BACKTEST_CONTEXT.md

Cortex Investment Agent -- Backtest Context Reference Auto-generated scan, read-only. Last updated: 2026-02-25

1. 技術指標完整清單

系統有兩層指標計算：

Pipeline 層 (data_pipeline/indicators.py)：寫入 technical_indicators 表，共 47 欄位
Snapshot 層 (scripts/daily_analysis_snapshot.py)：讀取 Pipeline 結果 + OHLC 原始資料，計算衍生欄位後寫入 daily_technical_snapshot 表（180+ 欄位）

1.1 MA (Simple Moving Average)

項目	值
計算方式	`close.rolling(window=N).mean()` (pandas)
週期	5, 10, 20, 50, 60, 100, 120, 200, 240
Pipeline 輸出欄位	`ma_5`, `ma_10`, `ma_20`, `ma_50`, `ma_60`, `ma_100`, `ma_120`, `ma_200`, `ma_240`
Snapshot 映射	`ma_5` → `ma5`, `ma_20` → `ma20`, `ma_60` → `ma60`, `ma_120` → `ma120`, `ma_240` → `ma240`
Snapshot 衍生欄位	`ma_alignment` ("bullish"/"bearish"/"mixed"), `price_position` ("above_all"/"below_all"/"mixed"), `ma60_slope` ("positive"/"negative"/"flat")

MA period taxonomy contract

每個 ma_<N> 欄位都是真正的 N 日 SMA — 沒有 alias。TW 市場慣用的 60/120/240 (季線/半年線/年線) 與 US 慣用的 50/100/200 並存。

ma_50 is true 50-day MA.
ma_60 is true 60-day MA (TW 季線).
ma_100 is true 100-day MA.
ma_120 is true 120-day MA (TW 半年線).
ma_200 is true 200-day MA.
ma_240 is true 240-day MA (TW 年線).

Pre-2026-04 bug history: snapshot 的 ma60 欄位實際存的是 50 日 MA (因為 writer aliased iv["ma_50"])。已於 migration 041 修正並 backfill 全部歷史資料。ma_10 仍只存在 technical_indicators，未載入 snapshot。

1.2 EMA (Exponential Moving Average)

項目	值
計算方式	`close.ewm(span=N, adjust=False).mean()` (pandas)
週期	5, 10, 20, 50, 100, 200
Pipeline 輸出欄位	`ema_5`, `ema_10`, `ema_20`, `ema_50`, `ema_100`, `ema_200`
Snapshot 映射	無（EMA 未載入 daily_technical_snapshot）

1.3 RSI (Relative Strength Index)

項目	值
計算方式	Pipeline: `ta.rsi(close, length=14)` (pandas_ta)；Snapshot: 手動 gain/loss rolling(14)
週期	14
Pipeline 輸出欄位	`rsi`
Snapshot 映射	`rsi` → `rsi_value`
Snapshot 衍生欄位	`rsi_status` ("overbought"/"oversold"/"neutral"), `rsi_divergence` ("bullish_warning"/"bearish_warning"/"none"), `rsi_slope_3` (3-day diff)

1.4 MACD (Moving Average Convergence Divergence)

項目	值
計算方式	Pipeline: `ta.macd(close, fast=12, slow=26, signal=9)` (pandas_ta)；Snapshot: 手動 EMA12-EMA26
參數	fast=12, slow=26, signal=9
Pipeline 輸出欄位	`macd`, `macd_signal`, `macd_hist`
Snapshot 映射	`macd` → `macd_dif`, `macd_signal` → `macd_dea`, `macd_hist` → `macd_histogram`
Snapshot 衍生欄位	`macd_cross_signal` ("golden_cross"/"death_cross"/"none"), `macd_divergence` ("bullish"/"bearish"/"none"), `macd_hist_slope` (3-day diff), `prev_macd_histogram`

1.5 KD (Stochastic Oscillator)

項目	值
計算方式	Pipeline: `ta.stoch(high, low, close, k=14, d=3, smooth_k=3)`；Snapshot: 手動 rolling(9) + ewm
參數	k_period=14, d_period=3, smooth=3（Pipeline）; k_period=9, d_period=3（Snapshot）
Pipeline 輸出欄位	`k`, `d`
Snapshot 映射	`k` → `kdj_k`, `d` → `kdj_d`
Snapshot 衍生欄位	`kdj_j` = 3k - 2d, `kdj_signal` ("high_passivization"/"low_passivization"/"neutral")

1.6 BB (Bollinger Bands)

項目	值
計算方式	Pipeline: `ta.bbands(close, length=20, std=2)`；Snapshot: 手動 rolling(20).mean() ± 2*std
參數	period=20, std=2
Pipeline 輸出欄位	`bb_upper`, `bb_middle`, `bb_lower`
Snapshot 映射	直接 1:1
Snapshot 衍生欄位	`bb_percent_b` = (close-lower)/(upper-lower), `bb_bandwidth` = (upper-lower)/middle, `bb_status` ("expanding"/"contracting"/"squeezing"/"normal"), `prev_bb_bandwidth`

1.7 ATR (Average True Range)

項目	值
計算方式	Pipeline: `ta.atr(high, low, close, length=14)`；Snapshot: 手動 TR rolling(14).mean()
週期	14
Pipeline 輸出欄位	`atr`
Snapshot 映射	`atr` → `atr_value`
Snapshot 衍生欄位	`atr_status` ("high_volatility"/"normal"), `atr_stop_distance` (= atr * 2.0)

