Time Series / Date functionality

pandas has proven very successful as a tool for working with time series data, especially in the financial data analysis space. With the 0.8 release, we have further improved the time series API in pandas by leaps and bounds. Using the new NumPy datetime64 dtype, we have consolidated a large number of features from other Python libraries like scikits.timeseries as well as created a tremendous amount of new functionality for manipulating time series data.

In working with time series data, we will frequently seek to:

  • generate sequences of fixed-frequency dates and time spans
  • conform or convert time series to a particular frequency
  • compute “relative” dates based on various non-standard time increments (e.g. 5 business days before the last business day of the year), or “roll” dates forward or backward

pandas provides a relatively compact and self-contained set of tools for performing the above tasks.

Create a range of dates:

# 72 hours starting with midnight Jan 1st, 2011
In [1]: rng = date_range('1/1/2011', periods=72, freq='H')

In [2]: rng[:5]

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-01 00:00:00, ..., 2011-01-01 04:00:00]
Length: 5, Freq: H, Timezone: None

Index pandas objects with dates:

In [3]: ts = Series(randn(len(rng)), index=rng)

In [4]: ts.head()

2011-01-01 00:00:00    0.469112
2011-01-01 01:00:00   -0.282863
2011-01-01 02:00:00   -1.509059
2011-01-01 03:00:00   -1.135632
2011-01-01 04:00:00    1.212112
Freq: H, dtype: float64

Change frequency and fill gaps:

# to 45 minute frequency and forward fill
In [5]: converted = ts.asfreq('45Min', method='pad')

In [6]: converted.head()

2011-01-01 00:00:00    0.469112
2011-01-01 00:45:00    0.469112
2011-01-01 01:30:00   -0.282863
2011-01-01 02:15:00   -1.509059
2011-01-01 03:00:00   -1.135632
Freq: 45T, dtype: float64

Resample:

# Daily means
In [7]: ts.resample('D', how='mean')

2011-01-01   -0.319569
2011-01-02   -0.337703
2011-01-03    0.117258
Freq: D, dtype: float64

Time Stamps vs. Time Spans

Time-stamped data is the most basic type of timeseries data that associates values with points in time. For pandas objects it means using the points in time to create the index

In [8]: dates = [datetime(2012, 5, 1), datetime(2012, 5, 2), datetime(2012, 5, 3)]

In [9]: ts = Series(np.random.randn(3), dates)

In [10]: type(ts.index)
pandas.tseries.index.DatetimeIndex

In [11]: ts

2012-05-01   -0.410001
2012-05-02   -0.078638
2012-05-03    0.545952
dtype: float64

However, in many cases it is more natural to associate things like change variables with a time span instead.

For example:

In [12]: periods = PeriodIndex([Period('2012-01'), Period('2012-02'),
   ....:                        Period('2012-03')])
   ....: 

In [13]: ts = Series(np.random.randn(3), periods)

In [14]: type(ts.index)
pandas.tseries.period.PeriodIndex

In [15]: ts

2012-01   -1.219217
2012-02   -1.226825
2012-03    0.769804
Freq: M, dtype: float64

Starting with 0.8, pandas allows you to capture both representations and convert between them. Under the hood, pandas represents timestamps using instances of Timestamp and sequences of timestamps using instances of DatetimeIndex. For regular time spans, pandas uses Period objects for scalar values and PeriodIndex for sequences of spans. Better support for irregular intervals with arbitrary start and end points are forth-coming in future releases.

Converting to Timestamps

To convert a Series or list-like object of date-like objects e.g. strings, epochs, or a mixture, you can use the to_datetime function. When passed a Series, this returns a Series (with the same index), while a list-like is converted to a DatetimeIndex:

In [16]: to_datetime(Series(['Jul 31, 2009', '2010-01-10', None]))

0   2009-07-31 00:00:00
1   2010-01-10 00:00:00
2                   NaT
dtype: datetime64[ns]

In [17]: to_datetime(['2005/11/23', '2010.12.31'])

<class 'pandas.tseries.index.DatetimeIndex'>
[2005-11-23 00:00:00, 2010-12-31 00:00:00]
Length: 2, Freq: None, Timezone: None

If you use dates which start with the day first (i.e. European style), you can pass the dayfirst flag:

In [18]: to_datetime(['04-01-2012 10:00'], dayfirst=True)

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-01-04 10:00:00]
Length: 1, Freq: None, Timezone: None

In [19]: to_datetime(['14-01-2012', '01-14-2012'], dayfirst=True)

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-01-14 00:00:00, 2012-01-14 00:00:00]
Length: 2, Freq: None, Timezone: None

Warning

You see in the above example that dayfirst isn’t strict, so if a date can’t be parsed with the day being first it will be parsed as if dayfirst were False.

Invalid Data

Pass coerce=True to convert invalid data to NaT (not a time):

In [20]: to_datetime(['2009-07-31', 'asd'])
array(['2009-07-31', 'asd'], dtype=object)

In [21]: to_datetime(['2009-07-31', 'asd'], coerce=True)

<class 'pandas.tseries.index.DatetimeIndex'>
[2009-07-31 00:00:00, NaT]
Length: 2, Freq: None, Timezone: None

Take care, to_datetime may not act as you expect on mixed data:

In [22]: to_datetime([1, '1'])
array([1, '1'], dtype=object)

Epoch Timestamps

It’s also possible to convert integer or float epoch times. The default unit for these is nanoseconds (since these are how Timestamps are stored). However, often epochs are stored in another unit which can be specified:

Typical epoch stored units

In [23]: to_datetime([1349720105, 1349806505, 1349892905,
   ....:              1349979305, 1350065705], unit='s')
   ....: 

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-10-08 18:15:05, ..., 2012-10-12 18:15:05]
Length: 5, Freq: None, Timezone: None

In [24]: to_datetime([1349720105100, 1349720105200, 1349720105300,
   ....:              1349720105400, 1349720105500 ], unit='ms')
   ....: 

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-10-08 18:15:05.100000, ..., 2012-10-08 18:15:05.500000]
Length: 5, Freq: None, Timezone: None

These work, but the results may be unexpected.

In [25]: to_datetime([1])

<class 'pandas.tseries.index.DatetimeIndex'>
[1970-01-01 00:00:00.000000001]
Length: 1, Freq: None, Timezone: None

In [26]: to_datetime([1, 3.14], unit='s')

<class 'pandas.tseries.index.DatetimeIndex'>
[1970-01-01 00:00:01, 1970-01-01 00:00:03.140000]
Length: 2, Freq: None, Timezone: None

Note

Epoch times will be rounded to the nearest nanosecond.

Generating Ranges of Timestamps

To generate an index with time stamps, you can use either the DatetimeIndex or Index constructor and pass in a list of datetime objects:

In [27]: dates = [datetime(2012, 5, 1), datetime(2012, 5, 2), datetime(2012, 5, 3)]

In [28]: index = DatetimeIndex(dates)

In [29]: index # Note the frequency information

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-05-01 00:00:00, ..., 2012-05-03 00:00:00]
Length: 3, Freq: None, Timezone: None

In [30]: index = Index(dates)

In [31]: index # Automatically converted to DatetimeIndex

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-05-01 00:00:00, ..., 2012-05-03 00:00:00]
Length: 3, Freq: None, Timezone: None

Practically, this becomes very cumbersome because we often need a very long index with a large number of timestamps. If we need timestamps on a regular frequency, we can use the pandas functions date_range and bdate_range to create timestamp indexes.

