10分钟学会Pandas

由xiaoshao创建，最终由xiaoshao更新于2023-11-26 16:58 被浏览 1159 用户

10分钟学会Pandas

Pandas最初被作为金融数据分析工具而开发出来，在金融领域被广泛使用。Pandas纳入了大量库和一些标准的数据模型，提供了高效地操作大型数据集所需的工具、函数和方法。

本文是针对pandas新手的快速入门学习指南。你可以在 AI量化平台-编写策略 里，一步一步的学习和实践。

# 导入库
import pandas as pd
import numpy as np

Pandas数据结构

主要数据结构：Series和DataFrame，Series是一种类似于一维数组的对象，它由一组数据以及一组与之相关的数据标签（即索引）组成。DataFrame是一个表格型 的数据结构，可以简单地理解为是由多个具有相同的索引的Series组成。DataFrame是金融数据分析中特别常用的数据结构，也是我们平台最常遇到的数据结构。

# 创建Series,以日期为索引
dates = pd.date_range('2/6/2017', periods=5, freq='D')
s = pd.Series([1,3,5,np.nan,6,],index=dates)
s

2017-02-06    1.0
2017-02-07    3.0
2017-02-08    5.0
2017-02-09    NaN
2017-02-10    6.0

Freq: D, dtype: float64

# 创建DataFrame，以日期为索引
df = pd.DataFrame(np.random.randn(5,4), index=dates, columns=list('ABCD'))
df

df = DataSource('bar1d_CN_STOCK_A').read(instruments=['000005.SZA'],start_date='2017-01-06',end_date='2017-01-16',fields=['open','high','low','close'])
df = df.set_index('date')

# 选择某一列数据，它会返回一个Series，等同于df.close：
df['close']
df

查看数据

# 显示DataFrame前几行（默认是5）的数据
df.head()

# 显示DataFrame后行的数据
df.tail(2)

显示索引、列名以及底层的numpy数据

# 数据索引
df.index

# output:
DatetimeIndex(['2017-01-06', '2017-01-09', '2017-01-10', '2017-01-11',
​                '2017-01-12', '2017-01-13', '2017-01-16'],
​              dtype='datetime64[ns]', name='date', freq=None)

# 数据列名
df.columns

# output:
Index(['instrument', 'open', 'high', 'low', 'close'], dtype='object')

# 底层numpy数据
df.values

# output:
array([['000005.SZA', 64.50249481201172, 66.07798767089844,
​        64.2244644165039, 65.15122985839844],
​       ['000005.SZA', 64.50249481201172, 65.89263916015625,
​        64.40982055664062, 65.52193450927734],
​       ['000005.SZA', 65.52193450927734, 65.61460876464844,
​        64.96587371826172, 64.96587371826172],
​       ['000005.SZA', 64.96587371826172, 64.96587371826172,
​        63.57573699951172, 63.57573699951172],
​       ['000005.SZA', 63.20503234863281, 63.76108932495117,
​        62.74165344238281, 62.927005767822266],
​       ['000005.SZA', 62.927005767822266, 63.20503234863281,
​        61.907569885253906, 61.907569885253906],
​       ['000005.SZA', 62.18559646606445, 62.27827072143555,
​        56.810386657714844, 59.683345794677734]], dtype=object)

# 数据类型
df.dtypes

# output:

instrument         object
open               float32
high                float32
low                  float32
close               float32
dtype:             object

转置、排序、偏移

# 转置
df.T

# 按轴进行排序
df.sort_index(axis=1, ascending=False)

# 按值进行排序
df.sort_values(by='close',ascending=False) # ascending默认升序，ascending=False降序

shift数据偏移（可理解为回溯）

# 利用pct_change计算每日收益
df = DataSource('bar1d_CN_STOCK_A').read(instruments=['000001.SZA'],start_date='2010-01-01',end_date='2016-12-31',fields=['open','high','low','close']).set_index('date')['close']
rets = df.pct_change()
rets.head()

date
2010-01-04         NaN
2010-01-05   -0.017292
2010-01-06   -0.017167
2010-01-07   -0.010917
2010-01-08   -0.002208
Name: close, dtype: float32

# 计算每日收益,等价于pct_change函数
daily_returns = df / df.shift(1) -1 # 正数表示向前偏移，负数表示向后偏移
daily_returns.head()

date
2010-01-04         NaN
2010-01-05   -0.017292
2010-01-06   -0.017167
2010-01-07   -0.010917
2010-01-08   -0.002208
Name: close, dtype: float32

数据选择

选取

df = DataSource('bar1d_CN_STOCK_A').read(instruments=['000005.SZA'],start_date='2017-01-06',end_date='2017-02-10',fields=['open','high','low','close'])
df = df.set_index('date')