1.8 ADX (Average Directional Index)

項目	值
計算方式	Pipeline: `ta.adx(high, low, close, length=14)` 取 `ADX_14`；Snapshot: 手動 DI → DX → rolling(14)
週期	14
Pipeline 輸出欄位	`adx`
Snapshot 映射	`adx` → `adx_value`
Snapshot 衍生欄位	`adx_status` ("trending"/"non_trending"), `adx_slope` (3-day diff), `adx_yesterday`, `adx_rising_days` (連續上升天數), `plus_di`, `minus_di`

1.9 OBV (On-Balance Volume)

項目	值
計算方式	Pipeline: `ta.obv(close, volume)`；Snapshot: 手動 sign(diff)*volume cumsum
Pipeline 輸出欄位	`obv`
Snapshot 映射	OBV 原始值不寫入 snapshot
Snapshot 衍生欄位	`obv_trend` ("new_high"/"new_low"/"flat"), `obv_slope` ("positive"/"negative"/"flat")

1.10 Williams %R

項目	值
計算方式	`ta.willr(high, low, close, length=14)` (pandas_ta)
週期	14
Pipeline 輸出欄位	`willr`
Snapshot 映射	無（未載入 daily_technical_snapshot）

1.11 Volume MA

項目	值
計算方式	`volume.rolling(window=N).mean()` (pandas)
週期	5, 20
Pipeline 輸出欄位	`volume_ma_5`, `volume_ma_20`
Snapshot 映射	`volume_ma_20` → `avg_volume_20`（`volume_ma_5` 未載入）
Snapshot 衍生欄位	`volume_ratio` = current_volume / avg_volume_20, `volume_status` ("aggressive_buying"/"aggressive_selling"/"normal"), `volume_divergence` ("PRICE_UP_VOL_LOW"/"PRICE_DOWN_VOL_HIGH"/None)

1.12 VWAP (Volume Weighted Average Price)

項目	值
計算方式	手動：`(typical_price * volume).cumsum() / volume.cumsum()`
Pipeline 輸出欄位	`vwap`
Snapshot 映射	直接 1:1 → `vwap`

1.13 Ichimoku Cloud

項目	值
計算方式	`ta.ichimoku(high, low, close, tenkan=9, kijun=26, senkou=52)` (pandas_ta)
參數	tenkan=9, kijun=26, senkou=52
Pipeline 輸出欄位	`ichimoku_tenkan`, `ichimoku_kijun`
Snapshot 映射	無（未載入 daily_technical_snapshot）

1.14 CCI (Commodity Channel Index)

項目	值
計算方式	Pipeline: `ta.cci(high, low, close, length=20)`；Agent: 手動 (TP-SMA(TP))/(0.015*MAD)
週期	20
Pipeline 輸出欄位	`cci`
Snapshot 映射	無

1.15 MFI (Money Flow Index)

項目	值
計算方式	`ta.mfi(high, low, close, volume, length=14)` (pandas_ta)
週期	14
Pipeline 輸出欄位	`mfi`
Snapshot 映射	無

1.16 CHOP (Choppiness Index) -- Pipeline Only

項目	值
計算方式	`ta.chop(high, low, close, length=14)` (pandas_ta)
週期	14
Pipeline 輸出欄位	`chop`
Snapshot 映射	`chop` → `chop_value`
Snapshot 衍生欄位	`chop_status` ("extreme_squeeze" if >70 / "choppy" if >61.8 / "trending" if <38.2 / else "neutral"), `chop_slope` (3-day diff)

1.17 WaveTrend (WT1, WT2) -- Pipeline Only

項目	值
計算方式	手動：HLC3 → EMA(n1=10) → deviation → CI → `wt1` = EMA(CI, n2=21), `wt2` = SMA(wt1, signal=4)
參數	n1=10, n2=21, signal_period=4
Pipeline 輸出欄位	`wt1`, `wt2`
Snapshot 映射	`wt1` → `wt1_value`, `wt2` → `wt2_value`
Snapshot 衍生欄位	`wt_signal` ("deep_bullish_cross"/"bullish_cross"/"deep_bearish_cross"/"bearish_cross"/"none"), `wt_status` ("extreme_overbought"/"extreme_oversold"/"overbought"/"oversold"/"neutral"), `wt_divergence` ("bearish_warning"/"bullish_warning")

1.18 Aroon -- Pipeline Only

項目	值
計算方式	`ta.aroon(high, low, length=25)` (pandas_ta)
週期	25
Pipeline 輸出欄位	`aroon_up`, `aroon_down`, `aroon_osc`
Snapshot 映射	→ `aroon_up_value`, `aroon_down_value`, `aroon_osc_value`
Snapshot 衍生欄位	`aroon_signal` ("strong_uptrend"/"strong_downtrend"/"neutral")

1.19 KC (Keltner Channel) -- Pipeline Only

項目	值
計算方式	`ta.kc(high, low, close, length=20, scalar=1.5)` (pandas_ta)
參數	period=20, multiplier=1.5
Pipeline 輸出欄位	`kc_upper`, `kc_middle`, `kc_lower`
Snapshot 映射	直接 1:1
Snapshot 衍生欄位 (TTM Squeeze)	`squeeze_on` (bool: bb_upper < kc_upper AND bb_lower > kc_lower), `squeeze_off` (bool: 昨日 squeeze_on 今日 off), `squeeze_duration` (連續 squeeze 天數)

1.20 Supertrend -- Pipeline Only

項目	值
計算方式	`ta.supertrend(high, low, close, length=10, multiplier=3.0)` (pandas_ta)
參數	period=10, multiplier=3.0
Pipeline 輸出欄位	`supertrend` (數值), `supertrend_direction` ("bullish"/"bearish")
Snapshot 映射	→ `supertrend_value`, `supertrend_direction`
Snapshot 衍生欄位	`distance_to_supertrend_pct` = (close-supertrend)/close*100, `supertrend_flip_recent` (bool: 5日內方向翻轉), `supertrend_flip_type` ("bullish"/"bearish")