In [32]: index = date_range('2000-1-1', periods=1000, freq='M')

In [33]: index

<class 'pandas.tseries.index.DatetimeIndex'>
[2000-01-31 00:00:00, ..., 2083-04-30 00:00:00]
Length: 1000, Freq: M, Timezone: None

In [34]: index = bdate_range('2012-1-1', periods=250)

In [35]: index

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-01-02 00:00:00, ..., 2012-12-14 00:00:00]
Length: 250, Freq: B, Timezone: None

Convenience functions like date_range and bdate_range utilize a variety of frequency aliases. The default frequency for date_range is a calendar day while the default for bdate_range is a business day

In [36]: start = datetime(2011, 1, 1)

In [37]: end = datetime(2012, 1, 1)

In [38]: rng = date_range(start, end)

In [39]: rng

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-01 00:00:00, ..., 2012-01-01 00:00:00]
Length: 366, Freq: D, Timezone: None

In [40]: rng = bdate_range(start, end)

In [41]: rng

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-03 00:00:00, ..., 2011-12-30 00:00:00]
Length: 260, Freq: B, Timezone: None

date_range and bdate_range makes it easy to generate a range of dates using various combinations of parameters like start, end, periods, and freq:

In [42]: date_range(start, end, freq='BM')

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-31 00:00:00, ..., 2011-12-30 00:00:00]
Length: 12, Freq: BM, Timezone: None

In [43]: date_range(start, end, freq='W')

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-02 00:00:00, ..., 2012-01-01 00:00:00]
Length: 53, Freq: W-SUN, Timezone: None

In [44]: bdate_range(end=end, periods=20)

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-12-05 00:00:00, ..., 2011-12-30 00:00:00]
Length: 20, Freq: B, Timezone: None

In [45]: bdate_range(start=start, periods=20)

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-03 00:00:00, ..., 2011-01-28 00:00:00]
Length: 20, Freq: B, Timezone: None

The start and end dates are strictly inclusive. So it will not generate any dates outside of those dates if specified.

DatetimeIndex

One of the main uses for DatetimeIndex is as an index for pandas objects. The DatetimeIndex class contains many timeseries related optimizations:

  • A large range of dates for various offsets are pre-computed and cached under the hood in order to make generating subsequent date ranges very fast (just have to grab a slice)
  • Fast shifting using the shift and tshift method on pandas objects
  • Unioning of overlapping DatetimeIndex objects with the same frequency is very fast (important for fast data alignment)
  • Quick access to date fields via properties such as year, month, etc.
  • Regularization functions like snap and very fast asof logic

DatetimeIndex objects has all the basic functionality of regular Index objects and a smorgasbord of advanced timeseries-specific methods for easy frequency processing.

Note

While pandas does not force you to have a sorted date index, some of these methods may have unexpected or incorrect behavior if the dates are unsorted. So please be careful.

DatetimeIndex can be used like a regular index and offers all of its intelligent functionality like selection, slicing, etc.

In [46]: rng = date_range(start, end, freq='BM')

In [47]: ts = Series(randn(len(rng)), index=rng)

In [48]: ts.index

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-31 00:00:00, ..., 2011-12-30 00:00:00]
Length: 12, Freq: BM, Timezone: None

In [49]: ts[:5].index

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-31 00:00:00, ..., 2011-05-31 00:00:00]
Length: 5, Freq: BM, Timezone: None

In [50]: ts[::2].index

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-31 00:00:00, ..., 2011-11-30 00:00:00]
Length: 6, Freq: 2BM, Timezone: None

Partial String Indexing

You can pass in dates and strings that parse to dates as indexing parameters:

In [51]: ts['1/31/2011']
-1.2812473076599531

In [52]: ts[datetime(2011, 12, 25):]

2011-12-30    0.687738
Freq: BM, dtype: float64

In [53]: ts['10/31/2011':'12/31/2011']

2011-10-31    0.149748
2011-11-30   -0.732339
2011-12-30    0.687738
Freq: BM, dtype: float64

To provide convenience for accessing longer time series, you can also pass in the year or year and month as strings:

In [54]: ts['2011']

2011-01-31   -1.281247
2011-02-28   -0.727707
2011-03-31   -0.121306
2011-04-29   -0.097883
2011-05-31    0.695775
2011-06-30    0.341734
2011-07-29    0.959726
2011-08-31   -1.110336
2011-09-30   -0.619976
2011-10-31    0.149748
2011-11-30   -0.732339
2011-12-30    0.687738
Freq: BM, dtype: float64

In [55]: ts['2011-6']

2011-06-30    0.341734
Freq: BM, dtype: float64

This type of slicing will work on a DataFrame with a DateTimeIndex as well. Since the partial string selection is a form of label slicing, the endpoints will be included. This would include matching times on an included date. Here’s an example:

In [56]: dft = DataFrame(randn(100000,1),columns=['A'],index=date_range('20130101',periods=100000,freq='T'))

In [57]: dft

                            A
2013-01-01 00:00:00  0.176444
2013-01-01 00:01:00  0.403310
2013-01-01 00:02:00 -0.154951
2013-01-01 00:03:00  0.301624
2013-01-01 00:04:00 -2.179861
2013-01-01 00:05:00 -1.369849
2013-01-01 00:06:00 -0.954208
2013-01-01 00:07:00  1.462696
2013-01-01 00:08:00 -1.743161
2013-01-01 00:09:00 -0.826591
2013-01-01 00:10:00 -0.345352
2013-01-01 00:11:00  1.314232
2013-01-01 00:12:00  0.690579
2013-01-01 00:13:00  0.995761
2013-01-01 00:14:00  2.396780
                          ...
[100000 rows x 1 columns]

In [58]: dft['2013']

                            A
2013-01-01 00:00:00  0.176444
2013-01-01 00:01:00  0.403310
2013-01-01 00:02:00 -0.154951
2013-01-01 00:03:00  0.301624
2013-01-01 00:04:00 -2.179861
2013-01-01 00:05:00 -1.369849
2013-01-01 00:06:00 -0.954208
2013-01-01 00:07:00  1.462696
2013-01-01 00:08:00 -1.743161
2013-01-01 00:09:00 -0.826591
2013-01-01 00:10:00 -0.345352
2013-01-01 00:11:00  1.314232
2013-01-01 00:12:00  0.690579
2013-01-01 00:13:00  0.995761
2013-01-01 00:14:00  2.396780
                          ...
[100000 rows x 1 columns]

This starts on the very first time in the month, and includes the last date & time for the month

In [59]: dft['2013-1':'2013-2']

                            A
2013-01-01 00:00:00  0.176444
2013-01-01 00:01:00  0.403310
2013-01-01 00:02:00 -0.154951
2013-01-01 00:03:00  0.301624
2013-01-01 00:04:00 -2.179861
2013-01-01 00:05:00 -1.369849
2013-01-01 00:06:00 -0.954208
2013-01-01 00:07:00  1.462696
2013-01-01 00:08:00 -1.743161
2013-01-01 00:09:00 -0.826591
2013-01-01 00:10:00 -0.345352
2013-01-01 00:11:00  1.314232
2013-01-01 00:12:00  0.690579
2013-01-01 00:13:00  0.995761
2013-01-01 00:14:00  2.396780
                          ...
[84960 rows x 1 columns]

