¶

interpret() -> Interpretation

Read the block structure as effective complexity and local event rates.

Source code in src/probcal/isotonic.py

def interpret(self) -> Interpretation:
    """Read the block structure as effective complexity and local event rates."""
    self._check_fitted()
    flats = int(np.sum(np.isclose(np.diff(self.block_mean_), 0.0)))
    messages = [
        (
            f"{self.n_blocks_} pooled blocks: each step level is the empirical event "
            "rate of a score region the data could not subdivide further"
        ),
        (
            f"block count = effective complexity actually estimated from the data "
            f"(range of levels: {self.block_mean_[0]:.4g} to {self.block_mean_[-1]:.4g})"
        ),
    ]
    if flats:
        messages.append(
            f"{flats} adjacent block pairs share a level: expect tied predictions there"
        )
    messages.append(
        "output range is limited to the span of block levels; targets outside it are "
        "unattainable (relevant for interval_inverse)"
    )
    return Interpretation(
        method=type(self).__name__,
        param_names=("n_blocks",),
        param_values=(float(self.n_blocks_),),
        messages=tuple(messages),
    )

CenteredIsotonicCalibrator ¶

CenteredIsotonicCalibrator()

Bases: IsotonicCalibrator

Centered isotonic regression (CIR): strictly increasing where data permit.

Post-processes the PAVA solution by collapsing each block to its weight-centered score coordinate and interpolating linearly through the points (Oron & Flournoy, 2017). Removes the step function's tied predictions — preferred when downstream ranking must be strict.

ATTRIBUTE	DESCRIPTION
`n_blocks_`	Number of pooled blocks — the effective complexity estimated from the data. Inherited from the PAVA fit. TYPE: `int`
`block_mean_`	Event rate of each pooled block (the interpolation y-values). TYPE: `ndarray`
`block_first_s_, block_last_s_`	Score range covered by each block (inherited; not used for prediction, which interpolates through `block_center_s_` instead). TYPE: `ndarray`
`block_center_s_`	Weight-centered score coordinate of each block — the interpolation x-values that make CIR strictly increasing. TYPE: `ndarray`

References

Oron & Flournoy (2017).

Source code in src/probcal/isotonic.py

def __init__(self) -> None:
    super().__init__(interpolation="linear")

interpret ¶

interpret() -> Interpretation

Isotonic reading plus the strictness property CIR adds.

Source code in src/probcal/isotonic.py

def interpret(self) -> Interpretation:
    """Isotonic reading plus the strictness property CIR adds."""
    base = super().interpret()
    return Interpretation(
        method=type(self).__name__,
        param_names=base.param_names,
        param_values=base.param_values,
        messages=base.messages
        + (
            "centered isotonic interpolation is strictly increasing wherever block "
            "levels differ: distinct scores keep distinct predictions (no ties)",
        ),
    )

binning ¶

Binning calibrators: histogram binning and scaling-binning.

Theory: docs/concepts/methods-nonparametric.md.

References

Zadrozny & Elkan (2001); Kumar, Liang & Ma (2019) — full records in the documentation.

HistogramBinningCalibrator ¶

HistogramBinningCalibrator(n_bins: int = 10, strategy: str = 'mass', shrinkage: str | None = 'jeffreys')

Bases: BaseCalibrator

Histogram binning: per-bin event rates with optional Jeffreys shrinkage.

PARAMETER	DESCRIPTION
`n_bins`	Requested number of bins `B` — the bias–variance dial. TYPE: `int` DEFAULT: `10`
`strategy`	`"mass"` (equal-mass, recommended default: lower estimator bias, no empty bins; with `sample_weight` the bins carry equal weight) or `"width"` (equal-width over [0, 1]). TYPE: `(mass, width)` DEFAULT: `"mass"`
`shrinkage`	`"jeffreys"` replaces the raw rate `k/n` with `(k + 1/2)/(n + 1)` — the posterior mean under the Beta(1/2, 1/2) prior — keeping small bins away from 0 and 1. TYPE: `(jeffreys, None)` DEFAULT: `"jeffreys"`

ATTRIBUTE	DESCRIPTION
`bin_rate_`	Calibrated value per (non-degenerate) bin. TYPE: `ndarray`
`is_monotone_`	Computed after fitting: binning does not assume monotonicity, so the flag reports whether the fitted rates happen to be non-decreasing. TYPE: `bool`

References

Zadrozny & Elkan (2001).

Source code in src/probcal/binning.py

def __init__(
    self, n_bins: int = 10, strategy: str = "mass", shrinkage: str | None = "jeffreys"
) -> None:
    self.n_bins = n_bins
    self.strategy = strategy
    self.shrinkage = shrinkage

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 10.0, for either strategy ("mass" or "width").

interpret ¶

interpret() -> Interpretation

Read bin rates as local event frequencies and B as the complexity dial.

Source code in src/probcal/binning.py

def interpret(self) -> Interpretation:
    """Read bin rates as local event frequencies and B as the complexity dial."""
    self._check_fitted()
    messages = [
        (
            f"{len(self.bin_rate_)} bins ({self.strategy} strategy): each calibrated "
            "value is the (shrunken) empirical event rate of its score bin"
        ),
        "B controls bias-variance: few bins are stable but coarse, many are sharp but noisy",
    ]
    if self.shrinkage == "jeffreys":
        messages.append("Jeffreys shrinkage (k+1/2)/(n+1) keeps sparse bins away from 0 and 1")
    if not self.is_monotone_:
        messages.append(
            "fitted bin rates are not monotone: binning does not enforce ranking "
            "preservation — read inversions as noise, not signal"
        )
    return Interpretation(
        method=type(self).__name__,
        param_names=("n_bins",),
        param_values=(float(len(self.bin_rate_)),),
        messages=tuple(messages),
    )

ScalingBinningCalibrator ¶

ScalingBinningCalibrator(n_bins: int = 10)

Bases: BaseCalibrator

Scaling-binning (Kumar–Liang–Ma): Platt stage, then bin the fitted values.

