1108.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "#导入依赖包\n",
    "%matplotlib inline\n",
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "from sklearn.metrics import mean_absolute_error\n",
    "from scipy.spatial.distance import pdist, squareform"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "#导入所有数据，比赛中由于有训练赛的存在，可以先导入data2文件夹中200天基金数据，做快速检验\n",
    "#data1文件夹放的是复赛的数据，data2文件夹放的是初赛的数据\n",
    "PATH='data1'\n",
    "trading_date=pd.read_csv(PATH+'/trading_date.csv',header=None)\n",
    "\n",
    "train_correlation=pd.read_csv(PATH+'/train_correlation.csv',index_col=0)\n",
    "train_fund_benchmark_return=pd.read_csv(PATH+'/train_fund_benchmark_return.csv',index_col=0)\n",
    "train_fund_return=pd.read_csv(PATH+'/train_fund_return.csv',index_col=0)\n",
    "train_index_return=pd.read_csv(PATH+'/train_index_return.csv',encoding='gbk',index_col=0)\n",
    "\n",
    "test_correlation=pd.read_csv(PATH+'/test_correlation.csv',index_col=0)\n",
    "test_fund_benchmark_return=pd.read_csv(PATH+'/test_fund_benchmark_return.csv',index_col=0)\n",
    "test_fund_return=pd.read_csv(PATH+'/test_fund_return.csv',index_col=0)\n",
    "test_index_return=pd.read_csv(PATH+'/test_index_return.csv',encoding='gbk',index_col=0)\n",
    "#合并train和test文件\n",
    "test_correlation=pd.concat([train_correlation.iloc[:,-100:],test_correlation],axis=1)\n",
    "test_fund_benchmark_return=pd.concat([train_fund_benchmark_return.iloc[:,-100:],test_fund_benchmark_return],axis=1)\n",
    "test_fund_return=pd.concat([train_fund_return.iloc[:,-100:],test_fund_return],axis=1)\n",
    "test_index_return=pd.concat([train_index_return.iloc[:,-100:],test_index_return],axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# #根据缺少的假期补全，保证每周有5天交易日，方便提取每周特征，但是没有提升\n",
    "# for i in pd.date_range('2015-09-30','2017-12-14'):\n",
    "#     if str(i.date()) not in test_correlation.columns and i.weekday()<5:\n",
    "#         test_correlation[i.date().isoformat()]=np.nan\n",
    "# test_correlation=test_correlation[sorted(test_correlation.columns.tolist())]\n",
    "# test_correlation=test_correlation.fillna(method='ffill', axis=1).fillna(method='bfill', axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# test_index_return=test_index_return.iloc[[0,2,4,17,18,19,21,22,25,26,-8,-7,-6,-5,-3,-2,-1],:]\n",
    "#筛选出沪深300、中证全债和商品指数3个重要指数\n",
    "test_index_return=test_index_return.iloc[[0,-7,-1],:]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(3, 300)"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_index_return.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "#utility\n",
    "def cal_error(l1,l2):\n",
    "\tnp.abs(l1-l2)\n",
    "def c_indexs(d,df,indexs): \n",
    "    d={}\n",
    "    for i in indexs:\n",
    "        for j in indexs:\n",
    "            d['%s-%s' % (i,j)]=np.abs(df.loc[i,:]-df.loc[j,:])\n",
    "    return d\n",
    "\n",
    "def select_df(start,len):\n",
    "\treturn df.loc[:,start:].iloc[:,:len]\n",
    "\n",
    "def generate_dateset(dfs,static_indexs):\n",
    "    r=dfs[0]\n",
    "    for df in dfs[1:]:\n",
    "#         use index merge data\n",
    "        pass\n",
    "    # add static_indexs\n",
    "    return r\n",
    "\n",
    "#给定窗口的df的相关系数\n",
    "#这个还有更简单的代码，但是为了保证万一产生的误差，提交的这版本就没用到c_indexs_corr_\n",
    "def c_indexs_corr(df,windows=[3,7,15,30,61]):\n",
    "    l=[]\n",
    "    for w in windows:\n",
    "        for m in ['pearson']:#, 'spearman']:\n",
    "            tmp=df.iloc[:,-w:].copy()\n",
    "            co=tmp.T.corr(method=m)\n",
    "            d={}\n",
    "            for i in co.index:\n",
    "                for j in co.columns:\n",
    "                    d['%s-%s' % (i,j)]=co.loc[i,j]\n",
    "            d=pd.Series(d)\n",
    "            d=d.reset_index()\n",
    "            l.append(d)\n",
    "    return l\n",
    "#备用，优化代码\n",
    "def c_indexs_corr_(df,windows=[3,7,15,30,61]):\n",
    "    l=[]\n",
    "    for w in windows:\n",
    "        for m in ['pearson']:#, 'spearman']:\n",
    "            tmp=df.iloc[:,-w:].copy()\n",
    "            l.append(1-pdist(tmp,metric='correlation'))\n",
    "    return pd.DataFrame(l).T\n",
    "\n",
    "#给定窗口的指定指数的求和、标准差，这个特征没有用，都是常数，但是没有删掉，特征重要性为0\n",
    "def index_status(df,windows=[1,5,10,30,60]):\n",
    "    l=[]\n",
    "    for w in windows:\n",
    "        l.append(df.iloc[0,-w:].sum())\n",
    "#         l.append(df.iloc[1,-w:].sum())\n",
    "#         l.append(df.iloc[2,-w:].sum())\n",
    "    for w in windows:\n",
    "        if w!=1:\n",
    "            l.append(df.iloc[0,-w:].std())\n",
    "#             l.append(df.iloc[1,-w:].std())\n",
    "#             l.append(df.iloc[2,-w:].std())\n",
    "    r=[]\n",
    "    for i in l:\n",
    "        if PATH=='data1':\n",
    "            r.append(np.ones(261003)*i)\n",
    "        if PATH=='data2':\n",
    "            r.append(np.ones(19900)*i)\n",
    "    return pd.DataFrame(r).T"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def find_cluster_distance(df):\n",
    "    from sklearn.cluster import AffinityPropagation\n",
    "    clustering = AffinityPropagation(damping=0.83).fit((df+1).cumprod(axis=1))\n",
    "    clus_m=pd.DataFrame({'class':clustering.labels_})\n",
    "    tmp=pd.DataFrame(clustering.cluster_centers_).reset_index()\n",
    "    clus_m=pd.merge(left=clus_m,right=tmp,left_on='class',right_on='index')\n",
    "    clus_m.drop(['index'],inplace=True,axis=1)\n",
    "    distances = pdist(clus_m.iloc[:,1:].values, metric='euclidean')\n",
    "    class_distances = pdist(pd.concat([clus_m.iloc[:,0],clus_m.iloc[:,0]],axis=1).values, metric='euclidean')\n",
    "    result=pd.DataFrame({'AP_dis':distances,'AP_C':class_distances})\n",
    "    result['AP_C']=result['AP_C'].map(lambda x:True if x==0 else False)\n",