This specifies a stop time that includes all of the times on the last day

In [60]: dft['2013-1':'2013-2-28']

                            A
2013-01-01 00:00:00  0.176444
2013-01-01 00:01:00  0.403310
2013-01-01 00:02:00 -0.154951
2013-01-01 00:03:00  0.301624
2013-01-01 00:04:00 -2.179861
2013-01-01 00:05:00 -1.369849
2013-01-01 00:06:00 -0.954208
2013-01-01 00:07:00  1.462696
2013-01-01 00:08:00 -1.743161
2013-01-01 00:09:00 -0.826591
2013-01-01 00:10:00 -0.345352
2013-01-01 00:11:00  1.314232
2013-01-01 00:12:00  0.690579
2013-01-01 00:13:00  0.995761
2013-01-01 00:14:00  2.396780
                          ...
[84960 rows x 1 columns]

This specifies an exact stop time (and is not the same as the above)

In [61]: dft['2013-1':'2013-2-28 00:00:00']

                            A
2013-01-01 00:00:00  0.176444
2013-01-01 00:01:00  0.403310
2013-01-01 00:02:00 -0.154951
2013-01-01 00:03:00  0.301624
2013-01-01 00:04:00 -2.179861
2013-01-01 00:05:00 -1.369849
2013-01-01 00:06:00 -0.954208
2013-01-01 00:07:00  1.462696
2013-01-01 00:08:00 -1.743161
2013-01-01 00:09:00 -0.826591
2013-01-01 00:10:00 -0.345352
2013-01-01 00:11:00  1.314232
2013-01-01 00:12:00  0.690579
2013-01-01 00:13:00  0.995761
2013-01-01 00:14:00  2.396780
                          ...
[83521 rows x 1 columns]

We are stopping on the included end-point as its part of the index

In [62]: dft['2013-1-15':'2013-1-15 12:30:00']

                            A
2013-01-15 00:00:00  0.501288
2013-01-15 00:01:00 -0.605198
2013-01-15 00:02:00  0.215146
2013-01-15 00:03:00  0.924732
2013-01-15 00:04:00 -2.228519
2013-01-15 00:05:00  1.517331
2013-01-15 00:06:00 -1.188774
2013-01-15 00:07:00  0.251617
2013-01-15 00:08:00 -0.775668
2013-01-15 00:09:00  0.521086
2013-01-15 00:10:00  2.030114
2013-01-15 00:11:00 -0.250333
2013-01-15 00:12:00 -1.158353
2013-01-15 00:13:00  0.685205
2013-01-15 00:14:00 -0.089428
                          ...
[751 rows x 1 columns]

Warning

The following selection will raise a KeyError; otherwise this selection methodology would be inconsistent with other selection methods in pandas (as this is not a slice, nor does it resolve to one)

dft['2013-1-15 12:30:00']

To select a single row, use .loc

In [63]: dft.loc['2013-1-15 12:30:00']

A    0.193284
Name: 2013-01-15 12:30:00, dtype: float64

Datetime Indexing

Indexing a DateTimeIndex with a partial string depends on the “accuracy” of the period, in other words how specific the interval is in relation to the frequency of the index. In contrast, indexing with datetime objects is exact, because the objects have exact meaning. These also follow the sematics of including both endpoints.

These datetime objects are specific hours, minutes, and seconds even though they were not explicity specified (they are 0).

In [64]: dft[datetime(2013, 1, 1):datetime(2013,2,28)]

                            A
2013-01-01 00:00:00  0.176444
2013-01-01 00:01:00  0.403310
2013-01-01 00:02:00 -0.154951
2013-01-01 00:03:00  0.301624
2013-01-01 00:04:00 -2.179861
2013-01-01 00:05:00 -1.369849
2013-01-01 00:06:00 -0.954208
2013-01-01 00:07:00  1.462696
2013-01-01 00:08:00 -1.743161
2013-01-01 00:09:00 -0.826591
2013-01-01 00:10:00 -0.345352
2013-01-01 00:11:00  1.314232
2013-01-01 00:12:00  0.690579
2013-01-01 00:13:00  0.995761
2013-01-01 00:14:00  2.396780
                          ...
[83521 rows x 1 columns]

With no defaults.

In [65]: dft[datetime(2013, 1, 1, 10, 12, 0):datetime(2013, 2, 28, 10, 12, 0)]

                            A
2013-01-01 10:12:00 -0.246733
2013-01-01 10:13:00 -1.429225
2013-01-01 10:14:00 -1.265339
2013-01-01 10:15:00  0.710986
2013-01-01 10:16:00 -0.818200
2013-01-01 10:17:00  0.543542
2013-01-01 10:18:00  1.577713
2013-01-01 10:19:00 -0.316630
2013-01-01 10:20:00 -0.773194
2013-01-01 10:21:00 -1.615112
2013-01-01 10:22:00  0.965363
2013-01-01 10:23:00 -0.882845
2013-01-01 10:24:00 -1.861244
2013-01-01 10:25:00 -0.742435
2013-01-01 10:26:00 -1.937111
                          ...
[83521 rows x 1 columns]

Truncating & Fancy Indexing

A truncate convenience function is provided that is equivalent to slicing:

In [66]: ts.truncate(before='10/31/2011', after='12/31/2011')

2011-10-31    0.149748
2011-11-30   -0.732339
2011-12-30    0.687738
Freq: BM, dtype: float64

Even complicated fancy indexing that breaks the DatetimeIndex’s frequency regularity will result in a DatetimeIndex (but frequency is lost):

In [67]: ts[[0, 2, 6]].index

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-31 00:00:00, ..., 2011-07-29 00:00:00]
Length: 3, Freq: None, Timezone: None

DateOffset objects

In the preceding examples, we created DatetimeIndex objects at various frequencies by passing in frequency strings like ‘M’, ‘W’, and ‘BM to the freq keyword. Under the hood, these frequency strings are being translated into an instance of pandas DateOffset, which represents a regular frequency increment. Specific offset logic like “month”, “business day”, or “one hour” is represented in its various subclasses.

