sig*_*ker 14 python join time-series kdb pandas
kdb +有一个aj函数,通常用于沿时间列连接表.
这是一个我有交易和报价表的例子,我得到每笔交易的现行报价.
q)5# t
time sym price size
-----------------------------
09:30:00.439 NVDA 13.42 60511
09:30:00.439 NVDA 13.42 60511
09:30:02.332 NVDA 13.42 100
09:30:02.332 NVDA 13.42 100
09:30:02.333 NVDA 13.41 100
q)5# q
time sym bid ask bsize asize
-----------------------------------------
09:30:00.026 NVDA 13.34 13.44 3 16
09:30:00.043 NVDA 13.34 13.44 3 17
09:30:00.121 NVDA 13.36 13.65 1 10
09:30:00.386 NVDA 13.36 13.52 21 1
09:30:00.440 NVDA 13.4 13.44 15 17
q)5# aj[`time; t; q]
time sym price size bid ask bsize asize
-----------------------------------------------------
09:30:00.439 NVDA 13.42 60511 13.36 13.52 21 1
09:30:00.439 NVDA 13.42 60511 13.36 13.52 21 1
09:30:02.332 NVDA 13.42 100 13.34 13.61 1 1
09:30:02.332 NVDA 13.42 100 13.34 13.61 1 1
09:30:02.333 NVDA 13.41 100 13.34 13.51 1 1
如何使用pandas进行相同的操作？我正在使用交易和报价数据框,其中索引是datetime64.
In [55]: quotes.head()
Out[55]:
bid ask bsize asize
2012-09-06 09:30:00.026000 13.34 13.44 3 16
2012-09-06 09:30:00.043000 13.34 13.44 3 17
2012-09-06 09:30:00.121000 13.36 13.65 1 10
2012-09-06 09:30:00.386000 13.36 13.52 21 1
2012-09-06 09:30:00.440000 13.40 13.44 15 17
In [56]: trades.head()
Out[56]:
price size
2012-09-06 09:30:00.439000 13.42 60511
2012-09-06 09:30:00.439000 13.42 60511
2012-09-06 09:30:02.332000 13.42 100
2012-09-06 09:30:02.332000 13.42 100
2012-09-06 09:30:02.333000 13.41 100
我看到pandas有一个asof函数,但是没有在DataFrame上定义,只在Series对象上定义.我猜一个人可以循环遍历每个系列并逐个对齐它们,但我想知道是否有更好的方法？
Wes*_*ney 11
我ordered_merge不久前写了一个宣传不足的功能:
In [27]: quotes
Out[27]:
time bid ask bsize asize
0 2012-09-06 09:30:00.026000 13.34 13.44 3 16
1 2012-09-06 09:30:00.043000 13.34 13.44 3 17
2 2012-09-06 09:30:00.121000 13.36 13.65 1 10
3 2012-09-06 09:30:00.386000 13.36 13.52 21 1
4 2012-09-06 09:30:00.440000 13.40 13.44 15 17
In [28]: trades
Out[28]:
time price size
0 2012-09-06 09:30:00.439000 13.42 60511
1 2012-09-06 09:30:00.439000 13.42 60511
2 2012-09-06 09:30:02.332000 13.42 100
3 2012-09-06 09:30:02.332000 13.42 100
4 2012-09-06 09:30:02.333000 13.41 100
In [29]: ordered_merge(quotes, trades)
Out[29]:
time bid ask bsize asize price size
0 2012-09-06 09:30:00.026000 13.34 13.44 3 16 NaN NaN
1 2012-09-06 09:30:00.043000 13.34 13.44 3 17 NaN NaN
2 2012-09-06 09:30:00.121000 13.36 13.65 1 10 NaN NaN
3 2012-09-06 09:30:00.386000 13.36 13.52 21 1 NaN NaN
4 2012-09-06 09:30:00.439000 NaN NaN NaN NaN 13.42 60511
5 2012-09-06 09:30:00.439000 NaN NaN NaN NaN 13.42 60511
6 2012-09-06 09:30:00.440000 13.40 13.44 15 17 NaN NaN
7 2012-09-06 09:30:02.332000 NaN NaN NaN NaN 13.42 100
8 2012-09-06 09:30:02.332000 NaN NaN NaN NaN 13.42 100
9 2012-09-06 09:30:02.333000 NaN NaN NaN NaN 13.41 100
In [32]: ordered_merge(quotes, trades, fill_method='ffill')
Out[32]:
time bid ask bsize asize price size
0 2012-09-06 09:30:00.026000 13.34 13.44 3 16 NaN NaN
1 2012-09-06 09:30:00.043000 13.34 13.44 3 17 NaN NaN
2 2012-09-06 09:30:00.121000 13.36 13.65 1 10 NaN NaN
3 2012-09-06 09:30:00.386000 13.36 13.52 21 1 NaN NaN
4 2012-09-06 09:30:00.439000 13.36 13.52 21 1 13.42 60511
5 2012-09-06 09:30:00.439000 13.36 13.52 21 1 13.42 60511
6 2012-09-06 09:30:00.440000 13.40 13.44 15 17 13.42 60511
7 2012-09-06 09:30:02.332000 13.40 13.44 15 17 13.42 100
8 2012-09-06 09:30:02.332000 13.40 13.44 15 17 13.42 100
9 2012-09-06 09:30:02.333000 13.40 13.44 15 17 13.41 100
它可以很容易(对于熟悉代码的人)扩展为模仿KDB的"左连接".在这种情况下,我意识到向前填充贸易数据是不合适的; 只是说明功能.
正如您在问题中提到的,循环遍历每个列应该适合您:
df1.apply(lambda x: x.asof(df2.index))
我们可以创建一个更快的NaN-naive版本的DataFrame.asof来一次性完成所有列.但就目前而言,我认为这是最直接的方式.
pd.merge_asof(trades, quotes, on='time')
语义与q/kdb +中的功能非常相似.