1.21 Donchian Channel -- Pipeline Only

項目	值
計算方式	`ta.donchian(high, low, lower_length=N, upper_length=N)` (pandas_ta)，跑兩次 (20, 55)
參數	short=20, long=55
Pipeline 輸出欄位	`donchian_high_20`, `donchian_low_20`, `donchian_high_55`, `donchian_low_55`
Snapshot 映射	直接 1:1
Snapshot 衍生欄位	`breakout_20d` ("upward"/"downward"), `breakout_55d` ("upward"/"downward")

1.22 Snapshot-Only 衍生指標（不在 Pipeline 中）

以下指標僅在 daily_analysis_snapshot.py 中從 OHLC 原始資料計算，不存在於 technical_indicators 表：

CTI (Correlation Trend Indicator)

計算：np.corrcoef(close.tail(20), np.arange(20))[0,1]，clipped [-1, 1]
欄位：cti_score, cti_interpretation ("strong_trend"/"weak_trend"/"range"), cti_linearity ("high"/"medium"/"low"), cti_delta (placeholder，目前未實作)

Vortex Indicator (VI)

計算：手動 vm_plus/vm_minus rolling(14) / TR rolling(14)
欄位：vi_plus, vi_minus, vi_signal ("bullish_cross"/"bearish_cross"/"bullish"/"bearish"), vi_plus_prev, vi_minus_prev

DMI Derived

計算：從 plus_di, minus_di 衍生
欄位：di_spread = plus_di - minus_di, di_spread_slope ("widening"/"narrowing"/"stable"), adx_rising_days

Relative Strength vs Market

計算：stock_close / benchmark_close (benchmark: 0050.TW), 20d/60d percentage change
欄位：rs_eval_ready (bool), rs_vs_market_20d, rs_vs_market_60d, rs_slope ("improving"/"deteriorating"/"stable")

Slope Features (3-day delta)

rsi_slope_3：RSI[-1] - RSI[-4]
macd_hist_slope：histogram[-1] - histogram[-4]
adx_slope：ADX[-1] - ADX[-4]
chop_slope：CHOP[-1] - CHOP[-4]

2. JSON Schema 結構

build_agent_payload() 輸出 JSON 的完整欄位樹。此 payload 是 LLM 分析的輸入。

頂層結構

{
  "meta": { ... },
  "market_summary": { ... },
  "indicator_settings": { ... },
  "technical_indicators": { ... },
  "signals": { ... },
  "risk": { ... }
}

2.1 `meta`

{
  "symbol": "string (e.g. 2330.TW)",
  "trading_date": "ISO 8601 date string",
  "interval": "string (e.g. 1d)",
  "bar_state": "string",
  "data_quality": "string (complete/partial)",
  "missing_fields": ["string", ...],
  "data_source": "string",
  "notes": ["string", ...]
}

2.2 `market_summary`

{
  "current_price": float,
  "change_percent": float,
  "market_phase": "string",
  "price_bar": {
    "open": float,
    "high": float,
    "low": float,
    "close": float,
    "prev_close": float,
    "vwap": float,
    "is_adjusted": bool,
    "adjustment_type": "string"
  }
}

2.3 `indicator_settings`

{
  "moving_averages": { "type": "SMA/EMA", "periods": [5,10,20,50,100,200], "source": "close" },
  "macd": { "fast": 12, "slow": 26, "signal": 9, "source": "close" },
  "rsi": { "period": 14, "source": "close" },
  "kdj": { "k_period": 14, "d_period": 3, "j_period": 3, "source": "close" },
  "bollinger_bands": { "period": 20, "stddev": 2, "source": "close" },
  "atr": { "period": 14 },
  "adx": { "period": 14 },
  "cti": { "range": [-1.0, 1.0], "comment": "rolling corr(close, index)" },
  "choppiness": { "period": 14 },
  "wavetrend": { "n1": 10, "n2": 21, "signal_period": 4 },
  "aroon": { "period": 25 },
  "keltner": { "period": 20, "multiplier": 1.5 },
  "supertrend": { "period": 10, "multiplier": 3.0 },
  "donchian": { "short": 20, "long": 55 },
  "relative_strength": { "benchmark": "0050.TW" }
}