Fits Platt scaling first, then forms equal-mass bins of the fitted function values and outputs the mean of the fitted values within each bin. Achieves measurable calibration error with O(1/eps^2 + B) samples versus O(B/eps^2) for histogram binning.

PARAMETER	DESCRIPTION
`n_bins`	Requested number of equal-mass bins of the Platt-fitted values (equal weight under `sample_weight`). TYPE: `int` DEFAULT: `10`

ATTRIBUTE	DESCRIPTION
`platt_`	The fitted first-stage Platt calibrator. TYPE: `PlattCalibrator`
`edges_`	Interior quantile edges of the Platt-fitted values. TYPE: `ndarray`
`bin_value_`	Mean Platt-fitted value per bin (the calibrated output for that bin). TYPE: `ndarray`
`is_monotone_`	Derived from the fitted state: `True` when the Platt stage is increasing (`a > 0`) — the bin means are then non-decreasing by construction. A decreasing Platt stage makes the composite map decreasing, and `interval_inverse` refuses it. TYPE: `bool`

References

Kumar, Liang & Ma (2019).

Source code in src/probcal/binning.py

def __init__(self, n_bins: int = 10) -> None:
    self.n_bins = n_bins

is_monotone_ `property` ¶

is_monotone_: bool

True iff the Platt stage is increasing (see class docstring).

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 4.0: a Platt stage plus a bin count, still lightweight.

interpret ¶

interpret() -> Interpretation

Two-stage reading: Platt map, then the error-measurability discretization.

Source code in src/probcal/binning.py

def interpret(self) -> Interpretation:
    """Two-stage reading: Platt map, then the error-measurability discretization."""
    self._check_fitted()
    platt_interp = self.platt_.interpret()
    return Interpretation(
        method=type(self).__name__,
        param_names=platt_interp.param_names + ("n_bins",),
        param_values=platt_interp.param_values + (float(len(self.bin_value_)),),
        messages=platt_interp.messages
        + (
            (
                f"binning stage: {len(self.bin_value_)} equal-mass bins of the fitted "
                "Platt values; outputs are bin means, which makes the residual "
                "calibration error estimable with O(1/eps^2 + B) samples "
                "(vs O(B/eps^2) for histogram binning)"
            ),
        ),
    )

bayesian ¶

Bayesian-ensemble calibrators: BBQ and ENIR.

Theory: docs/concepts/methods-nonparametric.md.

References

Naeini, Cooper & Hauskrecht (2015); Naeini & Cooper (2016); Tibshirani, Hoefling & Tibshirani (2011) — full records in the documentation.

BBQCalibrator ¶

BBQCalibrator(min_bins: int | None = None, max_bins: int | None = None, monotone: bool = False)

Bases: BaseCalibrator

Bayesian Binning into Quantiles: model averaging over equal-mass binnings.

Considers equal-mass binning models over a range of bin counts, scores each by its Beta-Binomial log marginal likelihood under a per-bin Jeffreys Beta(1/2, 1/2) prior, and predicts with the posterior-weighted average of the models' (posterior-mean) bin rates.

Nothing in the model forces the bin rates to increase with the score, and on low-default data a sparse bin often dips below its neighbours, so the averaged map is usually not monotone (is_monotone_ = False) and refuses to invert. monotone=True projects each candidate's bin rates onto the non-decreasing cone (weighted PAVA, weights = posterior counts) before averaging; a non-negative mixture of non-decreasing step functions is non-decreasing, so the averaged map is then monotone by construction. The model weights are unchanged (they score the unconstrained binnings).

PARAMETER	DESCRIPTION
`min_bins`	Range of candidate bin counts; defaults to `[2, ceil(sqrt(n))]` capped at 50. Bins are equal-mass in the sample weights. TYPE: `int or None` DEFAULT: `None`
`max_bins`	Range of candidate bin counts; defaults to `[2, ceil(sqrt(n))]` capped at 50. Bins are equal-mass in the sample weights. TYPE: `int or None` DEFAULT: `None`
`monotone`	Isotonize each candidate's bin rates before averaging, so the map is guaranteed non-decreasing and `interval_inverse` / `calibrated_bands_to_raw` are available. Leaves fits that are already monotone in every candidate unchanged. TYPE: `bool` DEFAULT: `False`

ATTRIBUTE	DESCRIPTION
`bins_grid_`	Candidate bin counts. TYPE: `ndarray`
`model_weights_`	Posterior weights over the candidates (sum to 1). TYPE: `ndarray`
`model_edges_, model_rates_`	Interior edges and posterior-mean bin rates of each candidate binning (isotonized when `monotone=True`). TYPE: `list of numpy.ndarray`
`is_monotone_`	Whether the averaged step map is non-decreasing over the whole score domain — checked exactly at every bin edge of every candidate. TYPE: `bool`

References

Naeini, Cooper & Hauskrecht (2015).

Source code in src/probcal/bayesian.py

def __init__(
    self, min_bins: int | None = None, max_bins: int | None = None, monotone: bool = False
) -> None:
    self.min_bins = min_bins
    self.max_bins = max_bins
    self.monotone = monotone

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 40.0: a Bayesian model average over binnings.

interpret ¶

interpret() -> Interpretation

Read the posterior weights as uncertainty about the data's resolution.

