Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(quantlib): stabilize covariance weights and validate singleton variances

Rudy Celeklie270d0ebc9cb
diff --git a/agent/src/quantlib/portfolio.py b/agent/src/quantlib/portfolio.py
index a53bf9d57..e346f02a0 100644
--- a/agent/src/quantlib/portfolio.py
+++ b/agent/src/quantlib/portfolio.py
@@ -7,7 +7,6 @@ Implements Marcos López de Prado's (2016) Hierarchical Risk Parity algorithm:
"""
-
import numpy as np
import pandas as pd
from scipy.cluster.hierarchy import linkage
@@ -83,9 +82,13 @@ def inverse_variance_weights(cov: pd.DataFrame | np.ndarray) -> np.ndarray:
if c.ndim != 2 or c.shape[0] != c.shape[1]:
raise ValueError("Covariance matrix must be square 2-D")
variances = np.diag(c)
- if np.any(variances <= 0.0):
- raise ValueError("All asset variances on the diagonal must be strictly positive")
- inv_var = 1.0 / variances
+ if not np.isfinite(variances).all() or np.any(variances <= 0.0):
+ raise ValueError(
+ "All asset variances on the diagonal must be finite and strictly positive"
+ )
+ # Scaling by the smallest variance cancels during normalization and avoids
+ # overflow in 1 / variance for valid subnormal positive variances.
+ inv_var = variances.min() / variances
return inv_var / np.sum(inv_var)
@@ -130,20 +133,21 @@ def hierarchical_risk_parity(
n = cov_mat.shape[0]
if n == 0:
raise ValueError("Covariance matrix cannot be empty")
+ diag = np.diag(cov_mat)
+ if not np.isfinite(diag).all() or np.any(diag <= 0.0):
+ raise ValueError("Diagonal variances must be finite and strictly positive")
if n == 1:
return pd.Series([1.0], index=labels) if is_df else np.array([1.0])
- diag = np.diag(cov_mat)
- if np.any(diag <= 0.0):
- raise ValueError("Diagonal variances must be strictly positive")
-
if corr is None:
std = np.sqrt(diag)
corr_mat = cov_mat / np.outer(std, std)
else:
if is_df and isinstance(corr, pd.DataFrame):
if set(corr.index) != set(labels) or set(corr.columns) != set(labels):
- raise ValueError("Correlation matrix labels must match covariance labels")
+ raise ValueError(
+ "Correlation matrix labels must match covariance labels"
+ )
corr = corr.loc[labels, labels]
corr_mat = np.asarray(corr, dtype=float)
if corr_mat.shape != cov_mat.shape:
diff --git a/agent/tests/quantlib/test_covariance_weight_stability.py b/agent/tests/quantlib/test_covariance_weight_stability.py
new file mode 100644
index 000000000..48da97378
--- /dev/null
+++ b/agent/tests/quantlib/test_covariance_weight_stability.py
@@ -0,0 +1,31 @@
+import numpy as np
+import pytest
+from src.quantlib.portfolio import hierarchical_risk_parity, inverse_variance_weights
+
+
+def test_inverse_weights_remain_finite_for_small_positive_variances() -> None:
+ cov = np.diag([1e-310, 2e-310])
+ actual = inverse_variance_weights(cov)
+ assert np.isfinite(actual).all()
+ np.testing.assert_allclose(actual, [2 / 3, 1 / 3])
+
+
[email protected]("variance", [0.0, -1.0, float("nan"), float("inf")])
+def test_singleton_hrp_rejects_invalid_variance_before_assigning_weight(
+ variance: float,
+) -> None:
+ with pytest.raises(ValueError, match="strictly positive"):
+ hierarchical_risk_parity(np.array([[variance]]))
+
+
[email protected]("variance", [float("nan"), float("inf")])
+def test_inverse_weights_reject_nonfinite_variance(variance: float) -> None:
+ with pytest.raises(ValueError, match="strictly positive"):
+ inverse_variance_weights(np.diag([0.04, variance]))
+
+
+def test_valid_singleton_and_ordinary_inverse_weights_are_unchanged() -> None:
+ np.testing.assert_array_equal(hierarchical_risk_parity(np.array([[0.04]])), [1.0])
+ np.testing.assert_allclose(
+ inverse_variance_weights(np.diag([0.04, 0.16])), [0.8, 0.2]
+ )