Pandas.resample to a non-integer multiple frequency
Question:
I have to resample my dataset from a 10-minute interval to a 15-minute interval to make it in sync with another dataset. Based on my searches at stackoverflow I have some ideas how to proceed, but none of them deliver a clean and clear solution.
Problem
Problem set up
#%% Import modules
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
#%% make timestamps
periods = 12
startdate = '2010-01-01'
timestamp10min = pd.date_range(startdate, freq='10Min', periods=periods)
#%% Make DataFrame and fill it with some data
df = pd.DataFrame(index=timestamp10min)
y = -(np.arange(periods)-periods/2)**2
df['y'] = y
Desired output
Now I want the values that are already at the 10 minutes to be unchanged, and the values at **:15 and **:45 to be the mean of **:10, **:20 and **:40, **:50. The core of the problem is that 15 minutes is not a integer multiple of 10 minutes. Otherwise simply applying df.resample('10Min', how='mean') would have worked.
Possible solutions
-
Simply use the 15 minutes resampling and just live with the small introduced error.
-
Using two forms of resample, with close='left', label='left' and close='right' , label='right'. Afterwards I could average both resampled forms. The results will give me some error on the results, but smaller than the first method.
-
Resample everything to 5 minute data and then apply a rolling average. Something like that is apllied here: Pandas: rolling mean by time interval
-
Resample and average with a varying number of input: Use numpy.average with weights for resampling a pandas array
Therefore I would have to create a new Series with varying weight length. Were the weight should be alternating between 1 and 2.
-
Resample everything to 5 minute data and then apply linear interpolation. This method is close to method 3. Pandas data frame: resample with linear interpolation
Edit: @Paul H gave a workable solution along these lines, which is stille readable. Thanks!
All the methods are not really statisfying for me. Some lead to a small error, and other methods would be quite difficult to read for an outsider.
Implementation
The implementation of method 1, 2 and 5 together with the desired ouput. In combination with visualization.
#%% start plot
plt.figure()
plt.plot(df.index, df['y'], label='original')
#%% resample the data to 15 minutes and plot the result
close = 'left'; label='left'
dfresamplell = pd.DataFrame()
dfresamplell['15min'] = df.y.resample('15Min', how='mean', closed=close, label=label)
labelstring = 'close ' + close + ' label ' + label
plt.plot(dfresamplell.index, dfresamplell['15min'], label=labelstring)
close = 'right'; label='right'
dfresamplerr = pd.DataFrame()
dfresamplerr['15min'] = df.y.resample('15Min', how='mean', closed=close, label=label)
labelstring = 'close ' + close + ' label ' + label
plt.plot(dfresamplerr.index, dfresamplerr['15min'], label=labelstring)
#%% make an average
dfresampleaverage = pd.DataFrame(index=dfresamplell.index)
dfresampleaverage['15min'] = (dfresamplell['15min'].values+dfresamplerr['15min'].values[:-1])/2
plt.plot(dfresampleaverage.index, dfresampleaverage['15min'], label='average of both resampling methods')
#%% desired output
ydesired = np.zeros(periods/3*2)
i = 0
j = 0
k = 0
for val in ydesired:
if i+k==len(y): k=0
ydesired[j] = np.mean([y[i],y[i+k]])
j+=1
i+=1
if k==0: k=1;
else: k=0; i+=1
plt.plot(dfresamplell.index, ydesired, label='ydesired')
#%% suggestion of Paul H
dfreindex = df.reindex(pd.date_range(startdate, freq='5T', periods=periods*2))
dfreindex.interpolate(inplace=True)
dfreindex = dfreindex.resample('15T', how='first').head()
plt.plot(dfreindex.index, dfreindex['y'], label='method Paul H')
#%% finalize plot
plt.legend()
Implementation for angles
As a bonus I have added the code I will use for the interpolation of angles. This is done by using complex numbers. Because complex interpolation is not implemented (yet), I split the complex numbers into a real and a imaginary part. After averaging these numbers can be converted to angels again. For certain angels this is a better resampling method than simply averaging the two angels, for example: 345 and 5 degrees.
