BUG: Fix PeriodIndex +/- TimedeltaIndex

doc/source/whatsnew/v0.24.0.txt

@@ -729,6 +729,7 @@ Datetimelike
 - Bug in :class:`DatetimeIndex` where frequency was being set if original frequency was ``None`` (:issue:`22150`)
 - Bug in rounding methods of :class:`DatetimeIndex` (:meth:`~DatetimeIndex.round`, :meth:`~DatetimeIndex.ceil`, :meth:`~DatetimeIndex.floor`) and :class:`Timestamp` (:meth:`~Timestamp.round`, :meth:`~Timestamp.ceil`, :meth:`~Timestamp.floor`) could give rise to loss of precision (:issue:`22591`)
 - Bug in :func:`to_datetime` with an :class:`Index` argument that would drop the ``name`` from the result (:issue:`21697`)
+- Bug in :class:`PeriodIndex` where adding or subtracting a :class:`timedelta` or :class:`Tick` object produced incorrect results (:issue:`22988`)
 
 Timedelta
 ^^^^^^^^^

pandas/core/arrays/datetimelike.py

@@ -382,6 +382,11 @@ def _add_delta_tdi(self, other):
         if not len(self) == len(other):
             raise ValueError("cannot add indices of unequal length")
 
+        if isinstance(other, np.ndarray):
+            # ndarray[timedelta64]; wrap in TimedeltaIndex for op
+            from pandas import TimedeltaIndex
+            other = TimedeltaIndex(other)
+
         self_i8 = self.asi8
         other_i8 = other.asi8
         new_values = checked_add_with_arr(self_i8, other_i8,
@@ -632,11 +637,17 @@ def __add__(self, other):
                 return self._add_datelike(other)
             elif is_integer_dtype(other):
                 result = self._addsub_int_array(other, operator.add)
-            elif is_float_dtype(other) or is_period_dtype(other):
+            elif is_float_dtype(other):
                 # Explicitly catch invalid dtypes
                 raise TypeError("cannot add {dtype}-dtype to {cls}"
                                 .format(dtype=other.dtype,
                                         cls=type(self).__name__))
+            elif is_period_dtype(other):
+                # if self is a TimedeltaArray and other is a PeriodArray with
+                #  a timedelta-like (i.e. Tick) freq, this operation is valid.
+                #  Defer to the PeriodArray implementation.
+                # In remaining cases, this will end up raising TypeError.
+                return NotImplemented
             elif is_extension_array_dtype(other):
                 # Categorical op will raise; defer explicitly
                 return NotImplemented

pandas/core/arrays/datetimes.py

@@ -506,7 +506,7 @@ def _add_delta(self, delta):
         Parameters
         ----------
         delta : {timedelta, np.timedelta64, DateOffset,
-                 TimedelaIndex, ndarray[timedelta64]}
+                 TimedeltaIndex, ndarray[timedelta64]}
 
         Returns
         -------

pandas/core/arrays/period.py

@@ -1,5 +1,6 @@
 # -*- coding: utf-8 -*-
 from datetime import timedelta
+import operator
 import warnings
 
 import numpy as np
@@ -17,8 +18,8 @@
 from pandas.util._decorators import (cache_readonly, deprecate_kwarg)
 
 from pandas.core.dtypes.common import (
-    is_integer_dtype, is_float_dtype, is_period_dtype,
-    is_datetime64_dtype)
+    is_integer_dtype, is_float_dtype, is_period_dtype, is_timedelta64_dtype,
+    is_datetime64_dtype, _TD_DTYPE)
 from pandas.core.dtypes.dtypes import PeriodDtype
 from pandas.core.dtypes.generic import ABCSeries
 
@@ -355,24 +356,54 @@ def _add_offset(self, other):
         return self._time_shift(other.n)
 
     def _add_delta_td(self, other):
+        assert isinstance(self.freq, Tick)  # checked by calling function
         assert isinstance(other, (timedelta, np.timedelta64, Tick))
-        nanos = delta_to_nanoseconds(other)
-        own_offset = frequencies.to_offset(self.freq.rule_code)
 
-        if isinstance(own_offset, Tick):
-            offset_nanos = delta_to_nanoseconds(own_offset)
-            if np.all(nanos % offset_nanos == 0):
-                return self._time_shift(nanos // offset_nanos)
+        delta = self._check_timedeltalike_freq_compat(other)
 
