Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(backtest): optimizers see the window in position space, so shorts hedge (#1548)

Haozhe Wu809f9b831a13
diff --git a/agent/backtest/optimizers/base.py b/agent/backtest/optimizers/base.py
index ab90c9644..de1c353a8 100644
--- a/agent/backtest/optimizers/base.py
+++ b/agent/backtest/optimizers/base.py
@@ -71,13 +71,17 @@ class BaseOptimizer(ABC):
continue
signs = np.array([np.sign(pos.at[dt, c]) for c in active])
- # Exposed for subclasses (e.g. mean-variance style optimizers)
- # that need the direction of the position actually being sized,
- # not just its unsigned covariance/volatility structure -- a
- # short's expected return is -1 * the asset's raw drift, not the
- # raw drift itself.
- self._active_signs = signs
- ctx = self._build_context(window, active)
+ # Hand subclasses the window in POSITION space: a short's column is
+ # negated, so every context built from it describes what is actually
+ # being sized. ``mu`` becomes the position's expected return (a
+ # short earns the negative of its asset's drift) and ``cov`` becomes
+ # D Sigma D, whose cross terms flip sign for a long/short pair --
+ # signing only ``mu`` would score the numerator in position space
+ # and the variance in asset space, so a hedged pair would still
+ # read as correlated. Volatility-only contexts are unaffected
+ # (std(-r) == std(r)).
+ signed = window.mul(pd.Series(signs, index=window.columns), axis=1)
+ ctx = self._build_context(signed, active)
if ctx is None:
continue
@@ -103,7 +107,10 @@ class BaseOptimizer(ABC):
Return None to skip the date.
Args:
- window: Return window for active assets.
+ window: Return window for active assets, in POSITION space -- a
+ short's column is already negated by ``optimize``, so a mean
+ taken here is the position's expected return and a covariance
+ is the position covariance.
active: Active asset codes.
Returns:
diff --git a/agent/backtest/optimizers/mean_variance.py b/agent/backtest/optimizers/mean_variance.py
index 0fe763df0..dec299ac9 100644
--- a/agent/backtest/optimizers/mean_variance.py
+++ b/agent/backtest/optimizers/mean_variance.py
@@ -18,17 +18,16 @@ class MeanVarianceOptimizer(BaseOptimizer):
def _build_context(
self, window: pd.DataFrame, active: List[str]
) -> "Dict[str, Any] | None":
- """Mean vector and covariance.
-
- ``mu`` is the expected return of the position actually being sized
- (``sign(pos) * raw asset drift``), not the raw asset drift -- a
- short's expected return is the negative of the asset's own drift.
- Without this, a strong short candidate (very negative raw drift)
- scores as a bad "long" in the Sharpe objective below and is starved
- of capital relative to a weak short (near-zero drift).
+ """Mean vector and covariance of the positions being sized.
+
+ ``window`` arrives in position space (see ``BaseOptimizer.optimize``),
+ so ``mu`` is the expected return of the position and ``cov`` its
+ covariance. On raw asset returns a strong short candidate (very
+ negative drift) scored as a bad "long" in the Sharpe objective below
+ and was starved of capital relative to a weak short, and a long/short
+ pair that hedges read as correlated.
"""
- signs = self._active_signs
- mu = signs * window.mean().values
+ mu = window.mean().values
cov = window.cov().values
if np.isnan(cov).any() or np.isnan(mu).any():
return None
diff --git a/agent/backtest/optimizers/turnover_aware.py b/agent/backtest/optimizers/turnover_aware.py
index a5afdcb53..1646edb0c 100644
--- a/agent/backtest/optimizers/turnover_aware.py
+++ b/agent/backtest/optimizers/turnover_aware.py
@@ -97,14 +97,13 @@ class TurnoverAwareOptimizer(BaseOptimizer):
) -> "Dict[str, Any] | None":
"""Mean vector, covariance, and active codes for the current window.
- ``mu`` is the expected return of the position actually being sized
- (``sign(pos) * raw asset drift``), not the raw asset drift -- see
- ``MeanVarianceOptimizer._build_context`` for the same fix and its
- rationale; this optimizer's ``objective`` has the identical
- ``w @ mu`` return term.
+ ``window`` arrives in position space (see ``BaseOptimizer.optimize``),
+ so ``mu`` is the position's expected return and ``cov`` the position
+ covariance -- this optimizer's ``objective`` has the same ``w @ mu``
+ term and the same ``w @ cov @ w`` variance term as the mean-variance
+ one, and needs both in the same space.
