I am playing around with ML models to forecast a time series. I'd like to generate a sklearn pipeline like

from sklearn.preprocessing import FunctionTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler

pipe = Pipeline(
    steps=[
        ("gen", FunctionTransformer(generate_features)),
        ("imp_nan", SimpleImputer(strategy="median")),
        ("scale", StandardScaler()),
        ("reg", keras.wrappers.scikit_learn.KerasRegressor(build_model, epochs=35))
    ]
)

Where generate_features takes in my data X and produces a collection of features for a given timestamp. For example a 5 day rolling standard deviation. However, my target y could have a different shape.

For example, I may want to forecast 5 hours in the future (and so my target will have a date time index with period 5 hours), but my input data could have a frequency of 1 min (and so initially an index with period 1 min that gets transformed into a period of 5 hours). This leads to issues when trying to cross validate. For example, if I use

splitter = TimeSeriesSplit(n_splits=3)
grid = GridSearchCV(pipe, {"gen__kw_args": [{"windows": (2,4,6)},]}, cv=splitter)
grid.fit(X_train, y_train)

I immediately get an error telling me the shape of X_train and y_train are not the same. This makes sense because the splitter needs to be used on both X and y.

I want to do this because I'd like to perform a grid cv on the parameters of my generate_features function (as well as the parameters of my model...)

My Question

Is there a nice way to create a pipeline and perform cross validation on time series feature generation when our input shape of X could be different to our input shape of y?