如何在熊猫中迭代时间段

Question

Consider I have a pandas Series with a DatetimeIndex with daily frequency. I want to iterate over this Series with arbitrary frequency and an arbitrary look-back window. For example: Iterate half-yearly with a lookback window of 1y.

Something like this would be ideal:

for df_year in df.timegroup(freq='6m', lookback='1y'):
   # df_year will span one year of daily prices and be generated every 6 months

I know about TimeGrouper but haven't figured out how it could do this. Anyway, I could just code this manually but was hoping for a clever pandas one-liner.

Edit: This is getting a bit closer:

pd.rolling_apply(df, 252, lambda s: s.sum(), freq=pd.datetools.BMonthEnd())

This doesn't quite work, because it applies a lookback window of 252*BMonthEnd() while I'd like that to be independent and have a lookback window of 252 days every end of the month.

Answer 1

I think this is what you are looking for

Construct a series of a frequency. Using 1 for clarify here.

In [77]: i = pd.date_range('20110101','20150101',freq='B')

In [78]: s = Series(1,index=i)

In [79]: s
Out[79]: 
2011-01-03    1
2011-01-04    1
2011-01-05    1
2011-01-06    1
2011-01-07    1
             ..
2014-12-26    1
2014-12-29    1
2014-12-30    1
2014-12-31    1
2015-01-01    1
Freq: B, dtype: int64

In [80]: len(s)
Out[80]: 1044

Conform the index to another frequency. This makes every index element be the end-of-month here.

In [81]: s.index = s.index.to_period('M').to_timestamp('M')

In [82]: s
Out[82]: 
2011-01-31    1
2011-01-31    1
2011-01-31    1
2011-01-31    1
2011-01-31    1
             ..
2014-12-31    1
2014-12-31    1
2014-12-31    1
2014-12-31    1
2015-01-31    1
dtype: int64

Then its straightforward to resample to another frequency. This gives you the number of business days in the period in this case.

In [83]: s.resample('3M',how='sum')
Out[83]: 
2011-01-31    21
2011-04-30    64
2011-07-31    65
2011-10-31    66
2012-01-31    66
              ..
2014-01-31    66
2014-04-30    63
2014-07-31    66
2014-10-31    66
2015-01-31    44
Freq: 3M, dtype: int64

Answer 2

This solution provides a one liner using list comprehension. Starting from the left of the time series and iterating forward (backward iteration could also be done), the iteration returns a subset of the index equal to the loopback window and jumps in a step size equal to the frequency. Note that the very last period is likely a stub with a length less than the lookback window.

This method uses days rather than month or week offsets.

freq = 30      # Days
lookback = 60  # Days

idx = pd.date_range('2010-01-01', '2015-01-01')
[idx[(freq * n):(lookback + freq * n)] for n in range(int(len(idx) / freq))]

Out[86]: 
[<class 'pandas.tseries.index.DatetimeIndex'>
 [2010-01-01, ..., 2010-03-01]
 Length: 60, Freq: D, Timezone: None,
 <class 'pandas.tseries.index.DatetimeIndex'>
 [2010-01-31, ..., 2010-03-31]
 Length: 60, Freq: D, Timezone: None,
...
 Length: 60, Freq: D, Timezone: None,
 <class 'pandas.tseries.index.DatetimeIndex'>
 [2014-11-06, ..., 2015-01-01]
 Length: 57, Freq: D, Timezone: None]