"""
- signs = self._active_signs
- mu = signs * window.mean().values
+ mu = window.mean().values
cov = window.cov().values
if np.isnan(cov).any() or np.isnan(mu).any():
return None
diff --git a/agent/tests/test_mean_variance_optimizer.py b/agent/tests/test_mean_variance_optimizer.py
index 18d13bb19..ba636e711 100644
--- a/agent/tests/test_mean_variance_optimizer.py
+++ b/agent/tests/test_mean_variance_optimizer.py
@@ -53,6 +53,41 @@ class TestMeanVarianceOptimize:
assert abs(last["STRONG"]) > abs(last["WEAK"])
+ def test_hedged_pair_is_not_starved_by_asset_space_covariance(self) -> None:
+ """A long and a short of two positively correlated assets hedge each
+ other, so both legs must be funded.
+
+ Regression on the other half of the same bug: signing only ``mu`` left
+ the variance term in ASSET space, where the pair still reads +0.92
+ correlated and diversification looks worthless. Measured on this
+ fixture, the asset-covariance objective put 100% of the book on the
+ long and 0% on the short; the position covariance (D Sigma D, off
+ diagonal -9.4e-05 instead of +9.4e-05) splits it 0.514 / -0.486.
+ """
+ rng = np.random.default_rng(11)
+ n = 200
+ common = rng.normal(0, 0.01, n)
+ dates = pd.bdate_range("2025-01-01", periods=n)
+ ret = pd.DataFrame(
+ {
+ "LONG": 0.002 + common + rng.normal(0, 0.003, n),
+ "SHORT": -0.0015 + common + rng.normal(0, 0.003, n),
+ },
+ index=dates,
+ )
+ assert ret.corr().iloc[0, 1] > 0.9 # the pair really is correlated
+
+ pos = pd.DataFrame(0.0, index=dates, columns=["LONG", "SHORT"])
+ pos.iloc[150:, 0] = 1.0
+ pos.iloc[150:, 1] = -1.0
+
+ result = MeanVarianceOptimizer(lookback=150).optimize(ret, pos, dates)
+ last = result.iloc[-1]
+
+ assert last["LONG"] > 0 and last["SHORT"] < 0
+ assert abs(last["SHORT"]) > 0.3, "the hedging short must be funded"
+ assert abs(last["LONG"]) > 0.3
+
def test_single_asset_unchanged(self) -> None:
dates = pd.bdate_range("2025-01-01", periods=100)
ret = pd.DataFrame(
diff --git a/agent/tests/test_turnover_aware_optimizer.py b/agent/tests/test_turnover_aware_optimizer.py
index 7c8ad477d..41f630972 100644
--- a/agent/tests/test_turnover_aware_optimizer.py
+++ b/agent/tests/test_turnover_aware_optimizer.py
@@ -140,6 +140,42 @@ class TestTurnoverAwareOptimize:
last = result.iloc[-1]
assert abs(last["STRONG"]) > abs(last["WEAK"])
+ def test_hedged_pair_is_not_starved_by_asset_space_covariance(self) -> None:
+ """Same second half as the mean-variance optimizer: this objective's
+ ``lam * (w @ cov @ w)`` risk term needs the POSITION covariance, or a
+ long/short pair of correlated assets reads as correlated and the
+ hedging short is starved.
+
+ Measured on this fixture at ``risk_aversion=100`` (daily variances run
+ ~1e-4, so the default lam=1 makes the risk term negligible and either
+ covariance concentrates): asset covariance gives 0.972 / -0.028, the
+ position covariance 0.521 / -0.479.
+ """
+ rng = np.random.default_rng(11)
+ n = 200
+ common = rng.normal(0, 0.01, n)
+ dates = pd.bdate_range("2025-01-01", periods=n)
+ ret = pd.DataFrame(
+ {
+ "LONG": 0.002 + common + rng.normal(0, 0.003, n),
+ "SHORT": -0.0015 + common + rng.normal(0, 0.003, n),
+ },
+ index=dates,
+ )
+ pos = pd.DataFrame(0.0, index=dates, columns=["LONG", "SHORT"])
+ pos.iloc[150:, 0] = 1.0
+ pos.iloc[150:, 1] = -1.0
+
+ result = optimize(
+ ret, pos, dates, lookback=150, turnover_penalty=0.0,
+ risk_aversion=100.0,
+ )
+ last = result.iloc[-1]
+
+ assert last["LONG"] > 0 and last["SHORT"] < 0
+ assert abs(last["SHORT"]) > 0.3, "the hedging short must be funded"
+ assert abs(last["LONG"]) > 0.3
+
def test_short_window_and_nan_do_not_raise(self) -> None:
ret, pos, dates = _sample_data(n_days=80)
ret.iloc[10:20, 0] = np.nan