三大法人買賣超 API：用外資、投信、自營商資料做台股籌碼分析

常見資料單位

Long format response

{
  "data": [
    {
      "date": "2026-04-24",
      "symbol": "2330",
      "market": "twse",
      "investor_type": "foreign",
      "buy": 12500000,
      "sell": 9800000,
      "net_buy": 2700000,
      "unit": "shares",
      "currency": "TWD"
    },
    {
      "date": "2026-04-24",
      "symbol": "2330",
      "market": "twse",
      "investor_type": "investment_trust",
      "buy": 860000,
      "sell": 420000,
      "net_buy": 440000,
      "unit": "shares",
      "currency": "TWD"
    },
    {
      "date": "2026-04-24",
      "symbol": "2330",
      "market": "twse",
      "investor_type": "dealer",
      "buy": 520000,
      "sell": 610000,
      "net_buy": -90000,
      "unit": "shares",
      "currency": "TWD"
    }
  ],
  "meta": {
    "source": "tw-market-data",
    "from": "2026-04-24",
    "to": "2026-04-24",
    "unit": "shares",
    "timezone": "Asia/Taipei"
  }
}

Wide format response

如果使用者主要是做 pandas 分析，也可以提供 wide format。這種格式適合直接和 OHLCV 合併。

{
  "data": [
    {
      "date": "2026-04-24",
      "symbol": "2330",
      "market": "twse",
      "foreign_net_buy": 2700000,
      "investment_trust_net_buy": 440000,
      "dealer_net_buy": -90000,
      "total_institutional_net_buy": 3050000,
      "unit": "shares"
    }
  ],
  "meta": {
    "format": "wide",
    "unit": "shares"
  }
}

Long format 比較彈性，wide format 比較方便。資料產品可以兩者都支援，或在 API 參數中讓使用者選擇。

Factor schema 示意

{
  "date": "2026-04-24",
  "symbol": "2330",
  "factor_name": "foreign_net_buy_volume_ratio_5d",
  "value": 0.036,
  "window": 5,
  "data_used": [
    "institutional_flows",
    "daily_ohlcv"
  ],
  "unit": "ratio"
}

如果你要把籌碼因子放進量化策略，可以參考 台股量化交易入門。

若要接進回測系統，可以參考 Python 台股回測系統實作。

和成交量對齊

如果要計算買賣超佔成交量比重，必須確認兩邊單位一致。例如 institutional flow 是股數，成交量也應該是股數，而不是張數或金額。

和股價對齊

如果要分析法人買超後的價格行為，必須小心避免 look-ahead bias。當日買賣超資料若在盤後才完整取得，回測時不能用它決定同日收盤前的交易。

和財報資料對齊

法人籌碼可以和基本面因子搭配，例如觀察投信是否持續買超基本面改善的股票，或外資是否在財報公告後調整持股方向。

如果你要理解財報與基本面資料，可以參考 台股財報 API 教學。

建立籌碼特徵

flows["foreign_net_buy_5d"] = flows["foreign_net_buy"].rolling(5).sum()
flows["investment_trust_net_buy_5d"] = flows["investment_trust_net_buy"].rolling(5).sum()
flows["dealer_net_buy_5d"] = flows["dealer_net_buy"].rolling(5).sum()

flows["total_institutional_net_buy_5d"] = (
    flows["foreign_net_buy_5d"]
    + flows["investment_trust_net_buy_5d"]
    + flows["dealer_net_buy_5d"]
)

print(flows[[
    "date",
    "foreign_net_buy_5d",
    "investment_trust_net_buy_5d",
    "dealer_net_buy_5d",
    "total_institutional_net_buy_5d"
]].tail())

和 OHLCV 合併

ohlcv = pd.DataFrame([
    {
        "date": "2025-01-02",
        "symbol": "2330",
        "close": 1005,
        "volume": 25000000
    }
])

ohlcv["date"] = pd.to_datetime(ohlcv["date"])

data = ohlcv.merge(flows, on=["date", "symbol"], how="left")

data["foreign_net_buy_volume_ratio"] = (
    data["foreign_net_buy"] / data["volume"]
)

print(data.head())

這裡的重點是資料對齊，而不是策略本身。正式回測還需要考慮資料可得時間、交易成本、滑價與風控。

Agent tool schema 示意

{
  "name": "get_institutional_flows",
  "description": "Get institutional investor buy/sell flows for a Taiwan stock.",
  "parameters": {
    "symbol": "2330",
    "from": "2025-01-01",
    "to": "2025-12-31",
    "unit": "shares",
    "format": "wide"
  }
}

Agent response schema 示意

{
  "symbol": "2330",
  "data_used": [
    "institutional_flows",
    "daily_ohlcv"
  ],
  "summary": "Foreign investors were net buyers over the selected window, while dealer flows were mixed.",
  "flow_features": {
    "foreign_net_buy_5d": 12000000,
    "investment_trust_net_buy_5d": 1800000,
    "dealer_net_buy_5d": -600000
  },
  "risk_flags": [
    "requires_volume_normalization",
    "not_a_standalone_trading_signal",
    "requires_backtesting"
  ],
  "not_investment_advice": true
}