diff --git a/code/train/RollingRank.ipynb b/code/train/RollingRank.ipynb new file mode 100644 index 0000000..29e5326 --- /dev/null +++ b/code/train/RollingRank.ipynb @@ -0,0 +1,1840 @@ +{ + "cells": [ + { + "cell_type": "code", + "id": "79a7758178bafdd3", + "metadata": { + "jupyter": { + "source_hidden": true + }, + "ExecuteTime": { + "end_time": "2025-03-29T17:43:30.876671Z", + "start_time": "2025-03-29T17:43:30.425776Z" + } + }, + "source": [ + "# %load_ext autoreload\n", + "# %autoreload 2\n", + "\n", + "import pandas as pd\n", + "import warnings\n", + "\n", + "warnings.filterwarnings(\"ignore\")\n", + "\n", + "pd.set_option('display.max_columns', None)\n" + ], + "outputs": [], + "execution_count": 1 + }, + { + "cell_type": "code", + "id": "a79cafb06a7e0e43", + "metadata": { + "scrolled": true, + "ExecuteTime": { + "end_time": "2025-03-29T17:44:18.824363Z", + "start_time": "2025-03-29T17:43:30.876671Z" + } + }, + "source": [ + "from utils.utils import read_and_merge_h5_data\n", + "\n", + "print('daily data')\n", + "df = read_and_merge_h5_data('../../data/daily_data.h5', key='daily_data',\n", + " columns=['ts_code', 'trade_date', 'open', 'close', 'high', 'low', 'vol', 'pct_chg'],\n", + " df=None)\n", + "\n", + "print('daily basic')\n", + "df = read_and_merge_h5_data('../../data/daily_basic.h5', key='daily_basic',\n", + " columns=['ts_code', 'trade_date', 'turnover_rate', 'pe_ttm', 'circ_mv', 'volume_ratio',\n", + " 'is_st'], df=df, join='inner')\n", + "\n", + "print('stk limit')\n", + "df = read_and_merge_h5_data('../../data/stk_limit.h5', key='stk_limit',\n", + " columns=['ts_code', 'trade_date', 'pre_close', 'up_limit', 'down_limit'],\n", + " df=df)\n", + "print('money flow')\n", + "df = read_and_merge_h5_data('../../data/money_flow.h5', key='money_flow',\n", + " columns=['ts_code', 'trade_date', 'buy_sm_vol', 'sell_sm_vol', 'buy_lg_vol', 'sell_lg_vol',\n", + " 'buy_elg_vol', 'sell_elg_vol', 'net_mf_vol'],\n", + " df=df)\n", + "print('cyq perf')\n", + "df = read_and_merge_h5_data('../../data/cyq_perf.h5', key='cyq_perf',\n", + " columns=['ts_code', 'trade_date', 'his_low', 'his_high', 'cost_5pct', 'cost_15pct',\n", + " 'cost_50pct',\n", + " 'cost_85pct', 'cost_95pct', 'weight_avg', 'winner_rate'],\n", + " df=df)\n", + "print(df.info())" + ], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "daily data\n", + "daily basic\n", + "inner merge on ['ts_code', 'trade_date']\n", + "stk limit\n", + "left merge on ['ts_code', 'trade_date']\n", + "money flow\n", + "left merge on ['ts_code', 'trade_date']\n", + "cyq perf\n", + "left merge on ['ts_code', 'trade_date']\n", + "\n", + "RangeIndex: 8450470 entries, 0 to 8450469\n", + "Data columns (total 31 columns):\n", + " # Column Dtype \n", + "--- ------ ----- \n", + " 0 ts_code object \n", + " 1 trade_date datetime64[ns]\n", + " 2 open float64 \n", + " 3 close float64 \n", + " 4 high float64 \n", + " 5 low float64 \n", + " 6 vol float64 \n", + " 7 pct_chg float64 \n", + " 8 turnover_rate float64 \n", + " 9 pe_ttm float64 \n", + " 10 circ_mv float64 \n", + " 11 volume_ratio float64 \n", + " 12 is_st bool \n", + " 13 up_limit float64 \n", + " 14 down_limit float64 \n", + " 15 buy_sm_vol float64 \n", + " 16 sell_sm_vol float64 \n", + " 17 buy_lg_vol float64 \n", + " 18 sell_lg_vol float64 \n", + " 19 buy_elg_vol float64 \n", + " 20 sell_elg_vol float64 \n", + " 21 net_mf_vol float64 \n", + " 22 his_low float64 \n", + " 23 his_high float64 \n", + " 24 cost_5pct float64 \n", + " 25 cost_15pct float64 \n", + " 26 cost_50pct float64 \n", + " 27 cost_85pct float64 \n", + " 28 cost_95pct float64 \n", + " 29 weight_avg float64 \n", + " 30 winner_rate float64 \n", + "dtypes: bool(1), datetime64[ns](1), float64(28), object(1)\n", + "memory usage: 1.9+ GB\n", + "None\n" + ] + } + ], + "execution_count": 2 + }, + { + "cell_type": "code", + "id": "cac01788dac10678", + "metadata": { + "jupyter": { + "source_hidden": true + }, + "ExecuteTime": { + "end_time": "2025-03-29T17:44:28.421215Z", + "start_time": "2025-03-29T17:44:19.106345Z" + } + }, + "source": [ + "print('industry')\n", + "industry_df = read_and_merge_h5_data('../../data/industry_data.h5', key='industry_data',\n", + " columns=['ts_code', 'l2_code', 'in_date'],\n", + " df=None, on=['ts_code'], join='left')\n", + "\n", + "\n", + "def merge_with_industry_data(df, industry_df):\n", + " # 确保日期字段是 datetime 类型\n", + " df['trade_date'] = pd.to_datetime(df['trade_date'])\n", + " industry_df['in_date'] = pd.to_datetime(industry_df['in_date'])\n", + "\n", + " # 对 industry_df 按 ts_code 和 in_date 排序\n", + " industry_df_sorted = industry_df.sort_values(['in_date', 'ts_code'])\n", + "\n", + " # 对原始 df 按 ts_code 和 trade_date 排序\n", + " df_sorted = df.sort_values(['trade_date', 'ts_code'])\n", + "\n", + " # 使用 merge_asof 进行向后合并\n", + " merged = pd.merge_asof(\n", + " df_sorted,\n", + " industry_df_sorted,\n", + " by='ts_code', # 按 ts_code 分组\n", + " left_on='trade_date',\n", + " right_on='in_date',\n", + " direction='backward'\n", + " )\n", + "\n", + " # 获取每个 ts_code 的最早 in_date 记录\n", + " min_in_date_per_ts = (industry_df_sorted\n", + " .groupby('ts_code')\n", + " .first()\n", + " .reset_index()[['ts_code', 'l2_code']])\n", + "\n", + " # 填充未匹配到的记录（trade_date 早于所有 in_date 的情况）\n", + " merged['l2_code'] = merged['l2_code'].fillna(\n", + " merged['ts_code'].map(min_in_date_per_ts.set_index('ts_code')['l2_code'])\n", + " )\n", + "\n", + " # 保留需要的列并重置索引\n", + " result = merged.reset_index(drop=True)\n", + " return result\n", + "\n", + "\n", + "# 使用示例\n", + "df = merge_with_industry_data(df, industry_df)\n", + "# print(mdf[mdf['ts_code'] == '600751.SH'][['ts_code', 'trade_date', 'l2_code']])" + ], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "industry\n" + ] + } + ], + "execution_count": 3 + }, + { + "cell_type": "code", + "id": "c4e9e1d31da6dba6", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:44:28.620721Z", + "start_time": "2025-03-29T17:44:28.436697Z" + } + }, + "source": [ + "def calculate_indicators(df):\n", + " \"\"\"\n", + " 计算四个指标：当日涨跌幅、5日移动平均、RSI、MACD。\n", + " \"\"\"\n", + " df = df.sort_values('trade_date')\n", + " df['daily_return'] = (df['close'] - df['pre_close']) / df['pre_close'] * 100\n", + " # df['5_day_ma'] = df['close'].rolling(window=5).mean()\n", + " delta = df['close'].diff()\n", + " gain = delta.where(delta > 0, 0)\n", + " loss = -delta.where(delta < 0, 0)\n", + " avg_gain = gain.rolling(window=14).mean()\n", + " avg_loss = loss.rolling(window=14).mean()\n", + " rs = avg_gain / avg_loss\n", + " df['RSI'] = 100 - (100 / (1 + rs))\n", + "\n", + " # 计算MACD\n", + " ema12 = df['close'].ewm(span=12, adjust=False).mean()\n", + " ema26 = df['close'].ewm(span=26, adjust=False).mean()\n", + " df['MACD'] = ema12 - ema26\n", + " df['Signal_line'] = df['MACD'].ewm(span=9, adjust=False).mean()\n", + " df['MACD_hist'] = df['MACD'] - df['Signal_line']\n", + "\n", + " # 4. 情绪因子1：市场上涨比例（Up Ratio）\n", + " df['up_ratio'] = df['daily_return'].apply(lambda x: 1 if x > 0 else 0)\n", + " df['up_ratio_20d'] = df['up_ratio'].rolling(window=20).mean() # 过去20天上涨比例\n", + "\n", + " # 5. 情绪因子2：成交量变化率（Volume Change Rate）\n", + " df['volume_mean'] = df['vol'].rolling(window=20).mean() # 过去20天的平均成交量\n", + " df['volume_change_rate'] = (df['vol'] - df['volume_mean']) / df['volume_mean'] * 100 # 成交量变化率\n", + "\n", + " # 6. 情绪因子3：波动率（Volatility）\n", + " df['volatility'] = df['daily_return'].rolling(window=20).std() # 过去20天的日收益率标准差\n", + "\n", + " # 7. 情绪因子4：成交额变化率（Amount Change Rate）\n", + " df['amount_mean'] = df['amount'].rolling(window=20).mean() # 过去20天的平均成交额\n", + " df['amount_change_rate'] = (df['amount'] - df['amount_mean']) / df['amount_mean'] * 100 # 成交额变化率\n", + "\n", + " return df\n", + "\n", + "\n", + "def generate_index_indicators(h5_filename):\n", + " df = pd.read_hdf(h5_filename, key='index_data')\n", + " df['trade_date'] = pd.to_datetime(df['trade_date'], format='%Y%m%d')\n", + " df = df.sort_values('trade_date')\n", + "\n", + " # 计算每个ts_code的相关指标\n", + " df_indicators = []\n", + " for ts_code in df['ts_code'].unique():\n", + " df_index = df[df['ts_code'] == ts_code].copy()\n", + " df_index = calculate_indicators(df_index)\n", + " df_indicators.append(df_index)\n", + "\n", + " # 合并所有指数的结果\n", + " df_all_indicators = pd.concat(df_indicators, ignore_index=True)\n", + "\n", + " # 保留trade_date列，并将同一天的数据按ts_code合并成一行\n", + " df_final = df_all_indicators.pivot_table(\n", + " index='trade_date',\n", + " columns='ts_code',\n", + " values=['daily_return', 'RSI', 'MACD', 'Signal_line',\n", + " 'MACD_hist', 'up_ratio_20d', 'volume_change_rate', 'volatility',\n", + " 'amount_change_rate', 'amount_mean'],\n", + " aggfunc='last'\n", + " )\n", + "\n", + " df_final.columns = [f\"{col[1]}_{col[0]}\" for col in df_final.columns]\n", + " df_final = df_final.reset_index()\n", + "\n", + " return df_final\n", + "\n", + "\n", + "# 使用函数\n", + "h5_filename = '../../data/index_data.h5'\n", + "index_data = generate_index_indicators(h5_filename)\n", + "index_data = index_data.dropna()\n" + ], + "outputs": [], + "execution_count": 