    "    return result"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.linear_model import RidgeCV\n",
    "def index_coef(index_return,fund_return):\n",
    "    x=index_return.T\n",
    "    l=[]\n",
    "    for i in range(len(fund_return)):\n",
    "        ll=[]\n",
    "        y=fund_return.iloc[i,:]\n",
    "        regr=RidgeCV(cv=5)\n",
    "        regr.fit(x,y)\n",
    "        ll.extend(regr.coef_.tolist())\n",
    "        ll.append(regr.intercept_)\n",
    "        l.append(ll)\n",
    "    fund_index_coef=pd.DataFrame(l)\n",
    "    #correlation\n",
    "    distances = pdist(fund_index_coef.values, metric='correlation')\n",
    "    f1=1-distances\n",
    "#     #mahalanobis\n",
    "#     distances = pdist(fund_index_coef.values, metric='mahalanobis')\n",
    "#     f2=1-distances\n",
    "#     print('sdf')\n",
    "#     #euclidean\n",
    "#     distances = pdist(fund_index_coef.values, metric='euclidean')\n",
    "#     f3=distances\n",
    "    return f1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "# (test_fund_benchmark_return+1).cumprod().std()\n",
    "# test_fund_benchmark_return.std(axis=1)\n",
    "### 特征：比较基准（普通形式和累积形式）在一段时间(t)的标准差，反应出一段时间内基金操作的稳定性\n",
    "def get_cn2_values_by_n(series):\n",
    "    out=[]\n",
    "    for i in range(len(series)):\n",
    "        for j in [ii for ii in range(len(series)) if ii>i]:\n",
    "            out.append((series[i]-series[j])**2)\n",
    "    return out\n",
    "def get_return_std_corr(df,windows=[30,60,120]):\n",
    "    l=[]\n",
    "    for w in windows:\n",
    "        tmp=df.iloc[:,-w:].copy()\n",
    "        l.append(get_cn2_values_by_n(tmp.std(axis=1)))\n",
    "        l.append(get_cn2_values_by_n((tmp+1).cumprod().std(axis=1)))\n",
    "    return l"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "#计算correlation的时间窗口和滑动时间窗口的特征\n",
    "def f_corr(test_correlation,valid_period=16,off_set=0):\n",
    "    if off_set==0:\n",
    "        test_correlation_x=test_correlation.iloc[:,-valid_period*5:]\n",
    "    else:\n",
    "        test_correlation_x=test_correlation.iloc[:,-valid_period*5-off_set:-off_set]\n",
    "    feature_map=test_correlation_x.reset_index()\n",
    "    feature_map['ALL_median']=test_correlation_x.median(axis=1).values\n",
    "    #采用一周5天，按斐波那契数列增长\n",
    "    windows=[i*5 for i in [1,2,3,5,8,13]]\n",
    "    for w in windows:\n",
    "        if w>test_correlation_x.shape[1]:\n",
    "            break\n",
    "        val = 'FW'+str(int(w/5))\n",
    "        feature_map[val]=test_correlation_x.iloc[:,-w:].median(axis=1).values-feature_map['ALL_median']\n",
    "    #滑动窗口\n",
    "    for ii,jj in [(0,1),(1,2),(2,3),(3,4),(4,5),(5,6),(6,7),(7,8)]:\n",
    "        if jj*5>test_correlation_x.shape[1]:\n",
    "            break\n",
    "        val = 'WW'+str(jj)\n",
    "        if ii==0:\n",
    "            feature_map[val]=test_correlation_x.iloc[:,-jj*5:].median(axis=1).values-feature_map['ALL_median']\n",
    "        else:\n",
    "            feature_map[val]=test_correlation_x.iloc[:,-jj*5:-ii*5].median(axis=1).values-feature_map['ALL_median']\n",
    "    return feature_map.iloc[:,-15:]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "这部分代码多采用绝对位置，具体哪一天画个时间图会比较好读"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "#valid set\n",
    "#验证集选用test_correlation可用的最后一天作为label\n",
    "d1=test_fund_return.copy()\n",
    "d11=test_fund_benchmark_return.copy()\n",
    "d_index=test_index_return.copy()\n",
    "d2=test_correlation.copy()\n",
    "l=[]\n",
    "#label前61天以及更前日期的corr作为第一特征\n",
    "l=[d2.iloc[:,-1],d2.iloc[:,-1-61],d2.iloc[:,-1-61-20*1],d2.iloc[:,-1-61-20*2],d2.iloc[:,-1-61-20*3]]\n",
    "#label前61天到label前一天这些天可以计算fund（或fund_benchmark或差值）的窗口相关系数\n",
    "l.extend(c_indexs_corr(d1.iloc[:,-61-61:-61],[5,10,15,20,25,30,35,45,61]))\n",
    "l.extend(c_indexs_corr(d11.iloc[:,-61-61:-61],[5,10,15,20,25,30,35,45,61]))\n",
    "l.extend(c_indexs_corr(d1.iloc[:,-61-61:-61]-d11.iloc[:,-61-61:-61],[5,10,15,20,25,30,35,45,61]))\n",
    "#指数在最近61天内的统计特征，每个样本都一样，之前是考虑back_window这种扩充数据集的操作，但最后没有back_window。现为特征重要性为0的冗余特征，之后的代码删掉了这部分\n",
    "l.append(index_status(d_index.iloc[:,-61-61:-61]))\n",
    "#merge各个特征，（注：这里用了优化代码后其实可以一起concat，不用merge）\n",
    "df=pd.concat(l[:5],axis=1).reset_index()\n",
    "for c in l[5:-1]:\n",
    "    df=df.merge(c,on='index',how='left')\n",
    "df=df.set_index('index')\n",
    "df.reset_index(drop=True, inplace=True)\n",
    "df=pd.concat([df,l[-1]],axis=1)\n",
    "#比较基准（普通形式和累积形式）在一段时间(t)的标准差\n",
    "df=pd.concat([df,pd.DataFrame(get_return_std_corr(d11.iloc[:,-61-61:-61])).T],axis=1)\n",
    "df=pd.concat([df,pd.DataFrame(get_return_std_corr(d1.iloc[:,-61-61:-61])).T],axis=1)\n",
    "#计算correlation的时间窗口和滑动时间窗口的特征\n",
    "df=pd.concat([df,f_corr(d2.iloc[:,:-1-61])],axis=1)\n",
    "#计算return或return-benchmark_return的历史特征，然后再分两断计算每个基金的相关性\n",
    "df['dist1']=1-pdist(f_corr(d1.iloc[:,-61-61:-61]).iloc[:,:-8],metric='correlation')\n",
    "df['dist2']=1-pdist(f_corr(d1.iloc[:,-61-61:-61]).iloc[:,-8:],metric='correlation')\n",
    "df['dist3']=1-pdist(f_corr(d1.iloc[:,-61-61:-61]-d11.iloc[:,-61-61:-61]).iloc[:,:-8],metric='correlation')\n",
    "df['dist4']=1-pdist(f_corr(d1.iloc[:,-61-61:-61]-d11.iloc[:,-61-61:-61]).iloc[:,-8:],metric='correlation')\n",
    "# df=pd.concat([df,find_cluster_distance(d1.iloc[:,-61-100:-61])],axis=1)\n",
    "# distances = pdist(d1.iloc[:,-61-61:-61].values, metric='euclidean')\n",
    "# df['dis'] = distances\n",
    "# df['index_coef_corr1']=index_coef(d_index.iloc[:,-61-61:-61],d1.iloc[:,-61-61:-61])\n",
    "df.columns=range(len(df.columns))\n",
    "valid_data=df.copy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "#训练集构建时，因为最后提交的数据和最后可以利用的数据相差61天，不能够让训练数据和验证数据只相差1天，也得相差61天，这样才有一定的参考价值\n",
    "#下面的代码和上面基本一样只是向前移动了61天\n",
