KDB+ like asof join for timeseries data in pandas?

Question

kdb+ has an aj function that is usually used to join tables along time columns.

Here is an example where I have trade and quote tables and I get the prevailing quote for every trade.

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  

How can I do the same operation using pandas? I am working with trade and quote dataframes where the index is 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

I see that pandas has an asof function but that is not defined on the DataFrame, only on the Series object. I guess one could loop through each of the Series and align them one by one, but I am wondering if there is a better way?

Talos

this is also called rolling join – Mcloughlin

Answer 1

As you mentioned in the question, looping through each column should work for you:

df1.apply(lambda x: x.asof(df2.index))

We could potentially create a faster NaN-naive version of DataFrame.asof to do all the columns in one shot. But for now, I think this is the most straightforward way.

Stoss

Thanks. I am taking this approach for now. But a NaN-naive version would be very welcome! – Talos

Answer 2

I wrote an under-advertised ordered_merge function some time ago:

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

It could be easily (well, for someone who is familiar with the code) extended to be a "left join" mimicking KDB. I realize in this case that forward-filling the trade data is not appropriate; just illustrating the function.

Palmate

Thanks, this is very good to know. This is essentially uj (code.kx.com/wiki/Reference/uj) in KDB!. For the aj functionality, I am going with Chang's approach but I plan to take a serious stab at the code later. – Talos

Could this be generalized the case where the dataframe contains many Series together, for example if the data, in addition of timestamps, also had a stock ID column? (Thus we may have thousands of groups, and each of them is a Series). I suspect we'll need a mix of groupby() and ordered_merge, but I'm wrestling about how to do it... For sure, it would be wrong to simply ffill on the overall order of the dataframe (I don't want a group to spill into the next group by virtue of forward filling). – Lefthand

Answer 3

pandas 0.19 has introduced an asof join:

pd.merge_asof(trades, quotes, on='time')

The semantics are very similar to the functionality in q/kdb+.

Catalyze