TA-Lib是一个技术分析库，涵盖了150多种股票、期货交易软件中常用的技术分析指标，如MACD、RSI、KDJ、动量指标、布林带等等，而
pandas-ta则是一个基于pandas和ta-lib的高级技术分析工具，具有130多个指标和实用功能以及60多个TA-Lib中包含的蜡烛模式。本章节记录如何利用pandas-ta快速提取技术面特征。
安装
安装pandas-ta本身非常简单，直接pip一下就可以，如果想用ta-lib的一些特性，则还需要安装ta-lib本身
- Anaconda：使用Anaconda集成环境，可直接
conda install -c conda-forge ta-lib - Windows: 直接在www.lfd.uci.edu/~gohlke/pyt… 中下载离线包安装即可
- Linux: 需要编译安装C++原始环境库(prdownloads.sourceforge.net/ta-lib/ta-l…) ，然后pip安装
pip install TA-Lib
pandas-ta提供的主要函数有：
- 蜡烛图形态：Candles ，基于K线图的形态识别，如三只乌鸦等
- 周期特征：Cycles，如正弦波拟合等
- 动量特征：Momentum，如KDJ、RSI等
- 覆盖特征：Overlap，主要包含各种移动平均线系列，如EMA指数平滑
- 回报特征：Performance ，如百分比回报、log回报等
- 统计特征：Statistics，如熵、中位数、分位数、标准差等
- 趋势特征：Trend， 如阿隆指数等
- 波动率特征：Volatility，如布林带等
- 成交量特征：Volume，如资金流指数等
- 其他特征：如神奇九转(td_seq)等
准备K线数据
这里主要提取1min和1d两种粒度数据中的特征，日内1分钟数据未除权，日间1天数据使用后复权，保持序列的连贯性。
读取日内K线数据(未复权)
这里以聚宽数据为例，读取指定时间区间中1min数据，并生成对应的日内收益标签，详细功能见代码中的注释。
主要的收益计算为日内指标，分为有两类：
- 日内剩余时间段里的平均收益百分比
- 日内随后5min、10min的平均收益
# 日内：聚宽
def load_jqdata_kline(code='510050.SH', start_date='2021-11-01', end_date='2022-01-01'):
# 代码转换
if not code.endswith(('.XSHG', '.XSHE')):
ticker, exchange = code.split('.')
code = ticker + '.' + {'SZ': 'XSHE', 'BJ': 'XSHE', 'SH': 'XSHG'}.get(exchange.upper())
# 提取stock_kline中的1min数据
query = "select * from jqdata.stock_kline where code='{}' and volume !=0.0 and trade_day >='{}' and trade_day <'{}' and toDayOfWeek(trade_day) < 6".format(code, start_date, end_date)
data = client.execute(query)
# 提取stock_dayline中的1d数据,主要是复权因子
day_query = "select trade_day, pre_close, factor from jqdata.stock_dayline where code='{}' and volume !=0.0 and trade_day >='{}' and trade_day <'{}' and toDayOfWeek(trade_day) < 6".format(code, start_date, end_date)
day_data = client.execute(day_query)
# 获取复权后的昨收数据
pre_close = {i[0]: np.round(i[1] / i[2] , 3) for i in day_data}
df = pd.DataFrame(data, columns=['code', 'dtype', 'trade_day', 'trade_time', 'open', 'close', 'low', 'high', 'volume', 'money', 'avg'])
df = df.sort_values(by='trade_time').drop_duplicates(subset=['trade_time'], keep='first').reset_index(drop=True)
# 补充昨收数据
df['pre_close'] = df['trade_day'].apply(lambda x: pre_close.get(x, 0))
df['trade_day'] = df['trade_day'].astype(str)
# 当天收盘相对于昨收的百分比
df['daily_percent'] = (df['close'] - df['pre_close']) / df['pre_close']
