Question

我有一个熊猫数据框，其中有日期时间（不在索引中，因此更喜欢）。我想将其上采样（重新采样）到指定的时间范围，例如“ 10S”。并将字符串数据（即Acitivty / Action / EPIC等列）也保留在dataFrame中。

Ind TIME_STAMP          Activity    Action  Quantity    EPIC   Price    Sub-activity    Venue   Position
0   2018-08-22 08:01:36 Allocation  SELL    100.0       BB.    1.142200 CPTY     300AD  -427.0
1   2018-08-22 08:02:17 Allocation  BUY     15.0        BB.    1.152300 CPTY    ZDDD02  -388.0
2   2018-08-22 08:24:51 Allocation  SELL    60.0        BB.    1.165900 CPTY    666     -515.0
3   2018-08-22 09:07:59 NaN         NaN     NaN         NaN    1.167921 NaN             -515.0
4   2018-08-22 09:11:00 NaN         NaN     NaN         NaN    1.174500 NaN

我尝试了几种不同的方法，即dataFrame.asfreq（freq ='10S'）;和dataFrame.resample（'10S'，on ='TIME_STAMP'）

我真正想做的是 1）将数据升采样为10秒的块，并保留原始数据； 2）使用“ TIME_STAMP”列。和3）之后，可以使用.fillna（method ='pad'）

之类的填充方法来填充数值数据

Answer 1

想法是由GroupBy.cumcount创建帮助列，由unstack创建唯一的Datetimeindex，最后由stack重新创建：

print (df)
             TIME_STAMP    Activity Action  Quantity EPIC     Price  \
Ind                                                                   
0   2018-08-22 08:01:36  Allocation   SELL     100.0  BB.  1.142200   
1   2018-08-22 08:01:36  Allocation    BUY      15.0  BB.  1.152300   
2   2018-08-22 08:01:51  Allocation   SELL      60.0  BB.  1.165900   
3   2018-08-22 08:02:59         NaN    NaN       NaN  NaN  1.167921   
4   2018-08-22 08:02:59         NaN    NaN       NaN  NaN  1.174500   

    Sub-activity   Venue  Position  
Ind                                 
0           CPTY   300AD    -427.0  
1           CPTY  ZDDD02    -388.0  
2           CPTY     666    -515.0  
3            NaN  -515.0       NaN  
4            NaN     NaN       NaN

df = (df.set_index(['TIME_STAMP', df.groupby('TIME_STAMP').cumcount()])
        .unstack()
        .asfreq('10S', method ='pad')
        .stack()
        .reset_index(level=1, drop=True)
        .sort_index())
print (df)
                       Activity Action  Quantity EPIC   Price Sub-activity  \
TIME_STAMP                                                                   
2018-08-22 08:01:36  Allocation   SELL     100.0  BB.  1.1422         CPTY   
2018-08-22 08:01:36  Allocation    BUY      15.0  BB.  1.1523         CPTY   
2018-08-22 08:01:46  Allocation   SELL     100.0  BB.  1.1422         CPTY   
2018-08-22 08:01:46  Allocation    BUY      15.0  BB.  1.1523         CPTY   
2018-08-22 08:01:56  Allocation   SELL      60.0  BB.  1.1659         CPTY   
2018-08-22 08:02:06  Allocation   SELL      60.0  BB.  1.1659         CPTY   
2018-08-22 08:02:16  Allocation   SELL      60.0  BB.  1.1659         CPTY   
2018-08-22 08:02:26  Allocation   SELL      60.0  BB.  1.1659         CPTY   
2018-08-22 08:02:36  Allocation   SELL      60.0  BB.  1.1659         CPTY   
2018-08-22 08:02:46  Allocation   SELL      60.0  BB.  1.1659         CPTY   
2018-08-22 08:02:56  Allocation   SELL      60.0  BB.  1.1659         CPTY   

                      Venue  Position  
TIME_STAMP                             
2018-08-22 08:01:36   300AD    -427.0  
2018-08-22 08:01:36  ZDDD02    -388.0  
2018-08-22 08:01:46   300AD    -427.0  
2018-08-22 08:01:46  ZDDD02    -388.0  
2018-08-22 08:01:56     666    -515.0  
2018-08-22 08:02:06     666    -515.0  
2018-08-22 08:02:16     666    -515.0  
2018-08-22 08:02:26     666    -515.0  
2018-08-22 08:02:36     666    -515.0  
2018-08-22 08:02:46     666    -515.0  
2018-08-22 08:02:56     666    -515.0

Answer 2

这是强制执行计算的一种惊人的蛮力方式。我通过asfreq和resample生成1秒的时间序列；然后我做必要的计算；最后重新采样到实际的时间频率。

dataT = pd.merge_ordered(VWAP, dataT, on='TIME_STAMP') # this is to merge the market-data stream and order stream
dataT.Quantity = dataT.Quantity.astype('float')
dataT.Price = dataT.Price.astype('float')
dataT.TIME_STAMP = pd.to_datetime(dataT.TIME_STAMP)

dataT['Position'] = dataT.apply(lambda row : (row['Quantity'] if row['Action'] == 'BUY' else -row['Quantity']) if row['Activity'] == 'Allocation' else 0, axis =1).cumsum()
dataT['Price'] = dataT.apply(lambda row : (row['Price'] if pd.isnull(row['Price']) == False else row['VWAP']), axis =1)


dataY = dataT.set_index(['TIME_STAMP']).resample('1S').pad()
dataT = dataT.set_index(['TIME_STAMP']).asfreq(freq='1S', method=None)

dataT.Price = dataT.Price.fillna(method = 'pad')
dataT.Position = dataT.Position.fillna(method = 'ffill')

dataT.Price = dataY.Price
dataT.Quantity = dataT.Quantity.fillna(0)
dataT.Position = dataY.Position
dataT.EPIC = dataT.EPIC.fillna('BB.')

dataT['MtM_Trd'] = (dataT.Quantity * dataT.Price).fillna(0)
dataT['MtM_Pos'] = dataT.Position * dataT.Price
dataT['MtM_PnL'] = (dataT.Price.diff(periods = 1) * dataT.Position.shift(periods=1)).cumsum().fillna(0)

dataT = dataT.reset_index()
dataT['Ret_Cum'] = (dataT.MtM_PnL/dataT.MtM_Trd.abs().cumsum()) * 100
dataT['Ret_Ins'] = ((dataT.Price.diff(periods = 1) * dataT.Position.shift(periods=1))/dataT.MtM_Pos.abs()) * 100
dataT['Std_Cum'] = ((dataT.Ret_Cum**2).cumsum()/(dataT.index.values) - (dataT.Ret_Cum.cumsum()/dataT.index.values)**2).apply(np.sqrt)
dataT['Std_Ins'] = pd.rolling_std(dataT.Ret_Ins, window = 60).fillna(0)
dataT['Sharpe_Cum'] = (dataT.Ret_Cum/dataT.Std_Cum).fillna(0)
dataT['Sharpe_Ins'] = (pd.rolling_mean(dataT.Ret_Ins, window = 60)/pd.rolling_std(dataT.Ret_Ins, window = 60)).fillna(0)

dataT = dataT.resample('10S', on='TIME_STAMP').first()

使用“重复索引”

2 个答案: