fix: add jp kr closing auction phase windows (#1875)

2026-09-20 10:53:33 +08:00 · 2026-07-03 20:14:17 +08:00
parent d70af7eafb
commit 1a06568e4e
4 changed files with 93 additions and 11 deletions
--- a/docs/CHANGELOG.md
+++ b/docs/CHANGELOG.md
@@ -11,6 +11,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 - [新功能] 钉钉群机器人通知支持 — 支持通过 `DINGTALK_WEBHOOK_URL` 和 `DINGTALK_SECRET` 配置钉钉推送，并支持长文本自动切片以适配 20KB 限制。

 - [文档] 记录 Agent `/chat/stream` progress event 契约，说明新增 `stage_start`、`stage_done`、`pipeline_timeout`、`pipeline_budget_skipped` 的字段语义、Web 兼容边界、验证方式、回滚方式；其中 `pipeline_budget_skipped` 表示剩余预算不足、未启动下一阶段即跳过的语义；本变更不触及 provider/model/Base URL 或运行时配置迁移语义。
+- [修复] 日股/韩股 `market_phase` 补齐收盘集合竞价识别：JP 15:25-15:30 与 KR 15:20-15:30 现在会进入 `closing_auction`，避免临近收盘阶段仍被标记为普通 `intraday`；仅调整阶段标签和派生 `market_phase_summary`，不改变数据源、配置或交易日 fail-open/fail-closed 语义。
 - [修复] Discord 长报告推送按 2000 字符上限分片逐段发送，遇到 429 限流会按 `retry_after`/`Retry-After` 有限重试，避免中途失败后只收到前半段报告。
 - [改进] #1777 台股三大法人 fetcher（`TwInstitutionalFetcher`）增加缓存防击穿：并发同 (市场, 日期) 调用合并为单次上游请求，保护 TWSE T86 ~3 req/5s 限流额度；不同 key 仍并行；新增并发单次抓取、不同 key 各抓一次、HTTP 错误 fail-open 回归测试。
 - [修复] 修复桌面端启动时 `.env` 中 `WEBUI_PORT` 与 Electron 自动选择端口不一致会导致窗口继续等待旧端口并连接超时的问题。
--- a/docs/market-support.md
+++ b/docs/market-support.md
@@ -17,7 +17,7 @@
 - 日股/韩股日线和基础实时/近实时行情只走 `YfinanceFetcher`，不尝试 AkShare、Tushare、Efinance、Pytdx、Baostock 等 A 股专属数据源；yfinance 报价会尽量带上 `market`、`currency`、`data_quality`、`missing_fields` 等质量元数据。
 - 基本面复用既有 offshore yfinance 轻量路径；A 股专属资金流、龙虎榜、板块等能力按 `not_supported` 降级，offshore 基本面上下文也会标记 provider、as_of、data_quality 和缺失块。
 - 报告 Prompt 已增加日股/韩股市场语义，避免套用 A 股涨跌停、北向资金、龙虎榜、融资融券等概念。
- 交易日历注册 `jp: XTKS / Asia/Tokyo` 与 `kr: XKRX / Asia/Seoul`。若本地 `exchange-calendars` 版本缺少对应日历，既有 fail-open/fail-closed 语义保持不变。
+- 交易日历注册 `jp: XTKS / Asia/Tokyo` 与 `kr: XKRX / Asia/Seoul`。日股常规阶段可识别盘前、盘中、午休、15:25-15:30 收盘集合竞价、盘后与非交易日；韩股常规阶段可识别盘前、盘中、15:20-15:30 收盘集合竞价、盘后与非交易日。若本地 `exchange-calendars` 版本缺少对应日历，既有 fail-open/fail-closed 语义保持不变。