df['minutes_of_day'] = df.trade_time.dt.hour * 60 + df.trade_time.dt.minute
# 过滤尾盘集合竞价数据
df = df[~df.minutes_of_day.isin([898, 899])]
# 此刻买进，日内剩余时间段内的收益25%分位数
df['label_d0_25'] = ((df.groupby(['trade_day'])['avg'].transform(lambda row: row[::-1].shift(1).expanding(min_periods=1).quantile(0.25)[::-1]) - df['avg']) * 100).div(df['avg'])
# 此刻买进，日内剩余时间段内的收益50%分位数
df['label_d0_50'] = ((df.groupby(['trade_day'])['avg'].transform(lambda row: row[::-1].shift(1).expanding(min_periods=1).quantile(0.5)[::-1]) - df['avg']) * 100).div(df['avg'])
# 此刻买进，日内剩余时间段内的收益75%分位数
df['label_d0_75'] = ((df.groupby(['trade_day'])['avg'].transform(lambda row: row[::-1].shift(1).expanding(min_periods=1).quantile(0.75)[::-1]) - df['avg']) * 100).div(df['avg'])
# 此刻买进，日内5分钟时间段内的收益50%分位具体值
df['label_rolling_5_50_bps'] = (df.groupby(['trade_day'])['avg'].transform(lambda row: row[::-1].shift(1).rolling(5, min_periods=1).quantile(0.5)[::-1]) - df['avg']) * 100
# 此刻买进，日内5分钟时间段内的收益50%分位数
df['label_rolling_5_50'] = df['label_rolling_5_50_bps'].div(df['avg'])
# 此刻买进，日内10分钟时间段内的收益50%分位具体值
df['label_rolling_10_50_bps'] = df.groupby(['trade_day'])['avg'].transform(lambda row: row[::-1].shift(1).rolling(10, min_periods=1).quantile(0.5)[::-1]) - df['avg']
# 此刻买进，日内10分钟时间段内的收益50%分位数
df['label_rolling_10_50'] = df['label_rolling_10_50_bps'].div(df['avg'])
return df.reset_index(drop=True)
读取日间K线数据(后复权)
日线有两个，一个是聚宽数据，主要针对国内市场，还有一个是华盛通数据，针对香港和美国市场，都从clickhouse数据库中读取。
# 日线: 华盛通
def load_hstong_dayline(market='cn', code='300369.SZ', start_date='2018-11-01', end_date='2022-01-01'):
query = "select code, data_date, price_open, price_close, price_low, price_high, price_last, volume, deal from hstong.dayline_hfq_{} where code='{}' and volume !=0.0 and data_date >='{}' and data_date <='{}' and toDayOfWeek(data_date) < 6".format(market, code, start_date, end_date)
data = client.execute(query)
df = pd.DataFrame(data, columns=['code', 'trade_day', 'open', 'close', 'low', 'high', 'pre_close', 'volume', 'money'])
df = df.sort_values(by='trade_day').drop_duplicates(subset=['trade_day'], keep='first').reset_index(drop=True)
return df
# 日线：聚宽
def load_jqdata_dayline(code='510050.SH', start_date='2018-11-01', end_date='2022-01-01'):
if not code.endswith(('.XSHG', '.XSHE')):
ticker, exchange = code.split('.')
code = ticker + '.' + {'SZ': 'XSHE', 'BJ': 'XSHE', 'SH': 'XSHG'}.get(exchange.upper())
query = "select code, trade_day, open, close, low, high, pre_close, volume, money, factor from jqdata.stock_dayline where code='{}' and volume !=0.0 and trade_day >='{}' and trade_day <'{}' and toDayOfWeek(trade_day) < 6".format(code, start_date, end_date)
data = client.execute(query)
df = pd.DataFrame(data, columns=['code', 'trade_day', 'open', 'close', 'low', 'high', 'pre_close', 'volume', 'money', 'factor'])
df = df.sort_values(by='trade_day').drop_duplicates(subset=['trade_day'], keep='first').reset_index(drop=True)
df['trade_day'] = df['trade_day'].astype(str)
return df
我们以茅台(600519.SH)为例,读取'2020-01-01'至'2022-07-07'间的数据：
kline_df = load_jqdata_kline(code='600519.SH', start_date='2020-01-01', end_date='2022-07-07')
print(kline_df.iloc[-5:])
| code | dtype | trade_day | trade_time | open | close | low | high | volume | money | avg | pre_close | daily_percent | minutes_of_day | label_d0_25 | label_d0_50 | label_d0_75 | label_rolling_5_50_bps | label_rolling_5_50 | label_rolling_10_50_bps | label_rolling_10_50 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 600519.XSHG | STOCK | 2022-07-07 | 2022-07-07 14:54:00+08:00 | 1994.86 | 1993.66 | 1993.66 | 1994.86 | 10500 | 2.09366e+07 | 1993.99 | 2002 | -0.00416583 | 894 | -0.0850054 | -0.0308427 | -0.0132899 | -61.5 | -0.0308427 | -0.615 | -0.000308427 |
| 600519.XSHG | STOCK | 2022-07-07 | 2022-07-07 14:55:00+08:00 | 1993.58 | 1994.04 | 1993.57 | 1994.5 | 8900 | 1.77452e+07 | 1993.83 | 2002 | -0.00397602 | 895 | -0.115356 | -0.0386191 | -0.0228204 | -77 | -0.0386191 | -0.77 | -0.000386191 |
| 600519.XSHG | STOCK | 2022-07-07 | 2022-07-07 14:56:00+08:00 | 1993.99 | 1993.68 | 1993.51 | 1993.99 | 18700 | 3.72816e+07 | 1993.69 | 2002 | -0.00415584 | 896 | -0.146713 | -0.108342 | -0.0699708 | -216 | -0.108342 | -2.16 | -0.00108342 |
| 600519.XSHG | STOCK | 2022-07-07 | 2022-07-07 14:57:00+08:00 | 1993.29 | 1993 | 1993 | 1993.29 | 15900 | 3.16902e+07 | 1993.06 | 2002 | -0.0044955 | 897 | -0.153533 | -0.153533 | -0.153533 | -306 | -0.153533 | -3.06 | -0.00153533 |
| 600519.XSHG | STOCK | 2022-07-07 | 2022-07-07 15:00:00+08:00 | 1993.01 | 1990 | 1990 | 1993.01 | 47800 | 9.51379e+07 | 1990 | 2002 | -0.00599401 | 900 | nan | nan | nan | nan | nan | nan | nan |
构建日内特征(分钟级)
读取K线数据后，便进行最重要的特征提取工作，同时生成对应的标签
def make_tech_feature_intraday_jqdata(kline_df, index_col='trade_time'):
df_grouped = kline_df.set_index(index_col).groupby('trade_day', as_index=False)
rsi_5 = (df_grouped.apply(lambda row: row.ta.rsi(length=5) / 100)).fillna(0.5).to_frame().reset_index(drop=True)
cmo_3 = (df_grouped.apply(lambda row: row.ta.cmo(length=3) / 100)).fillna(0.0).to_frame().reset_index(drop=True)
natr_3 = (df_grouped.apply(lambda row: row.ta.natr(length=3))).fillna(0.0).to_frame().reset_index(drop=True)
rvi_5 = (df_grouped.apply(lambda row: row.ta.rvi(length=5) / 100)).fillna(0.5).to_frame().reset_index(drop=True)
try:
kdj_9_3 = (df_grouped.apply(lambda row: row.ta.kdj(9, 3) / 100)).fillna(0.5).reset_index(drop=True)