4 + }, + { + "cell_type": "code", + "id": "a735bc02ceb4d872", + "metadata": { + "jupyter": { + "source_hidden": true + }, + "ExecuteTime": { + "end_time": "2025-03-29T17:44:28.706766Z", + "start_time": "2025-03-29T17:44:28.650141Z" + } + }, + "source": [ + "import numpy as np\n", + "import talib\n", + "\n", + "\n", + "def get_rolling_factor(df):\n", + " old_columns = df.columns.tolist()[:]\n", + " # 按股票和日期排序\n", + " df = df.sort_values(by=['ts_code', 'trade_date'])\n", + " grouped = df.groupby('ts_code', group_keys=False)\n", + "\n", + " # df[\"gap_next_open\"] = (df[\"open\"].shift(-1) - df[\"close\"]) / df[\"close\"]\n", + "\n", + " df['return_skew'] = grouped['pct_chg'].rolling(window=5).skew().reset_index(0, drop=True)\n", + " df['return_kurtosis'] = grouped['pct_chg'].rolling(window=5).kurt().reset_index(0, drop=True)\n", + "\n", + " # 因子 1：短期成交量变化率\n", + " df['volume_change_rate'] = (\n", + " grouped['vol'].rolling(window=2).mean() /\n", + " grouped['vol'].rolling(window=10).mean() - 1\n", + " ).reset_index(level=0, drop=True) # 确保索引对齐\n", + "\n", + " # 因子 2：成交量突破信号\n", + " max_volume = grouped['vol'].rolling(window=5).max().reset_index(level=0, drop=True) # 确保索引对齐\n", + " df['cat_volume_breakout'] = (df['vol'] > max_volume)\n", + "\n", + " # 因子 3：换手率均线偏离度\n", + " mean_turnover = grouped['turnover_rate'].rolling(window=3).mean().reset_index(level=0, drop=True)\n", + " std_turnover = grouped['turnover_rate'].rolling(window=3).std().reset_index(level=0, drop=True)\n", + " df['turnover_deviation'] = (df['turnover_rate'] - mean_turnover) / std_turnover\n", + "\n", + " # 因子 4：换手率激增信号\n", + " df['cat_turnover_spike'] = (df['turnover_rate'] > mean_turnover + 2 * std_turnover)\n", + "\n", + " # 因子 5：量比均值\n", + " df['avg_volume_ratio'] = grouped['volume_ratio'].rolling(window=3).mean().reset_index(level=0, drop=True)\n", + "\n", + " # 因子 6：量比突破信号\n", + " max_volume_ratio = grouped['volume_ratio'].rolling(window=5).max().reset_index(level=0, drop=True)\n", + " df['cat_volume_ratio_breakout'] = (df['volume_ratio'] > max_volume_ratio)\n", + "\n", + " df['vol_spike'] = grouped.apply(\n", + " lambda x: pd.Series(x['vol'].rolling(20).mean(), index=x.index)\n", + " )\n", + " df['vol_std_5'] = df['vol'].pct_change().rolling(5).std()\n", + "\n", + " # 计算 ATR\n", + " df['atr_14'] = grouped.apply(\n", + " lambda x: pd.Series(talib.ATR(x['high'].values, x['low'].values, x['close'].values, timeperiod=14),\n", + " index=x.index)\n", + " )\n", + " df['atr_6'] = grouped.apply(\n", + " lambda x: pd.Series(talib.ATR(x['high'].values, x['low'].values, x['close'].values, timeperiod=6),\n", + " index=x.index)\n", + " )\n", + "\n", + " # 计算 OBV 及其均线\n", + " df['obv'] = grouped.apply(\n", + " lambda x: pd.Series(talib.OBV(x['close'].values, x['vol'].values), index=x.index)\n", + " )\n", + " df['maobv_6'] = grouped.apply(\n", + " lambda x: pd.Series(talib.SMA(x['obv'].values, timeperiod=6), index=x.index)\n", + " )\n", + "\n", + " df['rsi_3'] = grouped.apply(\n", + " lambda x: pd.Series(talib.RSI(x['close'].values, timeperiod=3), index=x.index)\n", + " )\n", + " # df['rsi_6'] = grouped.apply(\n", + " # lambda x: pd.Series(talib.RSI(x['close'].values, timeperiod=6), index=x.index)\n", + " # )\n", + " # df['rsi_9'] = grouped.apply(\n", + " # lambda x: pd.Series(talib.RSI(x['close'].values, timeperiod=9), index=x.index)\n", + " # )\n", + "\n", + " # 计算 return_10 和 return_20\n", + " df['return_5'] = grouped['close'].apply(lambda x: x / x.shift(5) - 1)\n", + " # df['return_10'] = grouped['close'].apply(lambda x: x / x.shift(10) - 1)\n", + " df['return_20'] = grouped['close'].apply(lambda x: x / x.shift(20) - 1)\n", + "\n", + " # df['avg_close_5'] = grouped['close'].apply(lambda x: x.rolling(window=5).mean() / x)\n", + "\n", + " # 计算标准差指标\n", + " df['std_return_5'] = grouped['close'].apply(lambda x: x.pct_change().rolling(window=5).std())\n", + " # df['std_return_15'] = grouped['close'].apply(lambda x: x.pct_change().rolling(window=15).std())\n", + " # df['std_return_25'] = grouped['close'].apply(lambda x: x.pct_change().rolling(window=25).std())\n", + " df['std_return_90'] = grouped['close'].apply(lambda x: x.pct_change().rolling(window=90).std())\n", + " df['std_return_90_2'] = grouped['close'].apply(lambda x: x.shift(10).pct_change().rolling(window=90).std())\n", + "\n", + " # 计算 EMA 指标\n", + " df['_ema_5'] = grouped['close'].apply(\n", + " lambda x: pd.Series(talib.EMA(x.values, timeperiod=5), index=x.index)\n", + " )\n", + " df['_ema_13'] = grouped['close'].apply(\n", + " lambda x: pd.Series(talib.EMA(x.values, timeperiod=13), index=x.index)\n", + " )\n", + " df['_ema_20'] = grouped['close'].apply(\n", + " lambda x: pd.Series(talib.EMA(x.values, timeperiod=20), index=x.index)\n", + " )\n", + " df['_ema_60'] = grouped['close'].apply(\n", + " lambda x: pd.Series(talib.EMA(x.values, timeperiod=60), index=x.index)\n", + " )\n", + "\n", + " # 计算 act_factor1, act_factor2, act_factor3, act_factor4\n", + " df['act_factor1'] = grouped['_ema_5'].apply(\n", + " lambda x: np.arctan((x / x.shift(1) - 1) * 100) * 57.3 / 50\n", + " )\n", + " df['act_factor2'] = grouped['_ema_13'].apply(\n", + " lambda x: np.arctan((x / x.shift(1) - 1) * 100) * 57.3 / 40\n", + " )\n", + " df['act_factor3'] = grouped['_ema_20'].apply(\n", + " lambda x: np.arctan((x / x.shift(1) - 1) * 100) * 57.3 / 21\n", + " )\n", + " df['act_factor4'] = grouped['_ema_60'].apply(\n", + " lambda x: np.arctan((x / x.shift(1) - 1) * 100) * 57.3 / 10\n", + " )\n", + "\n", + " # 根据 trade_date 截面计算排名\n", + " df['rank_act_factor1'] = df.groupby('trade_date', group_keys=False)['act_factor1'].rank(ascending=False, pct=True)\n", + " df['rank_act_factor2'] = df.groupby('trade_date', group_keys=False)['act_factor2'].rank(ascending=False, pct=True)\n", + " df['rank_act_factor3'] = df.groupby('trade_date', group_keys=False)['act_factor3'].rank(ascending=False, pct=True)\n", + "\n", + " df['log(circ_mv)'] = np.log(df['circ_mv'])\n", + "\n", + " def rolling_covariance(x, y, window):\n", + " return x.rolling(window).cov(y)\n", + "\n", + " def delta(series, period):\n", + " return series.diff(period)\n", + "\n", + " def rank(series):\n", + " return series.rank(pct=True)\n", + "\n", + " def stddev(series, window):\n", + " return series.rolling(window).std()\n", + "\n", + " window_high_volume = 5\n", + " window_close_stddev = 20\n", + " period_delta = 5\n", + " df['cov'] = rolling_covariance(df['high'], df['vol'], window_high_volume)\n", + " df['delta_cov'] = delta(df['cov'], period_delta)\n", + " df['_rank_stddev'] = rank(stddev(df['close'], window_close_stddev))\n", + " df['alpha_22_improved'] = -1 * df['delta_cov'] * df['_rank_stddev']\n", + "\n", + " df['alpha_003'] = np.where(df['high'] != df['low'],\n", + " (df['close'] - df['open']) / (df['high'] - df['low']),\n", + " 0)\n", + "\n", + " df['alpha_007'] = grouped.apply(lambda x: x['close'].rolling(5).corr(x['vol'])).reset_index(level=0, drop=True)\n", + " df['alpha_007'] = df.groupby('trade_date', group_keys=False)['alpha_007'].rank(ascending=True, pct=True)\n", + "\n", + " df['alpha_013'] = grouped['close'].transform(lambda x: x.rolling(5).sum() - x.rolling(20).sum())\n", + " df['alpha_013'] = df.groupby('trade_date', group_keys=False)['alpha_013'].rank(ascending=True, pct=True)\n", + "\n", + " df['cat_up_limit'] = (df['close'] == df['up_limit']) # 是否涨停（1表示涨停，0表示未涨停）\n", + " df['cat_down_limit'] = (df['close'] == df['down_limit']) # 是否跌停（1表示跌停，0表示未跌停）\n", + " df['up_limit_count_10d'] = grouped['cat_up_limit'].rolling(window=10, min_periods=1).sum().reset_index(level=0,\n", + " drop=True)\n", + " df['down_limit_count_10d'] = grouped['cat_down_limit'].rolling(window=10, min_periods=1).sum().reset_index(level=0,\n", + " drop=True)\n", + "\n", + " # 3. 最近连续涨跌停天数\n", + " def calculate_consecutive_limits(series):\n", + " \"\"\"\n", + " 计算连续涨停/跌停天数。\n", + " \"\"\"\n", + " consecutive_up = series * (series.groupby((series != series.shift()).cumsum()).cumcount() + 1)\n", + " consecutive_down = series * (series.groupby((series != series.shift()).cumsum()).cumcount() + 1)\n", + " return consecutive_up, consecutive_down\n", + "\n", + " # 连续涨停天数\n", + " df['consecutive_up_limit'] = grouped['cat_up_limit'].apply(\n", + " lambda x: calculate_consecutive_limits(x)[0]\n", + " ).reset_index(level=0, drop=True)\n", + "\n", + " df['vol_break'] = np.where((df['close'] > df['cost_85pct']) & (df['volume_ratio'] > 2), 1, 0)\n", + "\n", + " df['weight_roc5'] = grouped['weight_avg'].apply(lambda x: x.pct_change(5))\n", + "\n", + " def rolling_corr(group):\n", + " roc_close = group['close'].pct_change()\n", + " roc_weight = group['weight_avg'].pct_change()\n", + " return roc_close.rolling(10).corr(roc_weight)\n", + "\n", + " df['price_cost_divergence'] = grouped.apply(rolling_corr)\n", + "\n", + " df['smallcap_concentration'] = (1 / df['log(circ_mv)']) * (df['cost_85pct'] - df['cost_15pct'])\n", + "\n", + " # 16. 