From 03f8966856a8de677b3ef1aa95082ea9f3622ffa Mon Sep 17 00:00:00 2001
From: kevin9327 <5299031+kevin9327@users.noreply.github.com>
Date: Thu, 10 Sep 2026 07:50:04 +0900
Subject: [PATCH] =?UTF-8?q?fix(backtest):=20=E8=B4=A2=E5=8A=A1=E5=9B=A0?=
 =?UTF-8?q?=E5=AD=90=E7=9F=A9=E9=98=B5=E8=B7=AF=E5=BE=84=E4=B8=8D=E5=86=8D?=
 =?UTF-8?q?=E6=B2=BF=E7=94=A8=E4=B8=8A=E4=B8=80=E6=9C=9F=E7=9A=84=E8=BF=87?=
 =?UTF-8?q?=E6=9C=9F=E5=80=BC?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

矩阵侧逐列前向填充时跳过空值写入, 新一期财报缺某指标就继续沿用上一期,
同一行会混用两期报告 (pb 取新期净资产、roe 停在上一期); polars 侧
join_asof 只认最新一期整行, 该指标为 null。同一份配置两条路径给出不同
因子值。改为空值同样覆盖为 NaN, 与 attach_fundamental_factors 口径一致。
---
 backend/app/backtest/fundamentals.py      |  8 +++---
 backend/tests/test_fundamental_factors.py | 30 +++++++++++++++++++++++
 2 files changed, 35 insertions(+), 3 deletions(-)

diff --git a/backend/app/backtest/fundamentals.py b/backend/app/backtest/fundamentals.py
index 0b0e8aa..940d927 100644
--- a/backend/app/backtest/fundamentals.py
+++ b/backend/app/backtest/fundamentals.py
@@ -181,9 +181,11 @@ def build_fundamental_matrices(
             continue
         for column, target in raw_columns.items():
             value = snapshot[column][row_index]
-            if value is None or not np.isfinite(float(value)):
-                continue
-            target[start:, column_index] = float(value)
+            numeric = float("nan") if value is None else float(value)
+            # 新一期该指标为空时必须覆盖旧值为 NaN: 跳过写入会让同一行混用两期
+            # 报告 (bps 取新期、roe 停在上一期), 与 polars 侧 join_asof 只认
+            # 最新一期整行的口径不一致。
+            target[start:, column_index] = numeric if np.isfinite(numeric) else np.nan
 
     for name in requested:
         spec = FUNDAMENTAL_FACTORS[name]
diff --git a/backend/tests/test_fundamental_factors.py b/backend/tests/test_fundamental_factors.py
index 9e4bc8f..de52f1f 100644
--- a/backend/tests/test_fundamental_factors.py
+++ b/backend/tests/test_fundamental_factors.py
@@ -140,6 +140,36 @@ def test_matrix_field_matches_polars_attach_across_two_announcements():
             np.testing.assert_allclose(actual, value, rtol=1e-6)
 
 
+def test_matrix_field_clears_value_when_newer_report_lacks_metric():
+    """新一期财报缺该指标时不得继续沿用上一期值, 否则同一行混用两期报告。
+
+    矩阵路径逐列前向填充, 若跳过空值写入, 4-16 起 pb 已换到新期 bps=6,
+    roe 却仍停在上一期的 20 —— polars 侧 join_asof 只认最新一期整行 (roe 为
+    null), 两条路径给出不同的因子值。
+    """
+    panel = _daily_panel(date(2026, 4, 1), 20, ("600000.SH",))
+    snapshot = _snapshot_frame([
+        {"symbol": "600000.SH", "announce": "2026-04-05", "roe": 20.0, "bps": 4.0},
+        {"symbol": "600000.SH", "announce": "2026-04-15", "roe": None, "bps": 6.0},
+    ])
+    attached = attach_fundamental_factors(panel, snapshot, ["roe_latest", "pb_latest"]).sort("date")
+    market = build_market_data_matrix(panel)
+    matrices = build_fundamental_matrices(market, snapshot, ["roe_latest", "pb_latest"])
+    column = market.symbols.index("600000.SH")
+
+    # polars 口径: 新公告生效 (4-16) 后 roe 无值, pb 走新一期 bps
+    assert attached["roe_latest"].to_list()[15:] == [None] * 5
+    assert all(value is not None for value in attached["pb_latest"].to_list()[15:])
+
+    for name in ("roe_latest", "pb_latest"):
+        for row_index, value in enumerate(attached[name].to_list()):
+            actual = matrices[name][row_index, column]
+            if value is None:
+                assert np.isnan(actual), (name, row_index, actual)
+            else:
+                np.testing.assert_allclose(actual, value, rtol=1e-6)
+
+
 def test_bps_nonpositive_gives_null_pb():
     panel = _daily_panel(date(2026, 4, 1), 8, ("000001.SZ",))
     snapshot = _snapshot_frame([