Fix/optimized historical forecast with component specific lags (#2040)

* fix: properly call model._get_lags instead of model.lags.get to account for component specific lags

* updated changelog

* add unit tests

---------

Co-authored-by: dennisbader <dennis.bader@gmx.ch>

CHANGELOG.md

@@ -9,14 +9,16 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
 [Full Changelog](https://github.com/unit8co/darts/compare/0.26.0...master)
 
 ### For users of the library:
+
 **Improved**
 - Improvements to `TorchForecastingModel`:
   - Added callback `darts.utils.callbacks.TFMProgressBar` to customize at which model stages to display the progress bar. [#2020](https://github.com/unit8co/darts/pull/2020) by [Dennis Bader](https://github.com/dennisbader).
 - Improvements to documentation:
   - Adapted the example notebooks to properly apply data transformers and avoid look-ahead bias. [#2020](https://github.com/unit8co/darts/pull/2020) by [Samriddhi Singh](https://github.com/SimTheGreat).
 
 **Fixed**
-- Fixed a bug when trying to divide `pd.Timedelta` by a `pd.Offset` with an ambiguous conversion to `pd.Timedelta` when using encoders. [#2034](https://github.com/unit8co/darts/pull/2034) by [Antoine Madrona](https://github.com/madtoinou).
+- Fixed a bug when calling optimized `historical_forecasts()` for a `RegressionModel` trained with unequal component-specific lags. [#2040](https://github.com/unit8co/darts/pull/2040) by [Antoine Madrona](https://github.com/madtoinou).
+- Fixed a bug when using encoders with `RegressionModel` and series with a non-evenly spaced frequency (e.g. Month Begin). This raised an error during lagged data creation when trying to divide a pd.Timedelta by the ambiguous frequency. [#2034](https://github.com/unit8co/darts/pull/2034) by [Antoine Madrona](https://github.com/madtoinou).
 
 ### For developers of the library:
 

darts/models/forecasting/regression_model.py

@@ -223,6 +223,7 @@ def encode_year(idx):
         )
 
         # convert lags arguments to list of int
+        # lags attribute should always be accessed with self._get_lags(), not self.lags.get()
         self.lags, self.component_lags = self._generate_lags(
             lags=lags,
             lags_past_covariates=lags_past_covariates,
@@ -373,7 +374,7 @@ def _get_lags(self, lags_type: str):
         if lags_type in self.component_lags:
             return self.component_lags[lags_type]
         else:
-            return self.lags.get(lags_type)
+            return self.lags.get(lags_type, None)
 
     @property
     def _model_encoder_settings(

darts/tests/models/forecasting/test_historical_forecasts.py

@@ -952,6 +952,57 @@ def test_optimized_historical_forecasts_regression_with_encoders(self, config):
             assert (hfc.time_index == ohfc.time_index).all()
             np.testing.assert_array_almost_equal(hfc.all_values(), ohfc.all_values())
 
+    def test_optimized_historical_forecasts_regression_with_component_specific_lags(
+        self,
+    ):
+        horizon = 1
+        lags = 3
+        len_val_series = 10
+        series_train, series_val = (
+            self.ts_pass_train[:10],
+            self.ts_pass_val[:len_val_series],
+        )
+        model = LinearRegressionModel(
+            lags=lags,
+            lags_past_covariates={"default_lags": 2, "darts_enc_pc_dta_dayofweek": 1},
+            lags_future_covariates=[2, 3],
+            add_encoders={
+                "cyclic": {"future": ["month"]},
+                "datetime_attribute": {"past": ["dayofweek"]},
+            },
+        )
+        model.fit(series_train)
+        hist_fct = model.historical_forecasts(
+            series=series_val,
+            retrain=False,
+            enable_optimization=False,
+        )
+
+        opti_hist_fct = model._optimized_historical_forecasts(series=[series_val])
+
+        if not isinstance(hist_fct, list):
+            hist_fct = [hist_fct]
+            opti_hist_fct = [opti_hist_fct]
+
+        n_pred_series_expected = 1
+        n_pred_points_expected = len(series_val) - lags - horizon + 1
+        first_ts_expected = (
+            series_val.time_index[lags] + (horizon - 1) * series_val.freq
+        )
+        last_ts_expected = series_val.end_time()
+
+        # check length match between optimized and default hist fc
+        assert len(opti_hist_fct) == n_pred_series_expected
+        assert len(hist_fct) == len(opti_hist_fct)
+        # check hist fc start
+        assert opti_hist_fct[0].start_time() == first_ts_expected
+        # check hist fc end
+        assert opti_hist_fct[-1].end_time() == last_ts_expected
+        for hfc, ohfc in zip(hist_fct, opti_hist_fct):
+            assert len(ohfc) == n_pred_points_expected
+            assert (hfc.time_index == ohfc.time_index).all()
+            np.testing.assert_array_almost_equal(hfc.all_values(), ohfc.all_values())
+
     @pytest.mark.slow
     @pytest.mark.skipif(not TORCH_AVAILABLE, reason="requires torch")
     @pytest.mark.parametrize("model_config", models_torch_cls_kwargs)

darts/utils/historical_forecasts/optimized_historical_forecasts.py

@@ -93,17 +93,17 @@ def _optimized_historical_forecasts_regression_last_points_only(
 
         X, times = create_lagged_prediction_data(
             target_series=None
-            if len(model.lags.get("target", [])) == 0
+            if model._get_lags("target") is None
             else series_[hist_fct_tgt_start:hist_fct_tgt_end],
             past_covariates=None
             if past_covariates_ is None
             else past_covariates_[hist_fct_pc_start:hist_fct_pc_end],
             future_covariates=None
             if future_covariates_ is None
             else future_covariates_[hist_fct_fc_start:hist_fct_fc_end],
-            lags=model.lags.get("target", None),
-            lags_past_covariates=model.lags.get("past", None),
-            lags_future_covariates=model.lags.get("future", None),
+            lags=model._get_lags("target"),
+            lags_past_covariates=model._get_lags("past"),
+            lags_future_covariates=model._get_lags("future"),
             uses_static_covariates=model.uses_static_covariates,
             last_static_covariates_shape=model._static_covariates_shape,
             max_samples_per_ts=None,
@@ -238,17 +238,17 @@ def _optimized_historical_forecasts_regression_all_points(
 
         X, _ = create_lagged_prediction_data(
             target_series=None
-            if len(model.lags.get("target", [])) == 0
+            if model._get_lags("target") is None
             else series_[hist_fct_tgt_start:hist_fct_tgt_end],
             past_covariates=None
             if past_covariates_ is None
             else past_covariates_[hist_fct_pc_start:hist_fct_pc_end],
             future_covariates=None
             if future_covariates_ is None
             else future_covariates_[hist_fct_fc_start:hist_fct_fc_end],
-            lags=model.lags.get("target", None),
-            lags_past_covariates=model.lags.get("past", None),
-            lags_future_covariates=model.lags.get("future", None),
+            lags=model._get_lags("target"),
+            lags_past_covariates=model._get_lags("past"),
+            lags_future_covariates=model._get_lags("future"),
             uses_static_covariates=model.uses_static_covariates,
             last_static_covariates_shape=model._static_covariates_shape,
             max_samples_per_ts=None,