## kernels.MLDevelopmentSpec


What to fit: the estimator, its settings, the response and the design.


Usage

``` python
kernels.MLDevelopmentSpec(
    estimator,
    seed=0,
    n_estimators=100,
    max_depth=None,
    min_samples_leaf=None,
    learning_rate=None,
    response="incremental",
    origin="factor",
    calendar="none",
    zero_cells="observed",
    unsupported_factor="raise"
)
```


## Attributes


`estimator: str`  
`"random_forest"` (scikit-learn's `RandomForestRegressor`) or `"gradient_boosting"` (`GradientBoostingRegressor`). No default: the two answers are far apart (52% on one Schedule P triangle).

`seed: int`  
Passed to scikit-learn as `random_state`, unchanged: a whole number from 0 to `SEED_MAX`. `None` is refused, because a forest with no seed gives a different answer on every call.

`n_estimators: int`  
Trees in the forest, or boosting stages (100).

`max_depth: int or None`  
The deepest a tree may grow; `None` is unlimited for the forest and 3 for boosting, scikit-learn's defaults.

`min_samples_leaf: int or None`  
The forest only: the fewest rows in a leaf; `None` is 1. Refused with `gradient_boosting`, where it would change nothing here.

`learning_rate: float or None`  
Boosting only: the shrinkage of each stage, above 0; `None` is 0.1. Refused with `random_forest`, which has no such setting.

`response: str`  
`"incremental"` (the default) models each cell's increment; `"cumulative"` models its cumulative.

`origin: str`  
`"factor"` (the default): an indicator per origin. `"none"`: no origin term, so every origin shares one fitted pattern.

`calendar: str`  
`"none"` (the default) or `"trend"`: the calendar period, counted in development steps from the first origin's first cell, as one numeric column (chainladder-python's `valuation`).

`zero_cells: str`  
`"observed"` (the default) trains on every observed cell, a zero cumulative included. `"missing"` leaves out every cell whose cumulative is zero, as chainladder-python does.

`unsupported_factor: str`  
Under `zero_cells="missing"`, what to do at a development age whose cells are all zero, which leaves the model nothing to fit there: `"raise"` (the default) refuses; `"unity"` develops by a factor of 1 into that age (a fitted increment of 0).


## Attributes

| Name | Description |
|----|----|
| [settings](#settings) | The estimator's settings as scikit-learn gets them, `None` filled in. |

------------------------------------------------------------------------


#### settings


The estimator's settings as scikit-learn gets them, `None` filled in.


`settings: dict`


`max_depth` stays `None` for an unlimited forest; `min_samples_leaf` is `None` for boosting and `learning_rate` for the forest.