except:
kdj_9_3 = (df_grouped.apply(lambda row: row.ta.kdj(min(9, len(row)), 3).T.reset_index(drop=True).T / 100)).fillna(0.5).reset_index(drop=True)
kdj_9_3.columns = ['K_9_3', 'D_9_3', 'J_9_3']
adx_5 = (df_grouped.apply(lambda row: row.ta.adx(length=5) / 100)).clip(-10, 10).fillna(0.0).reset_index(drop=True)
try:
mfi_5 = (df_grouped.apply(lambda row: row.ta.mfi(length=5) / 100)).clip(-10, 10).fillna(0.5).to_frame().reset_index(drop=True)
except:
mfi_5 = (df_grouped.apply(lambda row: row.ta.mfi(length=min(5, len(row))).rename('MFI_5') / 100)).clip(-10, 10).fillna(0.5).to_frame().reset_index(drop=True)
pvi_5 = (df_grouped.apply(lambda row: row.ta.pvi(length=5) / 1000)).clip(-10, 10).fillna(1).to_frame().reset_index(drop=True)
nvi_5 = (df_grouped.apply(lambda row: row.ta.nvi(length=5) / 1000)).clip(-10, 10).fillna(1).to_frame().reset_index(drop=True)
willr_5 = (df_grouped.apply(lambda row: row.ta.willr(length=5) / 100)).clip(-1, 1).fillna(-0.5).to_frame().reset_index(drop=True)
try:
willr_10 = (df_grouped.apply(lambda row: row.ta.willr(length=10) / 100)).clip(-1, 1).fillna(-0.5).to_frame().reset_index(drop=True)
except:
willr_10 = (df_grouped.apply(lambda row: row.ta.willr(length=min(10, len(row))).rename('WILLR_10') / 100)).clip(-1, 1).fillna(-0.5).to_frame().reset_index(drop=True)
try:
cmf_10 = (df_grouped.apply(lambda row: row.ta.cmf(length=10) )).clip(-10, 10).fillna(0.0).to_frame().reset_index(drop=True)
except:
cmf_10 = (df_grouped.apply(lambda row: row.ta.cmf(length=min(10, len(row))).rename('CMF_10') )).clip(-10, 10).fillna(0.0).to_frame().reset_index(drop=True)
dpo_5 = (df_grouped.apply(lambda row: row.ta.dpo(length=5, lookahead=False) * 10 )).clip(-10, 10).fillna(0.0).to_frame().reset_index(drop=True)
log_return_5 = (df_grouped.apply(lambda row: row.ta.log_return(length=5) * 50 )).clip(-10, 10).fillna(0.0).to_frame().reset_index(drop=True)
zscore_5 = (df_grouped.apply(lambda row: row.ta.zscore(length=5) )).clip(-10, 10).fillna(0.0).to_frame().reset_index(drop=True)
pct_change_3 = (df_grouped.apply(lambda row: row['avg'].pct_change(periods=3).fillna(0) * 50 / 3)).clip(-10, 10).to_frame().rename(columns={'avg': 'pct_change_3'}).reset_index(drop=True)
pct_change_6 = (df_grouped.apply(lambda row: row['avg'].pct_change(periods=6).fillna(0) * 50 / 6)).clip(-10, 10).to_frame().rename(columns={'avg': 'pct_change_6'}).reset_index(drop=True)
rolling_std_5 = (df_grouped.apply(lambda row: row['avg'].rolling(5, min_periods=1).std(ddof=0))).clip(-10, 10).to_frame().rename(columns={'avg': 'rolling_std_5'}).reset_index(drop=True)
rolling_money_5 = (df_grouped.apply(lambda row: np.log1p(row['money'].rolling(5, min_periods=1).mean()) - np.log1p(row['money']) ) / 10.0).clip(-10, 10).to_frame().rename(columns={'money': 'rolling_money_5'}).reset_index(drop=True)
rolling_money_6_3 = (df_grouped.apply(lambda row: np.log1p(row['money'].rolling(6, min_periods=1).mean()) - np.log1p(row['money'].rolling(3, min_periods=1).mean()) ) / 10.0 ).clip(-10, 10).to_frame().rename(columns={'money': 'rolling_money_6_3'}).reset_index(drop=True)