    "#hyperparam days是训练集的back_windows的天数，线上结合线下发现越多的数据反而效果越差，只有最后一天的数据可以利用，体现了基金很难预测的特性\n",
    "days=1\n",
    "#train set\n",
    "d1=test_fund_return.copy()\n",
    "d11=test_fund_benchmark_return.copy()\n",
    "d_index=test_index_return.copy()\n",
    "d2=test_correlation.copy()\n",
    "train_data=[]\n",
    "for i in range(61,61+days):\n",
    "    l=[]\n",
    "    l=[d2.iloc[:,-1-i],d2.iloc[:,-1-61-i],d2.iloc[:,-1-61-20*1-i],d2.iloc[:,-1-61-20*2-i],d2.iloc[:,-1-61-20*3-i]]\n",
    "    l.extend(c_indexs_corr(d1.iloc[:,-61-i-61:-i-61],[5,10,15,20,25,30,35,45,61]))\n",
    "    l.extend(c_indexs_corr(d11.iloc[:,-61-i-61:-i-61],[5,10,15,20,25,30,35,45,61]))\n",
    "    l.extend(c_indexs_corr(d1.iloc[:,-61-i-61:-i-61]-d11.iloc[:,-61-i-61:-i-61],[5,10,15,20,25,30,35,45,61]))\n",
    "    l.append(index_status(d_index.iloc[:,-61-i-61:-i-61]))\n",
    "    df=pd.concat(l[:5],axis=1).reset_index()\n",
    "    for c in l[5:-1]:\n",
    "        df=df.merge(c,on='index',how='left')\n",
    "    df=df.set_index('index')\n",
    "    df.reset_index(drop=True, inplace=True)\n",
    "    df=pd.concat([df,l[-1]],axis=1)\n",
    "    df=pd.concat([df,pd.DataFrame(get_return_std_corr(d11.iloc[:,-61-i-61:-i-61])).T],axis=1)\n",
    "    df=pd.concat([df,pd.DataFrame(get_return_std_corr(d1.iloc[:,-61-i-61:-i-61])).T],axis=1)\n",
    "    ###\n",
    "    df=pd.concat([df,f_corr(d2.iloc[:,:-1-61-i])],axis=1)\n",
    "    df['dist1']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]).iloc[:,:-8],metric='correlation')\n",
    "    df['dist2']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]).iloc[:,-8:],metric='correlation')\n",
    "    df['dist3']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]-d11.iloc[:,-61-i-61:-i-61]).iloc[:,:-8],metric='correlation')\n",
    "    df['dist4']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]-d11.iloc[:,-61-i-61:-i-61]).iloc[:,-8:],metric='correlation')\n",
    "#     df=pd.concat([df,find_cluster_distance(d1.iloc[:,-i-61-100:-i-61])],axis=1)\n",
    "#     distances = pdist(d1.iloc[:,-61-i-61:-i-61].values, metric='euclidean')\n",
    "#     df['dis'] = distances\n",
    "#     df['index_coef_corr1']=index_coef(d_index.iloc[:,-61-i-61:-i-61],d1.iloc[:,-61-i-61:-i-61])\n",
    "    df.columns=range(len(df.columns))\n",
    "    train_data.append(df)\n",
    "train_data=pd.concat(train_data,axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((261003, 72), (261003, 72))"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "valid_data.shape,train_data.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(261003, 71)"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#测试集构建就是把数据从valid又向后推了61天\n",
    "#test\n",
    "d1=test_fund_return.copy()\n",
    "d11=test_fund_benchmark_return.copy()\n",
    "d_index=test_index_return.copy()\n",
    "d2=test_correlation.copy()\n",
    "l=[d2.iloc[:,-1],d2.iloc[:,-1-20*1],d2.iloc[:,-1-20*2],d2.iloc[:,-1-20*3]]\n",
    "l.extend(c_indexs_corr(d1.iloc[:,-61:],[5,10,15,20,25,30,35,45,61]))\n",
    "l.extend(c_indexs_corr(d11.iloc[:,-61:],[5,10,15,20,25,30,35,45,61]))\n",
    "l.extend(c_indexs_corr(d1.iloc[:,-61:]-d11.iloc[:,-61:],[5,10,15,20,25,30,35,45,61]))\n",
    "l.append(index_status(d_index.iloc[:,-61:]))\n",
    "test_df=pd.concat(l[:4],axis=1).reset_index()\n",
    "for c in l[4:-1]:\n",
    "    test_df=test_df.merge(c,on='index',how='left')\n",
    "test_df=test_df.set_index('index')\n",
    "test_df.reset_index(drop=True, inplace=True)\n",
    "test_df=pd.concat([test_df,l[-1]],axis=1)\n",
    "test_df=pd.concat([test_df,pd.DataFrame(get_return_std_corr(d11.iloc[:,-61:])).T],axis=1)\n",
    "test_df=pd.concat([test_df,pd.DataFrame(get_return_std_corr(d1.iloc[:,-61:])).T],axis=1)\n",
    "###\n",
    "test_df=pd.concat([test_df,f_corr(d2)],axis=1)\n",
    "test_df['dist1']=1-pdist(f_corr(d1.iloc[:,-61:]).iloc[:,:-8],metric='correlation')\n",
    "test_df['dist2']=1-pdist(f_corr(d1.iloc[:,-61:]).iloc[:,-8:],metric='correlation')\n",
    "test_df['dist3']=1-pdist(f_corr(d1.iloc[:,-61:]-d11.iloc[:,-61:]).iloc[:,:-8],metric='correlation')\n",
    "test_df['dist4']=1-pdist(f_corr(d1.iloc[:,-61:]-d11.iloc[:,-61:]).iloc[:,-8:],metric='correlation')\n",
    "# test_df=pd.concat([test_df,find_cluster_distance(d1.iloc[:,-100:])],axis=1)\n",
    "# distances = pdist(d1.ix[:,-61:].values, metric='euclidean')\n",
    "# test_df['dis'] = distances\n",
    "# test_df['index_coef_corr1']=index_coef(d_index.iloc[:,-61:],d1.iloc[:,-61:])\n",
    "test_df.columns=range(1,len(test_df.columns)+1)\n",
    "test_df.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# train_data.columns=['label','last1','last2','last3','last4',]+['return_%d' % i for i in range(1,6)]+['return_bmark_%d' % i for i in range(1,6)]\\\n",
    "# +['index_%d' % i for i in range(1,10)]#+['AP_dis','AP_class']+['dist']\n",
    "# valid_data.columns=['label','last1','last2','last3','last4',]+['return_%d' % i for i in range(1,6)]+['return_bmark_%d' % i for i in range(1,6)]\\\n",
    "# +['index_%d' % i for i in range(1,10)]#+['AP_dis','AP_class']+['index_coef_corr1']+['dist']\n",
    "# test_df.columns=['last1','last2','last3','last4',]+['return_%d' % i for i in range(1,6)]+['return_bmark_%d' % i for i in range(1,6)]\\\n",
    "# +['index_%d' % i for i in range(1,10)]#+['AP_dis','AP_class']+['index_coef_corr1']+['dist']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "# #这个特征太强了\n",
    "# # train_data.drop('dist',axis=1,inplace=True)\n",
    "# valid_data.drop(['index_coef_corr1','dist','AP_dis','AP_class'],axis=1,inplace=True)\n",
    "# test_df.drop(['index_coef_corr1','dist','AP_dis','AP_class'],axis=1,inplace=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Start training...\n",
      "Training until validation scores don't improve for 20 rounds.\n",
      "[100]\ttraining's l1: 0.088737\tvalid_1's l1: 0.0866515\n",