Class name Description
DateOffset Generic offset class, defaults to 1 calendar day
BDay business day (weekday)
CDay custom business day (experimental)
Week one week, optionally anchored on a day of the week
WeekOfMonth the x-th day of the y-th week of each month
LastWeekOfMonth the x-th day of the last week of each month
MonthEnd calendar month end
MonthBegin calendar month begin
BMonthEnd business month end
BMonthBegin business month begin
QuarterEnd calendar quarter end
QuarterBegin calendar quarter begin
BQuarterEnd business quarter end
BQuarterBegin business quarter begin
FY5253Quarter retail (aka 52-53 week) quarter
YearEnd calendar year end
YearBegin calendar year begin
BYearEnd business year end
BYearBegin business year begin
FY5253 retail (aka 52-53 week) year
Hour one hour
Minute one minute
Second one second
Milli one millisecond
Micro one microsecond

The basic DateOffset takes the same arguments as dateutil.relativedelta, which works like:

In [68]: d = datetime(2008, 8, 18)

In [69]: d + relativedelta(months=4, days=5)
datetime.datetime(2008, 12, 23, 0, 0)

We could have done the same thing with DateOffset:

In [70]: from pandas.tseries.offsets import *

In [71]: d + DateOffset(months=4, days=5)
Timestamp('2008-12-23 00:00:00', tz=None)

The key features of a DateOffset object are:

  • it can be added / subtracted to/from a datetime object to obtain a shifted date
  • it can be multiplied by an integer (positive or negative) so that the increment will be applied multiple times
  • it has rollforward and rollback methods for moving a date forward or backward to the next or previous “offset date”

Subclasses of DateOffset define the apply function which dictates custom date increment logic, such as adding business days:

class BDay(DateOffset):
    """DateOffset increments between business days"""
    def apply(self, other):
        ...
In [72]: d - 5 * BDay()
Timestamp('2008-08-11 00:00:00', tz=None)

In [73]: d + BMonthEnd()
Timestamp('2008-08-29 00:00:00', tz=None)

The rollforward and rollback methods do exactly what you would expect:

In [74]: d
datetime.datetime(2008, 8, 18, 0, 0)

In [75]: offset = BMonthEnd()

In [76]: offset.rollforward(d)
Timestamp('2008-08-29 00:00:00', tz=None)

In [77]: offset.rollback(d)
datetime.datetime(2008, 7, 31, 0, 0)

It’s definitely worth exploring the pandas.tseries.offsets module and the various docstrings for the classes.

Parametric offsets

Some of the offsets can be “parameterized” when created to result in different behavior. For example, the Week offset for generating weekly data accepts a weekday parameter which results in the generated dates always lying on a particular day of the week:

In [78]: d + Week()
datetime.datetime(2008, 8, 25, 0, 0)

In [79]: d + Week(weekday=4)
Timestamp('2008-08-22 00:00:00', tz=None)

In [80]: (d + Week(weekday=4)).weekday()
4

Another example is parameterizing YearEnd with the specific ending month:

In [81]: d + YearEnd()
Timestamp('2008-12-31 00:00:00', tz=None)

In [82]: d + YearEnd(month=6)
Timestamp('2009-06-30 00:00:00', tz=None)

Custom Business Days (Experimental)

The CDay or CustomBusinessDay class provides a parametric BusinessDay class which can be used to create customized business day calendars which account for local holidays and local weekend conventions.

In [83]: from pandas.tseries.offsets import CustomBusinessDay

# As an interesting example, let's look at Egypt where
# a Friday-Saturday weekend is observed.
In [84]: weekmask_egypt = 'Sun Mon Tue Wed Thu'

# They also observe International Workers' Day so let's
# add that for a couple of years
In [85]: holidays = ['2012-05-01', datetime(2013, 5, 1), np.datetime64('2014-05-01')]

In [86]: bday_egypt = CustomBusinessDay(holidays=holidays, weekmask=weekmask_egypt)

In [87]: dt = datetime(2013, 4, 30)

In [88]: print(dt + 2 * bday_egypt)
2013-05-05 00:00:00

In [89]: dts = date_range(dt, periods=5, freq=bday_egypt).to_series()

In [90]: print(dts)
2013-04-30   2013-04-30 00:00:00
2013-05-02   2013-05-02 00:00:00
2013-05-05   2013-05-05 00:00:00
2013-05-06   2013-05-06 00:00:00
2013-05-07   2013-05-07 00:00:00
Freq: C, dtype: datetime64[ns]

In [91]: print(Series(dts.weekday, dts).map(Series('Mon Tue Wed Thu Fri Sat Sun'.split())))
2013-04-30    Tue
2013-05-02    Thu
2013-05-05    Sun
2013-05-06    Mon
2013-05-07    Tue
dtype: object

Note

The frequency string ‘C’ is used to indicate that a CustomBusinessDay DateOffset is used, it is important to note that since CustomBusinessDay is a parameterised type, instances of CustomBusinessDay may differ and this is not detectable from the ‘C’ frequency string. The user therefore needs to ensure that the ‘C’ frequency string is used consistently within the user’s application.

Note

This uses the numpy.busdaycalendar API introduced in Numpy 1.7 and therefore requires Numpy 1.7.0 or newer.

Warning

There are known problems with the timezone handling in Numpy 1.7 and users should therefore use this experimental(!) feature with caution and at their own risk.

To the extent that the datetime64 and busdaycalendar APIs in Numpy have to change to fix the timezone issues, the behaviour of the CustomBusinessDay class may have to change in future versions.

Offset Aliases

A number of string aliases are given to useful common time series frequencies. We will refer to these aliases as offset aliases (referred to as time rules prior to v0.8.0).

Alias Description
B business day frequency
C custom business day frequency (experimental)
D calendar day frequency
W weekly frequency
M month end frequency
BM business month end frequency
MS month start frequency
BMS business month start frequency
Q quarter end frequency
BQ business quarter endfrequency
QS quarter start frequency
BQS business quarter start frequency
A year end frequency
BA business year end frequency
AS year start frequency
BAS business year start frequency
H hourly frequency
T minutely frequency
S secondly frequency
L milliseonds
U microseconds

Combining Aliases

As we have seen previously, the alias and the offset instance are fungible in most functions:

In [92]: date_range(start, periods=5, freq='B')

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-03 00:00:00, ..., 2011-01-07 00:00:00]
Length: 5, Freq: B, Timezone: None

In [93]: date_range(start, periods=5, freq=BDay())

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-03 00:00:00, ..., 2011-01-07 00:00:00]
Length: 5, Freq: B, Timezone: None

You can combine together day and intraday offsets:

In [94]: date_range(start, periods=10, freq='2h20min')

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-01 00:00:00, ..., 2011-01-01 21:00:00]
Length: 10, Freq: 140T, Timezone: None

In [95]: date_range(start, periods=10, freq='1D10U')

<class 'pandas.tseries.index.DatetimeIndex'>
[2011-01-01 00:00:00, ..., 2011-01-10 00:00:00.000090]
Length: 10, Freq: 86400000010U, Timezone: None

Anchored Offsets

For some frequencies you can specify an anchoring suffix:

Alias Description
W-SUN weekly frequency (sundays). Same as ‘W’
W-MON weekly frequency (mondays)
W-TUE weekly frequency (tuesdays)
W-WED weekly frequency (wednesdays)
W-THU weekly frequency (thursdays)
W-FRI weekly frequency (fridays)
W-SAT weekly frequency (saturdays)
(B)Q(S)-DEC quarterly frequency, year ends in December. Same as ‘Q’
(B)Q(S)-JAN quarterly frequency, year ends in January
(B)Q(S)-FEB quarterly frequency, year ends in February
(B)Q(S)-MAR quarterly frequency, year ends in March
(B)Q(S)-APR quarterly frequency, year ends in April
(B)Q(S)-MAY quarterly frequency, year ends in May
(B)Q(S)-JUN quarterly frequency, year ends in June
(B)Q(S)-JUL quarterly frequency, year ends in July
(B)Q(S)-AUG quarterly frequency, year ends in August
(B)Q(S)-SEP quarterly frequency, year ends in September
(B)Q(S)-OCT quarterly frequency, year ends in October
(B)Q(S)-NOV quarterly frequency, year ends in November
(B)A(S)-DEC annual frequency, anchored end of December. Same as ‘A’
(B)A(S)-JAN annual frequency, anchored end of January
(B)A(S)-FEB annual frequency, anchored end of February
(B)A(S)-MAR annual frequency, anchored end of March
(B)A(S)-APR annual frequency, anchored end of April
(B)A(S)-MAY annual frequency, anchored end of May
(B)A(S)-JUN annual frequency, anchored end of June
(B)A(S)-JUL annual frequency, anchored end of July
(B)A(S)-AUG annual frequency, anchored end of August
(B)A(S)-SEP annual frequency, anchored end of September
(B)A(S)-OCT annual frequency, anchored end of October
(B)A(S)-NOV annual frequency, anchored end of November

These can be used as arguments to date_range, bdate_range, constructors for DatetimeIndex, as well as various other timeseries-related functions in pandas.