rolling_volume_5 = (df_grouped.apply(lambda row: np.log1p(row['volume'].rolling(5, min_periods=1).mean()) - np.log1p(row['volume']) )/ 10.0).clip(-10, 10).to_frame().rename(columns={'volume': 'rolling_volume_5'}).reset_index(drop=True)
rolling_volume_6_3 = (df_grouped.apply(lambda row: np.log1p(row['volume'].rolling(6, min_periods=1).mean()) - np.log1p(row['volume'].rolling(3, min_periods=1).mean()) )/ 10.0).clip(-10, 10).to_frame().rename(columns={'volume': 'rolling_volume_6_3'}).reset_index(drop=True)
pct_volatility = (df_grouped.apply(lambda row: (row['high'] - row['low']) * 50 / row['avg'])).clip(-10, 10).fillna(0.0).to_frame().rename(columns={0: 'pct_volatility'}).reset_index(drop=True)
rolling_pct_volatility_5 = (df_grouped.apply(lambda row: (row['high'].rolling(5, min_periods=1).max() - row['low'].rolling(5, min_periods=1).min())* 20 / row['avg'])).clip(-10, 10).fillna(0.0).to_frame().rename(columns={0: 'rolling_pct_volatility_5'}).reset_index(drop=True)
rolling_pct_volatility_10 = (df_grouped.apply(lambda row: (row['high'].rolling(10, min_periods=1).max() - row['low'].rolling(10, min_periods=1).min()) * 20 / row['avg'])).clip(-10, 10).fillna(0.0).to_frame().rename(columns={0: 'rolling_pct_volatility_10'}).reset_index(drop=True)
pct_vwap_low = (df_grouped.apply(lambda row: (row.ta.vwap() - row['low']) * 50 / row['avg'])).clip(-10, 10).fillna(0.0).to_frame().rename(columns={0: 'pct_vwap_low'}).reset_index(drop=True)
pct_vwap_high = (df_grouped.apply(lambda row: (row['high'] - row.ta.vwap()) * 50 / row['avg'])).clip(-10, 10).fillna(0.0).to_frame().rename(columns={0: 'pct_vwap_high'}).reset_index(drop=True)
feature = pd.concat([rsi_5, cmo_3, natr_3, rvi_5, kdj_9_3, adx_5, mfi_5, pvi_5, nvi_5, willr_5, willr_10, cmf_10, dpo_5, log_return_5, zscore_5, pct_change_3, pct_change_6, rolling_std_5, rolling_money_5, rolling_money_6_3, rolling_volume_5, rolling_volume_6_3, pct_volatility, rolling_pct_volatility_5, rolling_pct_volatility_10, pct_vwap_low, pct_vwap_high, kline_df[['daily_percent']] * 50], axis=1)
feature.columns = ['intraday_%s' % i for i in feature.columns]
index = kline_df[index_col].apply(lambda x: int(x.timestamp()))
introday_data = pd.concat([kline_df[['code', 'trade_day', 'minutes_of_day']], feature], axis=1)
introday_data.index = index
if 'label_d0_25' in kline_df.columns:
label_data = kline_df[['code', 'trade_day', 'minutes_of_day', 'label_d0_25', 'label_d0_50', 'label_d0_75',
'label_rolling_5_50', 'label_rolling_10_50', 'label_rolling_5_50_bps', 'label_rolling_10_50_bps']]
label_data.index = index
else:
label_data = []
introday_data = introday_data[introday_data.minutes_of_day != 900]
label_data = label_data[label_data.minutes_of_day != 900]
return introday_data.reset_index(), label_data.reset_index()
同样的，我们展示上面茅台数据生成经过pandas-ta提取的特征：
introday_data, label_data = make_tech_feature_intraday_jqdata(kline_df, index_col='trade_time')
print(introday_data.iloc[-5:])
print(label_data.iloc[-5:])
