BUG: GroupBy.ffill()/bfill() do not return NaN values for NaN groups (#36790)

Author: smithto1

doc/source/whatsnew/v1.2.0.rst

@@ -445,6 +445,7 @@ Groupby/resample/rolling
 - Bug in :meth:`Rolling.sum()` returned wrong values when dtypes where mixed between float and integer and axis was equal to one (:issue:`20649`, :issue:`35596`)
 - Bug in :meth:`Rolling.count` returned ``np.nan`` with :class:`pandas.api.indexers.FixedForwardWindowIndexer` as window, ``min_periods=0`` and only missing values in window (:issue:`35579`)
 - Bug where :class:`pandas.core.window.Rolling` produces incorrect window sizes when using a ``PeriodIndex`` (:issue:`34225`)
+- Bug in :meth:`DataFrameGroupBy.ffill` and :meth:`DataFrameGroupBy.bfill` where a ``NaN`` group would return filled values instead of ``NaN`` when ``dropna=True`` (:issue:`34725`)
 - Bug in :meth:`RollingGroupby.count` where a ``ValueError`` was raised when specifying the ``closed`` parameter (:issue:`35869`)
 - Bug in :meth:`DataFrame.groupby.rolling` returning wrong values with partial centered window (:issue:`36040`).
 

pandas/_libs/groupby.pyx

@@ -344,7 +344,7 @@ def group_shift_indexer(int64_t[:] out, const int64_t[:] labels,
 @cython.boundscheck(False)
 def group_fillna_indexer(ndarray[int64_t] out, ndarray[int64_t] labels,
                          ndarray[uint8_t] mask, object direction,
-                         int64_t limit):
+                         int64_t limit, bint dropna):
     """
     Indexes how to fill values forwards or backwards within a group.
 
@@ -358,6 +358,7 @@ def group_fillna_indexer(ndarray[int64_t] out, ndarray[int64_t] labels,
     direction : {'ffill', 'bfill'}
         Direction for fill to be applied (forwards or backwards, respectively)
     limit : Consecutive values to fill before stopping, or -1 for no limit
+    dropna : Flag to indicate if NaN groups should return all NaN values
 
     Notes
     -----
@@ -381,7 +382,9 @@ def group_fillna_indexer(ndarray[int64_t] out, ndarray[int64_t] labels,
     with nogil:
         for i in range(N):
             idx = sorted_labels[i]
-            if mask[idx] == 1:  # is missing
+            if dropna and labels[idx] == -1:  # nan-group gets nan-values
+                curr_fill_idx = -1
+            elif mask[idx] == 1:  # is missing
                 # Stop filling once we've hit the limit
                 if filled_vals >= limit and limit != -1:
                     curr_fill_idx = -1

pandas/core/groupby/groupby.py

@@ -1866,6 +1866,7 @@ def _fill(self, direction, limit=None):
             result_is_index=True,
             direction=direction,
             limit=limit,
+            dropna=self.dropna,
         )
 
     @Substitution(name="groupby")

pandas/tests/groupby/test_missing.py

@@ -82,3 +82,37 @@ def test_fill_consistency():
     expected = df.groupby(level=0, axis=0).fillna(method="ffill")
     result = df.T.groupby(level=0, axis=1).fillna(method="ffill").T
     tm.assert_frame_equal(result, expected)
+
+
+@pytest.mark.parametrize("method", ["ffill", "bfill"])
+@pytest.mark.parametrize("dropna", [True, False])
+@pytest.mark.parametrize("has_nan_group", [True, False])
+def test_ffill_handles_nan_groups(dropna, method, has_nan_group):
+    # GH 34725
+
+    df_without_nan_rows = pd.DataFrame([(1, 0.1), (2, 0.2)])
+
+    ridx = [-1, 0, -1, -1, 1, -1]
+    df = df_without_nan_rows.reindex(ridx).reset_index(drop=True)
+
+    group_b = np.nan if has_nan_group else "b"
+    df["group_col"] = pd.Series(["a"] * 3 + [group_b] * 3)
+
+    grouped = df.groupby(by="group_col", dropna=dropna)
+    result = getattr(grouped, method)(limit=None)
+
+    expected_rows = {
+        ("ffill", True, True): [-1, 0, 0, -1, -1, -1],
+        ("ffill", True, False): [-1, 0, 0, -1, 1, 1],
+        ("ffill", False, True): [-1, 0, 0, -1, 1, 1],
+        ("ffill", False, False): [-1, 0, 0, -1, 1, 1],
+        ("bfill", True, True): [0, 0, -1, -1, -1, -1],
+        ("bfill", True, False): [0, 0, -1, 1, 1, -1],
+        ("bfill", False, True): [0, 0, -1, 1, 1, -1],
+        ("bfill", False, False): [0, 0, -1, 1, 1, -1],
+    }
+
+    ridx = expected_rows.get((method, dropna, has_nan_group))
+    expected = df_without_nan_rows.reindex(ridx).reset_index(drop=True)
+
+    tm.assert_frame_equal(result, expected)