BUG: Fix Rolling where duplicate datetimelike indexes are treated as consecutive rather than equal with closed='left' and closed='neither' (#54917)

* Add bugfix for rolling window with nonunique datetimelike index

* Run black

* Add entry to whatsnew

* Fix VariableOffsetWindowIndexer

* Simplify change in indexers.pyx

* Add test

Author: ihsansecer

doc/source/whatsnew/v2.2.0.rst

@@ -168,6 +168,7 @@ Performance improvements
 Bug fixes
 ~~~~~~~~~
 - Bug in :class:`AbstractHolidayCalendar` where timezone data was not propagated when computing holiday observances (:issue:`54580`)
+- Bug in :class:`pandas.core.window.Rolling` where duplicate datetimelike indexes are treated as consecutive rather than equal with ``closed='left'`` and ``closed='neither'`` (:issue:`20712`)
 
 Categorical
 ^^^^^^^^^^^

pandas/_libs/window/indexers.pyx

@@ -138,6 +138,8 @@ def calculate_variable_window_bounds(
                         break
             # end bound is previous end
             # or current index
+            elif index[end[i - 1]] == end_bound and not right_closed:
+                end[i] = end[i - 1] + 1
             elif (index[end[i - 1]] - end_bound) * index_growth_sign <= 0:
                 end[i] = i + 1
             else:

pandas/core/indexers/objects.py

@@ -262,7 +262,9 @@ def get_window_bounds(
             # end bound is previous end
             # or current index
             end_diff = (self.index[end[i - 1]] - end_bound) * index_growth_sign
-            if end_diff <= zero:
+            if end_diff == zero and not right_closed:
+                end[i] = end[i - 1] + 1
+            elif end_diff <= zero:
                 end[i] = i + 1
             else:
                 end[i] = end[i - 1]

pandas/tests/window/test_groupby.py

@@ -466,20 +466,23 @@ def test_groupby_rolling_subset_with_closed(self):
         # GH 35549
         df = DataFrame(
             {
-                "column1": range(6),
-                "column2": range(6),
-                "group": 3 * ["A", "B"],
-                "date": [Timestamp("2019-01-01")] * 6,
+                "column1": range(8),
+                "column2": range(8),
+                "group": ["A"] * 4 + ["B"] * 4,
+                "date": [
+                    Timestamp(date)
+                    for date in ["2019-01-01", "2019-01-01", "2019-01-02", "2019-01-02"]
+                ]
+                * 2,
             }
         )
         result = (
             df.groupby("group").rolling("1D", on="date", closed="left")["column1"].sum()
         )
         expected = Series(
-            [np.nan, 0.0, 2.0, np.nan, 1.0, 4.0],
-            index=MultiIndex.from_tuples(
-                [("A", Timestamp("2019-01-01"))] * 3
-                + [("B", Timestamp("2019-01-01"))] * 3,
+            [np.nan, np.nan, 1.0, 1.0, np.nan, np.nan, 9.0, 9.0],
+            index=MultiIndex.from_frame(
+                df[["group", "date"]],
                 names=["group", "date"],
             ),
             name="column1",
@@ -490,10 +493,14 @@ def test_groupby_subset_rolling_subset_with_closed(self):
         # GH 35549
         df = DataFrame(
             {
-                "column1": range(6),
-                "column2": range(6),
-                "group": 3 * ["A", "B"],
-                "date": [Timestamp("2019-01-01")] * 6,
+                "column1": range(8),
+                "column2": range(8),
+                "group": ["A"] * 4 + ["B"] * 4,
+                "date": [
+                    Timestamp(date)
+                    for date in ["2019-01-01", "2019-01-01", "2019-01-02", "2019-01-02"]
+                ]
+                * 2,
             }
         )
 
@@ -503,10 +510,9 @@ def test_groupby_subset_rolling_subset_with_closed(self):
             .sum()
         )
         expected = Series(
-            [np.nan, 0.0, 2.0, np.nan, 1.0, 4.0],
-            index=MultiIndex.from_tuples(
-                [("A", Timestamp("2019-01-01"))] * 3
-                + [("B", Timestamp("2019-01-01"))] * 3,
+            [np.nan, np.nan, 1.0, 1.0, np.nan, np.nan, 9.0, 9.0],
+            index=MultiIndex.from_frame(
+                df[["group", "date"]],
                 names=["group", "date"],
             ),
             name="column1",

pandas/tests/window/test_rolling.py

@@ -304,6 +304,76 @@ def test_datetimelike_nonunique_index_centering(
     tm.assert_equal(result, expected)
 
 
+@pytest.mark.parametrize(
+    "closed,expected",
+    [
+        ("left", [np.nan, np.nan, 1, 1, 1, 10, 14, 14, 18, 21]),
+        ("neither", [np.nan, np.nan, 1, 1, 1, 9, 5, 5, 13, 8]),
+        ("right", [0, 1, 3, 6, 10, 14, 11, 18, 21, 17]),
+        ("both", [0, 1, 3, 6, 10, 15, 20, 27, 26, 30]),
+    ],
+)
+def test_variable_window_nonunique(closed, expected, frame_or_series):
+    # GH 20712
+    index = DatetimeIndex(
+        [
+            "2011-01-01",
+            "2011-01-01",
+            "2011-01-02",
+            "2011-01-02",
+            "2011-01-02",
+            "2011-01-03",
+            "2011-01-04",
+            "2011-01-04",
+            "2011-01-05",
+            "2011-01-06",
+        ]
+    )
+
+    df = frame_or_series(range(10), index=index, dtype=float)
+    expected = frame_or_series(expected, index=index, dtype=float)
+
+    result = df.rolling("2D", closed=closed).sum()
+
+    tm.assert_equal(result, expected)
+
+
+@pytest.mark.parametrize(
+    "closed,expected",
+    [
+        ("left", [np.nan, np.nan, 1, 1, 1, 10, 15, 15, 18, 21]),
+        ("neither", [np.nan, np.nan, 1, 1, 1, 10, 15, 15, 13, 8]),
+        ("right", [0, 1, 3, 6, 10, 15, 21, 28, 21, 17]),
+        ("both", [0, 1, 3, 6, 10, 15, 21, 28, 26, 30]),
+    ],
+)
+def test_variable_offset_window_nonunique(closed, expected, frame_or_series):
+    # GH 20712
+    index = DatetimeIndex(
+        [
+            "2011-01-01",
+            "2011-01-01",
+            "2011-01-02",
+            "2011-01-02",
+            "2011-01-02",
+            "2011-01-03",
+            "2011-01-04",
+            "2011-01-04",
+            "2011-01-05",
+            "2011-01-06",
+        ]
+    )
+
+    df = frame_or_series(range(10), index=index, dtype=float)
+    expected = frame_or_series(expected, index=index, dtype=float)
+
+    offset = BusinessDay(2)
+    indexer = VariableOffsetWindowIndexer(index=index, offset=offset)
+    result = df.rolling(indexer, closed=closed, min_periods=1).sum()
+
+    tm.assert_equal(result, expected)
+
+
 def test_even_number_window_alignment():
     # see discussion in GH 38780
     s = Series(range(3), index=date_range(start="2020-01-01", freq="D", periods=3))