#%% make timestamps
periods = 24*6
startdate = '2010-01-01'
timestamp10min = pd.date_range(startdate, freq='10Min', periods=periods)
#%% Make DataFrame and fill it with some data
degrees = np.cumsum(np.random.randn(periods)*25) % 360
df = pd.DataFrame(index=timestamp10min)
df['deg'] = degrees
df['zreal'] = np.cos(df['deg']*np.pi/180)
df['zimag'] = np.sin(df['deg']*np.pi/180)
#%% suggestion of Paul H
dfreindex = df.reindex(pd.date_range(startdate, freq='5T', periods=periods*2))
dfreindex = dfreindex.interpolate()
dfresample = dfreindex.resample('15T', how='first')
#%% convert complex to degrees
def f(x):
return np.angle(x[0] + x[1]*1j, deg=True )
dfresample['degrees'] = dfresample[['zreal', 'zimag']].apply(f, axis=1)
#%% set all the values between 0-360 degrees
dfresample.loc[dfresample['degrees']<0] = 360 + dfresample.loc[dfresample['degrees']<0]
#%% wrong resampling
dfresample['deg'] = dfresample['deg'] % 360
#%% plot different sampling methods
plt.figure()
plt.plot(df.index, df['deg'], label='normal', marker='v')
plt.plot(dfresample.index, dfresample['degrees'], label='resampled according @Paul H', marker='^')
plt.plot(dfresample.index, dfresample['deg'], label='wrong resampling', marker='<')
plt.legend()
Answers:
Ok, here’s one way to do it.
- Make a list of the times you want to have filled in
- Make a combined index that includes the times you want and the times you already have
- Take your data and “forward fill it”
- Take your data and “backward fill it”
- Average the forward and backward fills
- Select only the rows you want
Note this only works since you want the values exactly halfway between the values you already have, time-wise. Note the last time comes out np.nan because you don’t have any later data.
times_15 = []
current = df.index[0]
while current < df.index[-2]:
current = current + dt.timedelta(minutes=15)
times_15.append(current)
combined = set(times_15) | set(df.index)
df = df.reindex(combined).sort_index(axis=0)
df['ff'] = df['y'].fillna(method='ffill')
df['bf'] = df['y'].fillna(method='bfill')
df['solution'] = df[['ff', 'bf']].mean(1)
df.loc[times_15, :]
I might be misunderstanding the problem, but does this work?
TL;DR version:
import numpy as np
import pandas
data = np.arange(0, 101, 8)
index_10T = pandas.DatetimeIndex(freq='10T', start='2012-01-01 00:00', periods=data.shape[0])
index_05T = pandas.DatetimeIndex(freq='05T', start=index_10T[0], end=index_10T[-1])
index_15T = pandas.DatetimeIndex(freq='15T', start=index_10T[0], end=index_10T[-1])
df1 = pandas.DataFrame(data=data, index=index_10T, columns=['A'])
print(df.reindex(index=index_05T).interpolate().loc[index_15T])
Long version
setup fake data
import numpy as np
import pandas
data = np.arange(0, 101, 8)
index_10T = pandas.DatetimeIndex(freq='10T', start='2012-01-01 00:00', periods=data.shape[0])
df1 = pandas.DataFrame(data=data, index=index_10T, columns=['A'])
print(df1)
A
2012-01-01 00:00:00 0
2012-01-01 00:10:00 8
2012-01-01 00:20:00 16
2012-01-01 00:30:00 24
2012-01-01 00:40:00 32
2012-01-01 00:50:00 40
2012-01-01 01:00:00 48
2012-01-01 01:10:00 56
2012-01-01 01:20:00 64
2012-01-01 01:30:00 72