      "[200]\ttraining's l1: 0.0838468\tvalid_1's l1: 0.0833385\n",
      "[300]\ttraining's l1: 0.0819195\tvalid_1's l1: 0.082372\n",
      "[400]\ttraining's l1: 0.0808446\tvalid_1's l1: 0.0820829\n",
      "[500]\ttraining's l1: 0.0801371\tvalid_1's l1: 0.0818835\n",
      "Early stopping, best iteration is:\n",
      "[496]\ttraining's l1: 0.0801593\tvalid_1's l1: 0.0818698\n",
      "The f2 of traindata is: 0.8508159429278003\n",
      "Start predicting...\n",
      "The mae of prediction is: 0.08414447595102456\n",
      "The f2 of prediction is: 0.8525843586672227\n"
     ]
    }
   ],
   "source": [
    "import lightgbm as lgb\n",
    "import pandas as pd\n",
    "from sklearn.metrics import mean_squared_error\n",
    "from sklearn.metrics import mean_absolute_error\n",
    "#评价指标\n",
    "def f2(preds, targs):\n",
    "    x=mean_absolute_error(targs,preds)\n",
    "    y=mean_absolute_error(targs,preds,sample_weight=[1/(1.5-i) for i in targs])\n",
    "    return  (2/(2+x+y))**2\n",
    "\n",
    "valid_data=valid_data.sample(frac=1)\n",
    "train_data=train_data.sample(frac=1)\n",
    "# create dataset for lightgbm\n",
    "lgb_train = lgb.Dataset(train_data.drop(0,axis=1), train_data.iloc[:,0],weight=[(1/abs(1.5-v)) for v in train_data.iloc[:,0]])\n",
    "lgb_eval = lgb.Dataset(valid_data.drop(0,axis=1), valid_data.iloc[:,0],weight=[(1/abs(1.5-v)) for v in valid_data.iloc[:,0]])\n",
    "# specify your configurations as a dict\n",
    "#调整好的参数\n",
    "params = {\n",
    "    'task': 'train',\n",
    "    'boosting_type': 'gbdt',\n",
    "    'objective': 'regression_l1',\n",
    "    'metric':'l1',\n",
    "    'num_leaves': 5,\n",
    "    'learning_rate': 0.01,\n",
    "    'feature_fraction': 0.8,\n",
    "    'bagging_fraction': 0.8,\n",
    "    'verbose': 0\n",
    "}\n",
    "\n",
    "print('Start training...')\n",
    "# train\n",
    "evals_result = {}\n",
    "gbm = lgb.train(params,\n",
    "                lgb_train,\n",
    "                num_boost_round=700,\n",
    "                valid_sets=[lgb_train,lgb_eval],\n",
    "                early_stopping_rounds=20,evals_result=evals_result,verbose_eval=100)\n",
    "\n",
    "# print('Save model...')\n",
    "# # save model to file\n",
    "# gbm.save_model('model.txt')\n",
    "y_pred = gbm.predict(train_data.drop(0,axis=1), num_iteration=gbm.best_iteration)\n",
    "print('The f2 of traindata is:', f2(y_pred,train_data.iloc[:,0]))\n",
    "print('Start predicting...')\n",
    "# predict\n",
    "y_pred = gbm.predict(valid_data.drop(0,axis=1), num_iteration=gbm.best_iteration)\n",
    "# eval\n",
    "print('The mae of prediction is:', mean_absolute_error(valid_data.iloc[:,0], y_pred))\n",
    "\n",
    "print('The f2 of prediction is:', f2(y_pred,valid_data.iloc[:,0]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "num_boost_round=gbm.best_iteration"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.figure.Figure at 0x130cb9a7860>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x130ff57b3c8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(12,6))\n",
    "lgb.plot_importance(gbm)\n",
    "plt.title(\"Featurertances\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "画出训练结果...\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x130a4a871d0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "print('画出训练结果...')\n",
    "plt.plot(range(len(evals_result['training']['l1'])),evals_result['training']['l1'])\n",
    "plt.plot(range(len(evals_result['valid_1']['l1'])),evals_result['valid_1']['l1'])\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### predict"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "# predict 因为我们刚训练的模型毕竟是61*2天前数据训练的模型，而我们希望用61天前数据训练的模型，照着test，往前推61天\n",
    "# 但是这个也是冗余代码，其实这个数据就是valid_data\n",
    "#train set\n",
    "d1=test_fund_return.copy()\n",
    "d11=test_fund_benchmark_return.copy()\n",
    "d_index=test_index_return.copy()\n",
    "d2=test_correlation.copy()\n",
    "train_data=[]\n",
    "for i in range(days):\n",
    "    l=[]\n",
    "    l=[d2.iloc[:,-1-i],d2.iloc[:,-1-61-i],d2.iloc[:,-1-61-20*1-i],d2.iloc[:,-1-61-20*2-i],d2.iloc[:,-1-61-20*3-i]]\n",
    "    l.extend(c_indexs_corr(d1.iloc[:,-61-i-61:-i-61],[5,10,15,20,25,30,35,45,61]))\n",
    "    l.extend(c_indexs_corr(d11.iloc[:,-61-i-61:-i-61],[5,10,15,20,25,30,35,45,61]))\n",
    "    l.extend(c_indexs_corr(d1.iloc[:,-61-i-61:-i-61]-d11.iloc[:,-61-i-61:-i-61],[5,10,15,20,25,30,35,45,61]))\n",
    "    l.append(index_status(d_index.iloc[:,-61-i-61:-i-61]))\n",
    "    df=pd.concat(l[:5],axis=1).reset_index()\n",
    "    for c in l[5:-1]:\n",
    "        df=df.merge(c,on='index',how='left')\n",
    "    df=df.set_index('index')\n",
    "    df.reset_index(drop=True, inplace=True)\n",
    "    df=pd.concat([df,l[-1]],axis=1)\n",
    "    df=pd.concat([df,pd.DataFrame(get_return_std_corr(d11.iloc[:,-61-i-61:-i-61])).T],axis=1)\n",
    "    df=pd.concat([df,pd.DataFrame(get_return_std_corr(d1.iloc[:,-61-i-61:-i-61])).T],axis=1)\n",
    "    df=pd.concat([df,f_corr(d2.iloc[:,:-1-61-i])],axis=1)\n",
    "    df['dist1']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]).iloc[:,:-8],metric='correlation')\n",
    "    df['dist2']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]).iloc[:,-8:],metric='correlation')\n",
    "    df['dist3']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]-d11.iloc[:,-61-i-61:-i-61]).iloc[:,:-8],metric='correlation')\n",
    "    df['dist4']=1-pdist(f_corr(d1.iloc[:,-61-i-61:-i-61]-d11.iloc[:,-61-i-61:-i-61]).iloc[:,-8:],metric='correlation')\n",
    "#     df=pd.concat([df,find_cluster_distance(d1.iloc[:,:-i-61])],axis=1)\n",
    "#     distances = pdist(d1.iloc[:,-61-i-61:-i-61].values, metric='euclidean')\n",
    "#     df['dis'] = distances\n",
