Fixes bug when using historical forecast with discrete lags < -1

darts/tests/utils/historical_forecasts/test_historical_forecasts.py

@@ -72,6 +72,8 @@
         {},
         (5, 3),
     ),
+    (LinearRegressionModel, {"lags": [-5]}, {}, (5, 1)),
+    (LinearRegressionModel, {"lags": [-5], "output_chunk_shift": 1}, {}, (5, 2)),
 ]
 if not isinstance(CatBoostModel, NotImportedModule):
     models_reg_no_cov_cls_kwargs.append((
@@ -132,6 +134,12 @@
         {},
         (6, 2),
     ),
+    (
+        LinearRegressionModel,
+        {"lags": [-3], "lags_past_covariates": 6, "lags_future_covariates": [0, 1]},
+        {},
+        (6, 2),
+    ),
 ]
 
 if TORCH_AVAILABLE:
@@ -665,7 +673,7 @@ def test_historical_forecasts_negative_rangeindex(self):
     def test_historical_forecasts(self, config):
         """Tests historical forecasts with retraining for expected forecast lengths and times"""
         forecast_horizon = 8
-        # if no fit and retrain=false, should fit at fist iteration
+        # if no fit and retrain=false, should fit at first iteration
         model_cls, kwargs, model_kwarg, bounds = config
         model = model_cls(**kwargs, **model_kwarg)
         # set train length to be the minimum required training length

darts/utils/historical_forecasts/utils.py

@@ -1020,7 +1020,13 @@ def _get_historical_forecast_boundaries(
     hist_fct_tgt_start, hist_fct_tgt_end = historical_forecasts_time_index
     if min_target_lag is not None:
         hist_fct_tgt_start += min_target_lag * freq
-    hist_fct_tgt_end -= 1 * freq
+
+    # target lag has a gap between the max lag and the present
+    if hasattr(model, "lags") and model._get_lags("target") and not overlap_end:
+        hist_fct_tgt_end += 1 * freq * model._get_lags("target")[-1]
+    else:
+        hist_fct_tgt_end -= 1 * freq
+
     # past lags are <= 0
     hist_fct_pc_start, hist_fct_pc_end = historical_forecasts_time_index
     if min_past_cov_lag is not None:

testing2.py

@@ -0,0 +1,59 @@
+from itertools import product
+
+from sklearn.preprocessing import MaxAbsScaler
+
+from darts.dataprocessing.transformers import Scaler
+from darts.models import LinearRegressionModel
+from darts.tests.utils.historical_forecasts.test_historical_forecasts import (
+    TestHistoricalforecast,
+)
+from darts.utils import timeseries_generation as tg
+
+tester = TestHistoricalforecast()
+
+sine_univariate1 = tg.sine_timeseries(length=50) * 2 + 1.5
+sine_univariate2 = tg.sine_timeseries(length=50, value_phase=1.5705) * 5 + 1.5
+sine_univariate3 = tg.sine_timeseries(length=50, value_phase=0.1963125) * -9 + 1.5
+
+params = product(
+    [
+        (
+            {
+                "series": sine_univariate1 - 11,
+            },
+            {"series": Scaler(scaler=MaxAbsScaler())},
+        ),
+        (
+            {
+                "series": sine_univariate3 + 2,
+                "past_covariates": sine_univariate1 * 3 + 3,
+            },
+            {"past_covariates": Scaler()},
+        ),
+        (
+            {
+                "series": sine_univariate3 + 5,
+                "future_covariates": sine_univariate1 * (-4) + 3,
+            },
+            {"future_covariates": Scaler(scaler=MaxAbsScaler())},
+        ),
+        (
+            {
+                "series": sine_univariate3 * 2 + 7,
+                "past_covariates": sine_univariate1 + 2,
+                "future_covariates": sine_univariate2 + 3,
+            },
+            {"series": Scaler(), "past_covariates": Scaler()},
+        ),
+    ],
+    [True, False],  # retrain
+    [True, False],  # last point only
+    [LinearRegressionModel],
+)
+
+i = 0
+
+for param in params:
+    if i == 2 or i == 3:
+        tester.test_historical_forecasts_with_scaler(params=param)
+    i += 1