Source code in src/probcal/bayesian.py

def interpret(self) -> Interpretation:
    """Read the posterior weights as uncertainty about the data's resolution."""
    self._check_fitted()
    wts = self.model_weights_
    top = np.argsort(wts)[::-1][:3]
    top_txt = ", ".join(f"B={int(self.bins_grid_[i])} (weight {wts[i]:.3f})" for i in top)
    return Interpretation(
        method=type(self).__name__,
        param_names=("n_models",),
        param_values=(float(len(self.bins_grid_)),),
        messages=(
            f"top-3 binning models by posterior weight: {top_txt}",
            "concentrated weight = the data speak clearly about their own resolution; "
            "diffuse weight = the averaging is doing real work",
        ),
    )

ENIRCalibrator ¶

ENIRCalibrator(max_solutions: int | None = 256)

Bases: BaseCalibrator

Ensemble of near-isotonic regressions (ENIR).

Computes the full nearly-isotonic solution path (modified PAVA of Tibshirani, Hoefling & Tibshirani, 2011) from the raw data (lambda = 0) to the fully isotonic fit, then averages the breakpoint solutions with BIC weights. The combined map may be non-monotone: is_monotone_ is False and consumers requiring order preservation should prefer a monotone calibrator. Fitting is quadratic in the number of unique scores and intended for m <= 50,000; above that, fit emits a single UserWarning stating the expected minutes.

PARAMETER	DESCRIPTION
`max_solutions`	Number of path solutions to keep for the ensemble, chosen by best (lowest) BIC; `None` keeps every breakpoint. Retention is what bounds memory: the path has up to `m` breakpoints, so keeping all of them costs `O(m^2)`. TYPE: `int or None` DEFAULT: `256`

ATTRIBUTE	DESCRIPTION
`path_lambdas_`	Breakpoints of the penalty parameter, starting at 0. All breakpoints are recorded, whether or not their solution is retained. TYPE: `numpy.ndarray of shape (T,)`
`path_solutions_`	Fitted values on the tie-aggregated score grid at the retained breakpoints, in breakpoint order. `K` is the number of retained breakpoints: at most `max_solutions`, and fewer when breakpoints are pruned (a breakpoint whose BIC weight is provably below 1e-15 relative is skipped without being scored, and is retained only when `max_solutions` is `None`). TYPE: `numpy.ndarray of shape (K, m)`
`kept_breakpoints_`	Indices into `path_lambdas_` of the retained breakpoints. TYPE: `numpy.ndarray of shape (K,)`
`model_weights_`	BIC weights over the retained solutions, renormalized to sum to 1. TYPE: `numpy.ndarray of shape (K,)`
`scores_`	Distinct calibration scores: the grid `path_solutions_` lives on. TYPE: `numpy.ndarray of shape (m,)`
`dropped_weight_`	BIC weight lost to retention — the weight of scored solutions that the `max_solutions` cap evicted, before renormalization; a `UserWarning` is raised above 1e-6. Pruned breakpoints do not count towards it: their weight is exactly 0 in double precision. TYPE: `float`

References

Naeini & Cooper (2016); Tibshirani, Hoefling & Tibshirani (2011).

Source code in src/probcal/bayesian.py

def __init__(self, max_solutions: int | None = 256) -> None:
    self.max_solutions = max_solutions

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 80.0: an ensemble over the full near-isotonic path.

interpret ¶

interpret() -> Interpretation

Read the path length and BIC weights; warn about non-monotonicity.

Source code in src/probcal/bayesian.py

def interpret(self) -> Interpretation:
    """Read the path length and BIC weights; warn about non-monotonicity."""
    self._check_fitted()
    wts = self.model_weights_
    top = np.argsort(wts)[::-1][:3]
    top_txt = ", ".join(
        f"lambda={self.path_lambdas_[self.kept_breakpoints_[i]]:.4g} (weight {wts[i]:.3f})"
        for i in top
    )
    return Interpretation(
        method=type(self).__name__,
        param_names=("n_path_solutions",),
        param_values=(float(len(self.path_lambdas_)),),
        messages=(
            f"top-3 path solutions by BIC weight: {top_txt}",
            "lambda trades monotonicity strictness against fit; BIC weights are model "
            "plausibility along the path",
            "the ensemble output may be non-monotone (is_monotone_ = False): consumers "
            "requiring order preservation should use a monotone calibrator",
        ),
    )

vennabers ¶

Venn–Abers calibrators: inductive (IVAP) and cross (CVAP).

Theory, validity guarantee scope, and the scalarization caveat: docs/concepts/methods-distribution-free.md. The guarantee attaches to the interval returned by :meth:VennAbersCalibrator.predict_interval; the scalar from predict_proba is the log-loss-minimax merger and is not itself covered by the validity theorem.

References

Vovk & Petej (2014) — full record in the documentation.

VennAbersCalibrator ¶

Bases: BaseCalibrator

Inductive Venn–Abers predictor (IVAP).

For a query score, two isotonic fits on the calibration set augmented with the query labeled 0 (resp. 1) yield the interval [p0, p1]; predict_proba scalarizes it as p1 / (1 - p0 + p1).

Both fits are precomputed at fit time by the Vovk & Petej (2014) cumulative- sum-diagram sweep, so prediction is a searchsorted gather rather than a pair of PAVA refits per unique query score.

Tied calibration scores are pooled into one point (weights summed), as isotonic regression requires, and a query whose score equals a calibration score joins that point; the output therefore does not depend on row order. A query labeled 1 (resp. 0) strictly between two calibration scores has the same fitted value as when merged into the next (resp. previous) point, so the two tables are read at searchsorted "left" (F1_) and "right" (F0_).