# 选择某一列数据，它会返回一个Series，等同于df.close：
df['close']

date
2017-01-06    65.151230
2017-01-09    65.521935
2017-01-10    64.965874
2017-01-11    63.575737
2017-01-12    62.927006
2017-01-13    61.907570
2017-01-16    59.683346
2017-01-17    60.054047
2017-01-18    60.702782
2017-01-19    59.683346
2017-01-20    60.239399
2017-01-23    60.517429
2017-01-24    60.332077
2017-01-25    60.424751
2017-01-26    60.795456
2017-02-03    60.517429
2017-02-06    60.795456
2017-02-07    61.166161
2017-02-08    61.073483
2017-02-09    61.444187
2017-02-10    61.536865
Name: close, dtype: float32

# 获取2017-02-06到2017-02-09的所有数据
df.loc['20170206':'20170209']

date	instrument	open	high	low	close
2017-02-06	000005.SZA	60.517429	60.795456	60.146725	60.795456
2017-02-07	000005.SZA	60.795456	61.629539	60.424751	61.166161
2017-02-08	000005.SZA	60.980808	61.166161	60.424751	61.073483
2017-02-09	000005.SZA	61.073483	61.536865	60.888130	61.444187

# 获取2017-02-06的所有数据
df.loc['2017-02-06':'2017-02-06']

# 获取2017-02-06的开盘价和收盘价数据
df.loc['2017-02-06':'2017-02-06'][['open','close']]

# 获取2017-02-06和2017-02-09间断两天的开盘价和收盘价数据
df.loc['2017-02-06':'2017-02-09'][['open','close']]

# 获取2017-02-06的开盘价，返回DataFrame
df.loc['2017-02-06':'2017-02-06'][['open']]

date	open
2017-02-06	60.517429

# 获取2017-02-06的收盘价，返回标量
df.loc['2017-02-06','close']

60.517429

布尔索引

# 获取最高价大于64的交易日数据
df[df.high > 64]

where 索引

# 选出大于60的元素
df[df > 60]

丢弃数据

# 丢弃行
df.drop([pd.Timestamp('2017-01-06')]) # 丢弃行

# 丢弃列 ，del df['open'],原地修改
df.drop('open',axis=1)  # 非原地修改

对齐、合并、赋值

# 通过索引，自动对齐两个DataFrame的数据
df1 = DataSource('bar1d_CN_STOCK_A').read(instruments = ['000005.SZA'],start_date='2017-02-01',end_date='2017-02-10',
                    fields=['open','high','low','close'], groupped_by_instrument=False).set_index('date')

df2 = DataSource('market_value_CN_STOCK_A').read(instruments = '000005.SZA',start_date='2017-01-01',end_date='2017-02-10',
                    fields=['market_cap','pe_ttm'], groupped_by_instrument=False).set_index('date')
# 取索引的交集进行合并
df1.merge(df2,left_index=True,right_index=True,how='inner')

# 取索引的并集进行合并
df1.merge(df2,left_index=True,right_index=True,how='outer')

# 通过标签赋值
df1.loc['2017-02-03','market_cap'] = 0

# 通过位置赋值
df1.iat[0,0] ='600519.SHA'
df1

# 通过where操作来赋值
df1[df1['market_cap'] == 0] = 111
df1.head(4)

缺失值处理

在pandas中，用np.nan来代表缺失值，这些值默认不会参与运算。

# 通过set_index设置某一列为索引
df2 = DataSource('bar1d_CN_STOCK_A').read(instruments = ['000005.SZA'],start_date = '2016-11-15',end_date = '2016-12-01',fields=['open','high','low','close'],
                     groupped_by_instrument=False).set_index('date')
df2

# 通过resset_index来将DataFrame的索引转化为列
df2.reset_index()

# 将2016-11-22以后的数据用缺失值代替
df2.loc['2016-11-22':,:] = np.nan
df2

# 剔除所有包含缺失值的行数据
df2.dropna(how='any')

# 填充缺失值
df2.fillna(value=154)

# 获取值是否为nan的布尔标记
pd.isnull(df2)

运算

# describe()能对数据做一个快速统计汇总
df2.describe()

运算过程中，通常不包含缺失值。

# 按列计算平均值
df2.mean()

open     70.211342
high     70.693253
low      69.488464
close    70.118668
dtype: float32

