BUG: GH15429 transform result of timedelta from datetime

The transform() operation needs to return a like-indexed. To facilitate this, transform starts with a copy of the original series. Then, after the computation for each group, sets the appropriate elements of the copied series equal to the result. At that point is does a type comparison, and discovers that the timedelta is not cast-able to a datetime.

Author: stephenrauch

doc/source/whatsnew/v0.20.0.txt

@@ -626,6 +626,7 @@ Bug Fixes
 
 
 - Bug in ``.read_csv()`` with ``parse_dates`` when multiline headers are specified (:issue:`15376`)
+- Bug in ``groupby.transform()`` that would coerce the resultant dtypes back to the original (:issue:`10972`)
 
 
 - Bug in ``DataFrame.boxplot`` where ``fontsize`` was not applied to the tick labels on both axes (:issue:`15108`)

pandas/core/groupby.py

@@ -31,7 +31,7 @@
                                  _ensure_object,
                                  _ensure_categorical,
                                  _ensure_float)
-from pandas.types.cast import _possibly_downcast_to_dtype
+from pandas.types.cast import _possibly_downcast_to_dtype, _find_common_type
 from pandas.types.missing import isnull, notnull, _maybe_fill
 
 from pandas.core.common import (_values_from_object, AbstractMethodError,
@@ -2906,8 +2906,15 @@ def transform(self, func, *args, **kwargs):
                 common_type = np.common_type(np.array(res), result)
                 if common_type != result.dtype:
                     result = result.astype(common_type)
-            except:
-                pass
+            except Exception as exc:
+                # date math can cause type of result to change
+                if i == 0 and (is_datetime64_dtype(result.dtype) or
+                               is_timedelta64_dtype(result.dtype)):
+                    try:
+                        dtype = res.dtype
+                    except Exception as exc:
+                        dtype = type(res)
+                    result = np.empty_like(result, dtype)
 
             indexer = self._get_index(name)
             result[indexer] = res

pandas/tests/groupby/test_transform.py

@@ -3,7 +3,7 @@
 import numpy as np
 import pandas as pd
 from pandas.util import testing as tm
-from pandas import Series, DataFrame, Timestamp, MultiIndex, concat
+from pandas import Series, DataFrame, Timestamp, MultiIndex, concat, date_range
 from pandas.types.common import _ensure_platform_int
 from .common import MixIn, assert_fp_equal
 
@@ -190,6 +190,43 @@ def test_transform_bug(self):
         expected = Series(np.arange(5, 0, step=-1), name='B')
         assert_series_equal(result, expected)
 
+    def test_transform_datetime_to_timedelta(self):
+        # GH 15429
+        # transforming a datetime to timedelta
+        df = DataFrame(dict(A=Timestamp('20130101'), B=np.arange(5)))
+        expected = pd.Series([
+            Timestamp('20130101') - Timestamp('20130101')] * 5, name='A')
+
+        # this does date math without changing result type in transform
+        base_time = df['A'][0]
+        result = df.groupby('A')['A'].transform(
+            lambda x: x.max() - x.min() + base_time) - base_time
+        assert_series_equal(result, expected)
+
+        # this does date math and causes the transform to return timedelta
+        result = df.groupby('A')['A'].transform(lambda x: x.max() - x.min())
+        assert_series_equal(result, expected)
+
+    def test_transform_datetime_to_numeric(self):
+        # GH 10972
+        # convert dt to float
+        df = DataFrame({
+            'a': 1, 'b': date_range('2015-01-01', periods=2, freq='D')})
+        result = df.groupby('a').b.transform(
+            lambda x: x.dt.dayofweek - x.dt.dayofweek.mean())
+
+        expected = Series([-0.5, 0.5], name='b')
+        assert_series_equal(result, expected)
+
+        # convert dt to int
+        df = DataFrame({
+            'a': 1, 'b': date_range('2015-01-01', periods=2, freq='D')})
+        result = df.groupby('a').b.transform(
+            lambda x: x.dt.dayofweek - x.dt.dayofweek.min())
+
+        expected = Series([0, 1], name='b')
+        assert_series_equal(result, expected)
+
     def test_transform_multiple(self):
         grouped = self.ts.groupby([lambda x: x.year, lambda x: x.month])