kernels.ConventionalCandidate
Settings held fixed when a candidate is refitted at successive dates.
Usage
kernels.ConventionalCandidate(
method="cl",
history_periods=None,
average="volume",
drop_high=0,
drop_low=0,
exclude=(),
expected_loss_ratio=None,
decay=None,
horizon=None,
unsupported_factor="raise",
exhausted_exclusions="raise",
zero_cells="observed",
preserve=1,
drop_above=None,
drop_below=None,
exclude_valuations=(),
trim_ties="origin",
n_iters=1,
trend=0.0,
tail=None
)history_periods selects the most recent usable paired origins at EACH age (under zero_cells="missing", the most recent paired origins, a pair left out for a zero cell keeping its place, as chainladder-python counts them). exclude identifies link ratios by (origin, FROM development lag in months). horizon is the final development lag in months, with no tail beyond it; omitted, it is the deepest observed age of this particular fit. Replays require an explicit horizon to keep the target fixed through time.
zero_cells says what a cumulative of exactly zero is. "observed" (the default) keeps it as data: the link ratio INTO a zero is 0 and is used, and the link ratio OUT of a zero is undefined and left out with reason "undefined_ratio". "missing" follows chainladder-python, which stores a zero cell as missing: a link ratio is used only when neither of its two cells is zero, and both the link into a zero and the link out of it are left out with reason "zero_cell". A zero on an origin’s latest diagonal is still its latest amount under either rule, so its chain-ladder ultimate is 0 (chainladder-python leaves that ultimate missing instead).
The link-ratio options (history_periods, exclude, exclude_valuations, drop_above, drop_below, drop_high, drop_low, preserve, trim_ties, exhausted_exclusions and zero_cells) are checked and applied by kernels.links, whose docstring gives the order they run in; link_rules holds them as one ~ibnr.kernels.links.LinkRules. drop_high and drop_low are counts, and True is stored as 1. trim_ties defaults to "origin", the tie rule of 0.7.2. average may also be "regression".
n_iters ("bf" and "gcc" only) is the number of Benktander iterations: 1 is the method itself, and each further one uses the last ultimate as the a priori ultimate, at most MAX_N_ITERS (10,000). trend ("gcc" only) is an annual rate: each origin’s losses are moved to the valuation date’s level by (1 + trend) a year from the end of the origin period before the loss ratios are pooled, and each pooled ratio is brought back to its origin’s level after.
tail (a ~ibnr.kernels.tail.TailSpec) extends the pattern past the last observed age: the link factors are estimated first, the tail is applied to them (replacing the links from its attachment age on, when that is before the last age), and every origin’s development to ultimate then includes the tail factor, the oldest origin’s too. Chain ladder, Bornhuetter-Ferguson, Benktander and Cape Cod all read it through the pattern. A tail cannot be combined with horizon, and its sigma and std_err are Mack settings, refused here.
Parameter Attributes
method: str = "cl"history_periods: int | None = Noneaverage: str = "volume"drop_high: int = 0drop_low: int = 0exclude: tuple[tuple[dt.date, int], …] = ()expected_loss_ratio: float | None = Nonedecay: float | None = Nonehorizon: int | None = Noneunsupported_factor: str = "raise"exhausted_exclusions: str = "raise"zero_cells: str = "observed"preserve: int = 1drop_above: float | None = Nonedrop_below: float | None = Noneexclude_valuations: tuple[dt.date, …] = ()trim_ties: str = "origin"n_iters: int = 1trend: float = 0.0tail: TailSpec | None = None
Attributes
| Name | Description |
|---|---|
| link_rules |
The link-ratio options as one ~ibnr.kernels.links.LinkRules.
|
link_rules
The link-ratio options as one ~ibnr.kernels.links.LinkRules.
link_rules: LinkRules
Built when read rather than stored, so vars(candidate) holds the fields alone and ConventionalCandidate(**vars(candidate)) works.