Francesco Bruzzesi

Hi @hrzn, I would like to propose a more general approach, which I find particularly useful for retraining. As a naive example to justify this, suppose that you want to...

Thanks for the feedback @hrzn, I completely agree that allowing for both implementations (`int` and `callable`) would lead to simplicity and flexibility at the same time, and require a contained...

> Do you think we need `counter: int` as a parameter? It seems to me this case would be covered by the case where `retrain `is an `int `already. Yes...

@chefPony > Am I missing something or you would still need to pass all the arguments to _retrain_wrapper inside historical_forecasts? Yes correct, that's exactly what I was thinking, passing all...

Adding my two cents on the discussion: > Not every classifier also implements a predict_proba method. Can be overcome by checking availability of `.predict_proba()` method with [sklearn.utils.metaestimators.available_if](https://scikit-learn.org/stable/modules/generated/sklearn.utils.metaestimators.available_if.html) > Can't we...

I probably wasn't clear in my reasoning. Suppse you have a `ZeroInflatedRegressor` that has a classifier which implements `.predict_proba()`. If I interpret this correctly: > I propose that we implement...

Sure! Let me try to write a pseudo code for what _could_ be an idea: ```py class ZeroInflatedRegressor(...): def __init__(self, classifier, regressor, predict_risk = False): ... def fit(self, X, y):...

Oooh you are right! If the model is _good_ then these `expected_impact` values should be multiplied by tiny probabilities and get close to zero! For the interface, as @koaning suggested,...

Hi, thank you for reporting the issue and sorry for the late reply! This seems to be unexpected/buggy behavior. My understanding is that some assumptions have been made in the...

It may be an overkill, but _a way_ to do it is to allow to specify the _index_ or _name_ of the argument that is the dataframe to track and...