diff --git a/backend/app/services/backtest.py b/backend/app/services/backtest.py
index d3c9640..4f4e942 100644
--- a/backend/app/services/backtest.py
+++ b/backend/app/services/backtest.py
@@ -118,6 +118,31 @@ _SIGNAL_COLS: dict[SignalKind, str] = {
 }
 
 
+def _build_max_hold_exits(entries: pd.DataFrame, max_hold_days: int) -> pd.DataFrame:
+    """为每个入场信号在 max_hold_days 个交易日后生成一个强制退出信号。
+
+    返回与 entries 同形状的布尔矩阵, 仅在「入场位之后第 max_hold_days 个交易日」置
+    True(不含入场位本身), 供调用方与用户 exits 做 OR。
+
+    两处易错点(见 #198):
+    - 必须从全 False 起步。若用 `entries.copy()` 起步会把入场位当成退出位,
+      导致入场当日即被强制平仓。
+    - 用单步定位写入 `iloc[row, col_loc]`。链式 `iloc[row][col] = True` 写入的是
+      临时行副本, 在 pandas Copy-on-Write 语义下不会落到原矩阵(pandas 3.x 直接报错),
+      强制退出信号会静默丢失。
+    """
+    out = pd.DataFrame(False, index=entries.index, columns=entries.columns)
+    n = len(entries)
+    for col in entries.columns:
+        col_loc = out.columns.get_loc(col)
+        entry_rows = np.where(entries[col].to_numpy())[0]
+        for i in entry_rows:
+            end_i = min(int(i) + max_hold_days, n - 1)
+            if end_i > i:
+                out.iloc[end_i, col_loc] = True
+    return out
+
+
 class BacktestService:
     def __init__(self, repo: KlineRepository) -> None:
         self.repo = repo
@@ -270,16 +295,10 @@ class BacktestService:
             if config.stop_loss_pct is not None:
                 pf_kwargs["sl_stop"] = abs(config.stop_loss_pct)
             if config.max_hold_days is not None:
-                # vectorbt 没有内置 max-hold;用时间退出近似:
-                # 在 max_hold_days 后强制 exit
-                exits_idx = entries.copy()
-                for col in entries.columns:
-                    entry_rows = np.where(entries[col].values)[0]
-                    for i in entry_rows:
-                        end_i = min(i + config.max_hold_days, len(entries) - 1)
-                        if end_i > i:
-                            exits_idx.iloc[end_i][col] = True
-                pf_kwargs["exits"] = (exits | exits_idx).astype(bool)
+                # vectorbt 没有内置 max-hold;用时间退出近似:入场后第 max_hold_days
+                # 个交易日强制 exit, 与用户 exits 做 OR(保留原有信号退出)。
+                forced_exits = _build_max_hold_exits(entries, config.max_hold_days)
+                pf_kwargs["exits"] = (exits | forced_exits).astype(bool)
 
             pf = vbt.Portfolio.from_signals(**pf_kwargs)
         except Exception as e:  # noqa: BLE001
diff --git a/backend/tests/backtest/test_max_hold_exits.py b/backend/tests/backtest/test_max_hold_exits.py
new file mode 100644
index 0000000..f1264f3
--- /dev/null
+++ b/backend/tests/backtest/test_max_hold_exits.py
@@ -0,0 +1,59 @@
+"""max_hold_days 强制退出矩阵回归测试(issue #198)。
+
+_build_max_hold_exits 是 /api/backtest/run 里 max_hold_days 强制平仓的纯逻辑,
+不依赖 vectorbt, 可独立断言。覆盖两处历史缺陷:
+1. 链式 `iloc[row][col] = True` 在 pandas CoW 下写入丢失 → 强制退出信号从不生效。
+2. 以 `entries.copy()` 起步 → 把入场位当退出位, 入场当日即被平仓。
+"""
+from __future__ import annotations
+
+import pandas as pd
+
+from app.services.backtest import _build_max_hold_exits
+
+
+def _entries(data: dict, n: int) -> pd.DataFrame:
+    return pd.DataFrame(data, index=pd.RangeIndex(n)).astype(bool)
+
+
+def test_forced_exit_placed_max_hold_days_after_entry():
+    """入场后第 max_hold_days 个交易日置强制退出(核心: 该单元格必须真的被写入)。"""
+    entries = _entries({"A": [True, False, False, False, False]}, 5)
+    out = _build_max_hold_exits(entries, 2)
+    assert out["A"].tolist() == [False, False, True, False, False]
+
+
+def test_does_not_mark_entry_bar_as_exit():
+    """回归: 强制退出矩阵不得包含入场位本身。"""
+    entries = _entries({"A": [True, False, False]}, 3)
+    out = _build_max_hold_exits(entries, 1)
+    assert out["A"].tolist() == [False, True, False]
+
+
+def test_end_index_clamped_to_last_row():
+    """入场后越界时 clamp 到最后一根 K。"""
+    entries = _entries({"A": [False, False, False, True, False]}, 5)
+    out = _build_max_hold_exits(entries, 5)  # 3+5 越界 → clamp 到 4
+    assert out["A"].tolist() == [False, False, False, False, True]
+
+
+def test_entry_on_last_row_produces_no_exit():
+    """入场即最后一根 K 时 end_i == i, 不产生退出(避免同根自相矛盾)。"""
+    entries = _entries({"A": [False, False, True]}, 3)
+    out = _build_max_hold_exits(entries, 2)
+    assert out["A"].tolist() == [False, False, False]
+
+
+def test_multiple_entries_single_column():
+    entries = _entries({"A": [True, False, True, False, False]}, 5)
+    out = _build_max_hold_exits(entries, 1)
+    assert out["A"].tolist() == [False, True, False, True, False]
+
+
+def test_multiple_columns_independent():
+    entries = _entries({"A": [True, False, False], "B": [False, True, False]}, 3)
+    out = _build_max_hold_exits(entries, 1)
+    assert out["A"].tolist() == [False, True, False]
+    assert out["B"].tolist() == [False, False, True]
+    assert list(out.columns) == ["A", "B"]
+    assert out.index.equals(entries.index)