2.4 `technical_indicators`

{
  "trend": {
    "moving_averages": {
      "ma5": float, "ma20": float, "ma60": float,
      "alignment": "bullish|bearish|mixed",
      "price_position": "above_all|below_all|mixed",
      "ma60_slope": "positive|negative|flat"
    },
    "cti": {
      "score": float (-1~1),
      "interpretation": "strong_trend|weak_trend|range",
      "linearity": "high|medium|low"
    },
    "adx": { "value": float, "status": "trending|non_trending" },
    "vortex": {
      "vi_plus": float, "vi_minus": float,
      "signal": "bullish_cross|bearish_cross|bullish|bearish"
    },
    "choppiness": { "value": float, "status": "extreme_squeeze|choppy|trending|neutral" },
    "aroon": {
      "up": float, "down": float, "oscillator": float,
      "signal": "strong_uptrend|strong_downtrend|neutral"
    },
    "supertrend": {
      "value": float, "direction": "bullish|bearish",
      "distance_pct": float, "flip_recent": bool, "flip_type": "bullish|bearish|null"
    },
    "relative_strength": {
      "eval_ready": bool, "rs_20d": float, "rs_60d": float,
      "slope": "improving|deteriorating|stable"
    }
  },
  "momentum": {
    "macd": {
      "dif": float, "dea": float, "histogram": float,
      "cross_signal": "golden_cross|death_cross|none",
      "divergence": "bullish|bearish|none"
    },
    "rsi": {
      "value": float, "status": "overbought|oversold|neutral",
      "divergence": "bullish_warning|bearish_warning|none"
    },
    "kdj": {
      "k": float, "d": float, "j": float,
      "signal": "high_passivization|low_passivization|neutral"
    },
    "wavetrend": {
      "wt1": float, "wt2": float,
      "signal": "deep_bullish_cross|bullish_cross|deep_bearish_cross|bearish_cross|none",
      "status": "extreme_overbought|extreme_oversold|overbought|oversold|neutral",
      "divergence": "bearish_warning|bullish_warning|null"
    },
    "di_derived": {
      "spread": float, "spread_slope": "widening|narrowing|stable",
      "adx_rising_days": int
    }
  },
  "volatility": {
    "bollinger_bands": {
      "upper": float, "middle": float, "lower": float,
      "percent_b": float, "bandwidth": float,
      "status": "expanding|contracting|squeezing|normal"
    },
    "atr": { "value": float, "status": "high_volatility|normal" },
    "keltner": { "upper": float, "middle": float, "lower": float },
    "squeeze": { "on": bool, "off": bool, "duration": int },
    "donchian": {
      "high_20": float, "low_20": float,
      "high_55": float, "low_55": float,
      "breakout_20d": "upward|downward|null",
      "breakout_55d": "upward|downward|null"
    }
  },
  "volume": {
    "current_volume": int, "average_volume_20": int,
    "volume_ratio": float, "status": "aggressive_buying|aggressive_selling|normal",
    "obv": { "trend": "new_high|new_low|flat", "slope": "positive|negative|flat" }
  },
  "slopes": {
    "rsi_slope_3": float, "macd_hist_slope": float,
    "adx_slope": float, "cti_delta": float|null, "chop_slope": float
  },
  "patterns": {
    "detected": [
      {
        "name": "string", "score": float, "reliability": "string",
        "trigger_price": float, "invalidation_price": float,
        "detected_at": "string", "notes": "string"
      }
    ],
    "support_resistance": {
      "major_support": float, "major_resistance": float,
      "low_20d_min": float, "low_60d_min": float,
      "distance_to_support_20d_pct": float, "distance_to_support_60d_pct": float,
      "distance_to_support_20d_price": float, "distance_to_support_60d_price": float,
      "historical_highs": {
        "all_time_high": {
          "price": float, "date": "string", "prev_price": float,
          "touched_intraday": bool, "is_new_intraday": bool, "is_new_close": bool
        },
        "high_52w": {
          "price": float, "date": "string", "prev_price": float,
          "touched_intraday": bool, "is_new_intraday": bool, "is_new_close": bool
        },
        "swing_high_20d": { "price": float, "date": "string" },
        "swing_high_60d": { "price": float, "date": "string" }
      },
      "notes": []
    },
    "breakout_context": {
      "is_new_all_time_high": bool, "is_new_52w_high": bool,
      "close_above_prev_ath": bool, "close_above_prev_52w_high": bool,
      "breakout_type": "string|null",
      "ath_eval_ready": bool, "high_52w_eval_ready": bool,
      "distance_to_all_time_high_pct": float,
      "distance_to_52w_high_pct": float,
      "price_discovery_mode": bool
    }
  }
}

2.5 `signals`

{
  "bias": "bullish|bearish|neutral",
  "confidence": float (0.0000 ~ 1.0000),
  "reasons": ["翻譯後的 ReasonCode 字串", ...],
  "warnings": ["翻譯後的 WarningCode 字串", ...],
  "volume_divergence": "PRICE_UP_VOL_LOW|PRICE_DOWN_VOL_HIGH|null",
  "market_structure_events": ["MarketEvent value string", ...]
}

2.6 `risk`

{
  "atr_value": float,
  "atr_stop_distance": float,
  "structure_stop": float,
  "invalidation_level": float,
  "rr_targets": [
    { "name": "to_major_resistance", "target_price": float }
  ],
  "position_sizing_hint": {
    "comment": "position size = risk_per_trade / stop_distance"
  }
}

atr_stop_distance = ATR * 2.0
structure_stop = major_support（全歷史最近、左右各 5 根嚴格確認的 adjusted 前低；不設到期）
invalidation_level = structure_stop；沒有已確認前低時為 null，不使用 ATR fallback
rr_targets 最多一筆 (to_major_resistance)，若 major_resistance 為 null 則空陣列

high_volume_breakdown 的前低另由完整歷史 low 序列計算，無固定日期上限。局部低點須嚴格低於左右各 5 根，右側確認完成後才可使用；取日期最近的一個。收盤嚴格低於該前低且有放量賣壓才成立，通道破底不能繞過前低要求。共用支撐與停損採同一個已確認前低規則；raw/adjusted 各自確認，決策使用 adjusted。