- 日内数据特征：
| trade_time | code | trade_day | minutes_of_day | intraday_RSI_5 | intraday_CMO_3 | intraday_NATR_3 | intraday_RVI_5 | intraday_K_9_3 | intraday_D_9_3 | intraday_J_9_3 | intraday_ADX_5 | intraday_DMP_5 | intraday_DMN_5 | intraday_MFI_5 | intraday_PVI_5 | intraday_NVI_5 | intraday_WILLR_5 | intraday_WILLR_10 | intraday_CMF_10 | intraday_DPO_5 | intraday_LOGRET_5 | intraday_ZS_5 | intraday_pct_change_3 | intraday_pct_change_6 | intraday_rolling_std_5 | intraday_rolling_money_5 | intraday_rolling_money_6_3 | intraday_rolling_volume_5 | intraday_rolling_volume_6_3 | intraday_pct_volatility | intraday_rolling_pct_volatility_5 | intraday_rolling_pct_volatility_10 | intraday_pct_vwap_low | intraday_pct_vwap_high | intraday_daily_percent |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1657176780 | 600519.XSHG | 2022-07-07 | 893 | 0.4868 | -0.064103 | 0.077846 | 0.472697 | 0.536684 | 0.41818 | 0.773693 | 0.404289 | 0.170589 | 0.053766 | 0.589231 | 0.999701 | 1.00016 | -0.377273 | -0.452 | 0.197077 | 4.1 | 0.0310894 | -0.381427 | 0.00802214 | 0.00689641 | 0.613697 | -0.010953 | -0.0104473 | -0.011178 | -0.0104575 | 0.03508 | 0.0220503 | 0.0250571 | 0.0622761 | -0.0271961 | -0.178072 |
| 1657176840 | 600519.XSHG | 2022-07-07 | 894 | 0.377828 | -0.418172 | 0.0721596 | 0.306806 | 0.382032 | 0.406131 | 0.333835 | 0.33549 | 0.144056 | 0.127672 | 0.353767 | 0.999656 | 1.00016 | -0.971429 | -0.931624 | -0.0242737 | -10 | -0.0225665 | -1.63965 | -0.00860476 | -0.00167135 | 0.50677 | -0.0136592 | -0.006727 | -0.0136515 | -0.00678157 | 0.0300904 | 0.0210633 | 0.0234705 | 0.0780107 | -0.0479203 | -0.208292 |
| 1657176900 | 600519.XSHG | 2022-07-07 | 895 | 0.428086 | -0.203659 | 0.0636436 | 0.536757 | 0.336506 | 0.382923 | 0.243674 | 0.269736 | 0.125327 | 0.123655 | 0.199639 | 0.999656 | 1.00009 | -0.779343 | -0.754545 | 0.0696795 | -7.7 | -0.0391013 | -0.75092 | -0.0102754 | 0.000668783 | 0.638041 | 0.00777374 | -0.00850169 | 0.00777846 | -0.00857828 | 0.0233219 | 0.0213659 | 0.0220681 | 0.0799701 | -0.0566481 | -0.198801 |
| 1657176960 | 600519.XSHG | 2022-07-07 | 896 | 0.390711 | -0.344547 | 0.051298 | 0.344622 | 0.25161 | 0.339152 | 0.0765271 | 0.222581 | 0.114705 | 0.122769 | 0.165302 | 0.999611 | 1.00009 | -0.922374 | -0.918182 | 0.0739293 | -10 | -0.0223156 | -0.902093 | -0.0146167 | -0.00330078 | 0.709123 | -0.0476034 | -0.0131696 | -0.0475763 | -0.0131647 | 0.012038 | 0.0219693 | 0.0220696 | 0.0807789 | -0.0687409 | -0.207792 |
| 1657177020 | 600519.XSHG | 2022-07-07 | 897 | 0.323935 | -0.563402 | 0.0455835 | 0.243084 | 0.16774 | 0.282015 | -0.0608087 | 0.242891 | 0.100978 | 0.197831 | 2.9012e-16 | 0.999611 | 0.999951 | -1 | -1 | -0.143719 | -10 | -0.0674407 | -1.39213 | -0.00777336 | -0.00818705 | 0.785147 | -0.0210747 | -0.0162005 | -0.0210629 | -0.0161831 | 0.00727525 | 0.027094 | 0.027094 | 0.0928892 | -0.0856139 | -0.224775 |
- 日内标签数据
| trade_time | code | trade_day | minutes_of_day | label_d0_25 | label_d0_50 | label_d0_75 | label_rolling_5_50 | label_rolling_10_50 | label_rolling_5_50_bps | label_rolling_10_50_bps |