筹码稳定性指数 (20日波动率)\n", + " df['weight_std20'] = grouped['weight_avg'].apply(lambda x: x.rolling(20).std())\n", + " df['cost_stability'] = df['weight_std20'] / grouped['weight_avg'].transform(lambda x: x.rolling(20).mean())\n", + "\n", + " # 17. 成本区间突破标记\n", + " df['high_cost_break_days'] = grouped.apply(lambda g: g['close'].gt(g['cost_95pct']).rolling(5).sum())\n", + "\n", + " # 20. 筹码-流动性风险\n", + " df['liquidity_risk'] = (df['cost_95pct'] - df['cost_5pct']) * (\n", + " 1 / grouped['vol'].transform(lambda x: x.rolling(10).mean()))\n", + "\n", + " # 7. 市值波动率因子\n", + " df['turnover_std'] = grouped['turnover_rate'].rolling(window=20).std().reset_index(level=0, drop=True)\n", + " df['mv_volatility'] = grouped.apply(lambda x: x['turnover_std'] / x['log(circ_mv)']).reset_index(level=0, drop=True)\n", + "\n", + " # 8. 市值成长性因子\n", + " df['volume_growth'] = grouped['vol'].pct_change(periods=20).reset_index(level=0, drop=True)\n", + " df['mv_growth'] = grouped.apply(lambda x: x['volume_growth'] / x['log(circ_mv)']).reset_index(level=0, drop=True)\n", + "\n", + " df[\"ar\"] = df[\"high\"].div(df[\"open\"]).rolling(3).sum() / df[\"open\"].div(df[\"low\"]).rolling(3).sum() * 100\n", + " # 计算 BR 指标\n", + " df[\"pre_close\"] = df[\"close\"].shift(1)\n", + " df[\"br_up\"] = (df[\"high\"] - df[\"pre_close\"]).clip(lower=0)\n", + " df[\"br_down\"] = (df[\"pre_close\"] - df[\"low\"]).clip(lower=0)\n", + " df[\"br\"] = df[\"br_up\"].rolling(3).sum() / df[\"br_down\"].rolling(3).sum() * 100\n", + " df['arbr'] = df['ar'] - df['br']\n", + " df.drop(columns=[\"pre_close\", \"br_up\", \"br_down\", 'ar', 'br'], inplace=True)\n", + "\n", + " df.drop(columns=['weight_std20'], inplace=True, errors='ignore')\n", + " new_columns = [col for col in df.columns.tolist()[:] if col not in old_columns]\n", + "\n", + " return df, new_columns\n", + "\n", + "\n", + "def get_simple_factor(df):\n", + " old_columns = df.columns.tolist()[:]\n", + " df = df.sort_values(by=['ts_code', 'trade_date'])\n", + "\n", + " alpha = 0.5\n", + " df['momentum_factor'] = df['volume_change_rate'] + alpha * df['turnover_deviation']\n", + " df['resonance_factor'] = df['volume_ratio'] * df['pct_chg']\n", + " df['log_close'] = np.log(df['close'])\n", + "\n", + " df['cat_vol_spike'] = df['vol'] > 2 * df['vol_spike']\n", + "\n", + " df['up'] = (df['high'] - df[['close', 'open']].max(axis=1)) / df['close']\n", + " df['down'] = (df[['close', 'open']].min(axis=1) - df['low']) / df['close']\n", + "\n", + " df['obv-maobv_6'] = df['obv'] - df['maobv_6']\n", + "\n", + " # 计算比值指标\n", + " df['std_return_5 / std_return_90'] = df['std_return_5'] / df['std_return_90']\n", + " # df['std_return_5 / std_return_25'] = df['std_return_5'] / df['std_return_25']\n", + "\n", + " # 计算标准差差值\n", + " df['std_return_90 - std_return_90_2'] = df['std_return_90'] - df['std_return_90_2']\n", + "\n", + " # df['cat_af1'] = df['act_factor1'] > 0\n", + " df['cat_af2'] = df['act_factor2'] > df['act_factor1']\n", + " df['cat_af3'] = df['act_factor3'] > df['act_factor2']\n", + " df['cat_af4'] = df['act_factor4'] > df['act_factor3']\n", + "\n", + " # 计算 act_factor5 和 act_factor6\n", + " df['act_factor5'] = df['act_factor1'] + df['act_factor2'] + df['act_factor3'] + df['act_factor4']\n", + " df['act_factor6'] = (df['act_factor1'] - df['act_factor2']) / np.sqrt(\n", + " df['act_factor1'] ** 2 + df['act_factor2'] ** 2)\n", + "\n", + " df['active_buy_volume_large'] = df['buy_lg_vol'] / df['net_mf_vol']\n", + " df['active_buy_volume_big'] = df['buy_elg_vol'] / df['net_mf_vol']\n", + " df['active_buy_volume_small'] = df['buy_sm_vol'] / df['net_mf_vol']\n", + "\n", + " df['buy_lg_vol_minus_sell_lg_vol'] = (df['buy_lg_vol'] - df['sell_lg_vol']) / df['net_mf_vol']\n", + " df['buy_elg_vol_minus_sell_elg_vol'] = (df['buy_elg_vol'] - df['sell_elg_vol']) / df['net_mf_vol']\n", + "\n", + " df['log(circ_mv)'] = np.log(df['circ_mv'])\n", + "\n", + " df['ctrl_strength'] = (df['cost_85pct'] - df['cost_15pct']) / (df['his_high'] - df['his_low'])\n", + "\n", + " df['low_cost_dev'] = (df['close'] - df['cost_5pct']) / (df['cost_50pct'] - df['cost_5pct'])\n", + "\n", + " df['asymmetry'] = (df['cost_95pct'] - df['cost_50pct']) / (df['cost_50pct'] - df['cost_5pct'])\n", + "\n", + " df['lock_factor'] = df['turnover_rate'] * (\n", + " 1 - (df['cost_95pct'] - df['cost_5pct']) / (df['his_high'] - df['his_low']))\n", + "\n", + " df['cat_vol_break'] = (df['close'] > df['cost_85pct']) & (df['volume_ratio'] > 2)\n", + "\n", + " df['cost_atr_adj'] = (df['cost_95pct'] - df['cost_5pct']) / df['atr_14']\n", + "\n", + " # 12. 小盘股筹码集中度\n", + " df['smallcap_concentration'] = (1 / df['log(circ_mv)']) * (df['cost_85pct'] - df['cost_15pct'])\n", + "\n", + " df['cat_golden_resonance'] = ((df['close'] > df['weight_avg']) &\n", + " (df['volume_ratio'] > 1.5) &\n", + " (df['winner_rate'] > 0.7))\n", + "\n", + " df['mv_turnover_ratio'] = df['turnover_rate'] / df['log(circ_mv)']\n", + "\n", + " df['mv_adjusted_volume'] = df['vol'] / df['log(circ_mv)']\n", + "\n", + " df['mv_weighted_turnover'] = df['turnover_rate'] * (1 / df['log(circ_mv)'])\n", + "\n", + " df['nonlinear_mv_volume'] = df['vol'] / df['log(circ_mv)']\n", + "\n", + " df['mv_volume_ratio'] = df['volume_ratio'] / df['log(circ_mv)']\n", + "\n", + " df['mv_momentum'] = df['turnover_rate'] * df['volume_ratio'] / df['log(circ_mv)']\n", + "\n", + " drop_columns = [col for col in df.columns if col.startswith('_')]\n", + " df.drop(columns=drop_columns, inplace=True, errors='ignore')\n", + "\n", + " new_columns = [col for col in df.columns.tolist()[:] if col not in old_columns]\n", + " return df, new_columns\n" + ], + "outputs": [], + "execution_count": 5 + }, + { + "cell_type": "code", + "id": "53f86ddc0677a6d7", + "metadata": { + "jupyter": { + "source_hidden": true + }, + "scrolled": true, + "ExecuteTime": { + "end_time": "2025-03-29T17:44:33.959917Z", + "start_time": "2025-03-29T17:44:28.720764Z" + } + }, + "source": [ + "from utils.factor import get_act_factor\n", + "\n", + "\n", + "def read_industry_data(h5_filename):\n", + " # 读取 H5 文件中所有的行业数据\n", + " industry_data = pd.read_hdf(h5_filename, key='sw_daily', columns=[\n", + " 'ts_code', 'trade_date', 'open', 'close', 'high', 'low', 'pe', 'pb', 'vol'\n", + " ]) # 假设 H5 文件的键是 'industry_data'\n", + " industry_data = industry_data.sort_values(by=['ts_code', 'trade_date'])\n", + " industry_data = industry_data.reindex()\n", + " industry_data['trade_date'] = pd.to_datetime(industry_data['trade_date'], format='%Y%m%d')\n", + "\n", + " grouped = industry_data.groupby('ts_code', group_keys=False)\n", + " industry_data['obv'] = grouped.apply(\n", + " lambda x: pd.Series(talib.OBV(x['close'].values, x['vol'].values), index=x.index)\n", + " )\n", + " industry_data['return_5'] = grouped['close'].apply(lambda x: x / x.shift(5) - 1)\n", + " industry_data['return_20'] = grouped['close'].apply(lambda x: x / x.shift(20) - 1)\n", + "\n", + " industry_data = get_act_factor(industry_data, cat=False)\n", + " industry_data = industry_data.sort_values(by=['trade_date', 'ts_code'])\n", + "\n", + " # # 计算每天每个 ts_code 的因子和当天所有 ts_code 的中位数的偏差\n", + " # factor_columns = ['obv', 'return_5', 'return_20', 'act_factor1', 'act_factor2', 'act_factor3', 'act_factor4'] # 因子列\n", + " # \n", + " # for factor in factor_columns:\n", + " # if factor in industry_data.columns:\n", + " # # 计算每天每个 ts_code 的因子值与当天所有 ts_code 的中位数的偏差\n", + " # industry_data[f'{factor}_deviation'] = industry_data.groupby('trade_date')[factor].transform(\n", + " # lambda x: x - x.mean())\n", + "\n", + " industry_data['return_5_percentile'] = industry_data.groupby('trade_date')['return_5'].transform(\n", + " lambda x: x.rank(pct=True))\n", + " industry_data['return_20_percentile'] = industry_data.groupby('trade_date')['return_20'].transform(\n", + " lambda x: x.rank(pct=True))\n", + " industry_data = industry_data.drop(columns=['open', 'close', 'high', 'low', 'pe', 'pb', 'vol'])\n", + "\n", + " industry_data = industry_data.rename(\n", + " columns={col: f'industry_{col}' for col in industry_data.columns if col not in ['ts_code', 'trade_date']})\n", + "\n", + " industry_data = industry_data.rename(columns={'ts_code': 'cat_l2_code'})\n", + " return industry_data\n", + "\n", + "\n", + "industry_df = read_industry_data('../../data/sw_daily.h5')\n" + ], + "outputs": [], + "execution_count": 6 + }, + { + "cell_type": "code", + "id": "dbe2fd8021b9417f", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:44:33.985360Z", + "start_time": "2025-03-29T17:44:33.975319Z" + } + }, + "source": [ + "origin_columns = df.columns.tolist()\n", + "origin_columns = [col for col in origin_columns if\n", + " col not in ['turnover_rate', 'pe_ttm', 'volume_ratio', 'vol', 'pct_chg', 'l2_code', 'winner_rate']]\n", + "origin_columns = [col for col in origin_columns if col not in index_data.columns]\n", + "origin_columns = [col for col in origin_columns if 'cyq' not in col]\n", + "print(origin_columns)" + ], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "['ts_code', 'open', 'close', 'high', 'low', 'circ_mv', 'is_st', 'up_limit', 'down_limit', 'buy_sm_vol', 'sell_sm_vol', 