3. 訊號產生邏輯

3.1 Bias 計算 — 純 Rule-Based

compute_bias_confidence_reasons_warnings() in daily_analysis_snapshot.py

起始分數：score = 0.5

14 個評分類別及其精確調整值

類別	條件	分數調整	ReasonCode / WarningCode
A. Trend Structure	ma_alignment == "bullish"	+0.20	MA_BULLISH_STACK
	ma_alignment == "bearish"	-0.20	MA_BEARISH_STACK
	close > ma60	±0	PRICE_ABOVE_MA60
	close <= ma60	±0	PRICE_BELOW_MA60
	MA entangled (\|ma5-ma20\|/close < 0.003)	±0	MA_ENTANGLED
B. Trend Strength	adx >= 25	+0.10	ADX_STRONG_TREND
	adx > adx_yesterday AND adx >= 20	+0.04	ADX_ACCELERATING
	plus_di > minus_di	±0	DI_BULLISH
	plus_di <= minus_di	±0	DI_BEARISH
C. Momentum	vi_signal == "bullish_cross"	+0.08	VORTEX_GOLDEN_CROSS
	vi_signal == "bearish_cross"	-0.08	VORTEX_DEAD_CROSS
	vi_signal == "bullish"	+0.04	(無)
	vi_signal == "bearish"	-0.04	(無)
	vortex diff widening (bullish)	+0.04	VORTEX_DIFF_WIDENING
	macd_hist > 0	+0.08	MACD_HIST_POSITIVE
	macd_hist < 0	-0.08	MACD_HIST_NEGATIVE
	\|hist\| > \|prev_hist\|	+0.03	MACD_HIST_EXPANDING
D. Volume	vol_ratio >= 1.2 + aggressive_buying	+0.08	VOLUME_EXPANSION_BUY
	vol_ratio >= 1.2 + aggressive_selling	-0.08	VOLUME_EXPANSION_SELL
	bullish MA + pullback + low volume	±0	LOW_VOLUME_PULLBACK
E. Reserved	inactive compatibility slot; 20-day breakout is scored by M	±0	(無新 emission)
E2. Breakout Highs	is_new_ath_close	+0.10	BREAKOUT_ALL_TIME_HIGH
	is_new_52w_high_close (elif)	+0.08	BREAKOUT_52W_HIGH
F. Squeeze/CTI	cti_score >= 0.6 (且 < 0.8)	+0.10	(無)
	cti_score <= -0.6 (且 > -0.8)	-0.10	(無)
	\|cti_score\| >= 0.8	±0	⚠️ CTI_EXTREME_LINEARITY
Warnings	rsi >= 70 (且 score > 0.5)	-0.05	⚠️ RSI_OVERBOUGHT
	rsi <= 30	±0	⚠️ RSI_OVERSOLD
	rsi_divergence == "bearish_warning"	-0.08	⚠️ DIVERGENCE_BEARISH
	rsi_divergence == "bullish_warning"	+0.06	⚠️ DIVERGENCE_BULLISH
G. Choppiness	chop_status in (extreme_squeeze, choppy)	score += (0.5-score)*0.30	⚠️ CHOP_CHOPPY
	chop_status == "trending"	+0.05	CHOP_TRENDING
H. WaveTrend	deep_bullish_cross	+0.10	WT_DEEP_BULLISH_CROSS
	deep_bearish_cross	-0.10	WT_DEEP_BEARISH_CROSS
	bullish_cross	+0.06	WT_BULLISH_CROSS
	bearish_cross	-0.06	WT_BEARISH_CROSS
	extreme_overbought	-0.05	⚠️ WT_EXTREME
	overbought	-0.03	⚠️ WT_OVERBOUGHT
	wt_divergence == "bearish_warning"	-0.06	⚠️ WT_DIVERGENCE_BEARISH
	wt_divergence == "bullish_warning"	+0.04	⚠️ WT_DIVERGENCE_BULLISH
I. Aroon	strong_uptrend AND score >= 0.5	+0.04	AROON_FRESH_UPTREND
	strong_downtrend AND score <= 0.5	-0.04	AROON_FRESH_DOWNTREND
	aroon_up < 50 AND adx > 25	±0	⚠️ AROON_STALE_TREND
J. TTM Squeeze	squeeze_off	+0.08	SQUEEZE_RELEASE
	squeeze_duration > 10	±0	⚠️ SQUEEZE_PROLONGED
K. DMI Derived	di_spread_slope == "widening" AND adx > 20	+0.04	DI_SPREAD_WIDENING
	adx_rising_days >= 3	+0.03	ADX_SUSTAINED_RISE
L. Supertrend	direction == "bullish"	+0.03	SUPERTREND_BULLISH
	direction == "bearish"	±0	⚠️ SUPERTREND_BEARISH
	flip == "bullish"	+0.06	SUPERTREND_FLIP_BULLISH
	flip == "bearish"	-0.06	SUPERTREND_FLIP_BEARISH
	\|dist_supertrend\| < 1.0	±0	⚠️ NEAR_SUPERTREND
M. Donchian	breakout_20d == "upward"	+0.05	DONCHIAN_BREAKOUT_20D
	breakout_20d == "downward"	-0.05	(無)
	breakout_55d == "upward"	+0.08	DONCHIAN_BREAKOUT_55D
	breakout_55d == "downward"	-0.08	(無)
N. RS vs Market	rs_slope == "rising" AND rs_20d > 2.0	+0.04	RS_OUTPERFORM
	rs_slope == "falling" AND rs_20d < -2.0	-0.04	RS_UNDERPERFORM + ⚠️ RS_LAGGING

回測的 high_20 是前一個 OHLC 交易日 snapshot 的 adjusted donchian_high_20；前一交易日缺 snapshot 時為 null，不取更早一筆，也不引用 major_resistance。欄位為舊資料與回測設定保留，但 Category E 不再加分或產生 BREAKOUT_ABOVE_RECENT_HIGH；相同的 20 日突破只由 Category M 的 breakout_20d 評分，避免同一證據重複加分。舊設定仍可解析 Category E，但新的 solo、ablation、forward selection、fallback、weight tuning 與 Layer 3 組合會排除這個 inactive 槽位，不能把 no-op 報成最佳化收益。

loader 保留原始 raw OHLC；直接比較支撐、壓力、布林或一目均衡表價格的條件使用 indicator_close。有此欄位但值無效時不觸發；僅舊的直接呼叫完全沒有此欄位時沿用 close。Phase 1 直接保留 loader 的欄位，既有 parquet 須重新產生才能採用新版資料。