|---|---|---|---|---|---|---|---|---|---|---|
| 1657176780 | 600519.XSHG | 2022-07-07 | 893 | -0.119272 | -0.0877 | -0.080684 | -0.0877 | -0.000877 | -175 | -1.75 |
| 1657176840 | 600519.XSHG | 2022-07-07 | 894 | -0.0850054 | -0.0308427 | -0.0132899 | -0.0308427 | -0.000308427 | -61.5 | -0.615 |
| 1657176900 | 600519.XSHG | 2022-07-07 | 895 | -0.115356 | -0.0386191 | -0.0228204 | -0.0386191 | -0.000386191 | -77 | -0.77 |
| 1657176960 | 600519.XSHG | 2022-07-07 | 896 | -0.146713 | -0.108342 | -0.0699708 | -0.108342 | -0.00108342 | -216 | -2.16 |
| 1657177020 | 600519.XSHG | 2022-07-07 | 897 | -0.153533 | -0.153533 | -0.153533 | -0.153533 | -0.00153533 | -306 | -3.06 |
简单计算技术因子与日内收益的相关性：
corr_data = []
for col in introday_data.columns.tolist()[4:]:
feature = introday_data[col]
label_d0_50 = feature.corr(label_data['label_d0_50'])
label_rolling_5_50 = feature.corr(label_data['label_rolling_5_50'])
label_rolling_10_50 = feature.corr(label_data['label_rolling_10_50'])
corr_data.append([col, label_d0_50, label_rolling_5_50, label_rolling_10_50])
corr_data = pd.DataFrame(corr_data, columns=['feature', 'd0_50', 'rolling_5_50', 'rolling_10_50'])
print(corr_data)
| feature | d0_50 | rolling_5_50 | rolling_10_50 |
|---|---|---|---|
| intraday_RSI_5 | 0.0402828 | 0.134321 | 0.103664 |
| intraday_CMO_3 | 0.0460717 | 0.163593 | 0.124605 |
| intraday_NATR_3 | -0.0044438 | 0.0184311 | 0.0159059 |
| intraday_RVI_5 | 0.0384155 | 0.139606 | 0.106858 |
| intraday_K_9_3 | 0.0281801 | 0.0653667 | 0.0561566 |
| intraday_D_9_3 | 0.019375 | 0.0322949 | 0.0315525 |
| intraday_J_9_3 | 0.0305751 | 0.0899986 | 0.0733925 |
| intraday_ADX_5 | 0.00178353 | 0.00119938 | -0.00362684 |
| intraday_DMP_5 | 0.023113 | 0.0693602 | 0.0517322 |
| intraday_DMN_5 | -0.0211941 | -0.0660362 | -0.0478488 |
| intraday_MFI_5 | 0.0187832 | 0.0573829 | 0.0454739 |
| intraday_PVI_5 | -0.00631178 | 0.00495645 | 0.0021088 |
| intraday_NVI_5 | -0.0138721 | 0.0016334 | -0.00108481 |
| intraday_WILLR_5 | 0.0508993 | 0.174907 | 0.135297 |
| intraday_WILLR_10 | 0.0437711 | 0.133309 | 0.106649 |
| intraday_CMF_10 | 0.0310067 | 0.0720305 | 0.060817 |
| intraday_DPO_5 | 0.0322019 | 0.101203 | 0.0789424 |
| intraday_LOGRET_5 | 0.0282654 | 0.101385 | 0.0758353 |
| intraday_ZS_5 | 0.048193 | 0.189238 | 0.140897 |
| intraday_pct_change_3 | 0.0111402 | 0.0316397 | 0.0211543 |
| intraday_pct_change_6 | 0.00922718 | 0.0199352 | 0.0154407 |
| intraday_rolling_std_5 | -0.00929529 | 0.0246064 | 0.0264212 |
| intraday_rolling_money_5 | -0.00175834 | -0.0044877 | -0.00455773 |
| intraday_rolling_money_6_3 | -0.00369373 | -0.00865403 | -0.009164 |
| intraday_rolling_volume_5 | -0.00169721 | -0.00431745 | -0.00442322 |
| intraday_rolling_volume_6_3 | -0.00365949 | -0.00860967 | -0.00912756 |
| intraday_pct_volatility | 0.00960336 | 0.0396089 | 0.0383 |
| intraday_rolling_pct_volatility_5 | 0.0118225 | 0.0395367 | 0.0425203 |
| intraday_rolling_pct_volatility_10 | 0.00827556 | 0.0335309 | 0.0334182 |