ATTRIBUTE	DESCRIPTION
`scores_`	Distinct calibration scores, sorted. TYPE: `numpy.ndarray of shape (m,)`
`F0_, F1_`	Fitted probabilities for a unit-weight query labeled 0 (resp. 1) inserted at each of the m+1 gaps between the distinct calibration scores. Both are non-decreasing, and `F0_ <= F1_` elementwise. TYPE: `numpy.ndarray of shape (m + 1,)`
`ties_pooled_`	`True` for every fit. `False` only for a payload re-saved by probcal 0.3.4/0.3.5 from an older fit (stored tables, unpooled lookup). TYPE: `bool`

Notes

With non-unit sample weights the query still enters at weight 1, which is the natural generalization but sits outside the validity theorem as proved; see the scope note in docs/concepts/methods-distribution-free.md.

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 60.0: a distribution-free interval predictor.

predict_interval ¶

predict_interval(s: object) -> ndarray

Venn–Abers intervals [p0, p1] for new scores.

PARAMETER	DESCRIPTION
`s`	Raw scores/probabilities in `[0, 1]`. TYPE: `array_like`

RETURNS	DESCRIPTION
`numpy.ndarray of shape (n, 2)`	Columns `p0` (lower) and `p1` (upper). The distribution-free validity guarantee attaches to this pair, not to the scalarized `predict_proba` output.

Source code in src/probcal/vennabers.py

def predict_interval(self, s: object) -> np.ndarray:
    """Venn–Abers intervals ``[p0, p1]`` for new scores.

    Parameters
    ----------
    s : array_like
        Raw scores/probabilities in ``[0, 1]``.

    Returns
    -------
    numpy.ndarray of shape (n, 2)
        Columns ``p0`` (lower) and ``p1`` (upper). The distribution-free
        validity guarantee attaches to this pair, not to the scalarized
        ``predict_proba`` output.
    """
    self._check_fitted()
    return self._interval(validate_scores(s))

interpret ¶

interpret() -> Interpretation

Report interval widths over the calibration scores — where to trust the map.

Source code in src/probcal/vennabers.py

def interpret(self) -> Interpretation:
    """Report interval widths over the calibration scores — where to trust the map."""
    self._check_fitted()
    intervals = self._interval(self.scores_)
    widths = intervals[:, 1] - intervals[:, 0]
    mean_w, max_w = float(widths.mean()), float(widths.max())
    return Interpretation(
        method=type(self).__name__,
        param_names=("mean_width", "max_width"),
        param_values=(mean_w, max_w),
        messages=(
            f"mean Venn–Abers interval width {mean_w:.4f}, maximum {max_w:.4f} over the "
            "distinct calibration scores: width is per-score calibration uncertainty",
            "the validity guarantee holds for the interval [p0, p1] from "
            "predict_interval(); the scalar from predict_proba() is the log-loss-minimax "
            "merger p1/(1-p0+p1) and is not itself covered by the guarantee",
        ),
    )

CrossVennAbersCalibrator ¶

CrossVennAbersCalibrator(cv: int = 5, random_state: int = 42)

Bases: BaseCalibrator

Cross Venn–Abers predictor (CVAP): fold-wise IVAPs, geometric-mean merge.

Splits the calibration data into cv stratified folds; each fold's IVAP is fitted on the remaining folds. The scalar output merges the fold-wise pairs by the log-loss rule of Vovk & Petej: GM(p1) / (GM(1 - p0) + GM(p1)). predict_interval returns the conservative envelope [min_k p0_k, max_k p1_k] (a probcal design choice — the paper defines only the scalar merge).

PARAMETER	DESCRIPTION
`cv`	Number of stratified folds; an integer `>= 2`. TYPE: `int` DEFAULT: `5`
`random_state`	Seed for the fold assignment. TYPE: `int` DEFAULT: `42`

ATTRIBUTE	DESCRIPTION
`ivaps_`	One fitted IVAP per fold, each trained on the other `cv - 1` folds. TYPE: `list of VennAbersCalibrator`

References

Vovk & Petej (2014).

Source code in src/probcal/vennabers.py

def __init__(self, cv: int = 5, random_state: int = 42) -> None:
    self.cv = cv
    self.random_state = random_state

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 60.0: same tier as IVAP, folded across cv splits.

predict_interval ¶

predict_interval(s: object) -> ndarray

Conservative fold envelope [min_k p0_k, max_k p1_k] for new scores.

PARAMETER	DESCRIPTION
`s`	Raw scores/probabilities in `[0, 1]`. TYPE: `array_like`

RETURNS	DESCRIPTION
`numpy.ndarray of shape (n, 2)`	Columns `p0` (lower) and `p1` (upper): the envelope across fold-wise IVAP intervals (a probcal design choice — the paper defines only the scalar merge, not an interval for CVAP).

Source code in src/probcal/vennabers.py

def predict_interval(self, s: object) -> np.ndarray:
    """Conservative fold envelope ``[min_k p0_k, max_k p1_k]`` for new scores.

    Parameters
    ----------
    s : array_like
        Raw scores/probabilities in ``[0, 1]``.

    Returns
    -------
    numpy.ndarray of shape (n, 2)
        Columns ``p0`` (lower) and ``p1`` (upper): the envelope across
        fold-wise IVAP intervals (a probcal design choice — the paper
        defines only the scalar merge, not an interval for CVAP).
    """
    self._check_fitted()
    p0, p1 = self._fold_pairs(validate_scores(s))
    return np.column_stack([p0.min(axis=0), p1.max(axis=0)])

interpret ¶

interpret() -> Interpretation

Report fold count and envelope widths over a probe grid.

