Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(backtest): seed the equity.csv drawdown peak with initial capital (#1818)

DrDLyuan7d80e9654fb2
diff --git a/agent/backtest/engines/base.py b/agent/backtest/engines/base.py
index a21d9342c..ab850a198 100644
--- a/agent/backtest/engines/base.py
+++ b/agent/backtest/engines/base.py
@@ -2278,7 +2278,9 @@ class BaseEngine(ABC):
# Equity curve
port_ret = equity_series.pct_change().fillna(0.0)
- peak = equity_series.cummax()
+ # Same high-water mark as calc_metrics: the account holds initial_capital
+ # before the first recorded bar, so a loss on that bar is a drawdown too.
+ peak = equity_series.cummax().clip(lower=float(self.initial_capital))
dd = (equity_series - peak) / peak.replace(0, 1)
eq_df = pd.DataFrame({
"ret": port_ret,
diff --git a/agent/tests/test_equity_artifact_drawdown.py b/agent/tests/test_equity_artifact_drawdown.py
new file mode 100644
index 000000000..905287245
--- /dev/null
+++ b/agent/tests/test_equity_artifact_drawdown.py
@@ -0,0 +1,76 @@
+"""The equity.csv drawdown column must agree with the reported max drawdown.
+
+``calc_metrics`` measures drawdown from ``initial_cash``, which the account holds
+before the first recorded bar. The equity artifact, which the Web UI plots as
+Drawdown%, took its high-water mark from the first recorded bar instead, so a
+loss on that bar never showed in it. That bar trades whenever a warm-up boundary
+is declared: the last warm-up signal fills at its open.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pandas as pd
+import pytest
+
+from backtest.engines.base import BaseEngine
+
+
+class _Engine(BaseEngine):
+ """Frictionless engine: only the drawdown bookkeeping is under test."""
+
+ def can_execute(self, symbol, direction, bar):
+ return True
+
+ def round_size(self, raw_size, price):
+ return float(raw_size)
+
+ def calc_commission(self, size, price, direction, is_open):
+ return 0.0
+
+ def apply_slippage(self, price, direction):
+ return price
+
+
+class _Loader:
+ def __init__(self, data_map):
+ self._data_map = data_map
+
+ def fetch(self, codes, start_date, end_date, fields=None, interval="1D"):
+ return self._data_map
+
+
+class _AlwaysLong:
+ def generate(self, data):
+ return {code: pd.Series(1.0, index=frame.index) for code, frame in data.items()}
+
+
+def test_equity_csv_drawdown_counts_a_loss_on_the_first_bar(tmp_path: Path) -> None:
+ dates = pd.bdate_range("2025-01-06", periods=5)
+ # Two warm-up bars, then the first evaluated bar opens at 100 and closes at 90.
+ frame = pd.DataFrame(
+ {
+ "open": [100.0, 100.0, 100.0, 95.0, 100.0],
+ "close": [100.0, 100.0, 90.0, 95.0, 100.0],
+ },
+ index=dates,
+ )
+ metrics = _Engine({"initial_cash": 100_000.0}).run_backtest(
+ {
+ "codes": ["AAA"],
+ "start_date": str(dates[0].date()),
+ "end_date": str(dates[-1].date()),
+ "warmup_bars": 2,
+ },
+ _Loader({"AAA": frame}),
+ _AlwaysLong(),
+ tmp_path,
+ )
+ equity = pd.read_csv(tmp_path / "artifacts" / "equity.csv", index_col="timestamp")
+
+ # The whole account bought at 100 and is marked at 90: 10% below the cash.
+ assert equity["equity"].iloc[0] == pytest.approx(90_000.0)
+ assert metrics["max_drawdown"] == pytest.approx(-0.10)
+ assert equity["drawdown"].tolist() == pytest.approx([-0.10, -0.05, 0.0])
+ assert equity["drawdown"].min() == pytest.approx(metrics["max_drawdown"])