diff --git a/aeon/transformations/collection/feature_based/_catch22.py b/aeon/transformations/collection/feature_based/_catch22.py
index b29bee8c44..fde03fb275 100644
--- a/aeon/transformations/collection/feature_based/_catch22.py
+++ b/aeon/transformations/collection/feature_based/_catch22.py
@@ -1904,6 +1904,21 @@ def _quantile(X, quant):
 _local_simple_mean3_stderr = Catch22._FC_LocalSimple_mean3_stderr
 
 
+@njit(fastmath=True, cache=True)
+def _zeros_promoted(series, n):
+    """Allocate a length-n zero array typed to match catch22's dtype rule.
+
+    Applies the same dtype-promotion rule used internally by
+    z_normalise_series_with_mean / _as_normalised_float (float32
+    input stays float32, every other dtype -- including integer input --
+    promotes to float64). Used both for typed placeholders whose real
+    values are assigned later, and for output buffers that must keep the
+    input's precision instead of silently upcasting to float64.
+    """
+    dtype = (series[:0] / 1).dtype
+    return np.zeros(n, dtype=dtype)
+
+
 @njit(fastmath=True, cache=True)
 def _transform_case_numba(
     X, f_idx, keep, outlier_norm, fft_case, ac_case, ac_tw, ac_nfft, stds
@@ -1920,7 +1935,7 @@ def _transform_case_numba(
     requested (numba's np.std can round differently from numpy's).
     """
     n_feats = len(f_idx)
-    c22 = np.zeros(n_feats * len(X))
+    c22 = _zeros_promoted(X[0], n_feats * len(X))
 
     f_count = -1
     for i in range(len(X)):
@@ -1934,7 +1949,10 @@ def _transform_case_numba(
         # typed placeholders; real values are assigned before first use. These
         # must be dtype-fixed (not aliases of series) so the kernel also
         # compiles for integer input arrays.
-        outlier_series = np.empty(0, np.float64)
+        # / 1 applies the same dtype promotion z_normalise_series_with_mean
+        # uses internally (float32 stays float32, everything else becomes
+        # float64), so this placeholder always matches the real value's dtype.
+        outlier_series = _zeros_promoted(series, 0)
         have_outlier = False
         fft = ac_tw[:0]
         have_fft = False
diff --git a/aeon/transformations/collection/feature_based/tests/test_catch22.py b/aeon/transformations/collection/feature_based/tests/test_catch22.py
index 63348ace1d..48c0c5954a 100644
--- a/aeon/transformations/collection/feature_based/tests/test_catch22.py
+++ b/aeon/transformations/collection/feature_based/tests/test_catch22.py
@@ -7,6 +7,7 @@
 from numpy import testing
 
 from aeon.datasets import load_basic_motions
+from aeon.testing.data_generation import make_example_3d_numpy
 from aeon.transformations.collection.feature_based import Catch22
 from aeon.utils.validation._dependencies import _check_soft_dependencies
 
@@ -1521,3 +1522,26 @@ def track_n(s):
     assert abs(np.std(passed_series["p"]) - 1.0) < 1e-10
     assert abs(np.mean(passed_series["n"])) < 1e-10
     assert abs(np.std(passed_series["n"]) - 1.0) < 1e-10
+
+
+def test_catch22_float32_output_dtype():
+    """Test Catch22 does not raise on float32 input with outlier features.
+
+    Regression test: the outlier_series placeholder in _transform_case_numba was
+    hardcoded to float64, which numba's nopython mode could not unify with a
+    float32 z-normalised series produced for features 13/14 (outlier_include /
+    outlier_include_n), raising a TypingError for float32 input.
+    """
+    X = make_example_3d_numpy(
+        n_cases=6, n_channels=1, n_timepoints=40, random_state=0, return_y=False
+    ).astype(np.float32)
+
+    # outlier_norm=True exercises the fixed code path (features 13/14)
+    c22 = Catch22(outlier_norm=True, replace_nans=True)
+    result = c22.fit_transform(X)
+    assert result.shape == (6, 22)
+
+    # outlier_norm=False takes the other branch -- should still work
+    c22_no_norm = Catch22(outlier_norm=False, replace_nans=True)
+    result_no_norm = c22_no_norm.fit_transform(X)
+    assert result_no_norm.shape == (6, 22)