'buy_lg_vol', 'sell_lg_vol', 'buy_elg_vol', 'sell_elg_vol', 'net_mf_vol', 'his_low', 'his_high', 'cost_5pct', 'cost_15pct', 'cost_50pct', 'cost_85pct', 'cost_95pct', 'weight_avg', 'in_date']\n" + ] + } + ], + "execution_count": 7 + }, + { + "cell_type": "code", + "id": "85c3e3d0235ffffa", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:27.764400Z", + "start_time": "2025-03-29T17:44:34.016244Z" + } + }, + "source": [ + "def filter_data(df):\n", + " # df = df.groupby('trade_date').apply(lambda x: x.nlargest(1000, 'act_factor1'))\n", + " df = df[~df['is_st']]\n", + " df = df[~df['ts_code'].str.endswith('BJ')]\n", + " df = df[~df['ts_code'].str.startswith('30')]\n", + " df = df[~df['ts_code'].str.startswith('68')]\n", + " df = df[~df['ts_code'].str.startswith('8')]\n", + " df = df[df['trade_date'] >= '20180101']\n", + " df = df.drop(columns=['in_date'])\n", + " df = df.reset_index(drop=True)\n", + " return df\n", + "\n", + "\n", + "df = filter_data(df)\n", + "# df = get_technical_factor(df)\n", + "# df = get_act_factor(df)\n", + "# df = get_money_flow_factor(df)\n", + "# df = get_alpha_factor(df)\n", + "# df = get_limit_factor(df)\n", + "# df = get_cyp_perf_factor(df)\n", + "# df = get_mv_factors(df)\n", + "df, _ = get_rolling_factor(df)\n", + "df, _ = get_simple_factor(df)\n", + "# df = df.merge(industry_df, on=['l2_code', 'trade_date'], how='left')\n", + "df = df.rename(columns={'l2_code': 'cat_l2_code'})\n", + "# df = df.merge(index_data, on='trade_date', how='left')\n", + "\n", + "print(df.info())" + ], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + "Index: 5102787 entries, 0 to 5102786\n", + "Columns: 115 entries, ts_code to mv_momentum\n", + "dtypes: bool(12), datetime64[ns](1), float64(98), int32(1), int64(1), object(2)\n", + "memory usage: 4.0+ GB\n", + "None\n" + ] + } + ], + "execution_count": 8 + }, + { + "cell_type": "code", + "id": "92d84ce15a562ec6", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:29.644814Z", + "start_time": "2025-03-29T17:46:28.384214Z" + } + }, + "source": [ + "from scipy.stats import ks_2samp, wasserstein_distance\n", + "from sklearn.metrics import roc_auc_score\n", + "from sklearn.model_selection import train_test_split\n", + "from sklearn.preprocessing import StandardScaler\n", + "\n", + "\n", + "def remove_shifted_features(train_data, test_data, feature_columns, ks_threshold=0.05, wasserstein_threshold=0.1,\n", + " importance_threshold=0.05):\n", + " dropped_features = []\n", + "\n", + " # **统计数据漂移**\n", + " numeric_columns = train_data.select_dtypes(include=['float64', 'int64']).columns\n", + " numeric_columns = [col for col in numeric_columns if col in feature_columns]\n", + " for feature in numeric_columns:\n", + " ks_stat, p_value = ks_2samp(train_data[feature], test_data[feature])\n", + " wasserstein_dist = wasserstein_distance(train_data[feature], test_data[feature])\n", + "\n", + " if p_value < ks_threshold and wasserstein_dist > wasserstein_threshold:\n", + " dropped_features.append(feature)\n", + "\n", + " print(f\"检测到 {len(dropped_features)} 个可能漂移的特征: {dropped_features}\")\n", + "\n", + " # **应用阈值进行最终筛选**\n", + " filtered_features = [f for f in feature_columns if f not in dropped_features]\n", + "\n", + " return filtered_features, dropped_features\n", + "\n" + ], + "outputs": [], + "execution_count": 9 + }, + { + "cell_type": "code", + "id": "f4f16d63ad18d1bc", + "metadata": { + "jupyter": { + "source_hidden": true + }, + "ExecuteTime": { + "end_time": "2025-03-29T17:46:29.655148Z", + "start_time": "2025-03-29T17:46:29.646857Z" + } + }, + "source": [ + "def create_deviation_within_dates(df, feature_columns):\n", + " groupby_col = 'cat_l2_code' # 使用 trade_date 进行分组\n", + " new_columns = {}\n", + " ret_feature_columns = feature_columns[:]\n", + "\n", + " # 自动选择所有数值型特征\n", + " num_features = [col for col in feature_columns if 'cat' not in col and 'index' not in col]\n", + "\n", + " # num_features = ['vol', 'pct_chg', 'turnover_rate', 'volume_ratio', 'cat_vol_spike', 'obv', 'maobv_6', 'return_5', 'return_10', 'return_20', 'std_return_5', 'std_return_15', 'std_return_90', 'std_return_90_2', 'act_factor1', 'act_factor2', 'act_factor3', 'act_factor4', 'act_factor5', 'act_factor6', 'rank_act_factor1', 'rank_act_factor2', 'rank_act_factor3', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small', 'alpha_022', 'alpha_003', 'alpha_007', 'alpha_013']\n", + " num_features = [col for col in num_features if 'cat' not in col and 'industry' not in col]\n", + " num_features = [col for col in num_features if 'limit' not in col]\n", + " num_features = [col for col in num_features if 'cyq' not in col]\n", + "\n", + " # 遍历所有数值型特征\n", + " for feature in num_features:\n", + " if feature == 'trade_date': # 不需要对 'trade_date' 计算偏差\n", + " continue\n", + "\n", + " # grouped_mean = df.groupby(['trade_date'])[feature].transform('mean')\n", + " # deviation_col_name = f'deviation_mean_{feature}'\n", + " # new_columns[deviation_col_name] = df[feature] - grouped_mean\n", + " # ret_feature_columns.append(deviation_col_name)\n", + "\n", + " grouped_mean = df.groupby(['trade_date', groupby_col])[feature].transform('mean')\n", + " deviation_col_name = f'deviation_mean_{feature}'\n", + " new_columns[deviation_col_name] = df[feature] - grouped_mean\n", + " ret_feature_columns.append(deviation_col_name)\n", + "\n", + " # 将新计算的偏差特征与原始 DataFrame 合并\n", + " df = pd.concat([df, pd.DataFrame(new_columns)], axis=1)\n", + "\n", + " # for feature in ['obv', 'return_20', 'act_factor1', 'act_factor2', 'act_factor3', 'act_factor4']:\n", + " # df[f'deviation_industry_{feature}'] = df[feature] - df[f'industry_{feature}']\n", + "\n", + " return df, ret_feature_columns\n" + ], + "outputs": [], + "execution_count": 10 + }, + { + "cell_type": "code", + "id": "40e6b68a91b30c79", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:30.303881Z", + "start_time": "2025-03-29T17:46:29.698776Z" + } + }, + "source": [ + "import pandas as pd\n", + "\n", + "\n", + "def remove_outliers_label_percentile(label: pd.Series, lower_percentile: float = 0.01, upper_percentile: float = 0.99,\n", + " log=True):\n", + " if not (0 <= lower_percentile < upper_percentile <= 1):\n", + " raise ValueError(\"Percentile values must satisfy 0 <= lower_percentile < upper_percentile <= 1.\")\n", + "\n", + " # Calculate lower and upper bounds based on percentiles\n", + " lower_bound = label.quantile(lower_percentile)\n", + " upper_bound = label.quantile(upper_percentile)\n", + "\n", + " # Filter out values outside the bounds\n", + " filtered_label = label[(label >= lower_bound) & (label <= upper_bound)]\n", + "\n", + " # Print the number of removed outliers\n", + " if log:\n", + " print(f\"Removed {len(label) - len(filtered_label)} outliers.\")\n", + " return filtered_label\n", + "\n", + "\n", + "def calculate_risk_adjusted_target(df, days=5):\n", + " df = df.sort_values(by=['ts_code', 'trade_date'])\n", + "\n", + " df['future_close'] = df.groupby('ts_code')['close'].shift(-days)\n", + " df['future_open'] = df.groupby('ts_code')['open'].shift(-1)\n", + " df['future_return'] = (df['future_close'] - df['future_open']) / df['future_open']\n", + "\n", + " df['future_volatility'] = df.groupby('ts_code')['future_return'].rolling(days, min_periods=1).std().reset_index(\n", + " level=0, drop=True)\n", + " sharpe_ratio = df['future_return'] * df['future_volatility']\n", + " sharpe_ratio.replace([np.inf, -np.inf], np.nan, inplace=True)\n", + "\n", + " return sharpe_ratio\n", + "\n", + "\n", + "def calculate_score(df, days=5, lambda_param=1.0):\n", + " def calculate_max_drawdown(prices):\n", + " peak = prices.iloc[0] # 初始化峰值\n", + " max_drawdown = 0 # 初始化最大回撤\n", + "\n", + " for price in prices:\n", + " if price > peak:\n", + " peak = price # 更新峰值\n", + " else:\n", + " drawdown = (peak - price) / peak # 计算当前回撤\n", + " max_drawdown = max(max_drawdown, drawdown) # 更新最大回撤\n", + "\n", + " return max_drawdown\n", + "\n", + " def compute_stock_score(stock_df):\n", + " stock_df = stock_df.sort_values(by=['trade_date'])\n", + " future_return = stock_df['future_return']\n", + " volatility = stock_df['close'].pct_change().rolling(days).std().shift(-days)\n", + " max_drawdown = stock_df['close'].rolling(days).apply(calculate_max_drawdown, raw=False).shift(-days)\n", + " score = future_return - lambda_param * max_drawdown\n", + "\n", + " return score\n", + "\n", + " scores = df.groupby('ts_code').apply(lambda x: compute_stock_score(x))\n", + " scores = scores.reset_index(level=0, drop=True)\n", + "\n", + " return scores\n", + "\n", + "\n", + "def remove_highly_correlated_features(df, feature_columns, threshold=0.9):\n", + " numeric_features = df[feature_columns].select_dtypes(include=[np.number]).columns.tolist()\n", + " if not numeric_features:\n", + " raise ValueError(\"No numeric features found in the provided data.