Data and preparation API

Bases: BaseModel

Names of the special (non-feature) columns in a pricing dataset.

Feature columns are everything else. .target and .exposure are required; the rest are optional. Unknown fields are rejected.

Bases: TransformerMixin, BaseEstimator

Bin numeric columns into actuarially-credible tariff bins.

Binned columns become ordered interval categoricals (DataFrame input) or integer bin codes (ndarray input). Non-target columns pass through unchanged. Finite bin edges are lower-inclusive and upper-exclusive. NaN is assigned to a reserved "Missing" category. Transform input must contain every fitted mapping column; arrays use x0, x1, etc. Unmatched column names raise ValueError.

Parameters

cols : list[str] | None Columns to bin. None -> all numeric columns found in fit. strategy : {"quantile", "tree"} "quantile" = (exposure-)weighted equal-frequency edges. "tree" = DecisionTreeRegressor(max_leaf_nodes=max_bins) on the target. max_bins : int Maximum number of bins per column (>=2). min_exposure : float | None Minimum total exposure per bin; smaller bins merged into neighbours. Requires a real exposure_col column and must be >= 1/366 (one day); raises otherwise (a count-unit fallback would change units). min_claims : float | None Minimum aggregate claim count per bin; smaller bins merge into neighbours. exposure_col, claim_count_col, target_col : str | None Configured exposure and claim-count columns must exist at fit and contain finite values; exposures must be at least one day (1/366, with day-count tolerance), and claim counts must be non-negative. Without exposure_col, binning is unweighted. For tree binning, the target is target_col / exposure_col (pure premium) when both are set, else target_col, else y. Supplied outcomes must be finite after numeric and unit conversion; missing outcomes are never treated as zero. monotonic : False | True | "increasing" | "decreasing" When set, bin means are smoothed to be monotonic in the target via isotonic_regression and adjacent bins with the same smoothed mean merge. No-op without a target. False = off, True / "increasing" = non-decreasing bin means, "decreasing" = non-increasing. Applied after the strategy and before the min_exposure small-bin merge. random_state : int Tree strategy reproducibility.

Attributes

mapping_ : dict[str, np.ndarray] Bin edges per binned column (length n_bins - 1). category_dtypes_ : dict[str, pd.CategoricalDtype] Stable ordered output vocabularies for DataFrame inputs. bin_cols_ : list[str] n_features_in_ : int feature_names_in_ : list[str] | None

Bases: TransformerMixin, BaseEstimator

Group categorical levels into credibility-stable groups.

Strategies

"rare" Levels below min_exposure / min_claims -> other_label. min_exposure requires a real exposure_col column and must be >= 1/366 (one day). "similarity" Nominal levels are stably risk-sorted; ordered categoricals retain their declared order. Under-credible groups merge with their nearest-risk neighbour, then the closest adjacent pairs merge until max_groups is met. Risks are aggregate claims / exposure, recomputed after each merge; ties go left.

mapping_ is {col: {original_level: group_label}}. Unknown levels at transform time map to other_label. Transform input must contain every fitted mapping column; arrays use x0, x1, etc. Unmatched column names raise ValueError.

Configured exposure and claim-count columns must exist at fit and contain finite values; exposures must be at least one day (1/366, with day-count tolerance), and claim counts must be non-negative. Without exposure_col, grouping uses unit row weights. target_col contains aggregate claims; when absent, y supplies rates, multiplied by exposure before pooling. Supplied outcomes must be finite after numeric and unit conversion; missing outcomes are never treated as zero.

Attributes

mapping_, category_dtypes_, group_cols_, n_features_in_, feature_names_in_.