diff --git a/tests/test_neutralize.py b/tests/test_neutralize.py
new file mode 100644
index 0000000..ff3395f
--- /dev/null
+++ b/tests/test_neutralize.py
@@ -0,0 +1,159 @@
+"""Cross-sectional and industry neutralisation tests."""
+from __future__ import annotations
+
+import torch
+import pytest
+
+from mlquant.data.synthetic import SyntheticConfig, make_synthetic_panel
+from mlquant.features.neutralize import neutralize_cs, neutralize_industry
+
+
+# =============================================================================
+# neutralize_cs
+# =============================================================================
+
+
+def test_neutralize_cs_basic():
+    """Per-date cross-sectional z-score: mean≈0, std≈1 on valid cells."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=50, n_dates=60, seed=1))
+    out = neutralize_cs(p.close, p.mask)
+
+    for t in range(p.mask.shape[0]):
+        valid = p.mask[t]
+        if valid.sum() < 2:
+            continue
+        vals = out[t][valid]
+        assert torch.abs(vals.mean()) < 1e-6, f"date {t}: mean {vals.mean().item():.6f}"
+        assert torch.abs(vals.std() - 1.0) < 0.15, f"date {t}: std {vals.std().item():.6f}"
+
+
+def test_neutralize_cs_mask():
+    """Masked cells are zero in output and unmasked ones are untouched."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=30, n_dates=20, seed=2))
+    out = neutralize_cs(p.close, p.mask)
+    # masked cells → 0
+    assert (out[~p.mask] == 0.0).all()
+    # valid cells ≠ 0
+    assert (out[p.mask] != 0.0).any()
+
+
+def test_neutralize_cs_const_values():
+    """Constant values across stocks → all-zero output (no cross-sectional variation)."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=10, n_dates=5, seed=3))
+    p.close[:] = 1.0
+    out = neutralize_cs(p.close, p.mask)
+    # z-score of identical values is NaN; clamped std gives 0
+    assert torch.allclose(out[p.mask], torch.zeros(1), atol=1e-6)
+
+
+def test_neutralize_cs_single_stock():
+    """Single valid stock per date → output is zero (std=0, clamped)."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=5, n_dates=5, seed=4))
+    mask = torch.zeros_like(p.mask)
+    mask[:, 0] = True  # only stock 0 is tradable
+    out = neutralize_cs(p.close, mask)
+    assert (out == 0.0).all()
+
+
+def test_neutralize_cs_all_masked():
+    """All stocks masked → output is all zero."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=10, n_dates=5, seed=5))
+    mask = torch.zeros_like(p.mask)
+    out = neutralize_cs(p.close, mask)
+    assert (out == 0.0).all()
+
+
+# =============================================================================
+# neutralize_industry
+# =============================================================================
+
+
+def _make_industry_onehot(n_stocks: int, n_industries: int, device: str = "cpu") -> torch.Tensor:
+    """Create a fixed per-stock industry one-hot matrix [1, N, K]."""
+    onehot = torch.zeros(1, n_stocks, n_industries, device=device)
+    for i in range(n_stocks):
+        onehot[0, i, i % n_industries] = 1.0
+    return onehot
+
+
+def test_neutralize_industry_basic():
+    """Residuals after industry neutralisation have per-date mean≈0."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=30, n_dates=20, seed=6))
+    onehot = _make_industry_onehot(30, 5)
+    out = neutralize_industry(p.close, p.mask, onehot.expand(p.close.shape[0], -1, -1))
+
+    for t in range(p.mask.shape[0]):
+        valid = p.mask[t]
+        if valid.sum() < 2:
+            continue
+        vals = out[t][valid]
+        assert torch.abs(vals.mean()) < 1e-5, f"date {t}: mean {vals.mean().item():.6f}"
+
+
+def test_neutralize_industry_with_size():
+    """Industry neutralisation with log_size regressor still produces mean≈0."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=30, n_dates=20, seed=7))
+    onehot = _make_industry_onehot(30, 5)
+    log_size = torch.log(p.close.abs().clamp_min(1e-12))
+    out = neutralize_industry(p.close, p.mask, onehot.expand(p.close.shape[0], -1, -1), log_size=log_size)
+
+    for t in range(p.mask.shape[0]):
+        valid = p.mask[t]
+        if valid.sum() <= 2:
+            continue
+        vals = out[t][valid]
+        assert torch.abs(vals.mean()) < 1e-5, f"date {t}: mean {vals.mean().item():.6f}"
+
+
+def test_neutralize_industry_mask():
+    """Masked cells are zero in output."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=20, n_dates=10, seed=8))
+    onehot = _make_industry_onehot(20, 4)
+    # mask out first 5 stocks each day
+    mask = p.mask.clone()
+    mask[:, :5] = False
+    out = neutralize_industry(p.close, mask, onehot.expand(p.close.shape[0], -1, -1))
+    assert (out[:, :5] == 0.0).all()
+    # valid cells ≠ 0 (unless degenerate)
+    assert (out[:, 5:].abs().sum() > 0)
+
+
+def test_neutralize_industry_single_stock():
+    """Single tradable stock → returned as-is (not enough for OLS)."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=10, n_dates=5, seed=9))
+    onehot = _make_industry_onehot(10, 3)
+    mask = torch.zeros_like(p.mask)
+    mask[:, 0] = True
+    out = neutralize_industry(p.close, mask, onehot.expand(p.close.shape[0], -1, -1))
+    # single stock → OLS skipped, output is zero (as per source: out initialized to zeros)
+    assert (out == 0.0).all()
+
+
+def test_neutralize_industry_perfect_collinear():
+    """All stocks in same industry → OLS still works (rank-deficient handled by lstsq)."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=10, n_dates=5, seed=10))
+    # all stocks same industry
+    onehot = torch.zeros(1, 10, 1)
+    onehot[:] = 1.0
+    out = neutralize_industry(p.close, p.mask, onehot.expand(p.close.shape[0], -1, -1))
+    # should not crash; residuals are z-scored per date
+    for t in range(p.mask.shape[0]):
+        valid = p.mask[t]
+        if valid.sum() < 2:
+            continue
+        vals = out[t][valid]
+        assert torch.abs(vals.mean()) < 1e-5, f"date {t}: mean {vals.mean().item():.6f}"
+
+
+def test_neutralize_industry_cross_sectional_output():
+    """Per-date residuals have std normalised to 1."""
+    p = make_synthetic_panel(SyntheticConfig(n_stocks=30, n_dates=20, seed=11))
+    onehot = _make_industry_onehot(30, 5)
+    out = neutralize_industry(p.close, p.mask, onehot.expand(p.close.shape[0], -1, -1))
+
+    for t in range(p.mask.shape[0]):
+        valid = p.mask[t]
+        if valid.sum() < 3:  # need at least 3 stocks for meaningful std
+            continue
+        vals = out[t][valid]
+        assert torch.abs(vals.std() - 1.0) < 0.15, f"date {t}: std {vals.std().item():.6f}"