2012-01-01 01:40:00 80
2012-01-01 01:50:00 88
2012-01-01 02:00:00 96
So then build a new 5-minute index and reindex the original dataframe
index_05T = pandas.DatetimeIndex(freq='05T', start=index_10T[0], end=index_10T[-1])
df2 = df.reindex(index=index_05T)
print(df2)
A
2012-01-01 00:00:00 0
2012-01-01 00:05:00 NaN
2012-01-01 00:10:00 8
2012-01-01 00:15:00 NaN
2012-01-01 00:20:00 16
2012-01-01 00:25:00 NaN
2012-01-01 00:30:00 24
2012-01-01 00:35:00 NaN
2012-01-01 00:40:00 32
2012-01-01 00:45:00 NaN
2012-01-01 00:50:00 40
2012-01-01 00:55:00 NaN
2012-01-01 01:00:00 48
2012-01-01 01:05:00 NaN
2012-01-01 01:10:00 56
2012-01-01 01:15:00 NaN
2012-01-01 01:20:00 64
2012-01-01 01:25:00 NaN
2012-01-01 01:30:00 72
2012-01-01 01:35:00 NaN
2012-01-01 01:40:00 80
2012-01-01 01:45:00 NaN
2012-01-01 01:50:00 88
2012-01-01 01:55:00 NaN
2012-01-01 02:00:00 96
and then linearly interpolate
print(df2.interpolate())
A
2012-01-01 00:00:00 0
2012-01-01 00:05:00 4
2012-01-01 00:10:00 8
2012-01-01 00:15:00 12
2012-01-01 00:20:00 16
2012-01-01 00:25:00 20
2012-01-01 00:30:00 24
2012-01-01 00:35:00 28
2012-01-01 00:40:00 32
2012-01-01 00:45:00 36
2012-01-01 00:50:00 40
2012-01-01 00:55:00 44
2012-01-01 01:00:00 48
2012-01-01 01:05:00 52
2012-01-01 01:10:00 56
2012-01-01 01:15:00 60
2012-01-01 01:20:00 64
2012-01-01 01:25:00 68
2012-01-01 01:30:00 72
2012-01-01 01:35:00 76
2012-01-01 01:40:00 80
2012-01-01 01:45:00 84
2012-01-01 01:50:00 88
2012-01-01 01:55:00 92
2012-01-01 02:00:00 96
build a 15-minute index and use that to pull out data:
index_15T = pandas.DatetimeIndex(freq='15T', start=index_10T[0], end=index_10T[-1])
print(df2.interpolate().loc[index_15T])
A
2012-01-01 00:00:00 0
2012-01-01 00:15:00 12
2012-01-01 00:30:00 24
2012-01-01 00:45:00 36
2012-01-01 01:00:00 48
2012-01-01 01:15:00 60
2012-01-01 01:30:00 72
2012-01-01 01:45:00 84
2012-01-01 02:00:00 96
In case someone is working with data without regularity at all, here is an adapted solution from the one provided by Paul H above.
If you don’t want to interpolate throughout the time-series, but only in those places where resample is meaningful, you may keep the interpolated column side by side and finish with a resample and dropna.
import numpy as np
import pandas
data = np.arange(0, 101, 3)
index_setup = pandas.date_range(freq='01T', start='2022-01-01 00:00', periods=data.shape[0])
df1 = pandas.DataFrame(data=data, index=index_setup, columns=['A'])
df1 = df1.sample(frac=0.2).sort_index()
print(df1)
A
2022-01-01 00:03:00 9
2022-01-01 00:06:00 18
2022-01-01 00:08:00 24
2022-01-01 00:18:00 54
2022-01-01 00:25:00 75
2022-01-01 00:27:00 81
2022-01-01 00:30:00 90
Notice resampling this DF without any regularity forces values to the floor index, without interpolating.
print(df1.resample('05T').mean())
A
2022-01-01 00:00:00 9.0
2022-01-01 00:05:00 24.0
2022-01-01 00:10:00 39.0
2022-01-01 00:15:00 51.0
2022-01-01 00:20:00 NaN
2022-01-01 00:25:00 79.5
A better solution can be achieved by interpolating in a small enough interval and then resampling. The result DF now has too much, but a dropna() brings it close to its original shape.
