Pandas Resampling
last modified November 17, 2026 Python 3.15
Resampling means changing the frequency of a time series — a sequence of values
indexed by dates or times. Downsampling combines many observations
into fewer, larger buckets, such as turning hourly readings into daily totals.
Upsampling does the opposite, creating more, smaller time slots from
fewer observations. Pandas provides the resample method for both,
and the examples below were verified with pandas 3.1.1.
Every example builds its own index with date_range, which creates a
regular sequence of dates, so nothing depends on an external file. Series with
a datetime index expose resample directly.
Frequency aliases
Resampling is controlled by a frequency alias, a short code for a
time interval. The common ones are 'D' for a day,
'W' for a week, 'M' for a month end and
'h' for an hour. In pandas 3.x the month-end alias
'M' is deprecated for resampling, so we use 'ME'
instead; it produces the same month-end dates. The snippet uses these aliases
to build small date ranges and print them.
import pandas as pd
print('D:')
print(pd.date_range('2026-01-01', periods=3, freq='D'))
print('W:')
print(pd.date_range('2026-01-01', periods=3, freq='W'))
print('ME:')
print(pd.date_range('2026-01-01', periods=3, freq='ME'))
print('h:')
print(pd.date_range('2026-01-01', periods=3, freq='h'))
Daily dates advance one day at a time. The weekly alias defaults to Sunday, so
the dates are successive Sundays starting on 4 January 2026. The month-end
alias gives the last day of each month, and the hourly alias advances by one
hour. These same strings are what we pass to resample.
$ python resample_aliases.py
D:
DatetimeIndex(['2026-01-01', '2026-01-02', '2026-01-03'],
dtype='datetime64[us]', freq='D')
W:
DatetimeIndex(['2026-01-04', '2026-01-11', '2026-01-18'],
dtype='datetime64[us]', freq='W-SUN')
ME:
DatetimeIndex(['2026-01-31', '2026-02-28', '2026-03-31'],
dtype='datetime64[us]', freq='ME')
h:
DatetimeIndex(['2026-01-01 00:00:00', '2026-01-01 01:00:00',
'2026-01-01 02:00:00'],
dtype='datetime64[us]', freq='h')
Downsampling with sum and mean
Here the source has twelve readings spaced six hours apart, four per day.
Calling resample('D') groups them by calendar day; on its own it
only defines the groups, so we finish with an aggregation method such as
sum or mean to produce values. The sum adds the four
readings in a day, and the mean averages them.
import pandas as pd
idx = pd.date_range('2026-01-01', periods=12, freq='6h')
s = pd.Series(range(1, 13), index=idx)
print('sum:')
print(s.resample('D').sum())
print()
print('mean:')
print(s.resample('D').mean())
For 1 January the four values are 1, 2, 3 and 4, which add up to 10 and average
2.5. The following days contain 5 to 8 and 9 to 12, giving totals of 26 and 42
and means of 6.5 and 10.5. The result is a new Series indexed one entry per
day, with Freq: D confirming the daily frequency.
$ python resample_downsample.py sum: 2026-01-01 10 2026-01-02 26 2026-01-03 42 Freq: D, dtype: int64 mean: 2026-01-01 2.5 2026-01-02 6.5 2026-01-03 10.5 Freq: D, dtype: float64
OHLC aggregation
The ohlc method is a special aggregation borrowed from financial
charts. Its name stands for open, high, low and close: the first value in each
bucket, the largest, the smallest and the last. Calling it after
resample returns a DataFrame with one column for each of those
four numbers.
import pandas as pd
idx = pd.date_range('2026-01-01', periods=12, freq='6h')
s = pd.Series(range(1, 13), index=idx)
print(s.resample('D').ohlc())
On 1 January the readings rise from 1 to 4, so the bucket opens at 1, reaches a high of 4, a low of 1 and closes at 4. Because the values increase throughout, open equals low and high equals close on every day. On data that moves up and down the four columns differ, which is why this view is popular for prices.
$ python resample_ohlc.py
open high low close
2026-01-01 1 4 1 4
2026-01-02 5 8 5 8
2026-01-03 9 12 9 12
Upsampling with asfreq and ffill
Upsampling inserts new time slots that have no observation. The
asfreq method creates those slots and leaves them as
NaN, the marker for a missing value. The ffill method
instead fills each new slot with the last known value — the initials stand for
"forward fill".
import pandas as pd
idx = pd.date_range('2026-01-01', periods=3, freq='D')
s = pd.Series([10, 20, 30], index=idx)
print('asfreq:')
print(s.resample('12h').asfreq())
print()
print('ffill:')
print(s.resample('12h').ffill())
Three daily values become six slots at twelve-hour intervals. With
asfreq the original values stay at midnight and the new noon slots
are NaN, which is why the dtype becomes float. With
ffill each noon slot repeats the value from midnight, so the
series is complete and stays as integers.
$ python resample_upsample.py asfreq: 2026-01-01 00:00:00 10.0 2026-01-01 12:00:00 NaN 2026-01-02 00:00:00 20.0 2026-01-02 12:00:00 NaN 2026-01-03 00:00:00 30.0 Freq: 12h, dtype: float64 ffill: 2026-01-01 00:00:00 10 2026-01-01 12:00:00 10 2026-01-02 00:00:00 20 2026-01-02 12:00:00 20 2026-01-03 00:00:00 30 Freq: 12h, dtype: int64
Closed and label
Every resampling bucket has two edges, and the closed argument
decides which edge belongs to the bucket: 'left', the default,
includes the start but not the end, while 'right' includes the end
but not the start. The label argument decides which edge is used
as the bucket's label. Changing both shifts values between adjacent days.
import pandas as pd
idx = pd.date_range('2026-01-01', periods=5, freq='6h')
s = pd.Series([1, 2, 3, 4, 5], index=idx)
print('default:')
print(s.resample('D').sum())
print()
print('closed=right, label=right:')
print(s.resample('D', closed='right', label='right').sum())
By default each day starts at midnight, so 1 January collects the four values
1 to 4 and totals 10, while the midnight reading on 2 January forms the next
day's total of 5. With closed='right' a midnight reading belongs to
the bucket ending at that midnight: the value 1 closes the first bucket, and
2, 3, 4 and 5 add up to 14 in the second. The label='right' setting
keeps each bucket labelled by its closing edge.
$ python resample_closed.py default: 2026-01-01 10 2026-01-02 5 Freq: D, dtype: int64 closed=right, label=right: 2026-01-01 1 2026-01-02 14 Freq: D, dtype: int64
Resampling, groupby and windowing
Resampling is related to groupby, but the grouping is done for us
by time. With groupby we supply a column of labels and decide how
rows are grouped; with resample the datetime index is divided into
regular, contiguous intervals automatically. Resampling also fills in empty
intervals, which a groupby on dates would simply omit.
It is also different from a rolling window, the topic of the Pandas windowing functions tutorial. A rolling window computes a statistic over a fixed number of recent rows, such as the last seven days, and produces one result per row. Resampling produces one result per bucket, combining every row inside it. Use resampling to change the series' frequency and rolling windows to smooth it without changing that frequency.
Source
Pandas Series.resample Documentation, Pandas date_range Documentation, and Pandas Time Series Documentation
In this article we have used frequency aliases, downsampled a time series with
sum, mean and ohlc, upsampled with
asfreq and ffill, and controlled bucket edges with
closed and label. For creating and parsing date values
see Pandas datetimes, and for smoothing without
changing frequency see
Pandas windowing functions.
Author
List all Pandas tutorials.