diff --git a/pandas/core/series.py b/pandas/core/series.py
index d6a982c65e9fd..589cf9662d81b 100644
--- a/pandas/core/series.py
+++ b/pandas/core/series.py
@@ -5890,32 +5890,36 @@ def _arith_method(self, other, op):
         self, other = self._align_for_op(other)
         return base.IndexOpsMixin._arith_method(self, other, op)
 
-    def _align_for_op(self, right, align_asobject: bool = False):
-        """align lhs and rhs Series"""
-        # TODO: Different from DataFrame._align_for_op, list, tuple and ndarray
-        # are not coerced here
-        # because Series has inconsistencies described in GH#13637
+    def _align_for_op(self, right, align_asobject=False, fill_value=np.nan):
+        """Align lhs and rhs Series for arithmetic operations"""
+
         left = self
 
-        if isinstance(right, Series):
-            # avoid repeated alignment
-            if not left.index.equals(right.index):
-                if align_asobject:
-                    if left.dtype not in (object, np.bool_) or right.dtype not in (
-                        object,
-                        np.bool_,
-                    ):
-                        pass
-                        # GH#52538 no longer cast in these cases
-                    else:
-                        # to keep original value's dtype for bool ops
-                        left = left.astype(object)
-                        right = right.astype(object)
-
-                left, right = left.align(right)
-
-        return left, right
+        if not isinstance(right, Series):
+            return left, right
+
+        if left.index.equals(right.index):
+            return left, right
 
+        if not (hasattr(left.index, "levels") or hasattr(right.index, "levels")):
+            if align_asobject:
+                if left.empty or right.empty:
+                    if left.dtype not in (object, np.bool_) or right.dtype not in (object, np.bool_):
+                        return left.iloc[0:0], right.iloc[0:0]
+            return left.align(right, join='outer', fill_value=fill_value)
+
+        if hasattr(left.index, "levels") and not hasattr(right.index, "levels"):
+            if left.empty or right.empty:
+                return left.iloc[0:0], right.iloc[0:0]
+            else:
+                first_level = left.index.get_level_values(0)
+                left = left.astype(object)
+                right = right.astype(object)
+                right_aligned = right.reindex(first_level, fill_value=fill_value)
+                return left, right_aligned
+
+        return left.align(right, join='outer', fill_value=fill_value)
+        
     def _binop(self, other: Series, func, level=None, fill_value=None) -> Series:
         """
         Perform generic binary operation with optional fill value.
diff --git a/pandas/tests/series/test_subtraction_nanindex.py b/pandas/tests/series/test_subtraction_nanindex.py
new file mode 100644
index 0000000000000..a97aec4ede490
--- /dev/null
+++ b/pandas/tests/series/test_subtraction_nanindex.py
@@ -0,0 +1,38 @@
+import pytest
+import pandas as pd
+import numpy as np
+import pandas.testing as tm
+
+def test_series_subtraction_with_nan_and_levels():
+    ix1 = pd.MultiIndex.from_arrays(
+        [
+            [np.nan, 81, 81, 82, 82],
+            [np.nan] * 5,
+            pd.to_datetime([np.nan, '2018-06-01', '2018-07-01', '2018-07-01', '2018-08-01'])
+        ],
+        names=['foo', 'bar', 'date']
+    )
+
+    s1 = pd.Series(
+        [np.nan, 25.058969, 22.519751, 20.847981, 21.625236],
+        index=ix1
+    )
+
+    ix2 = pd.Index([81, 82, 83, 84, 85, 86, 87], name='foo')
+    s2 = pd.Series(
+        [28.2800, 25.2500, 22.2200, 16.7660, 14.0087, 14.9480, 29.2900],
+        index=ix2
+    )
+
+    expected = pd.Series(
+        [np.nan, -3.221031, -5.760249, -4.402019, -3.624764],
+        index=ix1,
+        dtype='float64'
+    )
+
+    result = s1 - s2
+
+    result = result.astype('float64')
+
+    tm.assert_series_equal(result, expected)
+