\")\n", + "\n", + " corr_matrix = df[numeric_features].corr().abs()\n", + " upper = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k=1).astype(bool))\n", + " to_drop = [column for column in upper.columns if any(upper[column] > threshold)]\n", + " remaining_features = [col for col in feature_columns if col not in to_drop\n", + " or 'act' in col or 'af' in col]\n", + " return remaining_features\n", + "\n", + "\n", + "import pandas as pd\n", + "from sklearn.preprocessing import StandardScaler\n", + "\n", + "\n", + "def cross_sectional_standardization(df, features):\n", + " df_sorted = df.sort_values(by='trade_date') # 按时间排序\n", + " df_standardized = df_sorted.copy()\n", + "\n", + " for date in df_sorted['trade_date'].unique():\n", + " # 获取当前时间点的数据\n", + " current_data = df_standardized[df_standardized['trade_date'] == date]\n", + "\n", + " # 只对指定特征进行标准化\n", + " scaler = StandardScaler()\n", + " standardized_values = scaler.fit_transform(current_data[features])\n", + "\n", + " # 将标准化结果重新赋值回去\n", + " df_standardized.loc[df_standardized['trade_date'] == date, features] = standardized_values\n", + "\n", + " return df_standardized\n", + "\n", + "\n", + "import numpy as np\n", + "import pandas as pd\n", + "import statsmodels.api as sm\n", + "\n", + "\n", + "def mad_filter(df, features, n=3):\n", + " \"\"\" 使用中位数绝对偏差（MAD）去极值 \"\"\"\n", + " df = df.copy()\n", + " for col in features:\n", + " median = df[col].median()\n", + " mad = np.median(np.abs(df[col] - median))\n", + " upper = median + n * mad\n", + " lower = median - n * mad\n", + " df[col] = np.clip(df[col], lower, upper) # 截断极值\n", + " return df\n", + "\n", + "\n", + "from concurrent.futures import ProcessPoolExecutor\n", + "\n", + "\n", + "def neutralize_manual(df, features, industry_col, mkt_cap_col):\n", + " \"\"\" 手动实现简单回归以提升速度 \"\"\"\n", + "\n", + " for col in features:\n", + " residuals = []\n", + " for _, group in df.groupby(industry_col):\n", + " if len(group) > 1:\n", + " x = np.log(group[mkt_cap_col]) # 市值对数\n", + " y = group[col] # 因子值\n", + " beta = np.cov(y, x)[0, 1] / np.var(x) # 计算斜率\n", + " alpha = np.mean(y) - beta * np.mean(x) # 计算截距\n", + " resid = y - (alpha + beta * x) # 计算残差\n", + " residuals.extend(resid)\n", + " else:\n", + " residuals.extend(group[col]) # 样本不足时保留原值\n", + "\n", + " df[col] = residuals\n", + "\n", + " return df\n", + "\n", + "\n", + "import gc\n", + "\n", + "gc.collect()" + ], + "outputs": [ + { + "data": { + "text/plain": [ + "0" + ] + }, + "execution_count": 11, + "metadata": {}, + "output_type": "execute_result" + } + ], + "execution_count": 11 + }, + { + "cell_type": "code", + "id": "1c46817a-b5dd-4bec-8bb4-e6e80bfd9d66", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:30.320706Z", + "start_time": "2025-03-29T17:46:30.307889Z" + } + }, + "source": "# print(test_data.head()[['act_factor1', 'act_factor2', 'ts_code', 'trade_date']])", + "outputs": [], + "execution_count": 12 + }, + { + "cell_type": "code", + "id": "da2bb202843d9275", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:30.929501Z", + "start_time": "2025-03-29T17:46:30.343347Z" + } + }, + "source": [ + "from sklearn.preprocessing import StandardScaler\n", + "import lightgbm as lgb\n", + "import matplotlib.pyplot as plt\n", + "from sklearn.decomposition import PCA\n", + "\n", + "\n", + "def train_light_model(train_data_df, params, feature_columns, callbacks, evals,\n", + " print_feature_importance=True, num_boost_round=100,\n", + " validation_days=180, use_pca=False, split_date=None): # 新增参数：validation_days\n", + " # 确保数据按时间排序\n", + " train_data_df = train_data_df.sort_values(by='trade_date')\n", + "\n", + " numeric_columns = train_data_df.select_dtypes(include=['float64', 'int64']).columns\n", + " numeric_columns = [col for col in numeric_columns if col in feature_columns]\n", + " # X_train.loc[:, numeric_columns] = scaler.fit_transform(X_train[numeric_columns])\n", + " # X_val.loc[:, numeric_columns] = scaler.transform(X_val[numeric_columns])\n", + " train_data_df = cross_sectional_standardization(train_data_df, numeric_columns)\n", + "\n", + " # 去除标签为空的样本\n", + " train_data_df = train_data_df.dropna(subset=['label'])\n", + " print('原始训练集大小: ', len(train_data_df))\n", + "\n", + " # 按时间顺序划分训练集和验证集\n", + " if split_date is None:\n", + " all_dates = train_data_df['trade_date'].unique() # 获取所有唯一的 trade_date\n", + " split_date = all_dates[-validation_days] # 划分点为倒数第 validation_days 天\n", + " train_data_split = train_data_df[train_data_df['trade_date'] < split_date] # 训练集\n", + " val_data_split = train_data_df[train_data_df['trade_date'] >= split_date] # 验证集\n", + "\n", + " # 打印划分结果\n", + " print(f\"划分后的训练集大小: {len(train_data_split)}, 验证集大小: {len(val_data_split)}\")\n", + "\n", + " # 提取特征和标签\n", + " X_train = train_data_split[feature_columns]\n", + " y_train = train_data_split['label']\n", + "\n", + " X_val = val_data_split[feature_columns]\n", + " y_val = val_data_split['label']\n", + "\n", + " # 标准化数值特征\n", + " scaler = StandardScaler()\n", + "\n", + "\n", + " # 计算每个 trade_date 内的样本数（LTR 需要 group 信息）\n", + " train_groups = train_data_split.groupby('trade_date').size().tolist()\n", + " val_groups = val_data_split.groupby('trade_date').size().tolist()\n", + "\n", + " # 处理类别特征\n", + " categorical_feature = [col for col in feature_columns if 'cat' in col]\n", + "\n", + " pca = None\n", + " if use_pca:\n", + " pca = PCA(n_components=0.95) # 或指定 n_components=固定值（如 10）\n", + " numeric_features = [col for col in feature_columns if col not in categorical_feature]\n", + " numeric_pca = pca.fit_transform(X_train[numeric_features])\n", + " X_train = pd.concat([pd.DataFrame(numeric_pca, index=X_train.index), X_train[categorical_feature]], axis=1)\n", + "\n", + " numeric_pca = pca.transform(X_val[numeric_features])\n", + " X_val = pd.concat([pd.DataFrame(numeric_pca, index=X_val.index), X_val[categorical_feature]], axis=1)\n", + "\n", + " # 计算权重（基于时间）\n", + " # trade_date = train_data_split['trade_date'] # 交易日期\n", + " # weights = (trade_date - trade_date.min()).dt.days / (trade_date.max() - trade_date.min()).days + 1\n", + " # weights = train_data_split.groupby('trade_date')['std_return_5'].transform(\n", + " # lambda x: x / x.mean()\n", + " # )\n", + " ud = sorted(train_data_split[\"trade_date\"].unique().tolist())\n", + " date_weights = {date: weight * weight for date, weight in zip(ud, np.linspace(1, 10, len(ud)))}\n", + " params['weight'] = train_data_split[\"trade_date\"].map(date_weights).tolist()\n", + "\n", + " train_dataset = lgb.Dataset(\n", + " X_train, label=y_train, group=train_groups,\n", + " categorical_feature=categorical_feature\n", + " )\n", + "\n", + " # weights = val_data_split.groupby('trade_date')['std_return_5'].transform(\n", + " # lambda x: x / x.mean()\n", + " # )\n", + " val_dataset = lgb.Dataset(\n", + " X_val, label=y_val, group=val_groups,\n", + " categorical_feature=categorical_feature\n", + " )\n", + "\n", + " # 训练模型\n", + " model = lgb.train(\n", + " params, train_dataset, num_boost_round=num_boost_round,\n", + " valid_sets=[train_dataset, val_dataset], valid_names=['train', 'valid'],\n", + " callbacks=callbacks\n", + " )\n", + "\n", + " # 打印特征重要性（如果需要）\n", + " if print_feature_importance:\n", + " lgb.plot_metric(evals)\n", + " lgb.plot_importance(model, importance_type='split', max_num_features=20)\n", + " plt.show()\n", + "\n", + " return model, scaler, pca" + ], + "outputs": [], + "execution_count": 13 + }, + { + "cell_type": "code", + "id": "20b7836efae720a3", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:31.021203Z", + "start_time": "2025-03-29T17:46:30.953273Z" + } + }, + "source": [ + "# print('train data size: ', len(train_data))\n", + "\n", + "# label_gain = list(range(len(train_data['label'].unique())))\n", + "# label_gain = [gain * gain for gain in label_gain]\n", + "light_params = {\n", + " 'objective': 'lambdarank',\n", + " 'metric': 'lambdarank',\n", + " 'learning_rate': 0.1,\n", + " 'num_leaves': 1024,\n", + " 'min_data_in_leaf': 128,\n", + " 'max_depth': 16,\n", + " 'max_bin': 1024,\n", + " 'feature_fraction': 0.7,\n", + " 'bagging_fraction': 1,\n", + " 'bagging_freq': 5,\n", + " 'lambda_l1': 1,\n", + " 'lambda_l2': 1,\n", + " # 'boosting': 'dart',\n", + " 'verbosity': -1,\n", + " 'extra_trees': True,\n", + " 'max_position': 5,\n", + " 'ndcg_at': 1,\n", + " 'seed': 7\n", + "}\n", + "evals = {}\n", + "\n", + "gc.collect()" + ], + "outputs": [ + { + "data": { + "text/plain": [ + "0" + ] + }, + "execution_count": 14, + "metadata": {}, + "output_type": "execute_result" + } + ], + "execution_count": 14 + }, + { + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:46:35.629560Z", + "start_time": "2025-03-29T17:46:31.046580Z" + } + }, + "cell_type": "code", + "source": [ + "\n", + "days = 2\n", + "df['future_return'] = (df.groupby('ts_code')['close'].shift(-days) - df.groupby('ts_code')['open'].shift(-1)) / \\\n", + " df.groupby('ts_code')['open'].shift(-1)\n", + "df['future_volatility'] = (\n", + " df.groupby('ts_code')['future_return']\n", + " .transform(lambda x: x.rolling(days).std())\n", + ")\n", + "\n", + "df['future_score'] = (\n", + " 0.7 * df['future_return'] +\n", + " 0.3 * df['future_volatility']\n", + ")\n", + "\n", + "filter_index = df['future_return'].between(df['future_return'].quantile(0.01), df['future_return'].quantile(0.99))\n", + "filter_index = df['future_volatility'].between(df['future_volatility'].quantile(0.01),\n", + " df['future_volatility'].quantile(0.99)) | filter_index\n", + "\n", + "df['label'] = df.groupby('trade_date', group_keys=False)['future_score'].transform(\n", + " lambda x: pd.qcut(x, q=50, labels=False, duplicates='drop')\n", + ")\n" + ], + "id": "81d4570663ae21d7", + "outputs": [], + "execution_count": 15 + }, + { + "cell_type": "code", + "id": "8f134d435f71e9e2", + "metadata": { + "jupyter": { + "source_hidden": true + }, + "ExecuteTime": { + "end_time": "2025-03-29T17:46:35.745784Z", + "start_time": "2025-03-29T17:46:35.675465Z" + } + }, + "source": [ + "gc.collect()\n", + "\n", + "\n", + "def rolling_train_predict(df, train_days, test_days, industry_df, index_df, days=5, use_pca=False, validation_days=60, filter_index=None):\n", + " # 1. 