If I am using Pipeline for

models = [make_pipeline(

StandardScaler(),

RandomForestRegressor(random_state=123, n_estimators=1000, max_depth= 5))]

mlf = MLForecast(models=models,

freq='D',

num_threads=6)

If I view the data with

mlf.preprocess(data)

You don't see any change in the pipeline transformation or do you? My question is this: Can the transformation be seen using

mlf.preprocess(data)

or not? or the transformation is done when I use the

fit()

method

What transformation are you expecting? You didn't define any features (lags, lag_transforms, date_features)

For example if I am using this:

models = [make_pipeline(

StandardScaler(),

RandomForestRegressor(random_state=123, n_estimators=1000, max_depth= 5))]

The transformation would be the

StandardScaler(),

which I suppose should be applied to the target variable and that is not how it works. Or should I apply a

lag

, and Pipeline will be applied to that new transformation, which is the lag?

If you define it like that the pipeline will be applied to the features, not the target (that's also how it works with regular scikit-learn models). If you want to transform the target you should use a target transformation, e.g.

from mlforecast.target_transforms import LocalStandardScaler

mlf = MLForecast(
    models=models,
    freq='D',
    target_transforms=[LocalStandardScaler()],
    num_threads=6,
)

Perfect, I have been using it in this last way. Another question, what transformations can I use with make_pipeline? all the ones that are frequently used with sklearn?

Yes, every scikit-learn estimator should work. Note that if you want to see what the pipeline is doing you need to do it in two steps:

X, y = mlf.preprocess(data, return_X_y=True)  # this builds the features and applies target transformations
models[0][:-1].fit_transform(X)  # this applies the steps before the model in the pipeline

Thank you Jose