39. Pandas的时间序列数据-resample重采样
在pandas里对时序的频率的调整称之重新采样，即从一个时频调整为另一个时频的操作，可以借助resample的函数来完成。有upsampling和downsampling(高频变低频)两种。resample后的数据类型有类似'groupby'的接口函数可以调用得到相关数据信息。时序数据经resample后返回Resamper Object，而Resampler 是定义在pandas.core.resample模块里的一个类，可以通过dir查看该类的一些接口函数。
liao@liao:~/md$ python
Python 2.7.12 (default, Nov 12 2018, 14:36:49)
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import pandas.core.resample as pcr
>>> dir(pcr.Resampler)
['__bytes__', ......, '_wrap_result', 'agg', 'aggregate', 'apply', 'asfreq', 'ax', 'backfill', 'bfill', 'count', 'ffill', 'fillna', 'first', 'get_group', 'groups', 'indices', 'interpolate', 'last', 'max', 'mean', 'median', 'min', 'ndim', 'nearest', 'ngroups', 'nunique', 'obj', 'ohlc', 'pad', 'pipe', 'plot', 'prod', 'sem', 'size', 'std', 'sum', 'transform', 'var']
可以看出有mean、pad、ohlc、std、fisrt、fillna等接口函数可以对resample后的数据进行处理
39.1 downsampling 下(降)采用处理
以高频时间序列变低频时间粒度变大数据聚合，原来有100个时间点，假设变为低频的10个点，那么会将原数据每10个数据组成一组(bucket)，原来是100个时间点，100个数据，现在是10个时间点，应该有10个数据，那么这10个数据应该是什么呢？可以对每组里的数据的均值mean，或组里的第一个值first、或最后一个last，最为重采样后的数据来进行下一步处理或....。这就是要借助resample后的数据类型调用相应的接口函数来取得。 由于resample函数的参数众多，较为难理解，现在先做一个时序，如下图所示：
import numpy as np
import pandas as pd
c = 21
v = np.arange(1, c)
tx = pd.Series(v)
tx.index = pd.date_range('2018-12-01', periods = 20, freq = "d")
print "tx", "-" * 20, "\n", tx
程序执行结果：
tx --------------------
2018-12-01 1
2018-12-02 2
2018-12-03 3
2018-12-04 4
2018-12-05 5
2018-12-06 6
2018-12-07 7
2018-12-08 8
2018-12-09 9
2018-12-10 10
2018-12-11 11
2018-12-12 12
2018-12-13 13
2018-12-14 14
2018-12-15 15
2018-12-16 16
2018-12-17 17
2018-12-18 18
2018-12-19 19
2018-12-20 20
Freq: D, dtype: int64
程序的执行结果和图是一一对应的，即2018-12-01的数据为1。
好，现在对tx这个时序进行降采样，每4天为一个组进行分段segment，那么可以这样去分组(用数学的区域概念来描述)
- [2018-12-01，2018-12-05)为第一组，这样
2018-12-01可以落在这个区间里， - [2018-12-05， 2018-12-09)为第二组，
- [2018-12-09，2018-12-13)为第三组，
- [2018-12-13，2018-12-17)为第四组，
- [2018-12-17，2018-12-21)为第五组，第五组的日期2018-12-21尽管不在数据里，可以补齐。这样分组的特点的是左闭右开。
当然，也可采用左开右闭的区间描述这几个分组:
- (2018-11-27，2018-12-01]是第一分组，是为了让第一个时间
2018-12-01能落在第一个左开右闭的分组， - (2018-12-01, 2010-12-05]为第二组，
- (2018-12-05, 2010-12-09]为第三组，
- (2018-12-09, 2010-12-013]为第四组，
- (2018-12-13, 2010-12-17]为第五组，
- (2018-12-17, 2010-12-21]为第六组。 这里，多出来的一组是因为第一时间点要落在第一分组里的要求。
import numpy as np
import pandas as pd
v = np.arange(1, 21)
tx = pd.Series(v)
tx.index = pd.date_range('2018-12-01', periods = 20, freq = "d")
print "tx", "-" * 20, "\n", tx
tf = tx.resample("4d").sum()
print "tf closed using default", "-" * 5, "\n",tf
tf = tx.resample("4d", closed = "left").sum()
print "tf closed = 'left' ", "-" * 5, "\n",tf
tf = tx.resample("4d", closed = "right").sum()
print "tf closed = 'right' ", "-" * 5, "\n",tf
程序结果：
tx --------------------
2018-12-01 1
2018-12-02 2
....<省略>....
2018-12-19 19
2018-12-20 20
Freq: D, dtype: int64
tf closed using default -----
2018-12-01 10
2018-12-05 26
2018-12-09 42
2018-12-13 58
2018-12-17 74
dtype: int64
tf closed = 'left' -----
2018-12-01 10
2018-12-05 26
2018-12-09 42
2018-12-13 58
2018-12-17 74
dtype: int64
tf closed = 'right' -----
2018-11-27 1
2018-12-01 14
2018-12-05 30
2018-12-09 46
2018-12-13 62
2018-12-17 57
dtype: int64
从语句
tf = tx.resample("4d").sum()
print "tf closed using default", "-" * 5, "\n",tf
tf = tx.resample("4d", closed = "left").sum()
print "tf closed = 'left' ", "-" * 5, "\n",tf
的输出结果可以看出，resample函数默认closed参数值为left,即左闭右开。所以2018-12-01的输出值10 = 1 + 2 + 3 + 4。2018-12-02的输出值26 = 5 + 6 + 7 + 8。而当resample采用左开右闭时，第一区间里就只有2018-12-01这一天的数据据，所以和为1，奇怪的是第一项数据输出的index不是2018-12-01而是2018-11-27,而第二项输出的index却是2018-12-01,这是为什么？这里得看resample的第二个令人费解的参数label了，label参数是指输出时使用index是用区间的左界值还是右界值呢？例如(a, b]或[a, b)是用左界值a还右边界值b？
import numpy as np
import pandas as pd
v = np.arange(1, 21)
tx = pd.Series(v)
tx.index = pd.date_range('2018-12-01', periods = 20, freq = "d")
print "tx", "-" * 20, "\n", tx
tf = tx.resample("4d").sum()
print "tf closed using default", "-" * 5, "\n",tf
tf = tx.resample("4d", closed = "left").sum()
print "tf closed = 'left' ", "-" * 5, "\n",tf
tf = tx.resample("4d", closed = "right").sum()
print "tf closed = 'right' ", "-" * 5, "\n",tf
tf = tx.resample("4d", closed = "right", label = "right").sum()
print "tf closed = 'right' label = 'right'", "-" * 0, "\n",tf
程序执行结果：
tx --------------------
2018-12-01 1
....<省略>....