Source code in src/probcal/vennabers.py

def interpret(self) -> Interpretation:
    """Report fold count and envelope widths over a probe grid."""
    self._check_fitted()
    probe = np.linspace(0.01, 0.99, 99)
    env = self.predict_interval(probe)
    widths = env[:, 1] - env[:, 0]
    return Interpretation(
        method=type(self).__name__,
        param_names=("cv", "mean_envelope_width"),
        param_values=(float(self.cv), float(widths.mean())),
        messages=(
            f"{self.cv} stratified folds, one IVAP per fold; scalar output is the "
            "geometric-mean merge GM(p1)/(GM(1-p0)+GM(p1))",
            "predict_interval() returns the conservative fold envelope "
            "[min p0, max p1]; per-fold IVAP intervals carry the validity guarantee",
        ),
    )

spline ¶

Spline calibrator: penalized natural cubic splines on the logit scale.

Theory: docs/concepts/methods-nonparametric.md.

References

Lucena (2018); Hastie, Tibshirani & Friedman (2009), §5.2.1 — full records in the documentation.

SplineCalibrator ¶

SplineCalibrator(n_knots: int | None = None, lambdas: object = None, cv: int = 5, random_state: int = 42)

Bases: BaseCalibrator

Natural cubic spline calibration on the logit scale.

Models logit g(s) = sum_k theta_k N_k(logit s) with the natural cubic basis (linear beyond the boundary knots), fitted by penalized IRLS with a second-difference roughness penalty — the package's safeguarded Newton solver (:func:probcal._math.irls_logistic with a penalty matrix: step-halving, full-step convergence test, separation fallback). The penalty weight is chosen by K-fold cross-validated log loss within the calibration set.

PARAMETER	DESCRIPTION
`n_knots`	Number of knots (placed at equally spaced quantiles of the logit scores); defaults to `clip(ceil(n^(1/3)), 4, 12)`. TYPE: `int or None` DEFAULT: `None`
`lambdas`	Candidate penalty weights (finite, non-negative); defaults to `logspace(-4, 4, 17)`. TYPE: `array_like or None` DEFAULT: `None`
`cv`	Inner fold count for the lambda search; an integer `>= 2`. TYPE: `int` DEFAULT: `5`
`random_state`	Seed for the stratified fold assignment. TYPE: `int` DEFAULT: `42`

ATTRIBUTE	DESCRIPTION
`lambda_`	Selected penalty weight. TYPE: `float`
`edof_`	Effective degrees of freedom — trace of the smoother matrix at the fitted solution; the honest complexity measure. TYPE: `float`
`n_knots_`	Number of knots actually used. TYPE: `int`
`knots_`	Knot locations on the logit scale. TYPE: `ndarray`
`theta_`	Spline coefficients on the natural cubic basis. TYPE: `ndarray`
`converged_`	Whether the final penalized IRLS fit converged; if `False` a warning was raised at fit time and `interpret()` records it. TYPE: `bool`
`is_monotone_`	Checked exactly after fitting: the logit-scale slope is evaluated at every knot (which also gives both linear tails) and at each interior extremum of the piecewise-quadratic slope. The penalty does not enforce monotonicity, and a non-monotone fit is flagged with a warning. TYPE: `bool`

References

Lucena (2018); Hastie, Tibshirani & Friedman (2009), §5.2.1.

Source code in src/probcal/spline.py

def __init__(
    self,
    n_knots: int | None = None,
    lambdas: object = None,
    cv: int = 5,
    random_state: int = 42,
) -> None:
    self.n_knots = n_knots
    self.lambdas = lambdas
    self.cv = cv
    self.random_state = random_state

complexity_rank `property` ¶

complexity_rank: float

Parsimony rank 12.0: a penalized basis expansion, more flexible than binning.

interpret ¶

interpret() -> Interpretation

Read effective degrees of freedom as the honest complexity measure.

Source code in src/probcal/spline.py

def interpret(self) -> Interpretation:
    """Read effective degrees of freedom as the honest complexity measure."""
    self._check_fitted()
    messages = [
        (
            f"effective degrees of freedom {self.edof_:.2f} (trace of the smoother): "
            "values near 2 mean a parametric family would have sufficed; larger values "
            "mean the curvature is real"
        ),
        (
            f"penalty lambda = {self.lambda_:.4g} chosen by {self.cv}-fold "
            f"cross-validated log loss over {len(self.lambdas_grid_)} candidates; "
            f"{self.n_knots_} knots at logit-score quantiles"
        ),
        (
            "regions where the fitted curve runs steeper than the identity are locally "
            "underconfident score regions; shallower, locally overconfident"
        ),
    ]
    if not self.is_monotone_:
        messages.append("fitted curve is NOT monotone (warned at fit)")
    if not self.converged_:
        messages.append("penalized IRLS did not converge; the curve may be unreliable")
    return Interpretation(
        method=type(self).__name__,
        param_names=("edof", "lambda", "n_knots"),
        param_values=(self.edof_, self.lambda_, float(self.n_knots_)),
        messages=tuple(messages),
    )

segmented ¶

SegmentedCalibrator: empirical-Bayes shrunken per-segment logit offsets.

Theory and the DerSimonian-Laird method-of-moments derivation: docs/concepts/segmented.md.

References

DerSimonian & Laird (1986) — random-effects meta-analysis method-of-moments heterogeneity estimator, reused here across segments instead of studies.

SegmentedCalibrator ¶

SegmentedCalibrator(base: BaseCalibrator | None = None, *, unseen: str = 'global')

Bases: BaseCalibrator

Per-segment logit offsets on top of a shared base map, empirical-Bayes shrunk.