index_1min = pandas.date_range(freq='01T', start='2022-01-01 00:00', end='2022-01-01 23:59')
df2 = df1.reindex(index=index_1min)
df2['A_interp'] = df2['A'].interpolate(limit_direction='both')
print(df2.resample('05T').first().dropna())
A A_interp
2022-01-01 00:00:00 9.0 9.0
2022-01-01 00:05:00 21.0 15.0
2022-01-01 00:10:00 39.0 30.0
2022-01-01 00:15:00 51.0 45.0
2022-01-01 00:25:00 75.0 75.0
I have to resample my dataset from a 10-minute interval to a 15-minute interval to make it in sync with another dataset. Based on my searches at stackoverflow I have some ideas how to proceed, but none of them deliver a clean and clear solution.
Problem
Problem set up
#%% Import modules
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
#%% make timestamps
periods = 12
startdate = '2010-01-01'
timestamp10min = pd.date_range(startdate, freq='10Min', periods=periods)
#%% Make DataFrame and fill it with some data
df = pd.DataFrame(index=timestamp10min)
y = -(np.arange(periods)-periods/2)**2
df['y'] = y
Desired output
Now I want the values that are already at the 10 minutes to be unchanged, and the values at **:15 and **:45 to be the mean of **:10, **:20 and **:40, **:50. The core of the problem is that 15 minutes is not a integer multiple of 10 minutes. Otherwise simply applying df.resample('10Min', how='mean') would have worked.
Possible solutions
-
Simply use the 15 minutes resampling and just live with the small introduced error.
-
Using two forms of resample, with
close='left', label='left'andclose='right' , label='right'. Afterwards I could average both resampled forms. The results will give me some error on the results, but smaller than the first method. -
Resample everything to 5 minute data and then apply a rolling average. Something like that is apllied here: Pandas: rolling mean by time interval
-
Resample and average with a varying number of input: Use numpy.average with weights for resampling a pandas array
Therefore I would have to create a new Series with varying weight length. Were the weight should be alternating between 1 and 2. -
Resample everything to 5 minute data and then apply linear interpolation. This method is close to method 3. Pandas data frame: resample with linear interpolation
Edit: @Paul H gave a workable solution along these lines, which is stille readable. Thanks!
All the methods are not really statisfying for me. Some lead to a small error, and other methods would be quite difficult to read for an outsider.
Implementation
The implementation of method 1, 2 and 5 together with the desired ouput. In combination with visualization.
#%% start plot
plt.figure()
plt.plot(df.index, df['y'], label='original')
#%% resample the data to 15 minutes and plot the result
close = 'left'; label='left'
dfresamplell = pd.DataFrame()
dfresamplell['15min'] = df.y.resample('15Min', how='mean', closed=close, label=label)
labelstring = 'close ' + close + ' label ' + label
plt.plot(dfresamplell.index, dfresamplell['15min'], label=labelstring)
close = 'right'; label='right'
dfresamplerr = pd.DataFrame()
dfresamplerr['15min'] = df.y.resample('15Min', how='mean', closed=close, label=label)
labelstring = 'close ' + close + ' label ' + label
plt.plot(dfresamplerr.index, dfresamplerr['15min'], label=labelstring)
#%% make an average
dfresampleaverage = pd.DataFrame(index=dfresamplell.index)
dfresampleaverage['15min'] = (dfresamplell['15min'].values+dfresamplerr['15min'].values[:-1])/2
plt.plot(dfresampleaverage.index, dfresampleaverage['15min'], label='average of both resampling methods')
#%% desired output
ydesired = np.zeros(periods/3*2)
i = 0
j = 0
k = 0
for val in ydesired:
if i+k==len(y): k=0
ydesired[j] = np.mean([y[i],y[i+k]])
j+=1
i+=1
if k==0: k=1;
else: k=0; i+=1
plt.plot(dfresamplell.index, ydesired, label='ydesired')
#%% suggestion of Paul H
dfreindex = df.reindex(pd.date_range(startdate, freq='5T', periods=periods*2))
dfreindex.interpolate(inplace=True)
dfreindex = dfreindex.resample('15T', how='first').head()
plt.plot(dfreindex.index, dfreindex['y'], label='method Paul H')
#%% finalize plot
plt.legend()
Implementation for angles
As a bonus I have added the code I will use for the interpolation of angles. This is done by using complex numbers. Because complex interpolation is not implemented (yet), I split the complex numbers into a real and a imaginary part. After averaging these numbers can be converted to angels again. For certain angels this is a better resampling method than simply averaging the two angels, for example: 345 and 5 degrees.