# 按行计算平均值
df2.mean(1)

date
2016-11-15    70.109398
2016-11-16    70.711792
2016-11-17    69.808197
2016-11-18    69.947212
2016-11-21    70.063049
2016-11-22          NaN
2016-11-23          NaN
2016-11-24          NaN
2016-11-25          NaN
2016-11-28          NaN
2016-11-29          NaN
2016-11-30          NaN
2016-12-01          NaN
dtype: float32

# pandas会按照指定的索引对齐
dates = ['2016-11-15','2016-11-16','2016-11-17','2016-11-18','2016-11-23']
for i in range(len(dates)):
    dates[i] = pd.Timestamp(dates[i])
pd.DataFrame(df2,index=dates)

Apply 函数

# 通过apply函数求每个交易日高开低收价格的平均值
df3 = df2[['open','high','low','close']]
df3.apply(np.mean,axis=1)

date
2016-11-15    70.109398
2016-11-16    70.711792
2016-11-17    69.808197
2016-11-18    69.947212
2016-11-21    70.063049
2016-11-22          NaN
2016-11-23          NaN
2016-11-24          NaN
2016-11-25          NaN
2016-11-28          NaN
2016-11-29          NaN
2016-11-30          NaN
2016-12-01          NaN
dtype: float64

# 计算每日股价振幅，振幅=最高价-最低价
df3.apply(lambda x:x.max()-x.min())

open     1.575493
high     1.390144
low      0.648727
close    0.648735
dtype: float64

# 通过apply对各个元素调用字符串方法，比如将股票代码的后缀改成小写
temp = df2.ix['2016-11-15':'2016-11-21']
temp['instrument'].apply(lambda x : x.lower())

date
2016-11-15    000005.sza
2016-11-16    000005.sza
2016-11-17    000005.sza
2016-11-18    000005.sza
2016-11-21    000005.sza
Name: instrument,
dtype: object

rolling_apply的使用

# 获取数据
data = DataSource('bar1d_CN_STOCK_A').read(instruments = ['000001.SZA'],start_date = '2016-01-01',end_date = '2016-10-01',
           fields=['close'],groupped_by_instrument=False).set_index('date')

# 计算5日均线和20日均线
data['MA_' + str(5)] = data['close'].rolling(5).mean()
data['MA_' + str(20)] = data['close'].rolling(20).mean()
data.head(12)

合并

Concat 连接

# 通过concat()来连接pandas对象
a = df2.loc['2016-11-15':'2016-11-16']
b = df2.loc['2016-11-20':'2016-11-26']
pd.concat([a,b])

Join 合并

# 合并，和merge功能类似
df5 = DataSource('bar1d_CN_STOCK_A').read(instruments = ['000005.SZA'],start_date='2017-02-01',end_date='2017-02-10',
                    fields=['open','high','low','close'], groupped_by_instrument=False).set_index('date')

df6 = DataSource('market_value_CN_STOCK_A').read(instruments = ['000005.SZA'],start_date='2017-01-01',end_date='2017-02-10',
                    fields=['market_cap','pe_ttm'], groupped_by_instrument=False).set_index('date')

# 删除股票代码列
del df5['instrument']
del df6['instrument']

# 按日期轴合并数据
df5.join(df6)

Append 添加

# 通过append方法，将若干行添加到dataFrame后面
c = df1.loc['2017-02-03':'2017-02-06']
d = df1.loc['2017-02-08':'2017-02-10']
c.append(d)

数据透视表

也许大多数人都有在Excel中使用数据透视表的经历，其实Pandas也提供了一个类似的功能，名为pivot_table

df = pd.DataFrame({'A' : ['one', 'one', 'two', 'three'] * 3,
                   'B' : ['A', 'B', 'C'] * 4,
                   'C' : ['foo', 'foo', 'foo', 'bar', 'bar', 'bar'] * 2,
                   'D' : np.random.randn(12),
                   'E' : np.random.randn(12)})
df

# 我们可以轻松得到数据透视表
pd.pivot_table(df, values='D', index=['A', 'B'], columns=['C'])

# 查看每个交易日多只股票的收盘价数据（每列为股票）
df = D.history_data(['000001.SZA','000002.SZA'],'2016-12-15','2017-01-01',fields=['close'],groupped_by_instrument=False)
pd.pivot_table(df,values='close',index=['date'],columns=['instrument'])

# 查看不同股票每个交易日的收盘价数据（每列为交易日）
pd.pivot_table(df,values = 'close',index=['instrument'],columns=['date'])

分组

对于“group by”操作，我们通常是指以下一个或几个步骤： * 划分 按照某些标准将数据分为不同的组 * 应用 对每组数据分别执行一个函数 * 组合 将结果组合到一个数据结构

df = pd.DataFrame({'A' : ['foo', 'bar', 'foo', 'bar',
                          'foo', 'bar', 'foo', 'bar'],
                   'B' : ['one', 'one', 'two', 'three',
                          'two', 'two', 'one', 'three'],
                   'C' : np.random.randn(8),
                   'D' : np.random.randn(8)})
df