Fits one shared base calibrator on all data, then an offset-only logistic MLE (:func:probcal.offset.estimate_offset) of each segment's residual log-odds shift against the base map's predictions. Segments with few observations have a noisy, high-variance delta_hat; rather than use it directly (no pooling — overfits small segments) or discard it (complete pooling — ignores real heterogeneity), each segment's estimate is shrunk toward the across-segment mean by the classic empirical-Bayes/random-effects factor tau2 / (tau2 + se**2), where tau2 is the between-segment heterogeneity variance estimated by the DerSimonian-Laird (1986) method of moments. A small, noisy segment (large se) shrinks toward 0 (the base map); a large, precise segment (small se) keeps most of its own estimate.

Segments with only one outcome class have no offset MLE (estimate_offset raises); they are recorded as delta_hat=0.0, se=inf — fully shrunk, since an infinite-variance estimate carries no weight in the DerSimonian-Laird pooling and tau2 / (tau2 + inf) = 0.

fit and predict_proba add a keyword-only segments argument on top of the base signature (segments=None degrades to a single segment "__all__" at fit time, and to the plain base map — no segment-specific offset — at predict time), so the zero-argument protocol calls (SegmentedCalibrator().fit(s, y), cal.predict_proba(s)) still work. Because :class:~probcal.chain.Chain has no segments= slot, Chain([seg, ...]) always predicts through seg's global map (segments=None, delta=0) — the per-segment shift is not baked into a Chain; use SegmentedCalibrator directly (with segments=) when the per-segment offset must apply.

Segment labels are compared as strings (_coerce_segments calls .astype(str)): fit-time labels 0, 1 (int) become "0", "1", but predict-time labels 0.0, 1.0 (float) become "0.0", "1.0" — a mismatch that never raises (every label looks "unseen") and, under unseen="global", silently falls back to the base map for every row. Pass segments with the same representation (e.g. cast to str yourself) at fit and predict time. When every row of a predict_proba/inverse call is unseen and unseen="global", a UserWarning is raised naming the fitted segments_ — a partial overlap (some rows match, some are genuinely new segments) stays silent.

PARAMETER	DESCRIPTION
`base`	Unfitted calibrator cloned (:func:`~probcal.base.clone_unfitted`) and fitted on the pooled data at :meth:`fit` time. `None` (default) uses :class:`~probcal.parametric.BetaCalibrator`. TYPE: `BaseCalibrator or None` DEFAULT: `None`
`unseen`	Policy for a segment label at predict/inverse time that was not seen at fit time. `"global"` (default) applies `delta=0` (the base map); `"raise"` raises `ValueError`. TYPE: `('global', 'raise')` DEFAULT: `"global"`

ATTRIBUTE	DESCRIPTION
`base_`	The fitted clone of `base` (or a fitted `BetaCalibrator()`). TYPE: `BaseCalibrator`
`segments_`	Segment labels seen at fit time, sorted. TYPE: `tuple of str`
`n_, events_`	Per-segment observation count and weighted event count, aligned with `segments_`. TYPE: `ndarray`
`delta_hat_, se_`	Per-segment offset MLE and its standard error (:func:`~probcal.offset.estimate_offset`; `se=inf` for single-class segments). Invariant to the scale of `sample_weight`. TYPE: `ndarray`
`tau2_`	Between-segment heterogeneity variance (DerSimonian-Laird method-of-moments estimate); 0.0 when fewer than two segments have a finite `se` (complete pooling). TYPE: `float`
`shrink_`	Per-segment shrinkage factor `tau2 / (tau2 + se2)` in `[0, 1)`. TYPE:** `ndarray`
`delta_tilde_`	Per-segment shrunk offset, `delta_hat * shrink`: the offset actually applied at predict time. TYPE: `ndarray`
`is_monotone_`	`base_.is_monotone_` (read through, so it survives serialization). TYPE: `bool`

Examples:

>>> import numpy as np
>>> from probcal import SegmentedCalibrator, make_pd_portfolio
>>> d = make_pd_portfolio(n=900, random_state=0)
>>> segments = np.array(["a", "b", "c"])[np.arange(900) % 3]
>>> cal = SegmentedCalibrator().fit(d.scores, d.y, segments=segments)
>>> cal.segments_
('a', 'b', 'c')
>>> p_global = cal.predict_proba(d.scores)  # segments=None: the base map
>>> p_seg = cal.predict_proba(d.scores, segments=segments)
>>> p_global.shape == p_seg.shape == d.scores.shape
True

Source code in src/probcal/segmented.py

def __init__(self, base: BaseCalibrator | None = None, *, unseen: str = "global") -> None:
    self.base = base
    self.unseen = unseen

is_monotone_ `property` ¶

is_monotone_: bool

The fitted base map's is_monotone_ (True before fitting).

A per-segment logit shift is increasing, so it cannot break monotonicity; the base map alone decides.

complexity_rank `property` ¶

complexity_rank: float

The base map's parsimony rank (fitted base_, else the prototype).

affine_logit_coeffs_ `property` ¶

affine_logit_coeffs_: tuple[float, float] | None

(a, b + delta_tilde) only for a single fitted segment; else None.

With more than one segment the map is segment-dependent (there is no single affine map on the logit scale that fits every segment), so :meth:point_inverse (which relies on this property) is unavailable then — use :meth:interval_inverse with segment=.

fit ¶

fit(s: object, y: object, sample_weight: object = None, *, segments: object = None) -> SegmentedCalibrator

Fit the shared base map, then per-segment shrunk offsets.

PARAMETER	DESCRIPTION
`s`	Raw scores/probabilities in `[0, 1]`. TYPE: `array_like`
`y`	Binary outcomes in `{0, 1}`; both classes must be present overall (individual segments may be single-class — see class docstring). TYPE: `array_like`
`sample_weight`	Positive observation weights. TYPE: `array_like or None` DEFAULT: `None`
`segments`	Segment label per observation, same length as `s`. `None` (default) fits a single segment `"__all__"`. TYPE: `(array_like or None, keyword - only)` DEFAULT: `None`

RETURNS	DESCRIPTION
`SegmentedCalibrator`	The fitted calibrator.

