"Vibe-Trading: Your Personal Trading Agent"
Python ai-agentalgorithmic-tradingbacktestingfintechllmmcpmulti-agentpythonquantitative-financetrading
fix(shadow-account): overtrading span uses the true trading window
Lanre Shittu5696379043c3
diff --git a/agent/src/shadow_account/backtester.py b/agent/src/shadow_account/backtester.pyindex f2b63ba01..3cc4af237 100644--- a/agent/src/shadow_account/backtester.py+++ b/agent/src/shadow_account/backtester.py@@ -764,8 +764,14 @@ def _overtrading_pnl(median_hold, _ = profile.typical_holding_daysif median_hold <= 0:return 0.0+ # roundtrips is appended in sell-chronological order (pair_trades_fifo+ # closes each position as its sell/cover row is processed), so [-1]'s+ # sell_dt is the true latest sell, but [0]'s buy_dt is only the buy date+ # of whichever roundtrip happened to sell first -- not the true earliest+ # buy. A long hold mixed with short trades understates the span.span_days = (- pd.Timestamp(roundtrips[-1]["sell_dt"]) - pd.Timestamp(roundtrips[0]["buy_dt"])+ max(pd.Timestamp(rt["sell_dt"]) for rt in roundtrips)+ - min(pd.Timestamp(rt["buy_dt"]) for rt in roundtrips)).total_seconds() / 86400.0expected = max(1.0, span_days / max(2 * median_hold, 1.0))actual = len(roundtrips)diff --git a/agent/tests/test_shadow_attribution.py b/agent/tests/test_shadow_attribution.pyindex 92d640d02..75793ba7e 100644--- a/agent/tests/test_shadow_attribution.py+++ b/agent/tests/test_shadow_attribution.py@@ -12,7 +12,7 @@ from __future__ import annotationsimport pandas as pdimport pytest-from src.shadow_account.backtester import _compute_attribution+from src.shadow_account.backtester import _compute_attribution, _overtrading_pnlfrom src.shadow_account.models import ShadowProfile, ShadowRule@@ -126,3 +126,60 @@ def test_early_exit_trade_not_also_counted_as_overtrading() -> None:# 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)+++def test_overtrading_span_uses_the_true_earliest_buy_and_latest_sell() -> None:+ # roundtrips is appended in sell-chronological order, not buy order. A+ # long hold (bought first, sold well before the window's actual end)+ # mixed with several short, quick trades starting only after it must not+ # shrink the span to [roundtrips[0].buy_dt, roundtrips[-1].sell_dt] --+ # that drops the long hold's early buy_dt whenever something else closes+ # before it does, understating the trading window and spuriously+ # flagging a normal cadence as overtrading.+ long_hold = {+ "symbol": "AAA.US",+ "buy_dt": pd.Timestamp("2026-01-01"),+ "sell_dt": pd.Timestamp("2026-01-01") + pd.Timedelta(days=200),+ "qty": 10.0,+ "buy_price": 100.0,+ "sell_price": 105.0,+ "hold_days": 200.0,+ "pnl": 50.0,+ "pnl_pct": 0.5,+ }+ short_trades = []+ for i in range(14):+ buy = pd.Timestamp("2026-01-01") + pd.Timedelta(days=190 + i * 6)+ short_trades.append(+ {+ "symbol": f"S{i}.US",+ "buy_dt": buy,+ "sell_dt": buy + pd.Timedelta(days=2),+ "qty": 10.0,+ "buy_price": 20.0,+ "sell_price": 20.05,+ "hold_days": 2.0,+ "pnl": 0.5,+ "pnl_pct": 0.025,+ }+ )+ # pair_trades_fifo appends each roundtrip when its sell/cover row is+ # processed, so the real list order is sell-chronological -- reproduce+ # that here instead of the buy order the trades were built in.+ roundtrips = sorted([long_hold] + short_trades, key=lambda rt: rt["sell_dt"])++ profile = ShadowProfile(+ shadow_id="shadow_test",+ created_at="2026-01-01T00:00:00",+ journal_hash="test",+ source_market="us",+ profitable_roundtrips=len(roundtrips),+ total_roundtrips=len(roundtrips),+ date_range=("2026-01-01", "2026-12-31"),+ profile_text="test",+ rules=(),+ preferred_markets=("us",),+ typical_holding_days=(3.0, 4.0),+ )++ assert _overtrading_pnl(profile=profile, roundtrips=roundtrips) == 0.0