[MNT] [ENH] Several updates in direct statsforecast interface estimators

pyproject.toml

@@ -103,7 +103,7 @@ all_extras = [
   "seaborn>=0.11",
   "seasonal",
   "skpro<2.2.0,>=2",
-  'statsforecast<1.7.0,>=0.5.2; python_version < "3.12"',
+  'statsforecast<1.8.0,>=1.0.0; python_version < "3.12"',
   "statsmodels>=0.12.1",
   'stumpy>=1.5.1; python_version < "3.11"',
   'tbats>=1.1; python_version < "3.12"',
@@ -138,7 +138,7 @@ all_extras_pandas2 = [
   "seaborn>=0.11",
   "seasonal",
   "skpro<2.2.0,>=2",
-  'statsforecast<1.7.0,>=0.5.2; python_version < "3.12"',
+  'statsforecast<1.8.0,>=1.0.0; python_version < "3.12"',
   "statsmodels>=0.12.1",
   'stumpy>=1.5.1; python_version < "3.11"',
   'tbats>=1.1; python_version < "3.12"',

sktime/forecasting/arch/_statsforecast_arch.py

@@ -42,6 +42,7 @@ class StatsForecastGARCH(_GeneralisedStatsForecastAdapter):
         "ignores-exogeneous-X": False,
         "capability:pred_int": True,
         "capability:pred_int:insample": True,
+        "python_dependencies": ["statsforecast>=1.5.0"],
     }
 
     def __init__(
@@ -102,9 +103,16 @@ class StatsForecastARCH(_GeneralisedStatsForecastAdapter):
     """
 
     _tags = {
+        # packaging info
+        # --------------
+        "authors": ["eyjo"],
+        "maintainers": ["eyjo"],
+        # estimator type
+        # --------------
         "ignores-exogeneous-X": False,
         "capability:pred_int": True,
         "capability:pred_int:insample": True,
+        "python_dependencies": ["statsforecast>=1.5.0"],
     }
 
     def __init__(

sktime/forecasting/base/adapters/_generalised_statsforecast.py

@@ -8,8 +8,8 @@
 
 from sktime.forecasting.base import BaseForecaster
 
-__all__ = ["_GeneralisedStatsForecastAdapter"]
-__author__ = ["yarnabrina"]
+__all__ = ["_GeneralisedStatsForecastAdapter", "StatsForecastBackAdapter"]
+__author__ = ["yarnabrina", "arnaujc91", "luca-miniati"]
 
 
 class _GeneralisedStatsForecastAdapter(BaseForecaster):
@@ -18,7 +18,7 @@ class _GeneralisedStatsForecastAdapter(BaseForecaster):
     _tags = {
         # packaging info
         # --------------
-        "authors": ["yarnabrina"],
+        "authors": ["yarnabrina", "arnaujc91"],
         "maintainers": ["yarnabrina"],
         "python_version": ">=3.8",
         "python_dependencies": ["statsforecast"],
@@ -456,6 +456,7 @@ def __init__(self, estimator):
         self.prediction_intervals = None
 
     def __repr__(self):
+        """Representation dunder."""
         return "StatsForecastBackAdapter"
 
     def new(self):
@@ -545,6 +546,7 @@ def predict_in_sample(self, level=None):
         return self.format_pred_int("fitted", fitted, pred_int, coverage, level)
 
     def format_pred_int(self, y_pred_name, y_pred, pred_int, coverage, level):
+        """Convert prediction intervals into a StatsForecast-format dictionary."""
         pred_int_prefix = "fitted-" if y_pred_name == "fitted" else ""
 
         pred_int_no_lev = pred_int.droplevel(0, axis=1)

sktime/forecasting/statsforecast.py

@@ -188,6 +188,7 @@ class StatsForecastAutoARIMA(_GeneralisedStatsForecastAdapter):
         "ignores-exogeneous-X": False,
         "capability:pred_int": True,
         "capability:pred_int:insample": True,
+        "python_dependencies": ["statsforecast>=1.0.0"],
     }
 
     def __init__(
@@ -376,6 +377,7 @@ class StatsForecastAutoTheta(_GeneralisedStatsForecastAdapter):
         "ignores-exogeneous-X": True,
         "capability:pred_int": True,
         "capability:pred_int:insample": True,
+        "python_dependencies": ["statsforecast>=1.3.0"],
     }
 
     def __init__(
@@ -457,6 +459,8 @@ class StatsForecastAutoETS(_GeneralisedStatsForecastAdapter):
         Controlling state-space-equations.
     damped : bool
         A parameter that 'dampens' the trend.
+    phi : float, optional (default=None)
+        Smoothing parameter for trend damping. Only used when `damped=True`.
 