 兼容性与回退说明（针对结构化检测命中项）：

@@ -107,7 +107,7 @@ PY
 - 台股日线和基础实时/近实时行情只走 `YfinanceFetcher`，不尝试 AkShare、Tushare、Efinance、Pytdx、Baostock 等 A 股专属数据源。
 - 基本面复用既有 offshore yfinance 轻量路径；`institution` 区块额外消费台股三大法人资料并渲染到报告，A 股专属资金流、龙虎榜、板块等能力按 `not_supported` 降级。
 - 报告 Prompt 已增加台股市场语义（新台币、三大法人、TWSE/TPEx ±10% 涨跌停），并将三大法人净买卖超注入 LLM 分析上下文，避免套用 A 股北向资金、龙虎榜等概念。
- 交易日历注册 `tw: XTAI / Asia/Taipei`。TWSE 为 09:00–13:30 连续交易、无午休；收盘集合竞价 13:25–13:30 已按 5 分钟启发式窗口建模（`_CLOSING_AUCTION_WINDOW_MINUTES["tw"]=5`，`market_phase` 可返回 `closing_auction`），jp/kr 暂未建模。若本地 `exchange-calendars` 版本缺少对应日历，既有 fail-open/fail-closed 语义保持不变。
+- 交易日历注册 `tw: XTAI / Asia/Taipei`。TWSE 为 09:00–13:30 连续交易、无午休；收盘集合竞价 13:25–13:30 已按 5 分钟启发式窗口建模（`_CLOSING_AUCTION_WINDOW_MINUTES["tw"]=5`，`market_phase` 可返回 `closing_auction`）。JP/KR 也已按常规交易时段补齐收盘集合竞价窗口（JP 15:25-15:30、KR 15:20-15:30）。若本地 `exchange-calendars` 版本缺少对应日历，既有 fail-open/fail-closed 语义保持不变。
 - 主要指数提供加权指数 `^TWII` 与柜买指数 `^TWOII`。
 - 三大法人买卖超（institutional flows）资料层：`TwInstitutionalFetcher`（`data_provider/tw_institutional_fetcher.py`）提供上市（TWSE T86，legacy `rwd` 端点）/ 上柜（TPEx OpenAPI）每日外资·投信·自营商·三大法人买卖超（单位：**股数**；按日期+市场做单日全市场缓存再过滤个股，TPEx 民国年转西元有单测覆盖）。接口失败/限流/空响应/字段缺失一律 **fail-open** 返回无数据，不中断分析；仅对 `.TW`/`.TWO` 生效，不改动现有市场流程。资料来源为政府开放资料，采「政府资料开放授权条款第 1 版」(OGDL v1，允许商用与再散布，需标示来源)。
 - 三大法人 fetcher 已具备并发缓存防击穿和按 TWSE/TPEx 分流的熔断保护；TPEx OpenAPI 仅服务最新交易日，传入与服务日期不符的明确日期会 fail-open 返回无数据，避免错日资料静默进入报告。
--- a/src/core/trading_calendar.py
+++ b/src/core/trading_calendar.py
@@ -53,10 +53,18 @@ MARKET_TIMEZONE = {
 # P0 market phase baseline (Issue #1386). This is an intentionally small
 # regular-session inference layer; it does not change existing fail-open
 # trading-day filtering or effective-date behavior.
-# tw: TWSE/TPEx run a 13:25–13:30 closing call auction (5 min). Without an entry
-# here .get(market, 0) yields a zero-width window, so the tw closing auction was
-# never labelled CLOSING_AUCTION (the last tick stayed INTRADAY, then POSTMARKET).
-_CLOSING_AUCTION_WINDOW_MINUTES = {"cn": 3, "hk": 10, "us": 5, "tw": 5}
+# tw: TWSE/TPEx run a 13:25-13:30 closing call auction (5 min). JP/KR use
+# regular-session closing auction windows before the 15:30 close (JP 5 min,
+# KR 10 min). Without an entry here .get(market, 0) yields a zero-width
+# window, so the last regular-session minutes stay INTRADAY until POSTMARKET.
+_CLOSING_AUCTION_WINDOW_MINUTES = {
+    "cn": 3,
+    "hk": 10,
+    "us": 5,
+    "jp": 5,
+    "kr": 10,
+    "tw": 5,
+}
 _SUPPORTED_ANALYSIS_PHASES = {
    "auto",
    "premarket",
--- a/tests/test_trading_calendar.py
+++ b/tests/test_trading_calendar.py
@@ -115,12 +115,32 @@ class _NaiveTimestampCalendar(_FakeCalendar):
        return pd.Timestamp(datetime.combine(session.date(), time(self._close_hour, 0)))


-class _HalfHourCloseCalendar(_FakeCalendar):
-    """TWSE closes at 13:30 (half-hour); _FakeCalendar only models on-the-hour close."""
+class _CloseTimeCalendar(_FakeCalendar):
+    """Calendar with a minute-level close time; _FakeCalendar only models hour close."""
+
+    def __init__(
+        self,
+        *,
+        sessions,
+        close_time: time,
+        tz_name: str,
+        open_time: time = time(9, 30),
+        break_start: Optional[time] = None,
+        break_end: Optional[time] = None,
+    ):
+        super().__init__(
+            sessions=sessions,
+            close_hour=close_time.hour,
+            tz_name=tz_name,
+            open_time=open_time,
+            break_start=break_start,
+            break_end=break_end,
+        )
+        self._close_time = close_time

    def session_close(self, session: pd.Timestamp) -> pd.Timestamp:
        local_close = datetime.combine(
-            session.date(), time(13, 30), tzinfo=ZoneInfo(self._tz_name)
+            session.date(), self._close_time, tzinfo=ZoneInfo(self._tz_name)
        )
        return pd.Timestamp(local_close).tz_convert("UTC")

