Feat/hist fc start stride (#2560)

* improve hist fc start point

* add tests

* update documentation

* update changelog

* clean up code

* fix tests

* fix missed lines

* improve codecov

CHANGELOG.md

@@ -11,6 +11,8 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
 
 **Improved**
 
+- Improvements to `ForecastingModel`: Improved `start` handling for historical forecasts, backtest, residuals, and gridsearch. If `start` is not within the trainable / forecastable points, uses the closest valid start point that is a round multiple of `stride` ahead of start. Raises a ValueError, if no valid start point exists. This guarantees that all historical forecasts are `n * stride` points away from start, and will simplify many downstream tasks. [#2560](https://github.com/unit8co/darts/issues/2560) by [Dennis Bader](https://github.com/dennisbader).
+
 **Fixed**
 
 - Fixed a bug when using `darts.utils.data.tabularization.create_lagged_component_names()` with target `lags=None`, that did not return any lagged target label component names. [#2576](https://github.com/unit8co/darts/pull/2576) by [Dennis Bader](https://github.com/dennisbader).

darts/models/forecasting/forecasting_model.py

@@ -706,11 +706,12 @@ def historical_forecasts(
               or `retrain` is a Callable and the first trainable point is earlier than the first predictable point.
             - the first trainable point (given `train_length`) otherwise
 
+            Note: If `start` is not within the trainable / forecastable points, uses the closest valid start point that
+              is a round multiple of `stride` ahead of `start`. Raises a `ValueError`, if no valid start point exists.
             Note: If the model uses a shifted output (`output_chunk_shift > 0`), then the first predicted point is also
-            shifted by `output_chunk_shift` points into the future.
-            Note: Raises a ValueError if `start` yields a time outside the time index of `series`.
+              shifted by `output_chunk_shift` points into the future.
             Note: If `start` is outside the possible historical forecasting times, will ignore the parameter
-            (default behavior with ``None``) and start at the first trainable/predictable point.
+              (default behavior with ``None``) and start at the first trainable/predictable point.
         start_format
             Defines the `start` format. Only effective when `start` is an integer and `series` is indexed with a
             `pd.RangeIndex`.
@@ -1018,6 +1019,7 @@ def retrain_func(
                 historical_forecasts_time_index=historical_forecasts_time_index,
                 start=start,
                 start_format=start_format,
+                stride=stride,
                 show_warnings=show_warnings,
             )
 
@@ -1267,9 +1269,12 @@ def backtest(
               or `retrain` is a Callable and the first trainable point is earlier than the first predictable point.
             - the first trainable point (given `train_length`) otherwise
 
-            Note: Raises a ValueError if `start` yields a time outside the time index of `series`.
+            Note: If `start` is not within the trainable / forecastable points, uses the closest valid start point that
+              is a round multiple of `stride` ahead of `start`. Raises a `ValueError`, if no valid start point exists.
+            Note: If the model uses a shifted output (`output_chunk_shift > 0`), then the first predicted point is also
+              shifted by `output_chunk_shift` points into the future.
             Note: If `start` is outside the possible historical forecasting times, will ignore the parameter
-            (default behavior with ``None``) and start at the first trainable/predictable point.
+              (default behavior with ``None``) and start at the first trainable/predictable point.
         start_format
             Defines the `start` format. Only effective when `start` is an integer and `series` is indexed with a
             `pd.RangeIndex`.
@@ -1628,9 +1633,12 @@ def gridsearch(
               or `retrain` is a Callable and the first trainable point is earlier than the first predictable point.
             - the first trainable point (given `train_length`) otherwise
 
-            Note: Raises a ValueError if `start` yields a time outside the time index of `series`.
+            Note: If `start` is not within the trainable / forecastable points, uses the closest valid start point that
+              is a round multiple of `stride` ahead of `start`. Raises a `ValueError`, if no valid start point exists.
+            Note: If the model uses a shifted output (`output_chunk_shift > 0`), then the first predicted point is also
+              shifted by `output_chunk_shift` points into the future.
             Note: If `start` is outside the possible historical forecasting times, will ignore the parameter
-            (default behavior with ``None``) and start at the first trainable/predictable point.
+              (default behavior with ``None``) and start at the first trainable/predictable point.
         start_format
             Only used in expanding window mode. Defines the `start` format. Only effective when `start` is an integer
             and `series` is indexed with a `pd.RangeIndex`.
@@ -1924,9 +1932,12 @@ def residuals(
               or `retrain` is a Callable and the first trainable point is earlier than the first predictable point.
             - the first trainable point (given `train_length`) otherwise
 