Legacy Aliases

Note that prior to v0.8.0, time rules had a slightly different look. Pandas will continue to support the legacy time rules for the time being but it is strongly recommended that you switch to using the new offset aliases.

Legacy Time Rule Offset Alias
WEEKDAY B
EOM BM
W@MON W-MON
W@TUE W-TUE
W@WED W-WED
W@THU W-THU
W@FRI W-FRI
W@SAT W-SAT
W@SUN W-SUN
Q@JAN BQ-JAN
Q@FEB BQ-FEB
Q@MAR BQ-MAR
A@JAN BA-JAN
A@FEB BA-FEB
A@MAR BA-MAR
A@APR BA-APR
A@MAY BA-MAY
A@JUN BA-JUN
A@JUL BA-JUL
A@AUG BA-AUG
A@SEP BA-SEP
A@OCT BA-OCT
A@NOV BA-NOV
A@DEC BA-DEC
min T
ms L
us U

As you can see, legacy quarterly and annual frequencies are business quarter and business year ends. Please also note the legacy time rule for milliseconds ms versus the new offset alias for month start MS. This means that offset alias parsing is case sensitive.

Up- and downsampling

With 0.8, pandas introduces simple, powerful, and efficient functionality for performing resampling operations during frequency conversion (e.g., converting secondly data into 5-minutely data). This is extremely common in, but not limited to, financial applications.

See some cookbook examples for some advanced strategies

In [106]: rng = date_range('1/1/2012', periods=100, freq='S')

In [107]: ts = Series(randint(0, 500, len(rng)), index=rng)

In [108]: ts.resample('5Min', how='sum')

2012-01-01    25103
Freq: 5T, dtype: int64

The resample function is very flexible and allows you to specify many different parameters to control the frequency conversion and resampling operation.

The how parameter can be a function name or numpy array function that takes an array and produces aggregated values:

In [109]: ts.resample('5Min') # default is mean

2012-01-01    251.03
Freq: 5T, dtype: float64

In [110]: ts.resample('5Min', how='ohlc')

            open  high  low  close
2012-01-01   308   460    9    205
[1 rows x 4 columns]

In [111]: ts.resample('5Min', how=np.max)

2012-01-01    460
Freq: 5T, dtype: int64

Any function available via dispatching can be given to the how parameter by name, including sum, mean, std, max, min, median, first, last, ohlc.

For downsampling, closed can be set to ‘left’ or ‘right’ to specify which end of the interval is closed:

In [112]: ts.resample('5Min', closed='right')

2011-12-31 23:55:00    308.000000
2012-01-01 00:00:00    250.454545
Freq: 5T, dtype: float64

In [113]: ts.resample('5Min', closed='left')

2012-01-01    251.03
Freq: 5T, dtype: float64

For upsampling, the fill_method and limit parameters can be specified to interpolate over the gaps that are created:

# from secondly to every 250 milliseconds
In [114]: ts[:2].resample('250L')

2012-01-01 00:00:00           308
2012-01-01 00:00:00.250000    NaN
2012-01-01 00:00:00.500000    NaN
2012-01-01 00:00:00.750000    NaN
2012-01-01 00:00:01           204
Freq: 250L, dtype: float64

In [115]: ts[:2].resample('250L', fill_method='pad')

2012-01-01 00:00:00           308
2012-01-01 00:00:00.250000    308
2012-01-01 00:00:00.500000    308
2012-01-01 00:00:00.750000    308
2012-01-01 00:00:01           204
Freq: 250L, dtype: int64

In [116]: ts[:2].resample('250L', fill_method='pad', limit=2)

2012-01-01 00:00:00           308
2012-01-01 00:00:00.250000    308
2012-01-01 00:00:00.500000    308
2012-01-01 00:00:00.750000    NaN
2012-01-01 00:00:01           204
Freq: 250L, dtype: float64

Parameters like label and loffset are used to manipulate the resulting labels. label specifies whether the result is labeled with the beginning or the end of the interval. loffset performs a time adjustment on the output labels.

In [117]: ts.resample('5Min') # by default label='right'

2012-01-01    251.03
Freq: 5T, dtype: float64

In [118]: ts.resample('5Min', label='left')

2012-01-01    251.03
Freq: 5T, dtype: float64

In [119]: ts.resample('5Min', label='left', loffset='1s')

2012-01-01 00:00:01    251.03
dtype: float64

The axis parameter can be set to 0 or 1 and allows you to resample the specified axis for a DataFrame.

kind can be set to ‘timestamp’ or ‘period’ to convert the resulting index to/from time-stamp and time-span representations. By default resample retains the input representation.

convention can be set to ‘start’ or ‘end’ when resampling period data (detail below). It specifies how low frequency periods are converted to higher frequency periods.

Note that 0.8 marks a watershed in the timeseries functionality in pandas. In previous versions, resampling had to be done using a combination of date_range, groupby with asof, and then calling an aggregation function on the grouped object. This was not nearly convenient or performant as the new pandas timeseries API.

Time Span Representation

Regular intervals of time are represented by Period objects in pandas while sequences of Period objects are collected in a PeriodIndex, which can be created with the convenience function period_range.

Period

A Period represents a span of time (e.g., a day, a month, a quarter, etc). It can be created using a frequency alias:

In [120]: Period('2012', freq='A-DEC')
Period('2012', 'A-DEC')

In [121]: Period('2012-1-1', freq='D')
Period('2012-01-01', 'D')

In [122]: Period('2012-1-1 19:00', freq='H')
Period('2012-01-01 19:00', 'H')

Unlike time stamped data, pandas does not support frequencies at multiples of DateOffsets (e.g., ‘3Min’) for periods.

Adding and subtracting integers from periods shifts the period by its own frequency.