按照交易日期排序\n", + " unique_dates = df[df['trade_date'] >= '2020-01-01']['trade_date'].unique().tolist()\n", + " unique_dates = sorted(unique_dates)\n", + " n = len(unique_dates)\n", + "\n", + " # 2. 计算需要跳过的天数，使后续窗口对齐\n", + " extra_days = (n - train_days) % test_days\n", + " start_index = extra_days # 从此索引开始滚动\n", + "\n", + " predictions_list = []\n", + "\n", + " for start in range(start_index, n - train_days - test_days + 1, test_days):\n", + " gc.collect()\n", + "\n", + " train_dates = unique_dates[start: start + train_days]\n", + " test_dates = unique_dates[start + train_days: start + train_days + test_days]\n", + "\n", + " # 根据日期筛选数据\n", + " train_data = df[filter_index & df['trade_date'].isin(train_dates)]\n", + " test_data = df[filter_index & df['trade_date'].isin(test_dates)]\n", + "\n", + " train_data = train_data.sort_values('trade_date')\n", + " test_data = test_data.sort_values('trade_date')\n", + "\n", + " def select_pre_zt_stocks_dynamic(\n", + " stock_df,\n", + " ):\n", + " stock_df = stock_df.groupby('trade_date', group_keys=False).apply(\n", + " lambda x: x.nlargest(1000, 'return_20')\n", + " )\n", + " return stock_df\n", + "\n", + " train_data = select_pre_zt_stocks_dynamic(train_data)\n", + " test_data = select_pre_zt_stocks_dynamic(test_data)\n", + "\n", + " industry_df = industry_df.sort_values(by=['trade_date'])\n", + " # index_df = index_df.sort_values(by=['trade_date'])\n", + "\n", + " train_data = train_data.merge(industry_df, on=['cat_l2_code', 'trade_date'], how='left')\n", + " # train_data = train_data.merge(index_df, on='trade_date', how='left')\n", + " test_data = test_data.merge(industry_df, on=['cat_l2_code', 'trade_date'], how='left')\n", + " # test_data = test_data.merge(index_df, on='trade_date', how='left')\n", + "\n", + " train_data, test_data = train_data.replace([np.inf, -np.inf], np.nan), test_data.replace([np.inf, -np.inf],\n", + " np.nan)\n", + "\n", + " feature_columns = [col for col in train_data.columns if col in train_data.columns]\n", + " feature_columns = [col for col in feature_columns if col not in ['trade_date',\n", + " 'ts_code',\n", + " 'label']]\n", + " feature_columns = [col for col in feature_columns if 'future' not in col]\n", + " feature_columns = [col for col in feature_columns if 'label' not in col]\n", + " feature_columns = [col for col in feature_columns if 'score' not in col]\n", + " feature_columns = [col for col in feature_columns if 'gen' not in col]\n", + " feature_columns = [col for col in feature_columns if 'cat_l2_code' not in col]\n", + " feature_columns = [col for col in feature_columns if col not in origin_columns]\n", + " feature_columns = [col for col in feature_columns if not col.startswith('_')]\n", + "\n", + " numeric_columns = train_data.select_dtypes(include=['float64', 'int64']).columns\n", + " numeric_columns = [col for col in numeric_columns if col in feature_columns]\n", + " print('去极值')\n", + " train_data = mad_filter(train_data, numeric_columns) # 去极值\n", + " # print('中性化')\n", + " # train_data = neutralize_manual(train_data, numeric_columns, industry_col='cat_l2_code', mkt_cap_col='log(circ_mv)') # 中性化\n", + " print('去极值')\n", + " test_data = mad_filter(test_data, numeric_columns) # 去极值\n", + "\n", + " all_dates = train_data['trade_date'].unique() # 获取所有唯一的 trade_date\n", + " split_date = all_dates[-validation_days] # 划分点为倒数第 validation_days 天\n", + " train_data_split = train_data[train_data['trade_date'] < split_date] # 训练集\n", + " val_data_split = train_data[train_data['trade_date'] >= split_date] # 验证集\n", + "\n", + " feature_columns, _ = remove_shifted_features(train_data_split[train_data_split['label'] == train_data_split['label'].max()],\n", + " val_data_split[val_data_split['label'] == val_data_split['label'].max()],\n", + " feature_columns)\n", + "\n", + " feature_columns = remove_highly_correlated_features(train_data[train_data['label'] == train_data['label'].max()],\n", + " feature_columns)\n", + "\n", + " train_data = train_data.dropna(subset=feature_columns)\n", + " train_data = train_data.dropna(subset=['label'])\n", + " train_data = train_data.reset_index(drop=True)\n", + "\n", + " # print(test_data.tail())\n", + " test_data = test_data.dropna(subset=feature_columns)\n", + " # test_data = test_data.dropna(subset=['label'])\n", + " test_data = test_data.reset_index(drop=True)\n", + "\n", + " # print(len(train_data))\n", + " print(f\"最小日期: {train_data['trade_date'].min().strftime('%Y-%m-%d')}\")\n", + " print(f\"最大日期: {train_data['trade_date'].max().strftime('%Y-%m-%d')}\")\n", + " # print(len(test_data))\n", + " print(f\"最小日期: {test_data['trade_date'].min().strftime('%Y-%m-%d')}\")\n", + " print(f\"最大日期: {test_data['trade_date'].max().strftime('%Y-%m-%d')}\")\n", + "\n", + " cat_columns = [col for col in df.columns if col.startswith('cat')]\n", + " for col in cat_columns:\n", + " train_data[col] = train_data[col].astype('category')\n", + " test_data[col] = test_data[col].astype('category')\n", + "\n", + " label_gain = list(range(len(train_data['label'].unique())))\n", + " label_gain = [gain * gain for gain in label_gain]\n", + " light_params['label_gain'] = label_gain\n", + "\n", + " # ud = train_data[\"trade_date\"].unique()\n", + " # date_weights = {date: weight for date, weight in zip(ud, np.linspace(1, 2, len(unique_dates)))}\n", + " # light_params['weight'] = train_data[\"trade_date\"].map(date_weights).tolist()\n", + "\n", + " # print(f'feature_columns: {feature_columns}')\n", + " feature_contri = [2 if feat.startswith('act_factor') else 1 for feat in feature_columns]\n", + " light_params['feature_contri'] = feature_contri\n", + " model, _, _ = train_light_model(train_data.dropna(subset=['label']),\n", + " light_params, feature_columns,\n", + " [lgb.log_evaluation(period=100),\n", + " lgb.callback.record_evaluation(evals),\n", + " lgb.early_stopping(50, first_metric_only=True)\n", + " ], evals,\n", + " num_boost_round=3000, validation_days=validation_days,\n", + " print_feature_importance=False, use_pca=False)\n", + "\n", + " score_df = test_data.copy()\n", + " numeric_columns = score_df.select_dtypes(include=['float64', 'int64']).columns\n", + " numeric_columns = [col for col in numeric_columns if col in feature_columns]\n", + " score_df = cross_sectional_standardization(score_df, numeric_columns)\n", + " score_df['score'] = model.predict(score_df[feature_columns])\n", + " score_df = score_df.loc[score_df.groupby('trade_date')['score'].idxmax()]\n", + " score_df = score_df[['trade_date', 'score', 'ts_code']]\n", + " predictions_list.append(score_df)\n", + "\n", + " # m = 5\n", + " # all_data = []\n", + " # for i, trade_date in enumerate(sorted(score_df['trade_date'].unique().tolist())):\n", + " # # 提取当前日期的数据\n", + " # current_data = score_df[score_df['trade_date'] == trade_date]\n", + " # all_data.append(current_data)\n", + " #\n", + " # numeric_columns = [col for col in feature_columns if col in current_data.select_dtypes(include=['float64', 'int64']).columns]\n", + " # current_data = cross_sectional_standardization(current_data, numeric_columns)\n", + " # current_data['score'] = model.predict(current_data[feature_columns])\n", + " # daily_top_score = current_data.loc[[current_data['score'].idxmax()]]\n", + " # predictions_list.append(daily_top_score[['trade_date', 'score', 'ts_code']])\n", + " #\n", + " # if i % m == 0:\n", + " # train_data_split = pd.concat(all_data)\n", + " # train_data_split = train_data_split.dropna(subset=['label'])\n", + " #\n", + " # X_train = train_data_split[feature_columns]\n", + " # y_train = train_data_split['label']\n", + " #\n", + " # train_groups = train_data_split.groupby('trade_date').size().tolist()\n", + " # categorical_feature = [col for col in feature_columns if 'cat' in col]\n", + " #\n", + " # train_dataset = lgb.Dataset(\n", + " # X_train, label=y_train, group=train_groups,\n", + " # categorical_feature=categorical_feature\n", + " # )\n", + " #\n", + " # model = lgb.train(\n", + " # light_params, train_dataset, num_boost_round=36,\n", + " # init_model=model\n", + " # )\n", + " # all_data = []\n", + "\n", + " final_predictions = pd.concat(predictions_list, ignore_index=True)\n", + " return final_predictions\n" + ], + "outputs": [], + "execution_count": 16 + }, + { + "cell_type": "code", + "id": "63235069-dc59-48fb-961a-e80373e41a61", + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:52:27.309071Z", + "start_time": "2025-03-29T17:46:35.756531Z" + } + }, + "source": [ + "\n", + "gc.collect()\n", + "\n", + "final_predictions = rolling_train_predict(df.sort_values(['trade_date'], ascending=[True]), 500, 60, industry_df, None,\n", + " days=days, validation_days=120, filter_index=filter_index)\n", + "final_predictions.to_csv('predictions_test.tsv', index=False)\n" + ], + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "去极值\n", + "去极值\n", + "检测到 20 个可能漂移的特征: ['pct_chg', 'turnover_rate', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'act_factor3', 'log(circ_mv)', 'cov', 'delta_cov', 'alpha_22_improved', 'turnover_std', 'log_close', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small', 'buy_lg_vol_minus_sell_lg_vol', 'buy_elg_vol_minus_sell_elg_vol', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2020-03-11\n", + "最大日期: 2022-03-30\n", + "最小日期: 2022-03-31\n", + "最大日期: 2022-06-30\n", + "原始训练集大小: 402534\n", + "划分后的训练集大小: 307694, 验证集大小: 94840\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[15]\ttrain's ndcg@1: 0.662414\tvalid's ndcg@1: 0.613612\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 26 个可能漂移的特征: ['vol', 'pct_chg', 'turnover_rate', 'return_kurtosis', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'rsi_3', 'act_factor3', 'log(circ_mv)', 'cov', 'delta_cov', 'alpha_22_improved', 'turnover_std', 'resonance_factor', 'log_close', 'obv-maobv_6', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small', 'buy_lg_vol_minus_sell_lg_vol', 'buy_elg_vol_minus_sell_elg_vol', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2020-06-09\n", + "最大日期: 2022-06-30\n", + "最小日期: 2022-07-01\n", + "最大日期: 2022-09-23\n", + "原始训练集大小: 400974\n", + "划分后的训练集大小: 306295, 验证集大小: 94679\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[29]\ttrain's ndcg@1: 0.705891\tvalid's ndcg@1: 0.608979\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 14 个可能漂移的特征: ['turnover_rate', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'log(circ_mv)', 'delta_cov', 'alpha_22_improved', 'turnover_std', 'log_close', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small']\n", + "最小日期: 2020-09-03\n", + "最大日期: 2022-09-23\n", + "最小日期: 2022-09-26\n", + "最大日期: 2022-12-23\n", + "原始训练集大小: 398352\n", + "划分后的训练集大小: 303767, 验证集大小: 94585\n", + "Training until validation scores don't improve for 50 rounds\n", + "[100]\ttrain's ndcg@1: 0.827415\tvalid's ndcg@1: 0.630578\n", + "Early stopping, best iteration is:\n", + "[71]\ttrain's ndcg@1: 0.78919\tvalid's ndcg@1: 0.66253\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 18 个可能漂移的特征: ['vol', 'turnover_rate', 'return_kurtosis', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'log(circ_mv)', 'turnover_std', 'log_close', 'obv-maobv_6', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small', 'buy_elg_vol_minus_sell_elg_vol', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2020-12-04\n", + "最大日期: 2022-12-23\n", + "最小日期: 2022-12-26\n", + "最大日期: 2023-03-27\n", + "原始训练集大小: 395407\n", + "划分后的训练集大小: 305189, 验证集大小: 90218\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[49]\ttrain's ndcg@1: 0.808651\tvalid's ndcg@1: 0.637922\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 19 个可能漂移的特征: ['vol', 'pct_chg', 'turnover_rate', 'return_skew', 'return_kurtosis', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'rsi_3', 'act_factor3', 'resonance_factor', 'obv-maobv_6', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2021-03-08\n", + "最大日期: 2023-03-27\n", + "最小日期: 2023-03-28\n", + "最大日期: 2023-06-27\n", + "原始训练集大小: 393886\n", + "划分后的训练集大小: 303266, 验证集大小: 90620\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[8]\ttrain's ndcg@1: 0.622492\tvalid's ndcg@1: 0.606022\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 23 个可能漂移的特征: ['vol', 'pct_chg', 'turnover_rate', 'return_skew', 'return_kurtosis', 'vol_spike', 'atr_6', 'obv', 'maobv_6', 'rsi_3', 'act_factor3', 'log(circ_mv)', 'delta_cov', 'alpha_22_improved', 'turnover_std', 'resonance_factor', 'log_close', 'obv-maobv_6', 'active_buy_volume_large', 'active_buy_volume_big', 'active_buy_volume_small', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2021-06-04\n", + "最大日期: 2023-06-27\n", + "最小日期: 2023-06-28\n", + "最大日期: 2023-09-19\n", + "原始训练集大小: 393201\n", + "划分后的训练集大小: 300541, 验证集大小: 92660\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[3]\ttrain's ndcg@1: 0.623432\tvalid's ndcg@1: 0.604043\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 14 个可能漂移的特征: ['vol', 'pct_chg', 'return_kurtosis', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'act_factor3', 'log(circ_mv)', 'cov', 'alpha_22_improved', 'resonance_factor', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2021-08-30\n", + "最大日期: 2023-09-19\n", + "最小日期: 2023-09-20\n", + "最大日期: 2023-12-20\n", + "原始训练集大小: 386612\n", + "划分后的训练集大小: 296498, 验证集大小: 90114\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[12]\ttrain's ndcg@1: 0.663139\tvalid's ndcg@1: 0.652339\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 14 个可能漂移的特征: ['pct_chg', 'turnover_rate', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'act_factor3', 'log(circ_mv)', 'cov', 'alpha_22_improved', 'turnover_std', 'resonance_factor', 'log_close']\n", + "最小日期: 2021-12-01\n", + "最大日期: 2023-12-20\n", + "最小日期: 2023-12-21\n", + "最大日期: 2024-03-22\n", + "原始训练集大小: 379352\n", + "划分后的训练集大小: 293416, 验证集大小: 85936\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[5]\ttrain's ndcg@1: 0.634019\tvalid's ndcg@1: 0.638831\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 12 个可能漂移的特征: ['vol', 'turnover_rate', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'log(circ_mv)', 'cov', 'log_close', 'obv-maobv_6', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2022-03-03\n", + "最大日期: 2024-03-22\n", + "最小日期: 2024-03-25\n", + "最大日期: 2024-06-24\n", + "原始训练集大小: 379932\n", + "划分后的训练集大小: 290249, 验证集大小: 89683\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[34]\ttrain's ndcg@1: 0.759349\tvalid's ndcg@1: 0.64614\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 18 个可能漂移的特征: ['vol', 'pct_chg', 'turnover_rate', 'vol_spike', 'atr_6', 'obv', 'maobv_6', 'rsi_3', 'act_factor3', 'cov', 'delta_cov', 'alpha_22_improved', 'turnover_std', 'resonance_factor', 'log_close', 'obv-maobv_6', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2022-06-02\n", + "最大日期: 2024-06-24\n", + "最小日期: 2024-06-25\n", + "最大日期: 2024-09-18\n", + "原始训练集大小: 381303\n", + "划分后的训练集大小: 284860, 验证集大小: 96443\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[40]\ttrain's ndcg@1: 0.745108\tvalid's ndcg@1: 0.589855\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 21 个可能漂移的特征: ['vol', 'pct_chg', 'turnover_rate', 'return_kurtosis', 'vol_spike', 'atr_14', 'atr_6', 'obv', 'maobv_6', 'rsi_3', 'act_factor1', 'act_factor3', 'log(circ_mv)', 'cov', 'delta_cov', 'alpha_22_improved', 'resonance_factor', 'log_close', 'obv-maobv_6', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2022-08-26\n", + "最大日期: 2024-09-18\n", + "最小日期: 2024-09-19\n", + "最大日期: 2024-12-18\n", + "原始训练集大小: 379192\n", + "划分后的训练集大小: 285332, 验证集大小: 93860\n", + "Training until validation scores don't improve for 50 rounds\n", + "Early stopping, best iteration is:\n", + "[10]\ttrain's ndcg@1: 0.629259\tvalid's ndcg@1: 0.646921\n", + "Evaluated only: ndcg@1\n", + "去极值\n", + "去极值\n", + "检测到 18 个可能漂移的特征: ['vol', 'pct_chg', 'turnover_rate', 'return_skew', 'return_kurtosis', 'obv', 'maobv_6', 'rsi_3', 'act_factor1', 'log(circ_mv)', 'cov', 'delta_cov', 'alpha_22_improved', 'resonance_factor', 'log_close', 'obv-maobv_6', 'mv_adjusted_volume', 'nonlinear_mv_volume']\n", + "最小日期: 2022-11-28\n", + "最大日期: 2024-12-18\n", + "最小日期: 2024-12-19\n", + "最大日期: 2025-03-19\n", + "原始训练集大小: 371676\n", + "划分后的训练集大小: 289037, 验证集大小: 82639\n", + "Training until validation scores don't improve for 50 rounds\n", + "[100]\ttrain's ndcg@1: 0.844744\tvalid's ndcg@1: 0.580519\n", + "Early stopping, best iteration is:\n", + "[54]\ttrain's ndcg@1: 0.792906\tvalid's ndcg@1: 0.599084\n", + "Evaluated only: ndcg@1\n" + ] + } + ], + "execution_count": 17 + }, + { + "metadata": { + "ExecuteTime": { + "end_time": "2025-03-29T17:52:27.393208Z", + "start_time": "2025-03-29T17:52:27.388403Z" + } + }, + "cell_type": "code", + "source": "print('finish')", + "id": "10f15e935aa02a34", + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "finish\n" + ] + } + ], + "execution_count": 18 + }, + { + "cell_type": "code", + "id": "d86af99d15cb3bdd", + "metadata": { + "scrolled": true + }, + "source": [ + "import pandas as pd\n", + "\n", + "gc.collect()\n", + "def rolling_train_predict(df, train_days, test_days, industry_df, index_df, days=5, use_pca=False, validation_days=60):\n", + "\n", + " # 1. 