最終處理

score = max(0.0, min(1.0, score))  # clamp to [0, 1]

if score >= 0.6  → bias = "bullish"
if score <= 0.4  → bias = "bearish"
else             → bias = "neutral"

confidence = round(score, 4)  # confidence 就是 score 本身

3.2 Confidence 計算

Confidence 就是累積分數本身，沒有獨立的 confidence 公式或 sub-dimension breakdown。

confidence = round(score, 4) 其中 score 是從 0.5 開始經過 14 個類別累加/累減後 clamp 到 [0, 1] 的值。

3.3 Warnings 產生

Warnings 在評分過程中同步產生，由 WarningCode enum 標記。額外的 warnings 包括：

條件	WarningCode
close / major_resistance >= 0.98	NEAR_RESISTANCE
close / major_support <= 1.02	NEAR_SUPPORT
distance_to_ath_pct 在 -2.0% ~ 0 之間	NEAR_ALL_TIME_HIGH
avg_volume_20 < 500,000 OR avg_amount_20 < 10M OR flat_days_20 > 10	LOW_LIQUIDITY

Warning 去重規則：若 NEAR_ALL_TIME_HIGH 存在且 major_resistance 與 prev_ath_price 差距 < 0.5%，則移除 NEAR_RESISTANCE。

3.4 Signal Channel 分類

classify_channel() in api/routes/signals.py -- 嚴格 cascade（first match wins）：

共用品質門檻 `_passes_common_quality()`

bias, confidence, data_quality, macd_histogram, cti_score, volume_ratio 皆非 None
data_quality == "complete"

Channel 1: momentum_bullish

bias == "bullish" AND confidence >= 0.70
rsi_divergence != "bearish_warning"（RSI 背離否決）
trend quality score >= 2/5:
+1 if ma_alignment == "bullish" OR price_position == "above_all"
+1 if cti_score > 0.6
+1 if MACD histogram expanding (hist > 0 AND |hist| > |prev_hist|)
+1 if volume_ratio >= 1.2
+1 if has_high_pattern

Channel 2: momentum_bearish

bias == "bearish" AND confidence >= 0.70
rsi_divergence != "bullish_warning"
trend quality score >= 2/5（方向反轉）

Channel 3: reversal_bullish

confidence >= 0.40 AND price_position != "above_all"
CTI 否決：cti_score NOT < -0.8（過度線性下跌不可反轉）
reversal score >= 3/5:
+1 if rsi <= 35 OR rsi_divergence == "bullish_warning"
+1 if macd_histogram > prev_macd_histogram
+1 if close > ma5 AND ma5 >= ma20
+1 if volume_ratio >= 1.2 AND volume_status == "aggressive_buying"
+1 if has_high_pattern AND has_bottom_pattern

Channel 4: reversal_bearish

confidence >= 0.40 AND price_position != "below_all"
CTI 否決：cti_score NOT > 0.8
reversal score >= 3/5（方向反轉）

Channel 5: squeeze_bullish

bb_status in ("contracting", "squeezing") AND adx < 20
ma_alignment in ("bullish", "mixed") AND price_position != "below_all"
若 close > bb_upper 且 bandwidth_ratio >= 1.3 → "strong" breakout → 踢出 squeeze

Channel 6: squeeze_bearish

同 squeeze_bullish 但方向反轉
若 close < bb_lower 且 bandwidth_ratio >= 1.3 → 踢出 squeeze

Channel 7: neutral

以上皆不符合時的預設值

3.5 Entry/Exit 訊號

系統目前沒有明確的 entry/exit 訊號。只有方向性的 bias + confidence + signal channel。

沒有以下機制： - 明確的 entry_price / exit_price - Stop-loss / take-profit 具體建議（risk block 提供 invalidation_level 和 rr_targets 但非交易指令） - 持倉期間建議

4. 資料來源與覆蓋範圍

4.1 股票池

重要更正 (2026-02-25 DB 實查)：系統實際爬取的是台灣全市場（TWSE + TPEx），遠超 markets.yaml 定義的 20 支精選股。

4.1.1 設定層 vs 實際 DB

層級	來源	台股數量	美股數量	說明
`markets.yaml` 設定	手動清單	20 (.TW)	7	原始精選清單
`init.sql` 種子	SQL 初始化	25 (.TW)	7	含額外 2379, 2327, 2301, 2395, 2207
`taiwan_stocks.py` 同步	TWSE/TPEx API	全市場	—	動態同步上市櫃全股票
`stocks` 表實際	DB	2,019 (.TW) + 897 (.TWO)	7	共 2,923 支

4.1.2 各表實際資料量（DB 查詢 2026-02-25）

表	台股 .TW	櫃買 .TWO	美股	合計 tickers	合計 rows
`stocks` (master)	2,019	897	7	2,923	2,923
`stock_ohlc`	1,229	897	7	2,133	2,397,970
`daily_technical_snapshot`	1,228	897	0	2,125	382,385
`technical_indicators`	6	0	7	13	3,900
`daily_chip_snapshot`	1,228	897	0	2,125	769,250
`daily_financial_snapshot`	1,228	897	0	2,125	216,649

4.1.3 關鍵發現

美股無 snapshot：7 支美股有 OHLC 資料但 daily_technical_snapshot 為 0 筆。回測僅能對台股進行
technical_indicators 表僅 13 支：此表僅覆蓋 6 支台股 (2317, 2330, 2412, 2454, 2881, 2882) + 7 支美股。絕大多數股票的指標由 daily_technical_snapshot 在 snapshot 流程中直接從 OHLC 計算，不經過此表
台股包含 ETF：1,228 支 .TW 中約 136 支為 ETF（代號 00xxx），1,950 支為一般股票，20 支為其他（存託憑證等），19 支為 TWO 歸類問題
TWO (櫃買) 已覆蓋：897 支櫃買市場股票已完整爬取