In [123]: p = Period('2012', freq='A-DEC')

In [124]: p + 1
Period('2013', 'A-DEC')

In [125]: p - 3
Period('2009', 'A-DEC')

Taking the difference of Period instances with the same frequency will return the number of frequency units between them:

In [126]: Period('2012', freq='A-DEC') - Period('2002', freq='A-DEC')
10

PeriodIndex and period_range

Regular sequences of Period objects can be collected in a PeriodIndex, which can be constructed using the period_range convenience function:

In [127]: prng = period_range('1/1/2011', '1/1/2012', freq='M')

In [128]: prng

<class 'pandas.tseries.period.PeriodIndex'>
freq: M
[2011-01, ..., 2012-01]
length: 13

The PeriodIndex constructor can also be used directly:

In [129]: PeriodIndex(['2011-1', '2011-2', '2011-3'], freq='M')

<class 'pandas.tseries.period.PeriodIndex'>
freq: M
[2011-01, ..., 2011-03]
length: 3

Just like DatetimeIndex, a PeriodIndex can also be used to index pandas objects:

In [130]: Series(randn(len(prng)), prng)

2011-01   -0.253355
2011-02   -1.426908
2011-03    1.548971
2011-04   -0.088718
2011-05   -1.771348
2011-06   -0.989328
2011-07   -1.584789
2011-08   -0.288786
2011-09   -2.029806
2011-10   -0.761200
2011-11   -1.603608
2011-12    1.756171
2012-01    0.256502
Freq: M, dtype: float64

Frequency Conversion and Resampling with PeriodIndex

The frequency of Periods and PeriodIndex can be converted via the asfreq method. Let’s start with the fiscal year 2011, ending in December:

In [131]: p = Period('2011', freq='A-DEC')

In [132]: p
Period('2011', 'A-DEC')

We can convert it to a monthly frequency. Using the how parameter, we can specify whether to return the starting or ending month:

In [133]: p.asfreq('M', how='start')
Period('2011-01', 'M')

In [134]: p.asfreq('M', how='end')
Period('2011-12', 'M')

The shorthands ‘s’ and ‘e’ are provided for convenience:

In [135]: p.asfreq('M', 's')
Period('2011-01', 'M')

In [136]: p.asfreq('M', 'e')
Period('2011-12', 'M')

Converting to a “super-period” (e.g., annual frequency is a super-period of quarterly frequency) automatically returns the super-period that includes the input period:

In [137]: p = Period('2011-12', freq='M')

In [138]: p.asfreq('A-NOV')
Period('2012', 'A-NOV')

Note that since we converted to an annual frequency that ends the year in November, the monthly period of December 2011 is actually in the 2012 A-NOV period.

Period conversions with anchored frequencies are particularly useful for working with various quarterly data common to economics, business, and other fields. Many organizations define quarters relative to the month in which their fiscal year start and ends. Thus, first quarter of 2011 could start in 2010 or a few months into 2011. Via anchored frequencies, pandas works all quarterly frequencies Q-JAN through Q-DEC.

Q-DEC define regular calendar quarters:

In [139]: p = Period('2012Q1', freq='Q-DEC')

In [140]: p.asfreq('D', 's')
Period('2012-01-01', 'D')

In [141]: p.asfreq('D', 'e')
Period('2012-03-31', 'D')

Q-MAR defines fiscal year end in March:

In [142]: p = Period('2011Q4', freq='Q-MAR')

In [143]: p.asfreq('D', 's')
Period('2011-01-01', 'D')

In [144]: p.asfreq('D', 'e')
Period('2011-03-31', 'D')

Converting between Representations

Timestamped data can be converted to PeriodIndex-ed data using to_period and vice-versa using to_timestamp:

In [145]: rng = date_range('1/1/2012', periods=5, freq='M')

In [146]: ts = Series(randn(len(rng)), index=rng)

In [147]: ts

2012-01-31    0.020601
2012-02-29   -0.411719
2012-03-31    2.079413
2012-04-30   -1.077911
2012-05-31    0.099258
Freq: M, dtype: float64

In [148]: ps = ts.to_period()

In [149]: ps

2012-01    0.020601
2012-02   -0.411719
2012-03    2.079413
2012-04   -1.077911
2012-05    0.099258
Freq: M, dtype: float64

In [150]: ps.to_timestamp()

2012-01-01    0.020601
2012-02-01   -0.411719
2012-03-01    2.079413
2012-04-01   -1.077911
2012-05-01    0.099258
Freq: MS, dtype: float64

Remember that ‘s’ and ‘e’ can be used to return the timestamps at the start or end of the period:

In [151]: ps.to_timestamp('D', how='s')

2012-01-01    0.020601
2012-02-01   -0.411719
2012-03-01    2.079413
2012-04-01   -1.077911
2012-05-01    0.099258
Freq: MS, dtype: float64

Converting between period and timestamp enables some convenient arithmetic functions to be used. In the following example, we convert a quarterly frequency with year ending in November to 9am of the end of the month following the quarter end:

In [152]: prng = period_range('1990Q1', '2000Q4', freq='Q-NOV')

In [153]: ts = Series(randn(len(prng)), prng)

In [154]: ts.index = (prng.asfreq('M', 'e') + 1).asfreq('H', 's') + 9

In [155]: ts.head()

1990-03-01 09:00   -0.089851
1990-06-01 09:00    0.711329
1990-09-01 09:00    0.531761
1990-12-01 09:00    0.265615
1991-03-01 09:00   -0.174462
Freq: H, dtype: float64

Time Zone Handling

Using pytz, pandas provides rich support for working with timestamps in different time zones. By default, pandas objects are time zone unaware:

In [156]: rng = date_range('3/6/2012 00:00', periods=15, freq='D')

In [157]: print(rng.tz)
None

To supply the time zone, you can use the tz keyword to date_range and other functions:

In [158]: rng_utc = date_range('3/6/2012 00:00', periods=10, freq='D', tz='UTC')

In [159]: print(rng_utc.tz)
UTC

Timestamps, like Python’s datetime.datetime object can be either time zone naive or time zone aware. Naive time series and DatetimeIndex objects can be localized using tz_localize:

In [160]: ts = Series(randn(len(rng)), rng)

In [161]: ts_utc = ts.tz_localize('UTC')

In [162]: ts_utc

2012-03-06 00:00:00+00:00   -2.189293
2012-03-07 00:00:00+00:00   -1.819506
2012-03-08 00:00:00+00:00    0.229798
2012-03-09 00:00:00+00:00    0.119425
2012-03-10 00:00:00+00:00    1.808966
2012-03-11 00:00:00+00:00    1.015841
2012-03-12 00:00:00+00:00   -1.651784
2012-03-13 00:00:00+00:00    0.347674
2012-03-14 00:00:00+00:00   -0.773688
2012-03-15 00:00:00+00:00    0.425863
2012-03-16 00:00:00+00:00    0.579486
2012-03-17 00:00:00+00:00   -0.745396
2012-03-18 00:00:00+00:00    0.141880
2012-03-19 00:00:00+00:00   -1.077754
2012-03-20 00:00:00+00:00   -1.301174
Freq: D, dtype: float64

You can use the tz_convert method to convert pandas objects to convert tz-aware data to another time zone:

In [163]: ts_utc.tz_convert('US/Eastern')

2012-03-05 19:00:00-05:00   -2.189293
2012-03-06 19:00:00-05:00   -1.819506
2012-03-07 19:00:00-05:00    0.229798
2012-03-08 19:00:00-05:00    0.119425
2012-03-09 19:00:00-05:00    1.808966
2012-03-10 19:00:00-05:00    1.015841
2012-03-11 20:00:00-04:00   -1.651784
2012-03-12 20:00:00-04:00    0.347674
2012-03-13 20:00:00-04:00   -0.773688
2012-03-14 20:00:00-04:00    0.425863
2012-03-15 20:00:00-04:00    0.579486
2012-03-16 20:00:00-04:00   -0.745396
2012-03-17 20:00:00-04:00    0.141880
2012-03-18 20:00:00-04:00   -1.077754
2012-03-19 20:00:00-04:00   -1.301174
Freq: D, dtype: float64

