feat(factory): marginal Sharpe contribution (selection objective)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/fxhnt/domain/diversification/marginal.py

@@ -0,0 +1,27 @@
+"""Marginal contribution of a candidate sleeve to the book: the increase in annualized Sharpe from blending
+it in at `weight`. Positive = the sleeve improves the risk-adjusted book (the selection objective). NaN if
+the series are too short to estimate."""
+from __future__ import annotations
+
+import numpy as np
+
+_ANN = np.sqrt(252.0)
+
+
+def _sharpe(x: np.ndarray) -> float:
+    x = x[np.isfinite(x)]
+    if len(x) < 2 or x.std() == 0:
+        return float("nan")
+    return float(x.mean() / x.std() * _ANN)
+
+
+def marginal_sharpe(book: np.ndarray, sleeve: np.ndarray, *, weight: float = 0.2) -> float:
+    n = min(len(book), len(sleeve))
+    if n < 2:
+        return float("nan")
+    b, s = book[-n:], sleeve[-n:]
+    base = _sharpe(b)
+    blended = _sharpe((1.0 - weight) * b + weight * s)
+    if not (np.isfinite(base) and np.isfinite(blended)):
+        return float("nan")
+    return blended - base

tests/unit/test_marginal.py

@@ -0,0 +1,25 @@
+"""Marginal contribution = the change in book Sharpe from adding a sleeve at a small weight."""
+from __future__ import annotations
+
+import numpy as np
+
+from fxhnt.domain.diversification.marginal import marginal_sharpe
+
+
+def test_uncorrelated_positive_sleeve_adds_sharpe() -> None:
+    rng = np.random.default_rng(0)
+    book = rng.normal(0.0005, 0.01, 1000)
+    sleeve = rng.normal(0.0005, 0.01, 1000)       # independent, positive drift
+    assert marginal_sharpe(book, sleeve, weight=0.2) > 0.0
+
+
+def test_duplicate_book_adds_little() -> None:
+    rng = np.random.default_rng(1)
+    book = rng.normal(0.0005, 0.01, 1000)
+    add = marginal_sharpe(book, book.copy(), weight=0.2)   # same stream -> no diversification
+    indep = marginal_sharpe(book, rng.normal(0.0005, 0.01, 1000), weight=0.2)
+    assert indep > add
+
+
+def test_too_short_returns_nan() -> None:
+    assert np.isnan(marginal_sharpe(np.array([0.01]), np.array([0.01]), weight=0.2))