BUG: don't lose dtypes when concatenating empty array-likes

doc/source/release.rst

@@ -118,6 +118,7 @@ Bug Fixes
   - Bug in rolling skew/kurtosis when passed a Series with bad data (:issue:`5749`)
   - Bug in scipy ``interpolate`` methods with a datetime index (:issue:`5975`)
   - Bug in NaT comparison if a mixed datetime/np.datetime64 with NaT were passed (:issue:`5968`)
+  - Fixed bug with ``pd.concat`` losing dtype information if all inputs are empty (:issue:`5742`)
 
 pandas 0.13.0
 -------------

pandas/core/common.py

@@ -2326,20 +2326,23 @@ def _check_as_is(x):
 
 def _concat_compat(to_concat, axis=0):
     # filter empty arrays
-    to_concat = [x for x in to_concat if x.shape[axis] > 0]
-
-    # return the empty np array, if nothing to concatenate, #3121
-    if not to_concat:
-        return np.array([], dtype=object)
-
-    is_datetime64 = [x.dtype == _NS_DTYPE for x in to_concat]
-    if all(is_datetime64):
-        # work around NumPy 1.6 bug
-        new_values = np.concatenate([x.view(np.int64) for x in to_concat],
-                                    axis=axis)
-        return new_values.view(_NS_DTYPE)
-    elif any(is_datetime64):
-        to_concat = [_to_pydatetime(x) for x in to_concat]
+    nonempty = [x for x in to_concat if x.shape[axis] > 0]
+
+    # If all arrays are empty, there's nothing to convert, just short-cut to
+    # the concatenation, #3121.
+    #
+    # Creating an empty array directly is tempting, but the winnings would be
+    # marginal given that it would still require shape & dtype calculation and
+    # np.concatenate which has them both implemented is compiled.
+    if nonempty:
+        is_datetime64 = [x.dtype == _NS_DTYPE for x in nonempty]
+        if all(is_datetime64):
+            # work around NumPy 1.6 bug
+            new_values = np.concatenate([x.view(np.int64) for x in nonempty],
+                                        axis=axis)
+            return new_values.view(_NS_DTYPE)
+        elif any(is_datetime64):
+            to_concat = [_to_pydatetime(x) for x in nonempty]
 
     return np.concatenate(to_concat, axis=axis)
 

pandas/tests/test_frame.py

@@ -11909,6 +11909,23 @@ def test_to_csv_date_format(self):
 
                 assert_frame_equal(test, nat_frame)
 
+    def test_concat_empty_dataframe_dtypes(self):
+        df = DataFrame(columns=list("abc"))
+        df['a'] = df['a'].astype(np.bool_)
+        df['b'] = df['b'].astype(np.int32)
+        df['c'] = df['c'].astype(np.float64)
+
+        result = pd.concat([df, df])
+        self.assertEqual(result['a'].dtype, np.bool_)
+        self.assertEqual(result['b'].dtype, np.int32)
+        self.assertEqual(result['c'].dtype, np.float64)
+
+        result = pd.concat([df, df.astype(np.float64)])
+        self.assertEqual(result['a'].dtype, np.object_)
+        self.assertEqual(result['b'].dtype, np.float64)
+        self.assertEqual(result['c'].dtype, np.float64)
+
+
 def skip_if_no_ne(engine='numexpr'):
     if engine == 'numexpr':
         try:

pandas/tests/test_series.py

@@ -5441,6 +5441,15 @@ def test_numpy_unique(self):
         # it works!
         result = np.unique(self.ts)
 
+    def test_concat_empty_series_dtypes(self):
+        self.assertEqual(pd.concat([Series(dtype=np.float64)]).dtype, np.float64)
+        self.assertEqual(pd.concat([Series(dtype=np.int8)]).dtype, np.int8)
+        self.assertEqual(pd.concat([Series(dtype=np.bool_)]).dtype, np.bool_)
+
+        self.assertEqual(pd.concat([Series(dtype=np.bool_),
+                                    Series(dtype=np.int32)]).dtype, np.int32)
+
+
 
 class TestSeriesNonUnique(tm.TestCase):