-        # raise when input doesn't have freq
-        raise IncompatibleFrequency("Input has different freq from "
-                                    "{cls}(freq={freqstr})"
-                                    .format(cls=type(self).__name__,
-                                            freqstr=self.freqstr))
+        # Note: when calling parent class's _add_delta_td, it will call
+        #  delta_to_nanoseconds(delta).  Because delta here is an integer,
+        #  delta_to_nanoseconds will return it unchanged.
+        return DatetimeLikeArrayMixin._add_delta_td(self, delta)
+
+    def _add_delta_tdi(self, other):
+        assert isinstance(self.freq, Tick)  # checked by calling function
+
+        delta = self._check_timedeltalike_freq_compat(other)
+        return self._addsub_int_array(delta, operator.add)
 
     def _add_delta(self, other):
-        ordinal_delta = self._maybe_convert_timedelta(other)
-        return self._time_shift(ordinal_delta)
+        """
+        Add a timedelta-like, Tick, or TimedeltaIndex-like object
+        to self.
+
+        Parameters
+        ----------
+        other : {timedelta, np.timedelta64, Tick,
+                 TimedeltaIndex, ndarray[timedelta64]}
+
+        Returns
+        -------
+        result : same type as self
+        """
+        if not isinstance(self.freq, Tick):
+            # We cannot add timedelta-like to non-tick PeriodArray
+            raise IncompatibleFrequency("Input has different freq from "
+                                        "{cls}(freq={freqstr})"
+                                        .format(cls=type(self).__name__,
+                                                freqstr=self.freqstr))
+
+        # TODO: standardize across datetimelike subclasses whether to return
+        #  i8 view or _shallow_copy
+        if isinstance(other, (Tick, timedelta, np.timedelta64)):
+            new_values = self._add_delta_td(other)
+            return self._shallow_copy(new_values)
+        elif is_timedelta64_dtype(other):
+            # ndarray[timedelta64] or TimedeltaArray/index
+            new_values = self._add_delta_tdi(other)
+            return self._shallow_copy(new_values)
+        else:  # pragma: no cover
+            raise TypeError(type(other).__name__)
 
     @deprecate_kwarg(old_arg_name='n', new_arg_name='periods')
     def shift(self, periods):
@@ -428,14 +459,9 @@ def _maybe_convert_timedelta(self, other):
                 other, (timedelta, np.timedelta64, Tick, np.ndarray)):
             offset = frequencies.to_offset(self.freq.rule_code)
             if isinstance(offset, Tick):
-                if isinstance(other, np.ndarray):
-                    nanos = np.vectorize(delta_to_nanoseconds)(other)
-                else:
-                    nanos = delta_to_nanoseconds(other)
-                offset_nanos = delta_to_nanoseconds(offset)
-                check = np.all(nanos % offset_nanos == 0)
-                if check:
-                    return nanos // offset_nanos
+                # _check_timedeltalike_freq_compat will raise if incompatible
+                delta = self._check_timedeltalike_freq_compat(other)
+                return delta
         elif isinstance(other, DateOffset):
             freqstr = other.rule_code
             base = frequencies.get_base_alias(freqstr)
@@ -454,6 +480,58 @@ def _maybe_convert_timedelta(self, other):
         raise IncompatibleFrequency(msg.format(cls=type(self).__name__,
                                                freqstr=self.freqstr))
 
+    def _check_timedeltalike_freq_compat(self, other):
+        """
+        Arithmetic operations with timedelta-like scalars or array `other`
+        are only valid if `other` is an integer multiple of `self.freq`.
+        If the operation is valid, find that integer multiple.  Otherwise,
+        raise because the operation is invalid.
+
+        Parameters
+        ----------
+        other : timedelta, np.timedelta64, Tick,
+                ndarray[timedelta64], TimedeltaArray, TimedeltaIndex
+
+        Returns
+        -------
+        multiple : int or ndarray[int64]
+
+        Raises
+        ------
+        IncompatibleFrequency
+        """
+        assert isinstance(self.freq, Tick)  # checked by calling function
+        own_offset = frequencies.to_offset(self.freq.rule_code)
+        base_nanos = delta_to_nanoseconds(own_offset)
+
+        if isinstance(other, (timedelta, np.timedelta64, Tick)):
+            nanos = delta_to_nanoseconds(other)
+
+        elif isinstance(other, np.ndarray):
+            # numpy timedelta64 array; all entries must be compatible
+            assert other.dtype.kind == 'm'
+            if other.dtype != _TD_DTYPE:
+                # i.e. non-nano unit
+                # TODO: disallow unit-less timedelta64
+                other = other.astype(_TD_DTYPE)
+            nanos = other.view('i8')
+        else:
+            # TimedeltaArray/Index
+            nanos = other.asi8
+
+        if np.all(nanos % base_nanos == 0):
+            # nanos being added is an integer multiple of the
+            #  base-frequency to self.freq
+            delta = nanos // base_nanos
+            # delta is the integer (or integer-array) number of periods
+            # by which will be added to self.
+            return delta
+
+        raise IncompatibleFrequency("Input has different freq from "
+                                    "{cls}(freq={freqstr})"
+                                    .format(cls=type(self).__name__,
+                                            freqstr=self.freqstr))
+
 