    "#     df['index_coef_corr1']=index_coef(d_index.iloc[:,-61-i-61:-i-61],d1.iloc[:,-61-i-61:-i-61])\n",
    "    df.columns=range(len(df.columns))\n",
    "    train_data.append(df)\n",
    "train_data=pd.concat(train_data,axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Start training...\n",
      "[1]\ttraining's l1: 0.0945378\n",
      "Training until validation scores don't improve for 20 rounds.\n",
      "[2]\ttraining's l1: 0.0942047\n",
      "[3]\ttraining's l1: 0.0938756\n",
      "[4]\ttraining's l1: 0.0935517\n",
      "[5]\ttraining's l1: 0.0932329\n",
      "[6]\ttraining's l1: 0.0929198\n",
      "[7]\ttraining's l1: 0.0926112\n",
      "[8]\ttraining's l1: 0.0923102\n",
      "[9]\ttraining's l1: 0.0920125\n",
      "[10]\ttraining's l1: 0.0917196\n",
      "[11]\ttraining's l1: 0.0914278\n",
      "[12]\ttraining's l1: 0.0911443\n",
      "[13]\ttraining's l1: 0.0908507\n",
      "[14]\ttraining's l1: 0.0905624\n",
      "[15]\ttraining's l1: 0.0902822\n",
      "[16]\ttraining's l1: 0.090019\n",
      "[17]\ttraining's l1: 0.0897562\n",
      "[18]\ttraining's l1: 0.0894928\n",
      "[19]\ttraining's l1: 0.0892247\n",
      "[20]\ttraining's l1: 0.0889745\n",
      "[21]\ttraining's l1: 0.0887149\n",
      "[22]\ttraining's l1: 0.08847\n",
      "[23]\ttraining's l1: 0.0882305\n",
      "[24]\ttraining's l1: 0.0879905\n",
      "[25]\ttraining's l1: 0.0877565\n",
      "[26]\ttraining's l1: 0.0875267\n",
      "[27]\ttraining's l1: 0.0872974\n",
      "[28]\ttraining's l1: 0.0870745\n",
      "[29]\ttraining's l1: 0.0868422\n",
      "[30]\ttraining's l1: 0.0866181\n",
      "[31]\ttraining's l1: 0.0863983\n",
      "[32]\ttraining's l1: 0.0861811\n",
      "[33]\ttraining's l1: 0.0859735\n",
      "[34]\ttraining's l1: 0.0857731\n",
      "[35]\ttraining's l1: 0.0855677\n",
      "[36]\ttraining's l1: 0.0853598\n",
      "[37]\ttraining's l1: 0.0851563\n",
      "[38]\ttraining's l1: 0.0849661\n",
      "[39]\ttraining's l1: 0.0847663\n",
      "[40]\ttraining's l1: 0.0845757\n",
      "[41]\ttraining's l1: 0.084388\n",
      "[42]\ttraining's l1: 0.0842018\n",
      "[43]\ttraining's l1: 0.0840287\n",
      "[44]\ttraining's l1: 0.0838467\n",
      "[45]\ttraining's l1: 0.0836671\n",
      "[46]\ttraining's l1: 0.0834952\n",
      "[47]\ttraining's l1: 0.0833226\n",
      "[48]\ttraining's l1: 0.0831496\n",
      "[49]\ttraining's l1: 0.082981\n",
      "[50]\ttraining's l1: 0.0828146\n",
      "[51]\ttraining's l1: 0.0826486\n",
      "[52]\ttraining's l1: 0.0824818\n",
      "[53]\ttraining's l1: 0.0823214\n",
      "[54]\ttraining's l1: 0.0821639\n",
      "[55]\ttraining's l1: 0.082013\n",
      "[56]\ttraining's l1: 0.0818586\n",
      "[57]\ttraining's l1: 0.0817071\n",
      "[58]\ttraining's l1: 0.0815616\n",
      "[59]\ttraining's l1: 0.0814132\n",
      "[60]\ttraining's l1: 0.0812676\n",
      "[61]\ttraining's l1: 0.0811252\n",
      "[62]\ttraining's l1: 0.0809849\n",
      "[63]\ttraining's l1: 0.0808448\n",
      "[64]\ttraining's l1: 0.0807028\n",
      "[65]\ttraining's l1: 0.0805726\n",
      "[66]\ttraining's l1: 0.0804405\n",
      "[67]\ttraining's l1: 0.0803086\n",
      "[68]\ttraining's l1: 0.0801843\n",
      "[69]\ttraining's l1: 0.080053\n",
      "[70]\ttraining's l1: 0.079924\n",
      "[71]\ttraining's l1: 0.0797986\n",
      "[72]\ttraining's l1: 0.0796755\n",
      "[73]\ttraining's l1: 0.0795544\n",
      "[74]\ttraining's l1: 0.0794344\n",
      "[75]\ttraining's l1: 0.0793164\n",
      "[76]\ttraining's l1: 0.0792004\n",
      "[77]\ttraining's l1: 0.0790852\n",
      "[78]\ttraining's l1: 0.0789717\n",
      "[79]\ttraining's l1: 0.0788607\n",
      "[80]\ttraining's l1: 0.0787565\n",
      "[81]\ttraining's l1: 0.078648\n",
      "[82]\ttraining's l1: 0.0785385\n",
      "[83]\ttraining's l1: 0.0784338\n",
      "[84]\ttraining's l1: 0.0783275\n",
      "[85]\ttraining's l1: 0.0782225\n",
      "[86]\ttraining's l1: 0.0781267\n",
      "[87]\ttraining's l1: 0.0780265\n",
      "[88]\ttraining's l1: 0.0779337\n",
      "[89]\ttraining's l1: 0.0778327\n",
      "[90]\ttraining's l1: 0.0777344\n",
      "[91]\ttraining's l1: 0.0776415\n",
      "[92]\ttraining's l1: 0.0775497\n",
      "[93]\ttraining's l1: 0.0774576\n",
      "[94]\ttraining's l1: 0.0773665\n",
      "[95]\ttraining's l1: 0.0772743\n",
      "[96]\ttraining's l1: 0.0771876\n",
      "[97]\ttraining's l1: 0.0770974\n",
      "[98]\ttraining's l1: 0.0770091\n",
      "[99]\ttraining's l1: 0.0769219\n",
      "[100]\ttraining's l1: 0.0768389\n",
      "[101]\ttraining's l1: 0.0767595\n",
      "[102]\ttraining's l1: 0.0766773\n",
      "[103]\ttraining's l1: 0.076595\n",
      "[104]\ttraining's l1: 0.0765199\n",
      "[105]\ttraining's l1: 0.0764406\n",
      "[106]\ttraining's l1: 0.076363\n",
      "[107]\ttraining's l1: 0.0762882\n",
      "[108]\ttraining's l1: 0.0762149\n",
      "[109]\ttraining's l1: 0.0761434\n",
      "[110]\ttraining's l1: 0.0760691\n",
      "[111]\ttraining's l1: 0.0759958\n",
      "[112]\ttraining's l1: 0.0759286\n",
      "[113]\ttraining's l1: 0.0758613\n",
      "[114]\ttraining's l1: 0.0757931\n",
      "[115]\ttraining's l1: 0.0757258\n",
      "[116]\ttraining's l1: 0.0756555\n",
      "[117]\ttraining's l1: 0.0755894\n",
      "[118]\ttraining's l1: 0.0755246\n",
      "[119]\ttraining's l1: 0.0754584\n",
      "[120]\ttraining's l1: 0.0753955\n",
      "[121]\ttraining's l1: 0.0753299\n",
      "[122]\ttraining's l1: 0.0752687\n",
      "[123]\ttraining's l1: 0.0752086\n",
      "[124]\ttraining's l1: 0.0751514\n",
      "[125]\ttraining's l1: 0.0750908\n",
      "[126]\ttraining's l1: 0.0750338\n",
      "[127]\ttraining's l1: 0.0749773\n",
      "[128]\ttraining's l1: 0.0749183\n",
      "[129]\ttraining's l1: 0.0748604\n",
      "[130]\ttraining's l1: 0.0748019\n",
      "[131]\ttraining's l1: 0.0747437\n",