4.1.4 `markets.yaml` 精選 20 支台股

以下為系統原始設定的核心觀察清單（仍用於 RS benchmark）：

Ticker	名稱	Sector
2330.TW	台積電	半導體
2317.TW	鴻海	電子代工
2454.TW	聯發科	IC 設計
2412.TW	中華電	電信
2308.TW	台達電	電子零組件
2881.TW	富邦金	金融
2882.TW	國泰金	金融
2891.TW	中信金	金融
2303.TW	聯電	半導體
3711.TW	日月光投控	封測
1301.TW	台塑	塑化
1303.TW	南亞	塑化
2002.TW	中鋼	鋼鐵
2886.TW	兆豐金	金融
2884.TW	玉山金	金融
2892.TW	第一金	金融
5880.TW	合庫金	金融
3008.TW	大立光	光學
2357.TW	華碩	品牌 PC
2382.TW	廣達	伺服器/AI

4.1.5 美股 7 支 (Magnificent 7)

AAPL, MSFT, GOOGL, AMZN, NVDA, META, TSLA（僅有 OHLC，無 snapshot/chip/financial）

4.2 資料來源

項目	值
OHLCV 來源	`taiwan_stocks.py` 同步全市場（FinMind 為主要來源，yfinance 備援）
台股清單同步	`taiwan_stocks.py` 動態從 TWSE/TPEx API 拉取上市櫃完整名單
頻率	日線 (1d)
更新排程	cron `0 6 * * 1-5` (Asia/Taipei 工作日 06:00)
Rate Limit	60 requests/minute
重試	max_attempts=3, exponential backoff factor=2

4.3 資料欄位與日期範圍

OHLCV 表 (stock_ohlc)：ticker, timestamp, open, high, low, close, volume, adjusted_close

資料集	最早日期	最新日期	說明
`stock_ohlc` (TW/TWO)	2021-02-04	2026-02-24	⚠️ 非 2020-01-01，`markets.yaml` 的 `start_date: 2020-01-01` 為設定值，實際 DB 從 2021-02 起
`stock_ohlc` (US)	2021-02-04	2026-02-24	~1,269 交易日
`daily_technical_snapshot`	2023-11-08	2026-02-24	549 個交易日
snapshot 多數台股起始	2025-08-14	2026-02-24	1,738 支台股僅 ~126 天 snapshot
snapshot 早期批次	2024-02-02	2026-02-24	~316 支有 492 天（~2 年）
`daily_chip_snapshot`	2024-08-27	2026-02-24
`daily_financial_snapshot`	2025-09-18	2026-02-24	僅 ~5 個月，非常新
`technical_indicators`	2024-11-11	2026-01-30	僅 13 支股票

⚠️ 回測可用資料深度警告

問題	影響
OHLC 起始為 2021-02-04 而非 2020-01-01	無法回測 COVID crash (2020-01~03)，BACKTEST_SPEC 的 covid_crash 子區間不可用
多數 snapshot 僅從 2025-08-14 起	若回測依賴 snapshot 欄位（bias, confidence, channel），多數股票僅有 ~6 個月資料
316 支有 ~2 年 snapshot	這批是回測可用的核心股票池
Financial snapshot 僅 ~5 個月	Strategy E（三維度投票）的資料極淺
美股無 snapshot	Strategy A~F 全部無法對美股執行

4.4 前日資料處理

前日指標值：從 indicators_df.iloc[-2] 取得（不是從 DB 載入前日 snapshot）
前日收盤：通過 load_prev_close() SQL 查詢
cti_delta：placeholder，目前未實作
無 prev_day_snapshot 概念：不存在載入前一日完整 snapshot 的邏輯

4.5 分析維度覆蓋狀態

維度	狀態	DB 覆蓋	說明
Technical	ready	2,125 tickers / 382K rows	完整實作，180+ 欄位
Chip (籌碼)	ready	2,125 tickers / 769K rows	Engine A (法人流向 0.65) + Engine B (融資壓力 0.35)
Financial (財務)	ready	2,125 tickers / 217K rows	Engine A (營收) + B (獲利) + C (評價)，但僅 5 個月資料
Fundamental (基本面)	not_ready	0	placeholder，所有欄位標記 status="placeholder"

5. 目前缺少什麼（Backtesting Gaps）

5.1 無可量化的 Entry/Exit 規則

系統只產生 bias (bullish/bearish/neutral) + confidence (0~1)
沒有明確定義 "何時進場" 和 "何時出場" 的門檻
Signal channel 提供分類但非交易觸發條件
需要：定義 confidence 門檻（如 >= 0.7 做多）、持倉退出條件（如 bias 翻轉或 confidence 低於門檻）

5.2 Confidence 無法分解

Confidence 是一個累積分數，沒有 sub-dimension breakdown
無法分析 "哪個指標類別貢獻了多少信心"
需要：記錄各類別的分數貢獻（如 trend_score, momentum_score, volume_score）

5.3 無 Benchmark / 基準收益

RS vs Market 用 0050.TW 但僅用於相對強弱判斷
沒有 buy-and-hold 基準用於比較策略表現
需要：記錄 benchmark 每日收益供 alpha 計算

5.4 無持倉追蹤

系統為每日獨立 snapshot，沒有 "目前持有" 的概念
無法計算連續持倉的盈虧
需要：持倉狀態機 (idle → long → exit) 和對應的 PnL 追蹤