Under the hood, all timestamps are stored in UTC. Scalar values from a DatetimeIndex with a time zone will have their fields (day, hour, minute) localized to the time zone. However, timestamps with the same UTC value are still considered to be equal even if they are in different time zones:

In [164]: rng_eastern = rng_utc.tz_convert('US/Eastern')

In [165]: rng_berlin = rng_utc.tz_convert('Europe/Berlin')

In [166]: rng_eastern[5]
Timestamp('2012-03-10 19:00:00-0500', tz='US/Eastern')

In [167]: rng_berlin[5]
Timestamp('2012-03-11 01:00:00+0100', tz='Europe/Berlin')

In [168]: rng_eastern[5] == rng_berlin[5]
True

Like Series, DataFrame, and DatetimeIndex, Timestamps can be converted to other time zones using tz_convert:

In [169]: rng_eastern[5]
Timestamp('2012-03-10 19:00:00-0500', tz='US/Eastern')

In [170]: rng_berlin[5]
Timestamp('2012-03-11 01:00:00+0100', tz='Europe/Berlin')

In [171]: rng_eastern[5].tz_convert('Europe/Berlin')
Timestamp('2012-03-11 01:00:00+0100', tz='Europe/Berlin')

Localization of Timestamps functions just like DatetimeIndex and TimeSeries:

In [172]: rng[5]
Timestamp('2012-03-11 00:00:00', tz=None)

In [173]: rng[5].tz_localize('Asia/Shanghai')
Timestamp('2012-03-11 00:00:00+0800', tz='Asia/Shanghai')

Operations between TimeSeries in different time zones will yield UTC TimeSeries, aligning the data on the UTC timestamps:

In [174]: eastern = ts_utc.tz_convert('US/Eastern')

In [175]: berlin = ts_utc.tz_convert('Europe/Berlin')

In [176]: result = eastern + berlin

In [177]: result

2012-03-06 00:00:00+00:00   -4.378586
2012-03-07 00:00:00+00:00   -3.639011
2012-03-08 00:00:00+00:00    0.459596
2012-03-09 00:00:00+00:00    0.238849
2012-03-10 00:00:00+00:00    3.617932
2012-03-11 00:00:00+00:00    2.031683
2012-03-12 00:00:00+00:00   -3.303568
2012-03-13 00:00:00+00:00    0.695349
2012-03-14 00:00:00+00:00   -1.547376
2012-03-15 00:00:00+00:00    0.851726
2012-03-16 00:00:00+00:00    1.158971
2012-03-17 00:00:00+00:00   -1.490793
2012-03-18 00:00:00+00:00    0.283760
2012-03-19 00:00:00+00:00   -2.155508
2012-03-20 00:00:00+00:00   -2.602348
Freq: D, dtype: float64

In [178]: result.index

<class 'pandas.tseries.index.DatetimeIndex'>
[2012-03-06 00:00:00, ..., 2012-03-20 00:00:00]
Length: 15, Freq: D, Timezone: UTC

In some cases, localize cannot determine the DST and non-DST hours when there are duplicates. This often happens when reading files that simply duplicate the hours. The infer_dst argument in tz_localize will attempt to determine the right offset.

In [179]: rng_hourly = DatetimeIndex(['11/06/2011 00:00', '11/06/2011 01:00',
   .....:                             '11/06/2011 01:00', '11/06/2011 02:00',
   .....:                             '11/06/2011 03:00'])
   .....: 

In [180]: rng_hourly.tz_localize('US/Eastern')
---------------------------------------------------------------------------
AmbiguousTimeError                        Traceback (most recent call last)
<ipython-input-180-8c5fa6a37f5b> in <module>()
----> 1 rng_hourly.tz_localize('US/Eastern')
/home/docbuild/CI/pandas/pandas/tseries/index.pyc in tz_localize(self, tz, infer_dst)
   1602 
   1603         # Convert to UTC
-> 1604         new_dates = tslib.tz_localize_to_utc(self.asi8, tz, infer_dst=infer_dst)
   1605         new_dates = new_dates.view(_NS_DTYPE)
   1606 
/home/docbuild/CI/pandas/pandas/tslib.so in pandas.tslib.tz_localize_to_utc (pandas/tslib.c:28987)()
AmbiguousTimeError: Cannot infer dst time from Timestamp('2011-11-06 01:00:00', tz=None), try using the 'infer_dst' argument

In [181]: rng_hourly_eastern = rng_hourly.tz_localize('US/Eastern', infer_dst=True)

In [182]: rng_hourly_eastern.values

array(['2011-11-06T06:00:00.000000000+0200',
       '2011-11-06T07:00:00.000000000+0200',
       '2011-11-06T08:00:00.000000000+0200',
       '2011-11-06T09:00:00.000000000+0200',
       '2011-11-06T10:00:00.000000000+0200'], dtype='datetime64[ns]')

Time Deltas

Timedeltas are differences in times, expressed in difference units, e.g. days,hours,minutes,seconds. They can be both positive and negative. DateOffsets that are absolute in nature (Day, Hour, Minute, Second, Milli, Micro, Nano) can be used as timedeltas.

In [183]: from datetime import datetime, timedelta

In [184]: s = Series(date_range('2012-1-1', periods=3, freq='D'))

In [185]: td = Series([ timedelta(days=i) for i in range(3) ])

In [186]: df = DataFrame(dict(A = s, B = td))

In [187]: df

                    A                B
0 2012-01-01 00:00:00         00:00:00
1 2012-01-02 00:00:00 1 days, 00:00:00
2 2012-01-03 00:00:00 2 days, 00:00:00
[3 rows x 2 columns]

In [188]: df['C'] = df['A'] + df['B']

In [189]: df

                    A                B                   C
0 2012-01-01 00:00:00         00:00:00 2012-01-01 00:00:00
1 2012-01-02 00:00:00 1 days, 00:00:00 2012-01-03 00:00:00
2 2012-01-03 00:00:00 2 days, 00:00:00 2012-01-05 00:00:00
[3 rows x 3 columns]

In [190]: df.dtypes

A     datetime64[ns]
B    timedelta64[ns]
C     datetime64[ns]
dtype: object

In [191]: s - s.max()

0   -2 days, 00:00:00
1   -1 days, 00:00:00
2            00:00:00
dtype: timedelta64[ns]

In [192]: s - datetime(2011,1,1,3,5)

0   364 days, 20:55:00
1   365 days, 20:55:00
2   366 days, 20:55:00
dtype: timedelta64[ns]

In [193]: s + timedelta(minutes=5)

0   2012-01-01 00:05:00
1   2012-01-02 00:05:00
2   2012-01-03 00:05:00
dtype: datetime64[ns]

In [194]: s + Minute(5)

0   2012-01-01 00:05:00
1   2012-01-02 00:05:00
2   2012-01-03 00:05:00
dtype: datetime64[ns]

In [195]: s + Minute(5) + Milli(5)

0   2012-01-01 00:05:00.005000
1   2012-01-02 00:05:00.005000
2   2012-01-03 00:05:00.005000
dtype: datetime64[ns]

Getting scalar results from a timedelta64[ns] series

In [196]: y = s - s[0]

In [197]: y

0           00:00:00
1   1 days, 00:00:00
2   2 days, 00:00:00
dtype: timedelta64[ns]