| intraday_pct_vwap_low | 0.0097716 | -0.00987224 | -0.00514639 |
| intraday_pct_vwap_high | -0.00828736 | 0.015894 | 0.0109779 |
| intraday_daily_percent | 0.0473394 | 0.025758 | 0.0216949 |
可以发现，5min/10min的相关性还是有一些的，说明技术指标短期内可以起到一些预测作用，这也给我们做日内高频可转债提供了一个思路。
特征转存数据库
这里以日内特征为例，构建数据表，应用上述特征抽取函数，提取技术面特征，然后存储到Clickhouse数据库中
import pandas as pd
from tqdm import tqdm
import numpy as np
from clickhouse_driver import Client
storage_client = Client('10.0.16.11', password='******', settings={'use_numpy': True})
def make_cn_intraday_set(start_date='2018-01-01', end_date='2022-07-01'):
'''
主要股票：日内特征
'''
create_table_intraday_cn_data = '''
create table if not exists feature.intraday_cn_data
(
trade_time DateTime('Asia/Shanghai'), code String, trade_day Date, minutes_of_day Int16,
intraday_RSI_5 Float32, intraday_CMO_3 Float32, intraday_NATR_3 Float32, intraday_RVI_5 Float32, intraday_K_9_3 Float32, intraday_D_9_3 Float32, intraday_J_9_3 Float32, intraday_ADX_5 Float32, intraday_DMP_5 Float32, intraday_DMN_5 Float32, intraday_MFI_5 Float32, intraday_PVI_5 Float32, intraday_NVI_5 Float32, intraday_WILLR_5 Float32, intraday_WILLR_10 Float32, intraday_CMF_10 Float32, intraday_DPO_5 Float32, intraday_LOGRET_5 Float32, intraday_ZS_5 Float32, intraday_pct_change_3 Float32, intraday_pct_change_6 Float32, intraday_rolling_std_5 Float32, intraday_rolling_money_5 Float32, intraday_rolling_money_6_3 Float32, intraday_rolling_volume_5 Float32, intraday_rolling_volume_6_3 Float32, intraday_pct_volatility Float32, intraday_rolling_pct_volatility_5 Float32, intraday_rolling_pct_volatility_10 Float32, intraday_pct_vwap_low Float32, intraday_pct_vwap_high Float32, intraday_daily_percent Float32
)
ENGINE = ReplacingMergeTree()
ORDER BY (trade_day, trade_time, code)
'''
storage_client.execute(create_table_intraday_cn_data)
create_table_intraday_cn_label = '''
create table if not exists feature.intraday_cn_label
(
trade_time DateTime('Asia/Shanghai'), code String, trade_day Date, minutes_of_day Int16,
label_d0_25 Float32, label_d0_50 Float32, label_d0_75 Float32, label_rolling_5_50 Float32, label_rolling_10_50 Float32, label_rolling_5_50_bps Float32, label_rolling_10_50_bps Float32
)
ENGINE = ReplacingMergeTree()
ORDER BY (trade_day, trade_time, code)
'''
storage_client.execute(create_table_intraday_cn_label)
def make(code):
# 读取K线
kline_df = load_jqdata_kline(code=code, start_date=start_date, end_date=end_date)
if len(kline_df) < 1:
return
# 生成特征和标签
introday_data, label_data = make_tech_feature_intraday_jqdata(kline_df)
# 分批写入，降低性能压力
if len(introday_data) > 0:
for i in range(int(np.ceil(len(introday_data) / 2000))):
storage_client.insert_dataframe('INSERT INTO feature.intraday_cn_data VALUES', introday_data.iloc[i * 2000: (i+1)*2000])
storage_client.insert_dataframe('INSERT INTO feature.intraday_cn_label VALUES', label_data.iloc[i * 2000: (i+1)*2000])
# 获取指定的股票列表
tickers = get_jq_ticker(start_date, end_date)
# 遍历股票，生成特征
for ticker in tqdm(tickers):
make(ticker)
至此，我们完成了技术特征数据的提取、存储工作。
欢迎关注我的公众号“量化实战”，原创技术文章第一时间推送。