     References
     ----------
@@ -480,14 +484,20 @@ class StatsForecastAutoETS(_GeneralisedStatsForecastAdapter):
         "ignores-exogeneous-X": True,
         "capability:pred_int": True,
         "capability:pred_int:insample": True,
+        "python_dependencies": ["statsforecast>=1.3.2"],
     }
 
     def __init__(
-        self, season_length: int = 1, model: str = "ZZZ", damped: Optional[bool] = None
+        self,
+        season_length: int = 1,
+        model: str = "ZZZ",
+        damped: Optional[bool] = None,
+        phi: Optional[float] = None,
     ):
         self.season_length = season_length
         self.model = model
         self.damped = damped
+        self.phi = phi
 
         super().__init__()
 
@@ -502,6 +512,7 @@ def _get_statsforecast_params(self):
             "season_length": self.season_length,
             "model": self.model,
             "damped": self.damped,
+            "phi": self.phi,
         }
 
     @classmethod
@@ -572,6 +583,7 @@ class StatsForecastAutoCES(_GeneralisedStatsForecastAdapter):
         "ignores-exogeneous-X": True,
         "capability:pred_int": True,
         "capability:pred_int:insample": True,
+        "python_dependencies": ["statsforecast>=1.1.0"],
     }
 
     def __init__(self, season_length: int = 1, model: str = "Z"):
@@ -784,8 +796,9 @@ class StatsForecastMSTL(_GeneralisedStatsForecastAdapter):
         # estimator type
         # --------------
         "ignores-exogeneous-X": True,
-        "capability:pred_int": True,
-        "capability:pred_int:insample": True,
+        "capability:pred_int": False,
+        "capability:pred_int:insample": False,
+        "python_dependencies": ["statsforecast>=1.2.0"],
     }
 
     def __init__(
@@ -795,18 +808,27 @@ def __init__(
         stl_kwargs: Optional[Dict] = None,
         pred_int_kwargs: Optional[Dict] = None,
     ):
+        self.season_length = season_length
+        self.trend_forecaster = trend_forecaster
+        self.stl_kwargs = stl_kwargs
+        self.pred_int_kwargs = pred_int_kwargs
+
         super().__init__()
 
+        # adapter class sets probabilistic capability as true
+        # because level is present in statsforecast signature
+        # happens in _check_supports_pred_int method
+        # manually overriding this temporarily
+        self.set_tags(
+            **{"capability:pred_int": False, "capability:pred_int:insample": False}
+        )
+
         from sklearn.base import clone
 
-        self.trend_forecaster = trend_forecaster
-        self.season_length = season_length
         if trend_forecaster:
             self._trend_forecaster = clone(trend_forecaster)
         else:
             self._trend_forecaster = StatsForecastAutoETS(model="ZZN")
-        self.stl_kwargs = stl_kwargs
-        self.pred_int_kwargs = pred_int_kwargs
 
         # checks if trend_forecaster is already wrapped with
         # StatsForecastBackAdapter

sktime/forecasting/tests/test_statsforecast.py

@@ -14,6 +14,9 @@
 from sktime.tests.test_switch import run_test_for_class
 
 
+@pytest.mark.skip(
+    reason="probabilistic capability of StatsForecastMSTL is disabled, see #5703, #5920"
+)
 @pytest.mark.skipif(
     not run_test_for_class(StatsForecastMSTL),
     reason="run test only if softdeps are present and incrementally (if requested)",

sktime/tests/_config.py

@@ -189,6 +189,8 @@
     "LTSFDLinearForecaster": ["test_predict_time_index_in_sample_full"],
     "LTSFNLinearForecaster": ["test_predict_time_index_in_sample_full"],
     "WEASEL": ["test_multiprocessing_idempotent"],  # see 5658
+    # StatsForecastMSTL is failing in probabistic forecasts, see #5703, #5920
+    "StatsForecastMSTL": ["test_pred_int_tag", "test_statsforecast_mstl"],
 }
 
 # We use estimator tags in addition to class hierarchies to further distinguish