Series of timedeltas with NaT values are supported

In [198]: y = s - s.shift()

In [199]: y

0                NaT
1   1 days, 00:00:00
2   1 days, 00:00:00
dtype: timedelta64[ns]

Elements can be set to NaT using np.nan analagously to datetimes

In [200]: y[1] = np.nan

In [201]: y

0                NaT
1                NaT
2   1 days, 00:00:00
dtype: timedelta64[ns]

Operands can also appear in a reversed order (a singluar object operated with a Series)

In [202]: s.max() - s

0   2 days, 00:00:00
1   1 days, 00:00:00
2           00:00:00
dtype: timedelta64[ns]

In [203]: datetime(2011,1,1,3,5) - s

0   -364 days, 20:55:00
1   -365 days, 20:55:00
2   -366 days, 20:55:00
dtype: timedelta64[ns]

In [204]: timedelta(minutes=5) + s

0   2012-01-01 00:05:00
1   2012-01-02 00:05:00
2   2012-01-03 00:05:00
dtype: datetime64[ns]

Some timedelta numeric like operations are supported.

In [205]: td - timedelta(minutes=5, seconds=5, microseconds=5)

0          -00:05:05.000005
1           23:54:54.999995
2   1 days, 23:54:54.999995
dtype: timedelta64[ns]

min, max and the corresponding idxmin, idxmax operations are supported on frames

In [206]: A = s - Timestamp('20120101') - timedelta(minutes=5, seconds=5)

In [207]: B = s - Series(date_range('2012-1-2', periods=3, freq='D'))

In [208]: df = DataFrame(dict(A=A, B=B))

In [209]: df

                 A                 B
0        -00:05:05 -1 days, 00:00:00
1         23:54:55 -1 days, 00:00:00
2 1 days, 23:54:55 -1 days, 00:00:00
[3 rows x 2 columns]

In [210]: df.min()

A           -00:05:05
B   -1 days, 00:00:00
dtype: timedelta64[ns]

In [211]: df.min(axis=1)

0   -1 days, 00:00:00
1   -1 days, 00:00:00
2   -1 days, 00:00:00
dtype: timedelta64[ns]

In [212]: df.idxmin()

A    0
B    0
dtype: int64

In [213]: df.idxmax()

A    2
B    0
dtype: int64

min, max operations are supported on series; these return a single element timedelta64[ns] Series (this avoids having to deal with numpy timedelta64 issues). idxmin, idxmax are supported as well.

In [214]: df.min().max()

0   -00:05:05
dtype: timedelta64[ns]

In [215]: df.min(axis=1).min()

0   -1 days, 00:00:00
dtype: timedelta64[ns]

In [216]: df.min().idxmax()
'A'

In [217]: df.min(axis=1).idxmin()
0

You can fillna on timedeltas. Integers will be interpreted as seconds. You can pass a timedelta to get a particular value.

In [218]: y.fillna(0)

0           00:00:00
1           00:00:00
2   1 days, 00:00:00
dtype: timedelta64[ns]

In [219]: y.fillna(10)

0           00:00:10
1           00:00:10
2   1 days, 00:00:00
dtype: timedelta64[ns]

In [220]: y.fillna(timedelta(days=-1,seconds=5))

0          -23:59:55
1          -23:59:55
2   1 days, 00:00:00
dtype: timedelta64[ns]

Time Deltas & Reductions

Warning

A numeric reduction operation for timedelta64[ns] can return a single-element Series of dtype timedelta64[ns].

You can do numeric reduction operations on timedeltas.

In [221]: y2 = y.fillna(timedelta(days=-1,seconds=5))

In [222]: y2

0          -23:59:55
1          -23:59:55
2   1 days, 00:00:00
dtype: timedelta64[ns]

In [223]: y2.mean()

0   -07:59:56.666667
dtype: timedelta64[ns]

In [224]: y2.quantile(.1)
numpy.timedelta64(-86395000000000,'ns')

Time Deltas & Conversions

New in version 0.13.

string/integer conversion

Using the top-level to_timedelta, you can convert a scalar or array from the standard timedelta format (produced by to_csv) into a timedelta type (np.timedelta64 in nanoseconds). It can also construct Series.

Warning

This requires numpy >= 1.7

In [225]: to_timedelta('1 days 06:05:01.00003')
numpy.timedelta64(108301000030000,'ns')

In [226]: to_timedelta('15.5us')
numpy.timedelta64(15500,'ns')

In [227]: to_timedelta(['1 days 06:05:01.00003','15.5us','nan'])

0   1 days, 06:05:01.000030
1           00:00:00.000016
2                       NaT
dtype: timedelta64[ns]

In [228]: to_timedelta(np.arange(5),unit='s')

0   00:00:00
1   00:00:01
2   00:00:02
3   00:00:03
4   00:00:04
dtype: timedelta64[ns]

In [229]: to_timedelta(np.arange(5),unit='d')

0           00:00:00
1   1 days, 00:00:00
2   2 days, 00:00:00
3   3 days, 00:00:00
4   4 days, 00:00:00
dtype: timedelta64[ns]

frequency conversion

Timedeltas can be converted to other ‘frequencies’ by dividing by another timedelta, or by astyping to a specific timedelta type. These operations yield float64 dtyped Series.

In [230]: td = Series(date_range('20130101',periods=4))-Series(date_range('20121201',periods=4))

In [231]: td[2] += np.timedelta64(timedelta(minutes=5,seconds=3))

In [232]: td[3] = np.nan

In [233]: td

0   31 days, 00:00:00
1   31 days, 00:00:00
2   31 days, 00:05:03
3                 NaT
dtype: timedelta64[ns]

# to days
In [234]: td / np.timedelta64(1,'D')

0    31.000000
1    31.000000
2    31.003507
3          NaN
dtype: float64

In [235]: td.astype('timedelta64[D]')

0    31
1    31
2    31
3   NaN
dtype: float64

# to seconds
In [236]: td / np.timedelta64(1,'s')

0    2678400
1    2678400
2    2678703
3        NaN
dtype: float64

In [237]: td.astype('timedelta64[s]')

0    2678400
1    2678400
2    2678703
3        NaN
dtype: float64

Dividing or multiplying a timedelta64[ns] Series by an integer or integer Series yields another timedelta64[ns] dtypes Series.

In [238]: td * -1

0   -31 days, 00:00:00
1   -31 days, 00:00:00
2   -31 days, 00:05:03
3                  NaT
dtype: timedelta64[ns]

In [239]: td * Series([1,2,3,4])

0   31 days, 00:00:00
1   62 days, 00:00:00
2   93 days, 00:15:09
3                 NaT
dtype: timedelta64[ns]

Numpy < 1.7 Compatibility

Numpy < 1.7 has a broken timedelta64 type that does not correctly work for arithmetic. Pandas bypasses this, but for frequency conversion as above, you need to create the divisor yourself. The np.timetimedelta64 type only has 1 argument, the number of micro seconds.

The following are equivalent statements in the two versions of numpy.

from distutils.version import LooseVersion
if LooseVersion(np.__version__) <= '1.6.2':
    y / np.timedelta(86400*int(1e6))
    y / np.timedelta(int(1e6))
else:
    y / np.timedelta64(1,'D')
    y / np.timedelta64(1,'s')