      "[132]\ttraining's l1: 0.0746946\n",
      "[133]\ttraining's l1: 0.0746375\n",
      "[134]\ttraining's l1: 0.0745857\n",
      "[135]\ttraining's l1: 0.0745318\n",
      "[136]\ttraining's l1: 0.0744837\n",
      "[137]\ttraining's l1: 0.0744322\n",
      "[138]\ttraining's l1: 0.0743783\n",
      "[139]\ttraining's l1: 0.0743282\n",
      "[140]\ttraining's l1: 0.0742763\n",
      "[141]\ttraining's l1: 0.0742275\n",
      "[142]\ttraining's l1: 0.0741796\n",
      "[143]\ttraining's l1: 0.0741316\n",
      "[144]\ttraining's l1: 0.0740917\n",
      "[145]\ttraining's l1: 0.074044\n",
      "[146]\ttraining's l1: 0.0739982\n",
      "[147]\ttraining's l1: 0.0739493\n",
      "[148]\ttraining's l1: 0.0739056\n",
      "[149]\ttraining's l1: 0.0738594\n",
      "[150]\ttraining's l1: 0.0738221\n",
      "[151]\ttraining's l1: 0.0737818\n",
      "[152]\ttraining's l1: 0.0737402\n",
      "[153]\ttraining's l1: 0.0736941\n",
      "[154]\ttraining's l1: 0.0736507\n",
      "[155]\ttraining's l1: 0.0736105\n",
      "[156]\ttraining's l1: 0.0735701\n",
      "[157]\ttraining's l1: 0.0735329\n",
      "[158]\ttraining's l1: 0.0734941\n",
      "[159]\ttraining's l1: 0.0734578\n",
      "[160]\ttraining's l1: 0.0734183\n",
      "[161]\ttraining's l1: 0.0733828\n",
      "[162]\ttraining's l1: 0.0733449\n",
      "[163]\ttraining's l1: 0.0733098\n",
      "[164]\ttraining's l1: 0.073273\n",
      "[165]\ttraining's l1: 0.0732372\n",
      "[166]\ttraining's l1: 0.0731975\n",
      "[167]\ttraining's l1: 0.0731633\n",
      "[168]\ttraining's l1: 0.0731249\n",
      "[169]\ttraining's l1: 0.0730911\n",
      "[170]\ttraining's l1: 0.0730562\n",
      "[171]\ttraining's l1: 0.0730209\n",
      "[172]\ttraining's l1: 0.0729884\n",
      "[173]\ttraining's l1: 0.0729577\n",
      "[174]\ttraining's l1: 0.0729242\n",
      "[175]\ttraining's l1: 0.0728932\n",
      "[176]\ttraining's l1: 0.0728599\n",
      "[177]\ttraining's l1: 0.0728301\n",
      "[178]\ttraining's l1: 0.0727998\n",
      "[179]\ttraining's l1: 0.0727697\n",
      "[180]\ttraining's l1: 0.0727363\n",
      "[181]\ttraining's l1: 0.0727057\n",
      "[182]\ttraining's l1: 0.0726777\n",
      "[183]\ttraining's l1: 0.072649\n",
      "[184]\ttraining's l1: 0.0726212\n",
      "[185]\ttraining's l1: 0.0725899\n",
      "[186]\ttraining's l1: 0.0725584\n",
      "[187]\ttraining's l1: 0.0725317\n",
      "[188]\ttraining's l1: 0.0725011\n",
      "[189]\ttraining's l1: 0.0724718\n",
      "[190]\ttraining's l1: 0.072446\n",
      "[191]\ttraining's l1: 0.0724186\n",
      "[192]\ttraining's l1: 0.0723895\n",
      "[193]\ttraining's l1: 0.0723638\n",
      "[194]\ttraining's l1: 0.0723387\n",
      "[195]\ttraining's l1: 0.0723099\n",
      "[196]\ttraining's l1: 0.0722826\n",
      "[197]\ttraining's l1: 0.072258\n",
      "[198]\ttraining's l1: 0.0722337\n",
      "[199]\ttraining's l1: 0.072207\n",
      "[200]\ttraining's l1: 0.0721808\n",
      "[201]\ttraining's l1: 0.0721547\n",
      "[202]\ttraining's l1: 0.0721306\n",
      "[203]\ttraining's l1: 0.0721077\n",
      "[204]\ttraining's l1: 0.0720826\n",
      "[205]\ttraining's l1: 0.0720577\n",
      "[206]\ttraining's l1: 0.0720354\n",
      "[207]\ttraining's l1: 0.072013\n",
      "[208]\ttraining's l1: 0.071988\n",
      "[209]\ttraining's l1: 0.0719614\n",
      "[210]\ttraining's l1: 0.0719371\n",
      "[211]\ttraining's l1: 0.0719117\n",
      "[212]\ttraining's l1: 0.0718881\n",
      "[213]\ttraining's l1: 0.0718668\n",
      "[214]\ttraining's l1: 0.0718465\n",
      "[215]\ttraining's l1: 0.0718256\n",
      "[216]\ttraining's l1: 0.0718058\n",
      "[217]\ttraining's l1: 0.071785\n",
      "[218]\ttraining's l1: 0.0717621\n",
      "[219]\ttraining's l1: 0.0717424\n",
      "[220]\ttraining's l1: 0.0717208\n",
      "[221]\ttraining's l1: 0.0717021\n",
      "[222]\ttraining's l1: 0.071682\n",
      "[223]\ttraining's l1: 0.071657\n",
      "[224]\ttraining's l1: 0.0716377\n",
      "[225]\ttraining's l1: 0.0716174\n",
      "[226]\ttraining's l1: 0.0715985\n",
      "[227]\ttraining's l1: 0.0715783\n",
      "[228]\ttraining's l1: 0.0715587\n",
      "[229]\ttraining's l1: 0.07154\n",
      "[230]\ttraining's l1: 0.0715201\n",
      "[231]\ttraining's l1: 0.0714993\n",
      "[232]\ttraining's l1: 0.0714803\n",
      "[233]\ttraining's l1: 0.0714605\n",
      "[234]\ttraining's l1: 0.0714423\n",
      "[235]\ttraining's l1: 0.0714236\n",
      "[236]\ttraining's l1: 0.0714035\n",
      "[237]\ttraining's l1: 0.0713862\n",
      "[238]\ttraining's l1: 0.0713683\n",
      "[239]\ttraining's l1: 0.0713452\n",
      "[240]\ttraining's l1: 0.0713274\n",
      "[241]\ttraining's l1: 0.071308\n",
      "[242]\ttraining's l1: 0.0712905\n",
      "[243]\ttraining's l1: 0.0712723\n",
      "[244]\ttraining's l1: 0.0712529\n",
      "[245]\ttraining's l1: 0.0712361\n",
      "[246]\ttraining's l1: 0.0712178\n",
      "[247]\ttraining's l1: 0.0711966\n",
      "[248]\ttraining's l1: 0.0711811\n",
      "[249]\ttraining's l1: 0.0711596\n",
      "[250]\ttraining's l1: 0.0711428\n",
      "[251]\ttraining's l1: 0.0711252\n",
      "[252]\ttraining's l1: 0.071108\n",
      "[253]\ttraining's l1: 0.0710929\n",
      "[254]\ttraining's l1: 0.0710723\n",
      "[255]\ttraining's l1: 0.0710558\n",
      "[256]\ttraining's l1: 0.071039\n",
      "[257]\ttraining's l1: 0.0710223\n",
      "[258]\ttraining's l1: 0.0710025\n",
      "[259]\ttraining's l1: 0.0709863\n",
      "[260]\ttraining's l1: 0.0709706\n",
      "[261]\ttraining's l1: 0.0709515\n",
      "[262]\ttraining's l1: 0.0709369\n",
      "[263]\ttraining's l1: 0.0709225\n",
      "[264]\ttraining's l1: 0.0709048\n",
      "[265]\ttraining's l1: 0.0708913\n",
      "[266]\ttraining's l1: 0.0708761\n",
      "[267]\ttraining's l1: 0.07086\n",