Source code in src/probcal/segmented.py

def fit(
    self,
    s: object,
    y: object,
    sample_weight: object = None,
    *,
    segments: object = None,
) -> "SegmentedCalibrator":
    """Fit the shared base map, then per-segment shrunk offsets.

    Parameters
    ----------
    s : array_like
        Raw scores/probabilities in ``[0, 1]``.
    y : array_like
        Binary outcomes in ``{0, 1}``; both classes must be present
        overall (individual segments may be single-class — see class
        docstring).
    sample_weight : array_like or None
        Positive observation weights.
    segments : array_like or None, keyword-only
        Segment label per observation, same length as ``s``. ``None``
        (default) fits a single segment ``"__all__"``.

    Returns
    -------
    SegmentedCalibrator
        The fitted calibrator.
    """
    self._segments_arg = segments
    try:
        return super().fit(s, y, sample_weight)
    finally:
        self.__dict__.pop("_segments_arg", None)

predict_proba ¶

predict_proba(s: object, *, segments: object = None) -> ndarray

Calibrated probabilities, with an optional per-observation segment offset.

PARAMETER	DESCRIPTION
`s`	Raw scores/probabilities in `[0, 1]`. TYPE: `array_like`
`segments`	Segment label per observation, same length as `s`. `None` (default) returns the plain base map (`delta=0`); a label not in :attr:`segments_` is handled per `unseen`. TYPE: `(array_like or None, keyword - only)` DEFAULT: `None`

RETURNS	DESCRIPTION
`numpy.ndarray of shape (n,)`	Calibrated probabilities.

Source code in src/probcal/segmented.py

def predict_proba(self, s: object, *, segments: object = None) -> np.ndarray:
    """Calibrated probabilities, with an optional per-observation segment offset.

    Parameters
    ----------
    s : array_like
        Raw scores/probabilities in ``[0, 1]``.
    segments : array_like or None, keyword-only
        Segment label per observation, same length as ``s``. ``None``
        (default) returns the plain base map (``delta=0``); a label not
        in :attr:`segments_` is handled per ``unseen``.

    Returns
    -------
    numpy.ndarray of shape (n,)
        Calibrated probabilities.
    """
    self._check_fitted()
    s_arr = validate_scores(s)
    p0 = self.base_.predict_proba(s_arr)
    if segments is None:
        return p0
    deltas = self._lookup_deltas(_coerce_segments(segments, s_arr.shape[0]))
    return expit(logit(p0) + deltas)

interval_inverse ¶

interval_inverse(lo: float, hi: float, *, space: str = 'probability', buffer_logit: float = 0.0, segment: object = None) -> tuple[float, float]

Preimage of a calibrated interval, optionally for one fitted segment.

segment=None (default) uses the global map (delta=0), matching :meth:predict_proba's segments=None convention; otherwise the preimage is through base_ composed with that segment's delta_tilde (Chain([base_, LogitOffset(delta=...)])).

PARAMETER	DESCRIPTION
`lo`	Calibrated-probability bounds. TYPE: `float`
`hi`	Calibrated-probability bounds. TYPE: `float`
`space`	Scale of the returned raw bounds. TYPE: `(probability, logit)` DEFAULT: `"probability"`
`buffer_logit`	Logit-space shrinkage applied before inverting. TYPE: `(float, keyword - only)` DEFAULT: `0.0`
`segment`	Segment label to invert through; `None` for the global map. TYPE: `(str or None, keyword - only)` DEFAULT: `None`

RETURNS	DESCRIPTION
`tuple of float`	`(raw_lo, raw_hi)` on the requested scale.

Source code in src/probcal/segmented.py

def interval_inverse(
    self,
    lo: float,
    hi: float,
    *,
    space: str = "probability",
    buffer_logit: float = 0.0,
    segment: object = None,
) -> tuple[float, float]:
    """Preimage of a calibrated interval, optionally for one fitted segment.

    ``segment=None`` (default) uses the global map (``delta=0``),
    matching :meth:`predict_proba`'s ``segments=None`` convention;
    otherwise the preimage is through ``base_`` composed with that
    segment's ``delta_tilde`` (``Chain([base_, LogitOffset(delta=...)])``).

    Parameters
    ----------
    lo, hi : float
        Calibrated-probability bounds.
    space : {"probability", "logit"}, keyword-only
        Scale of the returned raw bounds.
    buffer_logit : float, keyword-only
        Logit-space shrinkage applied before inverting.
    segment : str or None, keyword-only
        Segment label to invert through; ``None`` for the global map.

    Returns
    -------
    tuple of float
        ``(raw_lo, raw_hi)`` on the requested scale.
    """
    self._check_fitted()
    target = self._segment_chain_or_base(segment)
    return target.interval_inverse(lo, hi, space=space, buffer_logit=buffer_logit)

point_inverse ¶

point_inverse(p: object, *, space: str = 'probability', segment: object = None) -> ndarray

Exact preimage of calibrated probabilities, optionally for one segment.

Same segment= convention as :meth:interval_inverse: inverts through base_ (composed with the segment's delta_tilde when segment is given), so this works for any number of fitted segments as long as base_ itself has an exact point inverse (base_.affine_logit_coeffs_ is not None) — unlike :attr:affine_logit_coeffs_ on self, which is only defined for a single fitted segment.

