"Vibe-Trading: Your Personal Trading Agent"
Python ai-agentalgorithmic-tradingbacktestingfintechllmmcpmulti-agentpythonquantitative-financetrading
fix(shadow-account): exclude already explained trades from overtrading pnl
Lanre Shittud6a13530fb58
diff --git a/agent/src/shadow_account/backtester.py b/agent/src/shadow_account/backtester.pyindex 3203c2402..f2b63ba01 100644--- a/agent/src/shadow_account/backtester.py+++ b/agent/src/shadow_account/backtester.py@@ -653,6 +653,7 @@ def _compute_attribution(noise = 0.0early = 0.0late = 0.0+ already_explained: set[int] = set()excluded_currencies: dict[str, int] = {}pool_roundtrips = [rt@@ -695,6 +696,7 @@ def _compute_attribution(impact += -pnlreason = "rule_violation"if impact != 0.0:+ already_explained.add(id(rt))counterfactuals.append({"symbol": rt["symbol"],"buy_dt": str(rt["buy_dt"]),@@ -705,7 +707,9 @@ def _compute_attribution("reason": reason,})- overtrading = _overtrading_pnl(profile=profile, roundtrips=pool_roundtrips)+ overtrading = _overtrading_pnl(+ profile=profile, roundtrips=pool_roundtrips, already_explained=already_explained+ )explained = noise + early + late + overtradingmissed = round(shadow_pnl - real_pnl - explained, 2)@@ -740,6 +744,7 @@ def _overtrading_pnl(*,profile: ShadowProfile,roundtrips: list[dict[str, Any]],+ already_explained: set[int] = frozenset(),) -> float:"""Excess-frequency PnL: trades beyond the shadow's expected budget.@@ -747,6 +752,12 @@ def _overtrading_pnl(compare against the user's actual roundtrip count over the same span.Excess trades' PnL is totaled with a negative sign (shadow would'veskipped them, so real PnL — positive or negative — is "noise").++ ``already_explained`` holds ``id()`` of roundtrips the caller already+ booked into noise/early/late (#17's mutual-exclusivity invariant): those+ trades' PnL already explains their gap from shadow behaviour, so they+ are excluded from the "extra" candidates here to avoid double-counting+ the same trade in two buckets."""if not roundtrips:return 0.0@@ -761,7 +772,8 @@ def _overtrading_pnl(if actual <= expected:return 0.0# Penalize the cheapest (lowest |pnl|) extras — those look like noise.- extras = sorted(roundtrips, key=lambda rt: abs(float(rt["pnl"])))+ candidates = [rt for rt in roundtrips if id(rt) not in already_explained]+ extras = sorted(candidates, key=lambda rt: abs(float(rt["pnl"])))extra_count = int(actual - expected)extra_pnl = sum(float(rt["pnl"]) for rt in extras[:extra_count])return -extra_pnldiff --git a/agent/tests/test_shadow_attribution.py b/agent/tests/test_shadow_attribution.pyindex a556dbac8..92d640d02 100644--- a/agent/tests/test_shadow_attribution.py+++ b/agent/tests/test_shadow_attribution.py@@ -103,3 +103,26 @@ def test_short_loser_is_noise_only() -> None:assert breakdown.noise_trades_pnl == pytest.approx(30.0)assert breakdown.early_exit_pnl == 0.0assert breakdown.late_exit_pnl == 0.0+++def test_early_exit_trade_not_also_counted_as_overtrading() -> None:+ # trade 1 is an early-exit winner (hold=1d < the 3d rule): its pnl is+ # already fully explained by early_exit_pnl. Trades 2-4 hold exactly the+ # rule's 3 days, so with expected == 1.0 (low budget) and actual == 4,+ # _overtrading_pnl must draw its "extra" candidates only from trades not+ # already booked into early_exit_pnl.+ roundtrips = [+ _rt("AAPL.US", 30.0, 1.0),+ _rt("AAPL.US", 40.0, 3.0),+ _rt("AAPL.US", 50.0, 3.0),+ _rt("AAPL.US", 60.0, 3.0),+ ]+ breakdown, _, _ = _compute_attribution(+ profile=_profile(), roundtrips=roundtrips, shadow_pnl=0.0,+ )+ assert breakdown.early_exit_pnl == pytest.approx(30.0 * (3 - 1) / 3)+ # Trade 1's pnl (30.0) must not also be folded into overtrading_pnl: the+ # buggy version summed the three cheapest of ALL four trades (30+40+50 =+ # 120), double-counting trade 1; the fix draws from the three trades not+ # already explained (40+50+60 = 150).+ assert breakdown.overtrading_pnl == pytest.approx(-150.0)