diff --git a/doc/source/release.rst b/doc/source/release.rst
index 5376e0396799e..101ec290a58cf 100644
--- a/doc/source/release.rst
+++ b/doc/source/release.rst
@@ -105,7 +105,7 @@ Improvements to existing features
     test to vbench (:issue:`4705` and :issue:`4722`)
   - Add ``axis`` and ``level`` keywords to ``where``, so that the ``other`` argument
     can now be an alignable pandas object.
-  - ``to_datetime`` with a format of 'YYYYMMDD' now parses much faster
+  - ``to_datetime`` with a format of '%Y%m%d' now parses much faster
 
 API Changes
 ~~~~~~~~~~~
diff --git a/pandas/tseries/tests/test_timeseries.py b/pandas/tseries/tests/test_timeseries.py
index c9e643e25b761..d7ca9d9b371d4 100644
--- a/pandas/tseries/tests/test_timeseries.py
+++ b/pandas/tseries/tests/test_timeseries.py
@@ -845,9 +845,19 @@ def test_to_datetime_format_YYYYMMDD(self):
         assert_series_equal(result, expected)
 
         # with NaT
+        expected = Series([Timestamp("19801222"),Timestamp("19801222")] + [Timestamp("19810105")]*5)
+        expected[2] = np.nan
         s[2] = np.nan
-        self.assertRaises(ValueError, to_datetime, s,format='%Y%m%d')
-        self.assertRaises(ValueError, to_datetime, s.apply(str),format='%Y%m%d')
+
+        result = to_datetime(s,format='%Y%m%d')
+        assert_series_equal(result, expected)
+
+        # string with NaT
+        s = s.apply(str)
+        s[2] = 'nat'
+        result = to_datetime(s,format='%Y%m%d')
+        assert_series_equal(result, expected)
+
 
     def test_to_datetime_format_microsecond(self):
         val = '01-Apr-2011 00:00:01.978'
diff --git a/pandas/tseries/tools.py b/pandas/tseries/tools.py
index cca4850c2c1bf..dd78bea385c61 100644
--- a/pandas/tseries/tools.py
+++ b/pandas/tseries/tools.py
@@ -106,8 +106,7 @@ def _convert_listlike(arg, box):
                 # shortcut formatting here
                 if format == '%Y%m%d':
                     try:
-                        carg = arg.astype(np.int64).astype(object)
-                        result = lib.try_parse_year_month_day(carg/10000,carg/100 % 100, carg % 100)
+                        result = _attempt_YYYYMMDD(arg)
                     except:
                         raise ValueError("cannot convert the input to '%Y%m%d' date format")
 
@@ -144,6 +143,43 @@ def _convert_listlike(arg, box):
 class DateParseError(ValueError):
     pass
 
+def _attempt_YYYYMMDD(arg):
+    """ try to parse the YYYYMMDD/%Y%m%d format, try to deal with NaT-like,
+        arg is a passed in as an object dtype, but could really be ints/strings with nan-like/or floats (e.g. with nan) """
+
+    def calc(carg):
+        # calculate the actual result
+        carg = carg.astype(object)
+        return lib.try_parse_year_month_day(carg/10000,carg/100 % 100, carg % 100)
+
+    def calc_with_mask(carg,mask):
+        result = np.empty(carg.shape, dtype='M8[ns]')
+        iresult = result.view('i8')
+        iresult[-mask] = tslib.iNaT
+        result[mask] = calc(carg[mask].astype(np.float64).astype(np.int64)).astype('M8[ns]')
+        return result
+
+    # try intlike / strings that are ints
+    try:
+        return calc(arg.astype(np.int64))
+    except:
+        pass
+
+    # a float with actual np.nan
+    try:
+        carg = arg.astype(np.float64)
+        return calc_with_mask(carg,com.notnull(carg))
+    except:
+        pass
+
+    # string with NaN-like
+    try:
+        mask = ~lib.ismember(arg, tslib._nat_strings)
+        return calc_with_mask(arg,mask)
+    except:
+        pass
+
+    return None
 
 # patterns for quarters like '4Q2005', '05Q1'
 qpat1full = re.compile(r'(\d)Q(\d\d\d\d)')
diff --git a/vb_suite/timeseries.py b/vb_suite/timeseries.py
index 999c3869daf62..353d7afc63cb3 100644
--- a/vb_suite/timeseries.py
+++ b/vb_suite/timeseries.py
@@ -154,16 +154,7 @@ def date_range(start=None, end=None, periods=None, freq=None):
 
 timeseries_to_datetime_YYYYMMDD = \
     Benchmark('to_datetime(strings,format="%Y%m%d")', setup,
-              start_date=datetime(2013, 9, 1))
-
-setup = common_setup + """
-rng = date_range('1/1/2000', periods=10000, freq='D')
-strings = Series(rng.year*10000+rng.month*100+rng.day,dtype=np.int64).apply(str)
-"""
-
-timeseries_to_datetime_YYYYMMDD_old = \
-    Benchmark('pandas.tslib.array_strptime(strings.values,"%Y%m%d")', setup,
-              start_date=datetime(2013, 9, 1))
+              start_date=datetime(2012, 7, 1))
 
 # ---- infer_freq
 # infer_freq