 PeriodArrayMixin._add_comparison_ops()
 PeriodArrayMixin._add_datetimelike_methods()

pandas/tests/arithmetic/test_period.py

@@ -446,26 +446,36 @@ def test_pi_add_sub_td64_array_non_tick_raises(self):
         with pytest.raises(period.IncompatibleFrequency):
             tdarr - rng
 
-    @pytest.mark.xfail(reason='op with TimedeltaIndex raises, with ndarray OK',
-                       strict=True)
     def test_pi_add_sub_td64_array_tick(self):
-        rng = pd.period_range('1/1/2000', freq='Q', periods=3)
+        # PeriodIndex + Timedelta-like is allowed only with
+        #   tick-like frequencies
+        rng = pd.period_range('1/1/2000', freq='90D', periods=3)
         tdi = pd.TimedeltaIndex(['-1 Day', '-1 Day', '-1 Day'])
         tdarr = tdi.values
 
-        expected = rng + tdi
+        expected = pd.period_range('12/31/1999', freq='90D', periods=3)
+        result = rng + tdi
+        tm.assert_index_equal(result, expected)
         result = rng + tdarr
         tm.assert_index_equal(result, expected)
+        result = tdi + rng
+        tm.assert_index_equal(result, expected)
         result = tdarr + rng
         tm.assert_index_equal(result, expected)
 
-        expected = rng - tdi
+        expected = pd.period_range('1/2/2000', freq='90D', periods=3)
+
+        result = rng - tdi
+        tm.assert_index_equal(result, expected)
         result = rng - tdarr
         tm.assert_index_equal(result, expected)
 
         with pytest.raises(TypeError):
             tdarr - rng
 
+        with pytest.raises(TypeError):
+            tdi - rng
+
     # -----------------------------------------------------------------
     # operations with array/Index of DateOffset objects
 
@@ -596,6 +606,56 @@ def test_pi_sub_intarray(self, box):
     # Timedelta-like (timedelta, timedelta64, Timedelta, Tick)
     # TODO: Some of these are misnomers because of non-Tick DateOffsets
 
+    def test_pi_add_timedeltalike_minute_gt1(self, three_days):
+        # GH#23031 adding a time-delta-like offset to a PeriodArray that has
+        # minute frequency with n != 1.  A more general case is tested below
+        # in test_pi_add_timedeltalike_tick_gt1, but here we write out the
+        # expected result more explicitly.
+        other = three_days
+        rng = pd.period_range('2014-05-01', periods=3, freq='2D')
+
+        expected = pd.PeriodIndex(['2014-05-04', '2014-05-06', '2014-05-08'],
+                                  freq='2D')
+
+        result = rng + other
+        tm.assert_index_equal(result, expected)
+
+        result = other + rng
+        tm.assert_index_equal(result, expected)
+
+        # subtraction
+        expected = pd.PeriodIndex(['2014-04-28', '2014-04-30', '2014-05-02'],
+                                  freq='2D')
+        result = rng - other
+        tm.assert_index_equal(result, expected)
+
+        with pytest.raises(TypeError):
+            other - rng
+
+    @pytest.mark.parametrize('freqstr', ['5ns', '5us', '5ms',
+                                         '5s', '5T', '5h', '5d'])
+    def test_pi_add_timedeltalike_tick_gt1(self, three_days, freqstr):
+        # GH#23031 adding a time-delta-like offset to a PeriodArray that has
+        # tick-like frequency with n != 1
+        other = three_days
+        rng = pd.period_range('2014-05-01', periods=6, freq=freqstr)
+
+        expected = pd.period_range(rng[0] + other, periods=6, freq=freqstr)
+
+        result = rng + other
+        tm.assert_index_equal(result, expected)
+
+        result = other + rng
+        tm.assert_index_equal(result, expected)
+
+        # subtraction
+        expected = pd.period_range(rng[0] - other, periods=6, freq=freqstr)
+        result = rng - other
+        tm.assert_index_equal(result, expected)
+
+        with pytest.raises(TypeError):
+            other - rng
+
     def test_pi_add_iadd_timedeltalike_daily(self, three_days):
         # Tick
         other = three_days

pandas/util/testing.py

@@ -805,6 +805,7 @@ def assert_index_equal(left, right, exact='equiv', check_names=True,
         Specify object name being compared, internally used to show appropriate
         assertion message
     """
+    __tracebackhide__ = True
 
     def _check_types(l, r, obj='Index'):
         if exact:
@@ -1048,6 +1049,8 @@ def assert_interval_array_equal(left, right, exact='equiv',
 
 
 def raise_assert_detail(obj, message, left, right, diff=None):
+    __tracebackhide__ = True
+
     if isinstance(left, np.ndarray):
         left = pprint_thing(left)
     elif is_categorical_dtype(left):