diff --git a/doc/source/whatsnew/v0.20.0.txt b/doc/source/whatsnew/v0.20.0.txt
index be487e165c602..9298d4b058d98 100644
--- a/doc/source/whatsnew/v0.20.0.txt
+++ b/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`)
diff --git a/pandas/core/groupby.py b/pandas/core/groupby.py
index 831ca3886773e..2c61a73d6814e 100644
--- a/pandas/core/groupby.py
+++ b/pandas/core/groupby.py
@@ -2890,32 +2890,32 @@ def transform(self, func, *args, **kwargs):
                     lambda: getattr(self, func)(*args, **kwargs))
 
         # reg transform
-        dtype = self._selected_obj.dtype
-        result = self._selected_obj.values.copy()
-
+        klass = self._selected_obj.__class__
+        results = []
         wrapper = lambda x: func(x, *args, **kwargs)
-        for i, (name, group) in enumerate(self):
+        for name, group in self:
             object.__setattr__(group, 'name', name)
             res = wrapper(group)
 
             if hasattr(res, 'values'):
                 res = res.values
 
-            # may need to astype
-            try:
-                common_type = np.common_type(np.array(res), result)
-                if common_type != result.dtype:
-                    result = result.astype(common_type)
-            except:
-                pass
-
             indexer = self._get_index(name)
-            result[indexer] = res
+            s = klass(res, indexer)
+            results.append(s)
 
-        result = _possibly_downcast_to_dtype(result, dtype)
-        return self._selected_obj.__class__(result,
-                                            index=self._selected_obj.index,
-                                            name=self._selected_obj.name)
+        from pandas.tools.concat import concat
+        result = concat(results).sort_index()
+
+        # we will only try to coerce the result type if
+        # we have a numeric dtype
+        dtype = self._selected_obj.dtype
+        if is_numeric_dtype(dtype):
+            result = _possibly_downcast_to_dtype(result, dtype)
+
+        result.name = self._selected_obj.name
+        result.index = self._selected_obj.index
+        return result
 
     def _transform_fast(self, func):
         """
diff --git a/pandas/tests/groupby/test_filters.py b/pandas/tests/groupby/test_filters.py
index 46ddb5a5318fb..de6757786a363 100644
--- a/pandas/tests/groupby/test_filters.py
+++ b/pandas/tests/groupby/test_filters.py
@@ -216,6 +216,7 @@ def test_filter_against_workaround(self):
         grouper = s.apply(lambda x: np.round(x, -1))
         grouped = s.groupby(grouper)
         f = lambda x: x.mean() > 10
+
         old_way = s[grouped.transform(f).astype('bool')]
         new_way = grouped.filter(f)
         assert_series_equal(new_way.sort_values(), old_way.sort_values())
diff --git a/pandas/tests/groupby/test_transform.py b/pandas/tests/groupby/test_transform.py
index cf5e9eb26ff13..51920ec642705 100644
--- a/pandas/tests/groupby/test_transform.py
+++ b/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])