BUG: Behavior with fallback between raise and coerce #46071

doc/source/whatsnew/v1.5.0.rst

@@ -827,7 +827,7 @@ Datetimelike
 - Bug in :meth:`DatetimeIndex.resolution` incorrectly returning "day" instead of "nanosecond" for nanosecond-resolution indexes (:issue:`46903`)
 - Bug in :class:`Timestamp` with an integer or float value and ``unit="Y"`` or ``unit="M"`` giving slightly-wrong results (:issue:`47266`)
 - Bug in :class:`.DatetimeArray` construction when passed another :class:`.DatetimeArray` and ``freq=None`` incorrectly inferring the freq from the given array (:issue:`47296`)
--
+- Bug in :func:`to_datetime` where ``infer_datetime_format`` fallback would not run if ``errors=coerce`` (:issue:`46071`)
 
 Timedelta
 ^^^^^^^^^

pandas/core/tools/datetimes.py

@@ -496,6 +496,9 @@ def _array_strptime_with_fallback(
         if "%Z" in fmt or "%z" in fmt:
             return _return_parsed_timezone_results(result, timezones, tz, name)
 
+        if infer_datetime_format and np.isnan(result).any():
+            return None
+
     return _box_as_indexlike(result, utc=utc, name=name)
 
 

pandas/tests/arrays/test_datetimes.py

@@ -639,3 +639,21 @@ def test_tz_localize_t2d(self):
 
         roundtrip = expected.tz_localize("US/Pacific")
         tm.assert_datetime_array_equal(roundtrip, dta)
+
+    @pytest.mark.parametrize(
+        "error",
+        ["coerce", "raise"],
+    )
+    def test_coerce_fallback(self, error):
+        # GH#46071
+        s = pd.Series(["6/30/2025", "1 27 2024"])
+        expected = pd.Series(
+            [pd.Timestamp("2025-06-30 00:00:00"), pd.Timestamp("2024-01-27 00:00:00")]
+        )
+
+        result = pd.to_datetime(s, errors=error, infer_datetime_format=True)
+
+        if error == "coerce":
+            assert result[1] is not pd.NaT
+
+        tm.assert_series_equal(expected, result)