&yarnabrina [BUG] correct ForecastX behaviour in case of multivaria…

…te `y` (#4719)

For multivariate `y`, the compositor `ForecastX` would always broadcast
over variables of `y`, even if wrapped forecasters were multivariate.

However, `ForecastX` itself should never broadcast over `y`, that should
be left to component forecasters.

Fixed here, this also fixes #4718

Includes a slightly extended test case from the bug report #4718 by
@yarnabrina

sktime/forecasting/compose/_pipeline.py

@@ -1229,6 +1229,7 @@ class ForecastX(BaseForecaster):
     _tags = {
         "X_inner_mtype": SUPPORTED_MTYPES,
         "y_inner_mtype": SUPPORTED_MTYPES,
+        "scitype:y": "both",
         "X-y-must-have-same-index": False,
         "fit_is_empty": False,
         "ignores-exogeneous-X": False,

sktime/forecasting/compose/tests/test_pipeline.py

@@ -480,3 +480,45 @@ def test_forecastx_logic():
 
     # compare that test and comparison case results are equal
     assert np.allclose(y_pred, y_pred_manual)
+
+
+def test_forecastx_attrib_broadcast():
+    """Test ForecastX broadcasting and forecaster attributes."""
+    from sktime.forecasting.compose import ForecastX
+    from sktime.forecasting.naive import NaiveForecaster
+
+    df = pd.DataFrame(
+        {
+            "a": ["series_1", "series_1", "series_1"],
+            "b": pd.to_datetime(["2024-01-01", "2024-01-02", "2024-01-03"]),
+            "c": [1, 2, 3],
+            "d": [4, 5, 6],
+            "e": [7, 8, 9],
+        }
+    )
+    df = df.set_index(["a", "b"])
+
+    model = ForecastX(NaiveForecaster(), NaiveForecaster())
+
+    model_1 = model.clone()
+    model_1.fit(df[["c"]], X=df[["d", "e"]], fh=[1, 2, 3])
+
+    assert hasattr(model_1, "forecaster_X_")
+    assert isinstance(model_1.forecaster_X_, NaiveForecaster)
+    assert model_1.forecaster_X_.is_fitted
+
+    assert hasattr(model_1, "forecaster_y_")
+    assert isinstance(model_1.forecaster_y_, NaiveForecaster)
+    assert model_1.forecaster_y_.is_fitted
+
+    model_2 = model.clone()
+    model_2.fit(df[["c", "d"]], X=df[["e"]], fh=[1, 2, 3])
+    assert hasattr(model_2, "forecaster_X_")
+
+    assert hasattr(model_2, "forecaster_X_")
+    assert isinstance(model_2.forecaster_X_, NaiveForecaster)
+    assert model_2.forecaster_X_.is_fitted
+
+    assert hasattr(model_2, "forecaster_y_")
+    assert isinstance(model_2.forecaster_y_, NaiveForecaster)
+    assert model_2.forecaster_y_.is_fitted