Example script of use:

import bambi as bmb
import numpy as np
import pandas as pd

rng = np.random.default_rng(0)

size = 2_000
x = rng.normal(size=size)
data = pd.DataFrame(
    {
        "x": x,
        "y": rng.normal(loc=x, size=size)
    }
)

bmb_model = bmb.Model("y ~ x", data)
idata = bmb_model.fit()

test_size = 100
test_x = rng.normal(size=test_size)
test_data = pd.DataFrame(
    {
        "x": test_x,
        "y": rng.normal(loc=test_x, size=test_size)
    }
)

predictions = bmb_model.predict_sklearn(idata, "y", test_data)

Thanks for moving this issue so quickly. I would like to add the following.

• I'm in favor of having a utility to predict using point estimates. I'm sure many will find this useful.
• I'm not sure if this should be a method or a function.
  • I hesitate because I don't know how good is to have two predict* methods.
• I'm not in favor of using predict_sklearn as the name. I know it could be super clear for those coming with a sklearn background, but I think it will be confusing for all the rest. I think predict_point_estimate is a better name because it's clearer, although it's quite long.

Apart from that, if it is a method in Model we don't need the name of the variable we want to predict. We can recover it from the model itself. We can use self.response.name instead of predict_variable in return idata.posterior_predictive[predict_variable].mean(("chain", "draw")).values

I still think that the best approach is to show how to work with InferenceData and not add new functions/method for this case. And if necessary improve InferenceData.

But following this proposal. I agree that having two predict methods is not good design. Having and utility function could work. Maybe something like compute_point, or aggregate_predictions or something similar. Or hopefully a more obvious name. It should be clear from the name that the function works with predictions. Otherwise people may want to use it to compute point estimates from posteriors parameters or whatever.

Yet another option is to add options or arguments to predict. But that will probably make that method a bit clunky and less clean than now.

What other point estimates we want? mean and median seems like immediate candidates, but what about other stuff like trimmed mean (or similar)...

BTW, i think most if not all the friction could be solved by adding a scikitlearn migration guide, or intro to bambi for scikitlearn users.

Hi @markgoodhead what do you think about my proposal of writing a scikitlearn migration guide? Are you interested in working on it?

It seems we don't have interest in implementing such a method. It would be nice to have a migration guide for those coming from scikit learn. See #729