      "[268]\ttraining's l1: 0.0708431\n",
      "[269]\ttraining's l1: 0.0708291\n",
      "[270]\ttraining's l1: 0.0708134\n",
      "[271]\ttraining's l1: 0.0707986\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[272]\ttraining's l1: 0.0707843\n",
      "[273]\ttraining's l1: 0.070769\n",
      "[274]\ttraining's l1: 0.0707544\n",
      "[275]\ttraining's l1: 0.0707387\n",
      "[276]\ttraining's l1: 0.0707239\n",
      "[277]\ttraining's l1: 0.0707123\n",
      "[278]\ttraining's l1: 0.0706956\n",
      "[279]\ttraining's l1: 0.0706816\n",
      "[280]\ttraining's l1: 0.0706691\n",
      "[281]\ttraining's l1: 0.0706552\n",
      "[282]\ttraining's l1: 0.0706417\n",
      "[283]\ttraining's l1: 0.0706268\n",
      "[284]\ttraining's l1: 0.0706111\n",
      "[285]\ttraining's l1: 0.070597\n",
      "[286]\ttraining's l1: 0.0705825\n",
      "[287]\ttraining's l1: 0.0705696\n",
      "[288]\ttraining's l1: 0.0705535\n",
      "[289]\ttraining's l1: 0.0705348\n",
      "[290]\ttraining's l1: 0.0705203\n",
      "[291]\ttraining's l1: 0.0705067\n",
      "[292]\ttraining's l1: 0.0704886\n",
      "[293]\ttraining's l1: 0.0704757\n",
      "[294]\ttraining's l1: 0.0704645\n",
      "[295]\ttraining's l1: 0.070452\n",
      "[296]\ttraining's l1: 0.0704389\n",
      "[297]\ttraining's l1: 0.0704265\n",
      "[298]\ttraining's l1: 0.0704146\n",
      "[299]\ttraining's l1: 0.0704021\n",
      "[300]\ttraining's l1: 0.0703851\n",
      "[301]\ttraining's l1: 0.0703714\n",
      "[302]\ttraining's l1: 0.0703594\n",
      "[303]\ttraining's l1: 0.0703465\n",
      "[304]\ttraining's l1: 0.0703301\n",
      "[305]\ttraining's l1: 0.0703142\n",
      "[306]\ttraining's l1: 0.0702984\n",
      "[307]\ttraining's l1: 0.0702866\n",
      "[308]\ttraining's l1: 0.0702771\n",
      "[309]\ttraining's l1: 0.0702648\n",
      "[310]\ttraining's l1: 0.070252\n",
      "[311]\ttraining's l1: 0.070237\n",
      "[312]\ttraining's l1: 0.070225\n",
      "[313]\ttraining's l1: 0.0702135\n",
      "[314]\ttraining's l1: 0.0702001\n",
      "[315]\ttraining's l1: 0.0701872\n",
      "[316]\ttraining's l1: 0.0701763\n",
      "[317]\ttraining's l1: 0.0701637\n",
      "[318]\ttraining's l1: 0.0701529\n",
      "[319]\ttraining's l1: 0.0701388\n",
      "[320]\ttraining's l1: 0.070127\n",
      "[321]\ttraining's l1: 0.0701188\n",
      "[322]\ttraining's l1: 0.0701081\n",
      "[323]\ttraining's l1: 0.070096\n",
      "[324]\ttraining's l1: 0.070086\n",
      "[325]\ttraining's l1: 0.0700727\n",
      "[326]\ttraining's l1: 0.0700644\n",
      "[327]\ttraining's l1: 0.0700544\n",
      "[328]\ttraining's l1: 0.0700439\n",
      "[329]\ttraining's l1: 0.0700321\n",
      "[330]\ttraining's l1: 0.0700217\n",
      "[331]\ttraining's l1: 0.0700088\n",
      "[332]\ttraining's l1: 0.0699992\n",
      "[333]\ttraining's l1: 0.0699881\n",
      "[334]\ttraining's l1: 0.0699795\n",
      "[335]\ttraining's l1: 0.0699674\n",
      "[336]\ttraining's l1: 0.0699555\n",
      "[337]\ttraining's l1: 0.0699466\n",
      "[338]\ttraining's l1: 0.0699372\n",
      "[339]\ttraining's l1: 0.0699296\n",
      "[340]\ttraining's l1: 0.0699201\n",
      "[341]\ttraining's l1: 0.0699083\n",
      "[342]\ttraining's l1: 0.069899\n",
      "[343]\ttraining's l1: 0.0698919\n",
      "[344]\ttraining's l1: 0.0698835\n",
      "[345]\ttraining's l1: 0.0698749\n",
      "[346]\ttraining's l1: 0.0698658\n",
      "[347]\ttraining's l1: 0.0698586\n",
      "[348]\ttraining's l1: 0.0698474\n",
      "[349]\ttraining's l1: 0.069836\n",
      "[350]\ttraining's l1: 0.0698279\n",
      "[351]\ttraining's l1: 0.0698183\n",
      "[352]\ttraining's l1: 0.06981\n",
      "[353]\ttraining's l1: 0.0698006\n",
      "[354]\ttraining's l1: 0.0697913\n",
      "[355]\ttraining's l1: 0.0697848\n",
      "[356]\ttraining's l1: 0.0697742\n",
      "[357]\ttraining's l1: 0.0697652\n",
      "[358]\ttraining's l1: 0.069757\n",
      "[359]\ttraining's l1: 0.0697477\n",
      "[360]\ttraining's l1: 0.069739\n",
      "[361]\ttraining's l1: 0.0697286\n",
      "[362]\ttraining's l1: 0.0697204\n",
      "[363]\ttraining's l1: 0.0697112\n",
      "[364]\ttraining's l1: 0.0697026\n",
      "[365]\ttraining's l1: 0.0696936\n",
      "[366]\ttraining's l1: 0.0696866\n",
      "[367]\ttraining's l1: 0.0696778\n",
      "[368]\ttraining's l1: 0.0696705\n",
      "[369]\ttraining's l1: 0.0696617\n",
      "[370]\ttraining's l1: 0.0696539\n",
      "[371]\ttraining's l1: 0.0696474\n",
      "[372]\ttraining's l1: 0.0696399\n",
      "[373]\ttraining's l1: 0.0696305\n",
      "[374]\ttraining's l1: 0.0696222\n",
      "[375]\ttraining's l1: 0.0696153\n",
      "[376]\ttraining's l1: 0.0696076\n",
      "[377]\ttraining's l1: 0.0696018\n",
      "[378]\ttraining's l1: 0.0695927\n",
      "[379]\ttraining's l1: 0.069585\n",
      "[380]\ttraining's l1: 0.0695785\n",
      "[381]\ttraining's l1: 0.0695704\n",
      "[382]\ttraining's l1: 0.0695623\n",
      "[383]\ttraining's l1: 0.0695549\n",
      "[384]\ttraining's l1: 0.0695477\n",
      "[385]\ttraining's l1: 0.0695402\n",
      "[386]\ttraining's l1: 0.069533\n",
      "[387]\ttraining's l1: 0.0695258\n",
      "[388]\ttraining's l1: 0.0695187\n",
      "[389]\ttraining's l1: 0.0695118\n",
      "[390]\ttraining's l1: 0.0695053\n",
      "[391]\ttraining's l1: 0.0694986\n",
      "[392]\ttraining's l1: 0.0694922\n",
      "[393]\ttraining's l1: 0.0694849\n",
      "[394]\ttraining's l1: 0.0694797\n",
      "[395]\ttraining's l1: 0.069473\n",
      "[396]\ttraining's l1: 0.0694661\n",
      "[397]\ttraining's l1: 0.0694594\n",
      "[398]\ttraining's l1: 0.0694527\n",
      "[399]\ttraining's l1: 0.0694459\n",
      "[400]\ttraining's l1: 0.0694398\n",
      "[401]\ttraining's l1: 0.0694334\n",
      "[402]\ttraining's l1: 0.0694278\n",
      "[403]\ttraining's l1: 0.0694215\n",
