[BUG] RegressionEnsembleModel not properly handling future_covariates

[chitxxx]

Describe the bug
.historical_forecasts() in RegressionEnsembleModel is bugged. It returns ValueError: The model has been trained without `future_covariates` variable, but the `future_covariates` parameter provided to `predict()` is not None. When the underlying models uses future_covariates.

To Reproduce
Steps to reproduce the behavior, preferably code snippet.

Expected behavior
I expect the model to feed in the ensemble models, with each underlying model using future_covariates, when using the backtest and historical_forecasts methods

System (please complete the following information):

• Python version: 3.9
• darts version 0.23.1

Additional context
Add any other context about the problem here.

[madtoinou]

Hi,

Can you please share a code snippet to reproduce the error? I am pretty sure that this bug should has been fixed by #1745, which be be part of the next release.

[philippGraf]

i also have an issue
python 3.8
darts 0.24.0

I guess the extreme_lags attributes is not correct:

import pandas as pd
from sklearn.ensemble import RandomForestRegressor

data = pd.DataFrame(
    {"a": [1] * 24 * 7 * 5, "b": [2] * 24 * 7 * 5},
    index=pd.date_range(start="2023-01-01", freq="1H", periods=24 * 7 * 5),
)

from darts import TimeSeries
from darts.models import RandomForest, RegressionModel
from darts.models.forecasting.regression_ensemble_model import RegressionEnsembleModel
from darts.dataprocessing.transformers import Scaler
from darts.metrics import r2_score, mape

TARGET = "a"
COVARIATES = ["b"]

data = TimeSeries.from_dataframe(data)

data = data.add_datetime_attribute("hour", one_hot=True)

COVARIATES = COVARIATES + [f"hour_{i}" for i in range(1, 25)]


def _run():
    model1 = RandomForest(
        lags_future_covariates=[0],
        n_estimators=500,
        add_encoders={
            "datetime_attribute": {"future": ["hour"]},
            "transformer": Scaler(),
            "position": {"future": "relative"},
        },
    )
    print("forest extreme lags: ", model1.extreme_lags)
    model2 = RegressionModel(
        lags_future_covariates=[0],
        add_encoders={
            # "datetime_attribute": {"future": ["hour"]},
            "transformer": Scaler(),
            "position": {"future": "relative"},
        },
    )
    print("linear extreme lags: ", model2.extreme_lags)

    class M(RegressionEnsembleModel):
        """so monkey it's sexy"""

        @property
        def extreme_lags(
            self,
        ):
            return (None, 0, None, None, 0, 0)

    model = M(
        forecasting_models=[model1, model2],
        regression_train_n_points=50,
        regression_model=RandomForestRegressor(),
    )

    print("ensemble extreme lags: ", model.extreme_lags)
    future_covariates = data[[x for x in data.columns if x in COVARIATES]]

    results = model.historical_forecasts(
        series=data[TARGET],
        future_covariates=future_covariates,
        forecast_horizon=1 * 24,
        stride=1 * 24,
        start=24 * 7 * 4,
        train_length=24 * 7 * 4,
        last_points_only=False,
    )

    return results


d = _run()

works. Without the monkey patch it crashes.

[madtoinou]

This indeed looks like a bug, could you please open another issue so that we can track it more easily?