按照交易日期排序\n", + " unique_dates = df[df['trade_date'] >= '2020-01-01']['trade_date'].unique().tolist()\n", + " unique_dates = sorted(unique_dates)\n", + " n = len(unique_dates)\n", + " \n", + " # 2. 计算需要跳过的天数，使后续窗口对齐\n", + " extra_days = (n - train_days) % test_days \n", + " start_index = extra_days # 从此索引开始滚动\n", + " \n", + " predictions_list = []\n", + "\n", + "\n", + " for start in range(start_index, n - train_days - test_days + 1, test_days):\n", + " gc.collect()\n", + "\n", + " train_dates = unique_dates[start : start + train_days]\n", + " test_dates = unique_dates[start + train_days : start + train_days + test_days]\n", + "\n", + " # 根据日期筛选数据\n", + " train_data = df[df['trade_date'].isin(train_dates)]\n", + " test_data = df[df['trade_date'].isin(test_dates)]\n", + "\n", + " train_data = train_data.sort_values('trade_date')\n", + " test_data = test_data.sort_values('trade_date')\n", + "\n", + " \n", + " def select_pre_zt_stocks_dynamic(\n", + " stock_df,\n", + " vol_spike_multiplier=1.5,\n", + " min_return=0.03, # 最小累计涨幅（例如 3%）\n", + " min_main_net_inflow=1e6, # 最小主力资金净流入（例如 100 万元）\n", + " window=30, # 计算历史均值的窗口大小\n", + " signal_days=1 # 异动信号需要连续出现的天数\n", + " ):\n", + " \n", + " # 排序数据\n", + " stock_df = stock_df.sort_values(by=['trade_date', 'ts_code'])\n", + " \n", + " # stock_df = stock_df[\n", + " # (stock_df['vol'] > vol_spike_multiplier * stock_df['avg_vol_20'])\n", + " # ]\n", + " cd1 = stock_df[\"close\"] > stock_df[\"close\"].shift(1)\n", + "\n", + " cd2 = stock_df[\"close\"] > stock_df[\"close\"].rolling(window=10).mean()\n", + "\n", + " cd3 = (stock_df[\"vol\"] > stock_df[\"vol\"].shift(1)) & (stock_df[\"vol\"] < 10 * stock_df[\"vol\"].shift(1))\n", + "\n", + " stock_df = stock_df[cd1 & cd2 & cd3]\n", + " stock_df = stock_df.groupby('trade_date', group_keys=False).apply(\n", + " lambda x: x.nlargest(1000, 'return_20')\n", + " )\n", + " \n", + " return stock_df\n", + " \n", + " train_data = select_pre_zt_stocks_dynamic(train_data)\n", + " test_data = select_pre_zt_stocks_dynamic(test_data)\n", + "\n", + " \n", + " # train_data, _ = get_simple_factor(train_data)\n", + " # test_data, _ = get_simple_factor(test_data)\n", + "\n", + " # df['future_return'] = (df.groupby('ts_code')['close'].shift(-days) - df.groupby('ts_code')['open'].shift(-1)) / \\\n", + " # df.groupby('ts_code')['open'].shift(-1)\n", + " \n", + " def symmetric_log_transform(values):\n", + " return np.sign(values) * np.log1p(np.abs(values))\n", + "\n", + " train_data['future_return'] = train_data.groupby('ts_code', group_keys=False)['close'].apply(lambda x: x.shift(-days) / x - 1)\n", + " train_data['future_score'] = calculate_score(train_data, days=days, lambda_param=0.3)\n", + " # train_data['future_score'] = symmetric_log_transform(train_data['future_score'])\n", + "\n", + " test_data['future_return'] = test_data.groupby('ts_code', group_keys=False)['close'].apply(lambda x: x.shift(-days) / x - 1)\n", + " test_data['future_score'] = calculate_score(test_data, days=days, lambda_param=0.3)\n", + " # test_data['future_score'] = symmetric_log_transform(test_data['future_score'])\n", + " \n", + " train_data['label'] = train_data.groupby('trade_date', group_keys=False)['future_score'].transform(\n", + " lambda x: pd.qcut(x, q=10, labels=False, duplicates='drop')\n", + " )\n", + " test_data['label'] = test_data.groupby('trade_date', group_keys=False)['future_score'].transform(\n", + " lambda x: pd.qcut(x, q=10, labels=False, duplicates='drop')\n", + " )\n", + " \n", + " industry_df = industry_df.sort_values(by=['trade_date'])\n", + " index_df = index_df.sort_values(by=['trade_date'])\n", + " \n", + " train_data = train_data.merge(industry_df, on=['cat_l2_code', 'trade_date'], how='left')\n", + " # train_data = train_data.merge(index_df, on='trade_date', how='left')\n", + " test_data = test_data.merge(industry_df, on=['cat_l2_code', 'trade_date'], how='left')\n", + " # test_data = test_data.merge(index_df, on='trade_date', how='left')\n", + " \n", + " train_data, test_data = train_data.replace([np.inf, -np.inf], np.nan), test_data.replace([np.inf, -np.inf], np.nan)\n", + " \n", + " feature_columns = [col for col in train_data.columns if col not in ['trade_date',\n", + " 'ts_code',\n", + " 'label']]\n", + " feature_columns = [col for col in feature_columns if 'future' not in col]\n", + " feature_columns = [col for col in feature_columns if 'score' not in col]\n", + " feature_columns = [col for col in feature_columns if col not in origin_columns]\n", + " feature_columns = [col for col in feature_columns if not col.startswith('_')]\n", + " # print(feature_columns)\n", + "\n", + " feature_columns_o = feature_columns[:]\n", + " train_data, feature_columns = create_deviation_within_dates(train_data, feature_columns_o)\n", + " test_data, _ = create_deviation_within_dates(test_data, feature_columns_o)\n", + " print(f'feature_columns size: {len(feature_columns)}')\n", + " \n", + " train_data = train_data.dropna(subset=feature_columns)\n", + " train_data = train_data.dropna(subset=['label'])\n", + " train_data = train_data.reset_index(drop=True)\n", + " \n", + " # print(test_data.tail())\n", + " # test_data = test_data.dropna(subset=feature_columns_new)\n", + " # test_data = test_data.dropna(subset=['label'])\n", + " test_data = test_data.reset_index(drop=True)\n", + " \n", + " # print(len(train_data))\n", + " print(f\"最小日期: {train_data['trade_date'].min().strftime('%Y-%m-%d')}\")\n", + " print(f\"最大日期: {train_data['trade_date'].max().strftime('%Y-%m-%d')}\")\n", + " # print(len(test_data))\n", + " print(f\"最小日期: {test_data['trade_date'].min().strftime('%Y-%m-%d')}\")\n", + " print(f\"最大日期: {test_data['trade_date'].max().strftime('%Y-%m-%d')}\")\n", + " \n", + " cat_columns = [col for col in df.columns if col.startswith('cat')]\n", + " for col in cat_columns:\n", + " train_data[col] = train_data[col].astype('category')\n", + " test_data[col] = test_data[col].astype('category')\n", + "\n", + "\n", + " feature_columns = remove_highly_correlated_features(train_data[train_data['label'] == 9], feature_columns)\n", + " feature_columns, _ = remove_shifted_features(train_data[train_data['label'] == 9], test_data[test_data['label'] == 9], feature_columns)\n", + " keep_columns = [col for col in train_data.columns if\n", + " col in feature_columns or col in ['ts_code', 'trade_date', 'label', 'future_return', 'future_score']]\n", + " train_data = train_data[keep_columns]\n", + "\n", + " label_gain = list(range(len(train_data['label'].unique())))\n", + " label_gain = [gain * 2 for gain in label_gain]\n", + " light_params['label_gain'] = label_gain\n", + " \n", + " ud = train_data[\"trade_date\"].unique()\n", + " date_weights = {date: weight for date, weight in zip(ud, np.linspace(1, 2, len(unique_dates)))}\n", + " light_params['weight'] = train_data[\"trade_date\"].map(date_weights).tolist()\n", + "\n", + " print(f'feature_columns: {feature_columns}')\n", + " model, scaler, pca = train_light_model(train_data.dropna(subset=['label']),\n", + " light_params, feature_columns,\n", + " [lgb.log_evaluation(period=100),\n", + " lgb.callback.record_evaluation(evals),\n", + " lgb.early_stopping(50, first_metric_only=True)\n", + " ], evals,\n", + " num_boost_round=3000, validation_days=validation_days,\n", + " print_feature_importance=False, use_pca=False)\n", + "\n", + " score_df = test_data.copy()\n", + " numeric_columns = score_df.select_dtypes(include=['float64', 'int64']).columns\n", + " numeric_columns = [col for col in numeric_columns if col in feature_columns]\n", + " score_df.loc[:, numeric_columns] = scaler.transform(score_df[numeric_columns])\n", + " if use_pca and pca is not None:\n", + " score_df.loc[:, numeric_columns] = pca.transform(score_df[numeric_columns])\n", + " score_df['score'] = model.predict(score_df[feature_columns])\n", + " # train_data['score'] = catboost_model.predict(train_data[feature_columns_new])\n", + " score_df = score_df.loc[score_df.groupby('trade_date')['score'].idxmax()]\n", + " # score_df = score_df[score_df['score'] > 0]\n", + " score_df = score_df[['trade_date', 'score', 'ts_code']]\n", + " predictions_list.append(score_df)\n", + " final_predictions = pd.concat(predictions_list, ignore_index=True)\n", + " return final_predictions\n", + "\n", + "\n", + "final_predictions = rolling_train_predict(df.sort_values(['trade_date'], ascending=[True]), 500, 60, industry_df, index_df, days=5, validation_days=100)\n", + "final_predictions.to_csv('predictions_test.tsv', index=False)\n" + ], + "outputs": [], + "execution_count": null + }, + { + "cell_type": "code", + "id": "7ed645f2-7755-496e-8a6d-c64adc9080ac", + "metadata": {}, + "source": [ + "print('finish')" + ], + "outputs": [], + "execution_count": null + }, + { + "cell_type": "code", + "id": "0dc75517-c857-4f1d-8815-e807400a6d33", + "metadata": {}, + "source": [], + "outputs": [], + "execution_count": null + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3 (ipykernel)", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.11.11" + } + }, + "nbformat": 4, + "nbformat_minor": 5 +}