Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(factors): propagate NaN through gtja191_004 instead of fabricating -1.0

Lanre Shittub03cd834c3d9
diff --git a/agent/src/factors/zoo/gtja191/alpha_004.py b/agent/src/factors/zoo/gtja191/alpha_004.py
index f49f9ec7d..6b827f1ce 100644
--- a/agent/src/factors/zoo/gtja191/alpha_004.py
+++ b/agent/src/factors/zoo/gtja191/alpha_004.py
@@ -60,4 +60,10 @@ def compute(panel: dict) -> pd.DataFrame:
res = np.where(cond_top, -1.0,
np.where(cond_bot, 1.0,
np.where(vol_strong, 1.0, -1.0)))
+ # NaN comparisons are False, not NaN, so a gap in any trailing window
+ # (a halt day inside the 8/20-day lookback, not just series warmup)
+ # would otherwise fall through to the innermost -1.0 branch instead of
+ # propagating NaN like every other operator in this zoo.
+ invalid = ma8.isna() | sd8.isna() | ma2.isna() | vmean20.isna()
+ res = np.where(invalid.to_numpy(), np.nan, res)
return pd.DataFrame(res, index=c.index, columns=c.columns).astype(float)
diff --git a/agent/tests/factors/test_gtja191_alpha004_nan_propagation.py b/agent/tests/factors/test_gtja191_alpha004_nan_propagation.py
new file mode 100644
index 000000000..19adbc56d
--- /dev/null
+++ b/agent/tests/factors/test_gtja191_alpha004_nan_propagation.py
@@ -0,0 +1,69 @@
+"""Regression: gtja191_004 must propagate NaN through a data gap, not
+fabricate a -1.0 signal.
+
+compute() picks a branch with nested np.where on comparisons like
+``upper < ma2``. A NaN comparison evaluates False, not NaN, so when any
+of the underlying rolling stats (ma8/sd8/ma2/vmean20) is NaN because a
+trading halt sits inside its lookback window, every comparison falls
+through to the innermost branch and returns a hard -1.0 instead of NaN.
+base.py's own NaN policy ("every operator propagates NaN; no silent
+fillna") requires the gap to stay NaN, and it must stay NaN for every
+bar whose window still touches the gap, not just the gap bar itself.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pandas as pd
+
+from src.factors.registry import Registry
+
+N_ROWS = 60
+GAP_ROW = 30 # past min_warmup_bars=20
+
+
+def _panel_with_one_gap() -> dict[str, pd.DataFrame]:
+ rng = np.random.default_rng(0)
+ idx = pd.date_range("2024-01-01", periods=N_ROWS, freq="D")
+ cols = ["SYM0", "SYM1"]
+
+ close = pd.DataFrame(
+ 100.0 + np.cumsum(rng.normal(0.0, 1.0, size=(N_ROWS, 2)), axis=0),
+ index=idx,
+ columns=cols,
+ )
+ volume = pd.DataFrame(
+ rng.integers(1_000, 100_000, size=(N_ROWS, 2)).astype(float),
+ index=idx,
+ columns=cols,
+ )
+
+ close.loc[idx[GAP_ROW], "SYM1"] = np.nan
+ volume.loc[idx[GAP_ROW], "SYM1"] = np.nan
+
+ return {"close": close, "volume": volume}
+
+
+def test_gap_stays_nan_through_the_full_affected_window():
+ panel = _panel_with_one_gap()
+ out = Registry().compute("gtja191_004", panel)
+
+ # The gap poisons every rolling window (up to 8/20 bars) still
+ # touching it, not just the gap row.
+ affected = out["SYM1"].iloc[GAP_ROW : GAP_ROW + 8]
+ assert affected.isna().all(), (
+ f"gtja191_004: every bar whose 8-day window touches the gap must "
+ f"stay NaN, got {affected.tolist()}"
+ )
+
+
+def test_unaffected_symbol_still_computes_real_values():
+ panel = _panel_with_one_gap()
+ out = Registry().compute("gtja191_004", panel)
+
+ unaffected = out["SYM0"].iloc[20:]
+ assert not unaffected.isna().any(), (
+ "gtja191_004: a symbol with no gap must not pick up stray NaN "
+ "from another symbol's column"
+ )
+ assert set(unaffected.unique()) <= {-1.0, 1.0}