-            Note: Raises a ValueError if `start` yields a time outside the time index of `series`.
+            Note: If `start` is not within the trainable / forecastable points, uses the closest valid start point that
+              is a round multiple of `stride` ahead of `start`. Raises a `ValueError`, if no valid start point exists.
+            Note: If the model uses a shifted output (`output_chunk_shift > 0`), then the first predicted point is also
+              shifted by `output_chunk_shift` points into the future.
             Note: If `start` is outside the possible historical forecasting times, will ignore the parameter
-            (default behavior with ``None``) and start at the first trainable/predictable point.
+              (default behavior with ``None``) and start at the first trainable/predictable point.
         start_format
             Defines the `start` format. Only effective when `start` is an integer and `series` is indexed with a
             `pd.RangeIndex`.

darts/tests/models/forecasting/test_backtesting.py

@@ -1,4 +1,5 @@
 import itertools
+import logging
 import random
 from itertools import product
 
@@ -733,8 +734,7 @@ def test_backtest_multiple_series(self):
         assert round(abs(error[0] - expected[0]), 4) == 0
         assert round(abs(error[1] - expected[1]), 4) == 0
 
-    @pytest.mark.skipif(not TORCH_AVAILABLE, reason="requires torch")
-    def test_backtest_regression(self):
+    def test_backtest_regression(self, caplog):
         np.random.seed(4)
 
         gaussian_series = gt(mean=2, length=50)
@@ -804,13 +804,26 @@ def test_backtest_regression(self):
         assert score > 0.9
 
         # Using a too small start value
-        with pytest.raises(ValueError):
-            RandomForest(lags=12).backtest(series=target, start=0, forecast_horizon=3)
+        warning_expected = (
+            "`start` position `{0}` corresponding to time `{1}` is before the first "
+            "predictable/trainable historical forecasting point for series at index: 0. Using the first historical "
+            "forecasting point `2000-01-15 00:00:00` that lies a round-multiple of `stride=1` ahead of `start`. "
+            "To hide these warnings, set `show_warnings=False`."
+        )
+        caplog.clear()
+        with caplog.at_level(logging.WARNING):
+            _ = RandomForest(lags=12).backtest(
+                series=target, start=0, forecast_horizon=3
+            )
+            assert warning_expected.format(0, target.start_time()) in caplog.text
+        caplog.clear()
 
-        with pytest.raises(ValueError):
-            RandomForest(lags=12).backtest(
+        with caplog.at_level(logging.WARNING):
+            _ = RandomForest(lags=12).backtest(
                 series=target, start=0.01, forecast_horizon=3
             )
+            assert warning_expected.format(0.01, target.start_time()) in caplog.text
+        caplog.clear()
 
         # Using RandomForest's start default value
         score = RandomForest(lags=12, random_state=0).backtest(
@@ -939,7 +952,6 @@ def test_gridsearch_metric_score(self):
 
         assert score == recalculated_score, "The metric scores should match"
 
-    @pytest.mark.skipif(not TORCH_AVAILABLE, reason="requires torch")
     def test_gridsearch_random_search(self):
         np.random.seed(1)
 
@@ -958,7 +970,6 @@ def test_gridsearch_random_search(self):
         assert isinstance(result[2], float)
         assert min(param_range) <= result[1]["lags"] <= max(param_range)
 
-    @pytest.mark.skipif(not TORCH_AVAILABLE, reason="requires torch")
     def test_gridsearch_n_random_samples_bad_arguments(self):
         dummy_series = get_dummy_series(ts_length=50)
 
@@ -981,7 +992,6 @@ def test_gridsearch_n_random_samples_bad_arguments(self):
                 params, dummy_series, forecast_horizon=1, n_random_samples=1.5
             )
 
-    @pytest.mark.skipif(not TORCH_AVAILABLE, reason="requires torch")
     def test_gridsearch_n_random_samples(self):
         np.random.seed(1)