# 分组并对每个分组应用sum函数
df.groupby('A').sum()

# 按多个列分组形成层级索引，然后应用函数
df.groupby(['A','B']).sum()

# 通过groupby函数来计算各股票和大盘指数的相关系数（按年划分）

# 获取数据
stocks = ['000005.SZA','000010.SZA','603998.SHA','603898.SHA']
df = DataSource('bar1d_CN_STOCK_A').read(instruments = stocks,start_date = '2010-01-01',end_date = '2016-12-31',fields=['close'],groupped_by_instrument=False)

# 每列转化为每日多只股票的收盘价数据
df = pd.pivot_table(df,values='close',index=['date'],columns=['instrument'])

# 计算日收益率
rets = df.pct_change().dropna()

# 定义求相关系数的函数
corr_func = lambda x:x.corrwith(x['000005.SZA'])

# 求每只股票和大盘相关性
by_year = rets.groupby(lambda x:x.year)
by_year.apply(corr_func)

时间序列

pandas在对频率转换进行重新采样时拥有着简单，强大而且高效的功能（例如把按秒采样的数据转换为按5分钟采样的数据）。这在金融领域很常见，但又不限于此。

# 按秒产生5000个时间、起始时间是2017-01-01的时间序列
rng = pd.date_range('1/1/2017', periods=5000, freq='S')
rng

DatetimeIndex(['2017-01-01 00 :00 :00 ', '2017-01-01 00 :00 :01 ',
​               '2017-01-01 00 :00 :02 ', '2017-01-01 00 :00 :03 ',
​               '2017-01-01 00 :00 :04 ', '2017-01-01 00 :00 :05 ',
​               '2017-01-01 00 :00 :06 ', '2017-01-01 00 :00 :07 ',
​               '2017-01-01 00 :00 :08 ', '2017-01-01 00 :00 :09 ',
​               ...
​               '2017-01-01 01 :23 :10 ', '2017-01-01 01 :23 :11 ',
​               '2017-01-01 01 :23 :12 ', '2017-01-01 01 :23 :13 ',
​               '2017-01-01 01 :23 :14 ', '2017-01-01 01 :23 :15 ',
​               '2017-01-01 01 :23 :16 ', '2017-01-01 01 :23 :17 ',
​               '2017-01-01 01 :23 :18 ', '2017-01-01 01 :23 :19 '],
​              dtype='datetime64[ns]', length=5000, freq='S')

# 重采样，计算每五分钟的总和
ts = pd.Series(np.random.randint(0,5000,len(rng)), index=rng)
ts.resample('5Min', how='sum')

2017-01-01 00:00:00    765477
2017-01-01 00:05:00    755782
2017-01-01 00:10:00    759474
2017-01-01 00:15:00    738611
2017-01-01 00:20:00    782165
2017-01-01 00:25:00    741331
2017-01-01 00:30:00    771220
2017-01-01 00:35:00    790782
2017-01-01 00:40:00    739392
2017-01-01 00:45:00    800815
2017-01-01 00:50:00    725954
2017-01-01 00:55:00    725840
2017-01-01 01:00:00    739548
2017-01-01 01:05:00    756409
2017-01-01 01:10:00    727530
2017-01-01 01:15:00    776613
2017-01-01 01:20:00    522169
Freq: 5T, dtype: int64

# 频率转化，将日K线数据装换成周K
df = DataSource('bar1d_CN_STOCK_A').read(instruments = ['000001.SZA'],start_date = '2016-11-01',end_date = '2016-12-31',fields=['close'],groupped_by_instrument=False).set_index('date')[['close']]
df.resample('W').ohlc()

绘图

数据可视化：可交互图表 （文档： 数据可视化）

# 绘制沪深300走势图
history_hs300 = DataSource('bar1d_CN_STOCK_A').read(instruments = ['000300.HIX'], start_date='2016-01-01', end_date='2017-01-01', fields=['close'])
T.plot(history_hs300[['date', 'close']].set_index('date'), chart_type='spline')

# 绘制多只股票走势图
data =DataSource('bar1d_CN_STOCK_A').read(instruments = ['000005.SZA','000010.SZA','603998.SHA','603898.SHA'], start_date='2016-01-01', end_date='2017-01-01', fields=['close'],groupped_by_instrument=False)
data = pd.pivot_table(data,values='close',index=['date'],columns=['instrument'])
T.plot(data, chart_type='spline')