2018-12-20 20
Freq: D, dtype: int64
tf closed using default -----
2018-12-01 10
2018-12-05 26
2018-12-09 42
2018-12-13 58
2018-12-17 74
dtype: int64
tf closed = 'left' -----
2018-12-01 10
2018-12-05 26
2018-12-09 42
2018-12-13 58
2018-12-17 74
dtype: int64
tf closed = 'right' -----
2018-11-27 1
2018-12-01 14
2018-12-05 30
2018-12-09 46
2018-12-13 62
2018-12-17 57
dtype: int64
tf closed = 'right' label = 'right'
2018-12-01 1
2018-12-05 14
2018-12-09 30
2018-12-13 46
2018-12-17 62
2018-12-21 57
dtype: int64
从语句
tf = tx.resample("4d", closed = "right", label = "right").sum()
print "tf closed = 'right' label = 'right'", "-" * 0, "\n",tf
的输出结果可以看到，第一项输出的index已经变成了2018-12-01了，求和为1,也是正确的，第二项2018-12-05的value为14即14 = 2 + 3 + 4 + 5也是对的，且有6组数据也是和之前分析是对的。
ohlc函数
在金融领域经常关系开盘、收盘和最高最低价，resample数据后可以进行这样的操作，pandas数据经resample后可以调用嗯ohlc函数得到汇总数据。
import numpy as np
import pandas as pd
v = np.arange(1, 21)
tx = pd.Series(v)
tx.index = pd.date_range('2018-12-01', periods = 20, freq = "d")
print "tx", "-" * 20, "\n", tx
tf = tx.resample("4d", closed = "right", label = "right").ohlc()
print "tf closed = 'right' label = 'right'", "-" * 0, "\n",tf
程序执行结果：
tx --------------------
2018-12-01 1
2018-12-02 2
....<省略>....
2018-12-19 19
2018-12-20 20
Freq: D, dtype: int64
tf closed = 'right' label = 'right'
open high low close
2018-12-01 1 1 1 1
2018-12-05 2 5 2 5
2018-12-09 6 9 6 9
2018-12-13 10 13 10 13
2018-12-17 14 17 14 17
2018-12-21 18 20 18 20
39.2 upsampling上(升)采样处理
低频变高频会出现大量的NaN数据，可以用method指定填充数据的方式。
import numpy as np
import pandas as pd
v = np.arange(1, 21)
#print v
t0 = pd.Series(v, index = pd.date_range('2018-12-01', periods = 20))
#print t0
print "first", "*" * 22
print t0.resample("6H").first()[:10]
print "bfill", "*" * 22
print t0.resample("6H").bfill()[:10]
print "ffill", "*" * 22
print t0.resample("6H").ffill()[:10]
print "interpolate", "*" * 16
print t0.resample("6H").interpolate()[:10]
程序执行结果如下：
first **********************
2018-12-01 00:00:00 1.0
2018-12-01 06:00:00 NaN
2018-12-01 12:00:00 NaN
2018-12-01 18:00:00 NaN
2018-12-02 00:00:00 2.0
2018-12-02 06:00:00 NaN
2018-12-02 12:00:00 NaN
2018-12-02 18:00:00 NaN
2018-12-03 00:00:00 3.0
2018-12-03 06:00:00 NaN
Freq: 6H, dtype: float64
bfill **********************
2018-12-01 00:00:00 1
2018-12-01 06:00:00 2
2018-12-01 12:00:00 2
2018-12-01 18:00:00 2
2018-12-02 00:00:00 2
2018-12-02 06:00:00 3
2018-12-02 12:00:00 3
2018-12-02 18:00:00 3
2018-12-03 00:00:00 3
2018-12-03 06:00:00 4
Freq: 6H, dtype: int32
ffill **********************
2018-12-01 00:00:00 1
2018-12-01 06:00:00 1
2018-12-01 12:00:00 1
2018-12-01 18:00:00 1
2018-12-02 00:00:00 2
2018-12-02 06:00:00 2
2018-12-02 12:00:00 2
2018-12-02 18:00:00 2
2018-12-03 00:00:00 3
2018-12-03 06:00:00 3
Freq: 6H, dtype: int32
interpolate ****************
2018-12-01 00:00:00 1.00
2018-12-01 06:00:00 1.25
2018-12-01 12:00:00 1.50
2018-12-01 18:00:00 1.75
2018-12-02 00:00:00 2.00
2018-12-02 06:00:00 2.25
2018-12-02 12:00:00 2.50
2018-12-02 18:00:00 2.75
2018-12-03 00:00:00 3.00
2018-12-03 06:00:00 3.25
Freq: 6H, dtype: float64
感谢Klang(金浪)智能数据看板klang.org.cn鼎力支持！