@@ -397,9 +417,9 @@ class InferMarketPhaseTestCase(unittest.TestCase):

    def test_tw_phase_boundaries_include_five_minute_closing_window(self):
        # TWSE: continuous 09:00-13:30, no lunch break, 13:25-13:30 closing auction.
-        fake_calendar = _HalfHourCloseCalendar(
+        fake_calendar = _CloseTimeCalendar(
            sessions=[date(2026, 3, 27)],
-            close_hour=13,  # unused: _HalfHourCloseCalendar hard-codes the 13:30 close
+            close_time=time(13, 30),
            tz_name="Asia/Taipei",
            open_time=time(9, 0),
            break_start=None,
@@ -420,6 +440,59 @@ class InferMarketPhaseTestCase(unittest.TestCase):
            with self.subTest(current_time=current_time):
                self.assertEqual(self._infer_with_calendar("tw", current_time, fake_calendar), expected)

+    def test_jp_phase_boundaries_include_lunch_and_five_minute_closing_window(self):
+        # TSE domestic stocks: 09:00-11:30, 12:30-15:30, closing auction 15:25-15:30.
+        fake_calendar = _CloseTimeCalendar(
+            sessions=[date(2026, 3, 27)],
+            close_time=time(15, 30),
+            tz_name="Asia/Tokyo",
+            open_time=time(9, 0),
+            break_start=time(11, 30),
+            break_end=time(12, 30),
+        )
+
+        tz = ZoneInfo("Asia/Tokyo")
+        cases = (
+            (datetime(2026, 3, 27, 8, 59, tzinfo=tz), trading_calendar.MarketPhase.PREMARKET),
+            (datetime(2026, 3, 27, 9, 0, tzinfo=tz), trading_calendar.MarketPhase.INTRADAY),
+            (datetime(2026, 3, 27, 11, 30, tzinfo=tz), trading_calendar.MarketPhase.LUNCH_BREAK),
+            (datetime(2026, 3, 27, 12, 30, tzinfo=tz), trading_calendar.MarketPhase.INTRADAY),
+            (datetime(2026, 3, 27, 15, 24, tzinfo=tz), trading_calendar.MarketPhase.INTRADAY),
+            (datetime(2026, 3, 27, 15, 25, tzinfo=tz), trading_calendar.MarketPhase.CLOSING_AUCTION),
+            (datetime(2026, 3, 27, 15, 29, tzinfo=tz), trading_calendar.MarketPhase.CLOSING_AUCTION),
+            (datetime(2026, 3, 27, 15, 30, tzinfo=tz), trading_calendar.MarketPhase.POSTMARKET),
+        )
+
+        for current_time, expected in cases:
+            with self.subTest(current_time=current_time):
+                self.assertEqual(self._infer_with_calendar("jp", current_time, fake_calendar), expected)
+
+    def test_kr_phase_boundaries_include_ten_minute_closing_window_without_lunch(self):
+        # KRX regular session: 09:00-15:30, closing auction 15:20-15:30.
+        fake_calendar = _CloseTimeCalendar(
+            sessions=[date(2026, 3, 27)],
+            close_time=time(15, 30),
+            tz_name="Asia/Seoul",
+            open_time=time(9, 0),
+            break_start=None,
+            break_end=None,
+        )
+
+        tz = ZoneInfo("Asia/Seoul")
+        cases = (
+            (datetime(2026, 3, 27, 8, 59, tzinfo=tz), trading_calendar.MarketPhase.PREMARKET),
+            (datetime(2026, 3, 27, 9, 0, tzinfo=tz), trading_calendar.MarketPhase.INTRADAY),
+            (datetime(2026, 3, 27, 12, 0, tzinfo=tz), trading_calendar.MarketPhase.INTRADAY),
+            (datetime(2026, 3, 27, 15, 19, tzinfo=tz), trading_calendar.MarketPhase.INTRADAY),
+            (datetime(2026, 3, 27, 15, 20, tzinfo=tz), trading_calendar.MarketPhase.CLOSING_AUCTION),
+            (datetime(2026, 3, 27, 15, 29, tzinfo=tz), trading_calendar.MarketPhase.CLOSING_AUCTION),
+            (datetime(2026, 3, 27, 15, 30, tzinfo=tz), trading_calendar.MarketPhase.POSTMARKET),
+        )
+
+        for current_time, expected in cases:
+            with self.subTest(current_time=current_time):
+                self.assertEqual(self._infer_with_calendar("kr", current_time, fake_calendar), expected)
+
    def test_unknown_market_and_calendar_failures_return_unknown(self):
        current_time = datetime(2026, 3, 27, 10, 0, tzinfo=ZoneInfo("Asia/Shanghai"))