PARAMETER	DESCRIPTION
`p`	Calibrated probabilities strictly inside `(0, 1)`. TYPE: `array_like`
`space`	Scale of the returned raw values. TYPE: `(probability, logit)` DEFAULT: `"probability"`
`segment`	Segment label to invert through; `None` for the global map. TYPE: `(str or None, keyword - only)` DEFAULT: `None`

RETURNS	DESCRIPTION
`ndarray`	Raw scores (or logits) whose calibrated probability equals `p`.

Source code in src/probcal/segmented.py

def point_inverse(
    self, p: object, *, space: str = "probability", segment: object = None
) -> np.ndarray:
    """Exact preimage of calibrated probabilities, optionally for one segment.

    Same ``segment=`` convention as :meth:`interval_inverse`: inverts
    through ``base_`` (composed with the segment's ``delta_tilde`` when
    ``segment`` is given), so this works for any number of fitted
    segments as long as ``base_`` itself has an exact point inverse
    (``base_.affine_logit_coeffs_`` is not ``None``) — unlike
    :attr:`affine_logit_coeffs_` on ``self``, which is only defined for
    a single fitted segment.

    Parameters
    ----------
    p : array_like
        Calibrated probabilities strictly inside ``(0, 1)``.
    space : {"probability", "logit"}, keyword-only
        Scale of the returned raw values.
    segment : str or None, keyword-only
        Segment label to invert through; ``None`` for the global map.

    Returns
    -------
    numpy.ndarray
        Raw scores (or logits) whose calibrated probability equals ``p``.
    """
    self._check_fitted()
    target = self._segment_chain_or_base(segment)
    return target.point_inverse(p, space=space)

interpret ¶

interpret() -> Interpretation

Per-segment shrinkage table plus the fitted heterogeneity variance.

Source code in src/probcal/segmented.py

def interpret(self) -> Interpretation:
    """Per-segment shrinkage table plus the fitted heterogeneity variance."""
    self._check_fitted()
    param_names = ("tau2", *(f"delta.{g}" for g in self.segments_))
    param_values = (self.tau2_, *(float(d) for d in self.delta_tilde_))
    messages = [
        f"tau2 = {self.tau2_:.4f}: between-segment heterogeneity variance "
        "(DerSimonian-Laird method of moments on the per-segment offset MLEs); "
        "tau2 = 0 means complete pooling (every delta_tilde = 0)",
    ]
    for g, n, ev, dh, se, dt, sh in zip(
        self.segments_,
        self.n_,
        self.events_,
        self.delta_hat_,
        self.se_,
        self.delta_tilde_,
        self.shrink_,
        strict=True,
    ):
        messages.append(
            f"segment {g!r}: n={int(n)}, events={ev:.1f}, delta_hat={dh:+.4f}, "
            f"se={se:.4f}, delta_tilde={dt:+.4f}, shrink={sh:.3f}"
        )
    policy = "use delta=0" if self.unseen == "global" else "raise ValueError"
    messages.append(
        f"unseen segments at predict/inverse time {policy} (unseen={self.unseen!r})"
    )
    return Interpretation(
        method=type(self).__name__,
        param_names=param_names,
        param_values=param_values,
        messages=tuple(messages),
    )

API: calibrators¶

base ¶

UnattainableTargetError ¶

BaseCalibrator ¶

affine_logit_coeffs_ property ¶

complexity_rank property ¶

fit ¶

predict_proba ¶

__sklearn_is_fitted__ ¶

__sklearn_tags__ ¶

interpret abstractmethod ¶

interval_inverse ¶

point_inverse ¶

get_params ¶

set_params ¶

to_dict ¶

from_dict classmethod ¶

shrink_interval ¶

validate_interval ¶

clone_unfitted ¶

parametric ¶

PlattCalibrator ¶

affine_logit_coeffs_ property ¶

complexity_rank property ¶

interpret ¶

TemperatureCalibrator ¶

affine_logit_coeffs_ property ¶

complexity_rank property ¶

interpret ¶

BetaCalibrator ¶

affine_logit_coeffs_ property ¶

complexity_rank property ¶

interpret ¶

isotonic ¶

IsotonicCalibrator ¶

complexity_rank property ¶

interpret ¶

CenteredIsotonicCalibrator ¶

interpret ¶

binning ¶

HistogramBinningCalibrator ¶

complexity_rank property ¶

interpret ¶

ScalingBinningCalibrator ¶

is_monotone_ property ¶

complexity_rank property ¶

interpret ¶

bayesian ¶

BBQCalibrator ¶

complexity_rank property ¶

interpret ¶

ENIRCalibrator ¶

complexity_rank property ¶

interpret ¶

vennabers ¶

VennAbersCalibrator ¶

complexity_rank property ¶

predict_interval ¶

interpret ¶

CrossVennAbersCalibrator ¶

complexity_rank property ¶

predict_interval ¶

interpret ¶

spline ¶

SplineCalibrator ¶

complexity_rank property ¶

interpret ¶

segmented ¶

SegmentedCalibrator ¶

is_monotone_ property ¶

complexity_rank property ¶

affine_logit_coeffs_ property ¶

fit ¶

predict_proba ¶

interval_inverse ¶

point_inverse ¶

interpret ¶

affine_logit_coeffs_ `property` ¶

complexity_rank `property` ¶

interpret `abstractmethod` ¶

from_dict `classmethod` ¶

affine_logit_coeffs_ `property` ¶

complexity_rank `property` ¶

affine_logit_coeffs_ `property` ¶

complexity_rank `property` ¶

affine_logit_coeffs_ `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

is_monotone_ `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

complexity_rank `property` ¶

is_monotone_ `property` ¶

complexity_rank `property` ¶

affine_logit_coeffs_ `property` ¶