ENH: support downcasting of nullable EAs in pd.to_numeric

pandas/core/tools/numeric.py

@@ -110,6 +110,21 @@ def to_numeric(arg, errors="raise", downcast=None):
     2    2.0
     3   -3.0
     dtype: float64
+
+    Downcasting of ``ExtensionDtype`` is supported:
+
+    >>> s = pd.Series([1, 2, 3], dtype="Int64")
+    >>> pd.to_numeric(s, downcast="integer")
+    0    1
+    1    2
+    2    3
+    dtype: Int8
+    >>> s = pd.Series([1.0, 2.1, 3.0], dtype="Float64")
+    >>> pd.to_numeric(s, downcast="float")
+    0    1.0
+    1    2.1
+    2    3.0
+    dtype: Float32
     """
     if downcast not in (None, "integer", "signed", "unsigned", "float"):
         raise ValueError("invalid downcasting method provided")
@@ -144,6 +159,8 @@ def to_numeric(arg, errors="raise", downcast=None):
     else:
         values = arg
 
+    # GH33013: for IntegerArray & FloatingArray extract non-null values for casting
+    # save mask to reconstruct the full array after casting
     if isinstance(values, NumericArray):
         mask = values._mask
         values = values._data[~mask]
@@ -196,6 +213,7 @@ def to_numeric(arg, errors="raise", downcast=None):
                     if values.dtype == dtype:
                         break
 
+    # GH33013: for IntegerArray & FloatingArray need to reconstruct masked array
     if mask is not None:
         data = np.zeros(mask.shape, dtype=values.dtype)
         data[~mask] = values

pandas/tests/tools/test_to_numeric.py

@@ -735,11 +735,20 @@ def test_to_numeric_from_nullable_string(values, expected):
         ([1.0, 1.1], "Float64", "integer", "Float64"),
         ([1, pd.NA], "Int64", "integer", "Int8"),
         ([450, 300], "Int64", "integer", "Int16"),
+        ([1, 1], "Float64", "integer", "Int8"),
+        ([np.iinfo(np.int64).max - 1, 1], "Int64", "integer", "Int64"),
         ([1, 1], "Int64", "signed", "Int8"),
         ([1.0, 1.0], "Float32", "signed", "Int8"),
         ([1.0, 1.1], "Float64", "signed", "Float64"),
         ([1, pd.NA], "Int64", "signed", "Int8"),
         ([450, -300], "Int64", "signed", "Int16"),
+        pytest.param(
+            [np.iinfo(np.uint64).max - 1, 1],
+            "UInt64",
+            "signed",
+            "UInt64",
+            marks=pytest.mark.xfail(reason="GH38798"),
+        ),
         ([1, 1], "Int64", "unsigned", "UInt8"),
         ([1.0, 1.0], "Float32", "unsigned", "UInt8"),
         ([1.0, 1.1], "Float64", "unsigned", "Float64"),
@@ -748,7 +757,6 @@ def test_to_numeric_from_nullable_string(values, expected):
         ([-1, -1], "Int32", "unsigned", "Int32"),
         ([1, 1], "Float64", "float", "Float32"),
         ([1, 1.1], "Float64", "float", "Float32"),
-        ([1, 1], "Float64", "integer", "Int8"),
     ),
 )
 def test_downcast_nullable_numeric(data, input_dtype, downcast, expected_dtype):