BUG: rolling.corr with MultiIndex columns

doc/source/whatsnew/v1.4.0.rst

@@ -408,6 +408,7 @@ Groupby/resample/rolling
 - Bug in :meth:`pandas.DataFrame.rolling` operation along rows (``axis=1``) incorrectly omits columns containing ``float16`` and ``float32`` (:issue:`41779`)
 - Bug in :meth:`Resampler.aggregate` did not allow the use of Named Aggregation (:issue:`32803`)
 - Bug in :meth:`Series.rolling` when the :class:`Series` ``dtype`` was ``Int64`` (:issue:`43016`)
+- Bug in :meth:`DataFrame.rolling.corr` when the :class:`DataFrame` columns was a :class:`MultiIndex` (:issue:`21157`)
 - Bug in :meth:`DataFrame.groupby.rolling` when specifying ``on`` and calling ``__getitem__`` would subsequently return incorrect results (:issue:`43355`)
 
 Reshaping

pandas/core/window/common.py

@@ -83,8 +83,24 @@ def dataframe_from_int_dict(data, frame_template):
                         # mypy needs to know columns is a MultiIndex, Index doesn't
                         # have levels attribute
                         arg2.columns = cast(MultiIndex, arg2.columns)
-                        result.index = MultiIndex.from_product(
-                            arg2.columns.levels + [result_index]
+                        # GH 21157: Equivalent to MultiIndex.from_product(
+                        #  [result_index], <unique combinations of arg2.columns.levels>,
+                        # )
+                        # A normal MultiIndex.from_product will produce too many
+                        # combinations.
+                        result_level = np.tile(
+                            result_index, len(result) // len(result_index)
+                        )
+                        arg2_levels = (
+                            np.repeat(
+                                arg2.columns.get_level_values(i),
+                                len(result) // len(arg2.columns),
+                            )
+                            for i in range(arg2.columns.nlevels)
+                        )
+                        result_names = list(arg2.columns.names) + [result_index.name]
+                        result.index = MultiIndex.from_arrays(
+                            [*arg2_levels, result_level], names=result_names
                         )
                         # GH 34440
                         num_levels = len(result.index.levels)

pandas/tests/window/test_pairwise.py

@@ -222,3 +222,18 @@ def test_cov_mulittindex(self):
         )
 
         tm.assert_frame_equal(result, expected)
+
+    def test_multindex_columns_pairwise_func(self):
+        # GH 21157
+        columns = MultiIndex.from_arrays([["M", "N"], ["P", "Q"]], names=["a", "b"])
+        df = DataFrame(np.ones((5, 2)), columns=columns)
+        result = df.rolling(3).corr()
+        expected = DataFrame(
+            np.nan,
+            index=MultiIndex.from_arrays(
+                [np.repeat(np.arange(5), 2), ["M", "N"] * 5, ["P", "Q"] * 5],
+                names=[None, "a", "b"],
+            ),
+            columns=columns,
+        )
+        tm.assert_frame_equal(result, expected)