5.5 Signal Scoring Weights 與 Snapshot Bias 斷裂

signal_scoring.py 定義了精細的權重系統 (momentum, reversal, squeeze weights)
但 daily_analysis_snapshot.py 的 compute_bias_confidence_reasons_warnings() 使用獨立的 hardcoded 分數調整
兩套邏輯各自獨立，signal channel 分類與 bias 計算用不同的權重
需要：統一或明確區分兩套系統的角色

5.6 Pattern 可靠性未整合

patterns.detected 有 score, reliability 欄位
但 patterns 未參與 bias/confidence 計算
需要：決定 pattern 是否應影響信號，以及如何影響

5.7 多維度整合缺失

Technical, Chip, Financial 三個維度各自產生 bias/confidence
沒有跨維度的綜合信號（如：技術面看多 + 籌碼面看多 → 強化信號）
需要：定義多維度融合公式或投票機制

5.8 歷史回測基礎設施

已有部分基礎（signal_backtest.py）： - Horizons: 1, 3, 5, 20 trading days - Deadzone: ±0.2% - Cooldown: 5 days per ticker - Squeeze volatility threshold: 5%

但缺少： - 完整的回測引擎（entry/exit 模擬） - 滑點和手續費模型 - 資金管理規則 - 績效指標計算 (Sharpe, max drawdown, win rate)

附錄：完整 Enum 值

SignalChannel (7)

momentum_bullish, momentum_bearish, reversal_bullish, reversal_bearish, squeeze_bullish, squeeze_bearish, neutral

Bias (3)

bullish, bearish, neutral

RiskFlag (11)

rsi_overbought, rsi_oversold, macd_divergence, rsi_divergence_bearish, rsi_divergence_bullish, bb_expanding, breakout_watch, volume_divergence, margin_buildup_risk, distribution_signs, foreign_distribution

QualityFlag (3)

complete, partial, low_liquidity

ReasonCode (48)

MA_BULLISH_STACK, MA_BEARISH_STACK, PRICE_ABOVE_MA60, PRICE_BELOW_MA60, MA_ENTANGLED, ADX_STRONG_TREND, ADX_ACCELERATING, DI_BULLISH, DI_BEARISH, VORTEX_GOLDEN_CROSS, VORTEX_DEAD_CROSS, VORTEX_DIFF_WIDENING, MACD_HIST_POSITIVE, MACD_HIST_NEGATIVE, MACD_HIST_EXPANDING, VOLUME_EXPANSION_BUY, VOLUME_EXPANSION_SELL, LOW_VOLUME_PULLBACK, PULLBACK_HOLD_MA5, BREAKOUT_ABOVE_RECENT_HIGH, SQUEEZE_ACTIVE, SQUEEZE_NARROW_BB, FOREIGN_NET_BUY, FOREIGN_NET_SELL, INST_CONSENSUS_BUY, INST_CONSENSUS_SELL, INSTITUTIONAL_STREAK_BUY, INSTITUTIONAL_STREAK_SELL, FLOW_WITH_VOLUME, SHORT_COVERING, SHORT_BUILDUP, MARGIN_CLEARING, BREAKOUT_52W_HIGH, BREAKOUT_ALL_TIME_HIGH, CHOP_TRENDING, WT_BULLISH_CROSS, WT_BEARISH_CROSS, WT_DEEP_BULLISH_CROSS, WT_DEEP_BEARISH_CROSS, AROON_FRESH_UPTREND, AROON_FRESH_DOWNTREND, SQUEEZE_RELEASE, DI_SPREAD_WIDENING, ADX_SUSTAINED_RISE, SUPERTREND_BULLISH, SUPERTREND_FLIP_BULLISH, SUPERTREND_FLIP_BEARISH, DONCHIAN_BREAKOUT_20D, DONCHIAN_BREAKOUT_55D, RS_OUTPERFORM, RS_UNDERPERFORM

WarningCode (27)

RSI_OVERBOUGHT, RSI_OVERSOLD, DIVERGENCE_BEARISH, DIVERGENCE_BULLISH, CTI_EXTREME_LINEARITY, CLIMAX_UNCONFIRMED, NEAR_RESISTANCE, NEAR_SUPPORT, VOLUME_DIVERGENCE, DISTRIBUTION_VOLUME, BREAKOUT_WATCH, BREAKOUT_CONFIRMED, DATA_PARTIAL, LOW_LIQUIDITY, NEAR_ALL_TIME_HIGH, CHOP_CHOPPY, WT_OVERBOUGHT, WT_OVERSOLD, WT_EXTREME, WT_DIVERGENCE_BEARISH, WT_DIVERGENCE_BULLISH, AROON_STALE_TREND, SQUEEZE_PROLONGED, SUPERTREND_BEARISH, NEAR_SUPERTREND, RS_LAGGING, MARGIN_INCREASING_FAST, FOREIGN_SELLING_STREAK, DATA_PARTIAL_CHIP, LOW_FLOAT_VOLUME, USING_STALE_MARGIN_DATA

MarketEvent (11)

TOUCHED_ALL_TIME_HIGH, NEW_ALL_TIME_HIGH, ATH_INTRADAY_ONLY, TOUCHED_52W_HIGH, NEW_52W_HIGH, HIGH_52W_INTRADAY_ONLY, LIMIT_UP_CLOSE, LIMIT_DOWN_CLOSE, GAP_UP, GAP_DOWN, NEAR_ALL_TIME_HIGH

ChipSignalType (9)

neutral, institutional_accumulation, short_squeeze, foreign_entry, cost_defense, institutional_distribution, retail_trap, margin_call_risk, fake_breakout