refactor: 以自实现指标替换 talib 与 technical 依赖

chanlun/indicators/ta.py 接口兼容 talib.abstract，实现代码实际用到的 SMA/MA/EMA/RSI/ATR/MACD/BBANDS；chanlun/pipeline/resample.py 替代 technical.util.resample_to_interval。调用点只改 import，逻辑未动。暖机长度与平滑种子按 TA-Lib 的约定实现，差一根 K 线就会让下游所有笔/线段/中枢整体位移。其中 MACD 需特别处理：TA-Lib 让快慢两条 EMA 在同一根 K 线出首值，因而快线的种子取 x[slow-fast:slow] 的均值，而非从 fastperiod-1 一路递推——两者在百元价位上相差约 0.17。 BBANDS 是有意的分歧：TA-Lib 用 sumsq/n - mean² 求方差，短窗口远离零时灾难性抵消（timeperiod=2 误差 8.7e-7），本实现用 rolling std，对 50 位精度基准误差为 0。项目实际使用的周期两者一致到 1e-10。顺带清理 12 个文件中 16 处从未调用的 talib/technical 导入。验证：9440 组随机对拨；真实 K 线端到端比对 add_indicators 全部 33 个指标列，NaN 模式一致、MACD 柱符号 100% 相同；屏蔽两个包后 60 个模块均可导入。新增 test_ta_compat.py 将输出逐 bar 钉在 TA-Lib 上，但该文件在 TA-Lib 缺失时静默跳过，改动 ta.py 需在装有 TA-Lib 的环境复跑。 Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-27 02:01:43 +08:00
parent 7f393b93ed
commit 206b27fe72
17 changed files with 456 additions and 25 deletions
@@ -0,0 +1,198 @@
+"""Pin chanlun.indicators.ta to TA-Lib's output, bar for bar.
+
+These indicators feed the Chan structure builders, so a one-bar shift in the
+warm-up or a different smoothing seed silently changes every downstream
+bi/seg/zs. Equality against the reference implementation is the only check
+that catches that.
+
+Skipped when talib is unavailable — which is the point of the replacement, so
+the suite still has to pass without it. Run in an environment that has talib
+whenever chanlun/indicators/ta.py changes.
+"""
+
+from __future__ import annotations
+
+import sys
+import unittest
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+
+from chanlun.indicators import ta  # noqa: E402
+from chanlun.pipeline.resample import resample_to_interval  # noqa: E402
+
+try:
+    import talib.abstract as reference
+except ImportError:  # pragma: no cover
+    reference = None
+
+requires_talib = unittest.skipIf(reference is None, "talib not installed")
+
+
+def make_ohlcv(n: int = 900, seed: int = 7) -> pd.DataFrame:
+    """Random walk with enough range for BBANDS(365) and EMA(208) to warm up."""
+    rng = np.random.default_rng(seed)
+    close = 100.0 + np.cumsum(rng.normal(0.0, 1.0, n))
+    spread = np.abs(rng.normal(0.0, 0.6, n)) + 0.05
+    high = close + spread
+    low = close - spread
+    open_ = np.concatenate([[close[0]], close[:-1]])
+    return pd.DataFrame(
+        {
+            "date": pd.date_range("2024-01-01", periods=n, freq="1min", tz="UTC"),
+            "open": open_,
+            "high": np.maximum.reduce([high, open_, close]),
+            "low": np.minimum.reduce([low, open_, close]),
+            "close": close,
+            "volume": rng.uniform(1.0, 100.0, n),
+        }
+    )
+
+
+class TAEquivalence(unittest.TestCase):
+    def setUp(self) -> None:
+        self.df = make_ohlcv()
+
+    def assertSameSeries(self, got, expected, label: str) -> None:
+        g = np.asarray(got, dtype=float)
+        e = np.asarray(expected, dtype=float)
+        self.assertEqual(g.shape, e.shape, f"{label}: shape")
+        np.testing.assert_array_equal(
+            np.isnan(g), np.isnan(e), err_msg=f"{label}: NaN warm-up differs"
+        )
+        mask = ~np.isnan(e)
+        np.testing.assert_allclose(
+            g[mask], e[mask], rtol=1e-9, atol=1e-8, err_msg=f"{label}: values differ"
+        )
+
+    @requires_talib
+    def test_sma(self) -> None:
+        for period in (5, 20, 90, 250):
+            self.assertSameSeries(
+                ta.SMA(self.df, timeperiod=period),
+                reference.SMA(self.df, timeperiod=period),
+                f"SMA({period})",
+            )
+
+    @requires_talib
+    def test_ma(self) -> None:
+        for period in (5, 10, 250):
+            self.assertSameSeries(
+                ta.MA(self.df, timeperiod=period),
+                reference.MA(self.df, timeperiod=period),
+                f"MA({period})",
+            )
+
+    @requires_talib
+    def test_ema(self) -> None:
+        for period in (5, 7, 10, 13, 24, 26, 30, 52, 104, 156, 208):
+            self.assertSameSeries(
+                ta.EMA(self.df, timeperiod=period),
+                reference.EMA(self.df, timeperiod=period),
+                f"EMA({period})",
+            )
+
+    @requires_talib
+    def test_rsi(self) -> None:
+        for period in (7, 14, 21):
+            self.assertSameSeries(
+                ta.RSI(self.df, timeperiod=period),
+                reference.RSI(self.df, timeperiod=period),
+                f"RSI({period})",
+            )
+
+    @requires_talib
+    def test_atr(self) -> None:
+        for period in (7, 14, 30):
+            self.assertSameSeries(
+                ta.ATR(self.df, timeperiod=period),
+                reference.ATR(self.df, timeperiod=period),
+                f"ATR({period})",
+            )
+
+    @requires_talib
+    def test_macd(self) -> None:
+        for fast, slow, signal in ((12, 26, 9), (26, 52, 9), (5, 35, 5)):
+            got = ta.MACD(self.df, fastperiod=fast, slowperiod=slow, signalperiod=signal)
+            exp = reference.MACD(self.df, fastperiod=fast, slowperiod=slow, signalperiod=signal)
+            for col in ("macd", "macdsignal", "macdhist"):
+                self.assertSameSeries(got[col], exp[col], f"MACD({fast},{slow},{signal}).{col}")
+
+    @requires_talib
+    def test_bbands(self) -> None:
+        cases = (
+            (365, 3.0, 3.0),
+            (120, 3.0, 3.0),
+            (41, 2.3, 2.3),
+            (41, 2.0, 2.0),
+            (26, 3.0, 3.0),
+            (20, 2.0, 2.0),
+            (14, 2.0, 2.0),
+        )
+        for period, up, dn in cases:
+            got = ta.BBANDS(self.df, timeperiod=period, nbdevup=up, nbdevdn=dn, matype=0)
+            exp = reference.BBANDS(self.df, timeperiod=period, nbdevup=up, nbdevdn=dn, matype=0)
+            for col in ("upperband", "middleband", "lowerband"):
+                self.assertSameSeries(got[col], exp[col], f"BBANDS({period},{up},{dn}).{col}")
+
+    @requires_talib
+    def test_bbands_is_more_accurate_than_talib_on_tiny_windows(self) -> None:
+        """A deliberate divergence, documented so nobody "fixes" it back.
+
+        TA-Lib derives the variance from sumsq/n - mean**2, which cancels
+        catastrophically when the window is short and prices are far from zero;
+        at timeperiod=2 it drifts ~1e-6. Rolling std is accurate there, so the
+        two disagree. No timeperiod below 14 is used in this codebase, and the
+        periods that are used agree to ~1e-10 (covered by test_bbands).
+        """
+        got = ta.BBANDS(self.df, timeperiod=2, nbdevup=1.0, nbdevdn=1.0, matype=0)["upperband"]
+        exp = reference.BBANDS(self.df, timeperiod=2, nbdevup=1.0, nbdevdn=1.0, matype=0)["upperband"]
+
+        window = self.df["close"].rolling(2)
+        truth = window.mean() + window.std(ddof=0)
+        ours = np.nanmax(np.abs((got - truth).to_numpy()))
+        theirs = np.nanmax(np.abs((exp - truth).to_numpy()))
+        self.assertLess(ours, 1e-9)
+        self.assertLess(ours, theirs)
+
+    @requires_talib
+    def test_matches_on_real_price_scale(self) -> None:
+        """Guard against tolerances that only hold near 100."""
+        df = self.df.copy()
+        for col in ("open", "high", "low", "close"):
+            df[col] *= 900.0
+        self.assertSameSeries(
+            ta.ATR(df, timeperiod=14), reference.ATR(df, timeperiod=14), "ATR@scale"
+        )
+        self.assertSameSeries(
+            ta.RSI(df, timeperiod=14), reference.RSI(df, timeperiod=14), "RSI@scale"
+        )
+
+
+class ResampleEquivalence(unittest.TestCase):
+    @unittest.skipIf(
+        __import__("importlib").util.find_spec("technical") is None,
+        "technical not installed",
+    )
+    def test_matches_technical(self) -> None:
+        from technical.util import resample_to_interval as ref_resample
+
+        df = make_ohlcv(600)
+        for interval in (5, 15, 60, "5m", "1h"):
+            got = resample_to_interval(df, interval)
+            exp = ref_resample(df, interval)
+            pd.testing.assert_frame_equal(got, exp, check_exact=False, rtol=1e-12)
+
+    def test_shapes_without_reference(self) -> None:
+        df = make_ohlcv(120)
+        out = resample_to_interval(df, 5)
+        self.assertEqual(list(out.columns), ["date", "open", "high", "low", "close", "volume"])
+        self.assertLessEqual(len(out), 120 // 5 + 1)
+        self.assertTrue((out["high"] >= out["low"]).all())
+
+
+if __name__ == "__main__":
+    unittest.main()