def traditional_signal(row, prev):
    # 买入:RSI 超卖 + KDJ 金叉 + MACD 绿柱缩短(准备翻红)
    buy = (row['RSI'] < 30 and
           row['K'] > row['D'] and prev['K'] <= prev['D'] and
           row['MACD'] > 0 and row['MACD'] < prev['MACD'])
    # 卖出:RSI 超买 + KDJ 死叉 + MACD 红柱缩短(准备翻绿)
    sell = (row['RSI'] > 70 and
            row['K'] < row['D'] and prev['K'] >= prev['D'] and
            row['MACD'] < 0 and row['MACD'] > prev['MACD'])
    return '买' if buy else ('卖' if sell else '持')

import numpy as np
import pandas as pd
import yfinance as yf
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import classification_report, accuracy_score

# 1. 拉取数据
df = yf.download('AAPL', start='2020-01-01', end='2026-06-27', progress=False)
df = df[['Open', 'High', 'Low', 'Close', 'Volume']].dropna()

# 2. 手工实现三大指标
def compute_rsi(close, period=14):
    delta = close.diff()
    gain = delta.clip(lower=0)
    loss = -delta.clip(upper=0)
    ag = gain.ewm(alpha=1/period, adjust=False).mean()
    al = loss.ewm(alpha=1/period, adjust=False).mean()
    rs = ag / (al + 1e-10)
    return 100 - 100 / (1 + rs)

def compute_kdj(high, low, close, n=9):
    low_n = low.rolling(n).min()
    high_n = high.rolling(n).max()
    rsv = (close - low_n) / (high_n - low_n + 1e-10) * 100
    k = rsv.ewm(alpha=1/3, adjust=False).mean()
    d = k.ewm(alpha=1/3, adjust=False).mean()
    j = 3 * k - 2 * d
    return k, d, j

def compute_macd(close, fast=12, slow=26, signal=9):
    ema_f = close.ewm(span=fast, adjust=False).mean()
    ema_s = close.ewm(span=slow, adjust=False).mean()
    dif = ema_f - ema_s
    dea = dif.ewm(span=signal, adjust=False).mean()
    macd = (dif - dea) * 2
    return dif, dea, macd

df['RSI']  = compute_rsi(df['Close'])
df['K'], df['D'], df['J'] = compute_kdj(df['High'], df['Low'], df['Close'])
df['DIF'], df['DEA'], df['MACD'] = compute_macd(df['Close'])

# 3. 特征工程
df['Return']        = df['Close'].pct_change()
df['VolChange']     = df['Volume'].pct_change()
df['RSI_Diff']      = df['RSI'].diff()
df['KDJ_Cross']     = ((df['K'] > df['D']).astype(int)
                       - (df['K'].shift(1) > df['D'].shift(1)).astype(int))
df['MACD_Bar_Chg']  = df['MACD'].diff()
df['DIF_Cross_DEA'] = ((df['DIF'] > df['DEA']).astype(int)
                       - (df['DIF'].shift(1) > df['DEA'].shift(1)).astype(int))

# 4. 标签:未来 5 日收益
fut_ret = df['Close'].pct_change(5).shift(-5)
df['Label'] = pd.cut(fut_ret, bins=[-np.inf, -0.01, 0.01, np.inf],
                     labels=[0, 1, 2]).astype(int)  # 0=卖 1=持 2=买

features = ['RSI', 'K', 'D', 'J', 'DIF', 'DEA', 'MACD',
            'Return', 'VolChange', 'RSI_Diff', 'KDJ_Cross',
            'MACD_Bar_Chg', 'DIF_Cross_DEA']
data = df[features + ['Label']].dropna()
X, y = data[features], data['Label']

# 5. 切分 + 训练(时间序列切分,严禁 shuffle)
split = int(len(X) * 0.7)
X_train, X_test = X.iloc[:split], X.iloc[split:]
y_train, y_test = y.iloc[:split], y.iloc[split:]

model = GradientBoostingClassifier(
    n_estimators=200, max_depth=4,
    learning_rate=0.05, subsample=0.8, random_state=42)
model.fit(X_train, y_train)

# 6. 评估
y_pred = model.predict(X_test)
print(f"准确率: {accuracy_score(y_test, y_pred):.2%}")
print(classification_report(y_test, y_pred, target_names=['卖', '持', '买']))

# 7. 特征重要性
imp = pd.Series(model.feature_importances_, index=features)
print(imp.sort_values(ascending=False).head(5))