      "[404]\ttraining's l1: 0.0694156\n",
      "[405]\ttraining's l1: 0.0694097\n",
      "[406]\ttraining's l1: 0.0694042\n",
      "[407]\ttraining's l1: 0.0693992\n",
      "[408]\ttraining's l1: 0.0693932\n",
      "[409]\ttraining's l1: 0.0693877\n",
      "[410]\ttraining's l1: 0.0693813\n",
      "[411]\ttraining's l1: 0.0693768\n",
      "[412]\ttraining's l1: 0.0693703\n",
      "[413]\ttraining's l1: 0.0693647\n",
      "[414]\ttraining's l1: 0.06936\n",
      "[415]\ttraining's l1: 0.0693532\n",
      "[416]\ttraining's l1: 0.0693478\n",
      "[417]\ttraining's l1: 0.0693418\n",
      "[418]\ttraining's l1: 0.0693355\n",
      "[419]\ttraining's l1: 0.0693297\n",
      "[420]\ttraining's l1: 0.0693246\n",
      "[421]\ttraining's l1: 0.0693185\n",
      "[422]\ttraining's l1: 0.0693125\n",
      "[423]\ttraining's l1: 0.0693078\n",
      "[424]\ttraining's l1: 0.0693035\n",
      "[425]\ttraining's l1: 0.0692983\n",
      "[426]\ttraining's l1: 0.069293\n",
      "[427]\ttraining's l1: 0.0692873\n",
      "[428]\ttraining's l1: 0.069282\n",
      "[429]\ttraining's l1: 0.0692775\n",
      "[430]\ttraining's l1: 0.0692722\n",
      "[431]\ttraining's l1: 0.0692667\n",
      "[432]\ttraining's l1: 0.0692612\n",
      "[433]\ttraining's l1: 0.0692559\n",
      "[434]\ttraining's l1: 0.0692511\n",
      "[435]\ttraining's l1: 0.0692458\n",
      "[436]\ttraining's l1: 0.0692405\n",
      "[437]\ttraining's l1: 0.0692356\n",
      "[438]\ttraining's l1: 0.0692305\n",
      "[439]\ttraining's l1: 0.0692248\n",
      "[440]\ttraining's l1: 0.0692205\n",
      "[441]\ttraining's l1: 0.0692154\n",
      "[442]\ttraining's l1: 0.0692098\n",
      "[443]\ttraining's l1: 0.0692048\n",
      "[444]\ttraining's l1: 0.0691993\n",
      "[445]\ttraining's l1: 0.069194\n",
      "[446]\ttraining's l1: 0.0691892\n",
      "[447]\ttraining's l1: 0.0691844\n",
      "[448]\ttraining's l1: 0.06918\n",
      "[449]\ttraining's l1: 0.0691737\n",
      "[450]\ttraining's l1: 0.0691693\n",
      "[451]\ttraining's l1: 0.0691644\n",
      "[452]\ttraining's l1: 0.0691588\n",
      "[453]\ttraining's l1: 0.0691539\n",
      "[454]\ttraining's l1: 0.0691501\n",
      "[455]\ttraining's l1: 0.0691449\n",
      "[456]\ttraining's l1: 0.0691409\n",
      "[457]\ttraining's l1: 0.0691364\n",
      "[458]\ttraining's l1: 0.0691315\n",
      "[459]\ttraining's l1: 0.0691262\n",
      "[460]\ttraining's l1: 0.0691213\n",
      "[461]\ttraining's l1: 0.0691168\n",
      "[462]\ttraining's l1: 0.0691114\n",
      "[463]\ttraining's l1: 0.0691076\n",
      "[464]\ttraining's l1: 0.0691033\n",
      "[465]\ttraining's l1: 0.0690984\n",
      "[466]\ttraining's l1: 0.0690945\n",
      "[467]\ttraining's l1: 0.0690905\n",
      "[468]\ttraining's l1: 0.0690857\n",
      "[469]\ttraining's l1: 0.0690821\n",
      "[470]\ttraining's l1: 0.0690785\n",
      "[471]\ttraining's l1: 0.0690743\n",
      "[472]\ttraining's l1: 0.0690699\n",
      "[473]\ttraining's l1: 0.0690654\n",
      "[474]\ttraining's l1: 0.0690608\n",
      "[475]\ttraining's l1: 0.0690566\n",
      "[476]\ttraining's l1: 0.0690524\n",
      "[477]\ttraining's l1: 0.0690482\n",
      "[478]\ttraining's l1: 0.0690438\n",
      "[479]\ttraining's l1: 0.0690403\n",
      "[480]\ttraining's l1: 0.0690361\n",
      "[481]\ttraining's l1: 0.0690312\n",
      "[482]\ttraining's l1: 0.0690277\n",
      "[483]\ttraining's l1: 0.0690238\n",
      "[484]\ttraining's l1: 0.0690191\n",
      "[485]\ttraining's l1: 0.0690148\n",
      "[486]\ttraining's l1: 0.0690107\n",
      "[487]\ttraining's l1: 0.0690062\n",
      "[488]\ttraining's l1: 0.0690009\n",
      "[489]\ttraining's l1: 0.0689976\n",
      "[490]\ttraining's l1: 0.068993\n",
      "[491]\ttraining's l1: 0.068989\n",
      "[492]\ttraining's l1: 0.068985\n",
      "[493]\ttraining's l1: 0.0689801\n",
      "[494]\ttraining's l1: 0.0689765\n",
      "[495]\ttraining's l1: 0.0689724\n",
      "[496]\ttraining's l1: 0.0689655\n",
      "Did not meet early stopping. Best iteration is:\n",
      "[496]\ttraining's l1: 0.0689655\n"
     ]
    }
   ],
   "source": [
    "#打乱数据，训练，利用上面选好的参数和轮数\n",
    "train_data=train_data.sample(frac=1)\n",
    "# create dataset for lightgbm\n",
    "lgb_train = lgb.Dataset(train_data.drop(0,axis=1), train_data.iloc[:,0],weight=[(1/abs(1.5-v)) for v in train_data.iloc[:,0]])\n",
    "# lgb_eval = lgb.Dataset(valid_data.drop(0,axis=1), valid_data.iloc[:,0],weight=[(1/abs(1.5-v)) for v in valid_data.iloc[:,0]])\n",
    "# specify your configurations as a dict\n",
    "print('Start training...')\n",
    "# train\n",
    "gbm = lgb.train(params,\n",
    "                lgb_train,\n",
    "                num_boost_round=num_boost_round,\n",
    "                valid_sets=[lgb_train],\n",
    "                early_stopping_rounds=20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Start predicting...\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "array([0.84832288, 0.7501522 , 0.73763672, ..., 0.64373057, 0.83726125,\n",
       "       0.75197965])"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "print('Start predicting...')\n",
    "y_pred = gbm.predict(test_df.values, num_iteration=gbm.best_iteration)\n",
    "y_pred"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "261003"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(y_pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "# pd.DataFrame({'id':test_correlation.index,'value':[math.tanh(i) for i in y_pred]}).to_csv('%s/lgbm_valid_data1_overfit_1029.csv' % PATH,index=False)\n",
    "pd.DataFrame({'id':test_correlation.index,'value':y_pred}).to_csv('%s/1108_lgbm_valid_data2_jump_validate_简单版特征.csv' % PATH,index=False)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.6.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}