Sampling at pre-defined time values

[amelio-vazquez-reina]

Consider a time-series, e.g.:

                                 some_quantity
timestamp                 
2014-10-06 18:00:40.063000-04:00      0.568000
2014-10-06 18:00:41.361000-04:00      3.014770
2014-10-06 18:00:42.896000-04:00      0.878154
2014-10-06 18:00:43.040000-04:00      0.723077
2014-10-06 18:00:44.791000-04:00      0.723077
2014-10-06 18:00:45.496000-04:00      0.309539
2014-10-06 18:00:45.799000-04:00      3.032000
2014-10-06 18:00:47.470000-04:00      3.014770
2014-10-06 18:00:48.092000-04:00      1.584616

and an independent set of timestamps, e.g.:

start_time = datetime.datetime(year   = 2014, 
                               month  = 10, 
                               day    = 6, 
                               hour   = 18, 
                               tzinfo = pytz.timezone('US/Eastern')

# Finish 400 seconds later
end_time    = start_time + datetime.timedelta(seconds=400)

new_timestamps = pd.date_range(start = start_time, 
                                  end   = end_time, 
                                  freq  = '2.5s')

Is there any way to sample the original time-series so that, for each of the new timestamps, we get the sum of values of the original time-series between the new time-stamps?

Another way to put it, is there any way to call resample by passing it a series of timestamps on which we want to do the sampling? e.g. something like my_series.resample(how='sum', new_timestamps)

[rockg]

While it is not using passed timestamps, I don't know if that's necessary to get what you are looking for. I think the below does what you want as your passed timestamps would be every 2.5s.

curve.resample('2500L', how='sum')
2014-10-06 18:00:40           3.582770
2014-10-06 18:00:42.500000    2.324308
2014-10-06 18:00:45           6.356309
2014-10-06 18:00:47.500000    1.584616
Freq: 2500L, dtype: float64

Unfortunately, 2.5s is not supported which I think it probably should be. The below is interpreted as 5 rather than 2.5 (I'm pretty sure there is already a request for this).

curve.resample('2.5s', how='sum')
2014-10-06 18:00:40    5.907078
2014-10-06 18:00:45    7.940925

[amelio-vazquez-reina]

Thank you @rockg I used new_timestamps with a constant 2.5 second separation as an example. I was hoping there would be a way of passing generic timestamps to pd.resample, but I guess there isn't. is there any way to label this as a feature request? Thanks!

[jreback]

@ribonoous

the soln by @rockg looks right. what exactly is the issue?
(aside from not being able to use 2.5s which is already in #8419 (its a bug/feature)

[amelio-vazquez-reina]

Thanks @jreback and @rockg.

@rockg's solution is fine for the example I provided, but I used this example only to convey the problem. In my case, I can't make assumptions about the timestamps (the timestamps may not even come at regular time intervals). They are timestamps arriving asynchronously at the ms scale, and they trigger specific events that we are trying to cross-reference with high precision. Let me know if any of this is still unclear. Thanks.

[mroeschke]

Seems like there hasn't been much interest in this feature over the years so closing