#%% make timestamps
periods = 24*6
startdate = '2010-01-01'
timestamp10min = pd.date_range(startdate, freq='10Min', periods=periods)
#%% Make DataFrame and fill it with some data
degrees = np.cumsum(np.random.randn(periods)*25) % 360
df = pd.DataFrame(index=timestamp10min)
df['deg'] = degrees
df['zreal'] = np.cos(df['deg']*np.pi/180)
df['zimag'] = np.sin(df['deg']*np.pi/180)
#%% suggestion of Paul H
dfreindex = df.reindex(pd.date_range(startdate, freq='5T', periods=periods*2))
dfreindex = dfreindex.interpolate()
dfresample = dfreindex.resample('15T', how='first')
#%% convert complex to degrees
def f(x):
return np.angle(x[0] + x[1]*1j, deg=True )
dfresample['degrees'] = dfresample[['zreal', 'zimag']].apply(f, axis=1)
#%% set all the values between 0-360 degrees
dfresample.loc[dfresample['degrees']<0] = 360 + dfresample.loc[dfresample['degrees']<0]
#%% wrong resampling
dfresample['deg'] = dfresample['deg'] % 360
#%% plot different sampling methods
plt.figure()
plt.plot(df.index, df['deg'], label='normal', marker='v')
plt.plot(dfresample.index, dfresample['degrees'], label='resampled according @Paul H', marker='^')
plt.plot(dfresample.index, dfresample['deg'], label='wrong resampling', marker='<')
plt.legend()
Ok, here’s one way to do it.
- Make a list of the times you want to have filled in
- Make a combined index that includes the times you want and the times you already have
- Take your data and “forward fill it”
- Take your data and “backward fill it”
- Average the forward and backward fills
- Select only the rows you want
Note this only works since you want the values exactly halfway between the values you already have, time-wise. Note the last time comes out np.nan because you don’t have any later data.
times_15 = []
current = df.index[0]
while current < df.index[-2]:
current = current + dt.timedelta(minutes=15)
times_15.append(current)
combined = set(times_15) | set(df.index)
df = df.reindex(combined).sort_index(axis=0)
df['ff'] = df['y'].fillna(method='ffill')
df['bf'] = df['y'].fillna(method='bfill')
df['solution'] = df[['ff', 'bf']].mean(1)
df.loc[times_15, :]
I might be misunderstanding the problem, but does this work?
TL;DR version:
import numpy as np
import pandas
data = np.arange(0, 101, 8)
index_10T = pandas.DatetimeIndex(freq='10T', start='2012-01-01 00:00', periods=data.shape[0])
index_05T = pandas.DatetimeIndex(freq='05T', start=index_10T[0], end=index_10T[-1])
index_15T = pandas.DatetimeIndex(freq='15T', start=index_10T[0], end=index_10T[-1])
df1 = pandas.DataFrame(data=data, index=index_10T, columns=['A'])
print(df.reindex(index=index_05T).interpolate().loc[index_15T])
Long version
setup fake data
import numpy as np
import pandas
data = np.arange(0, 101, 8)
index_10T = pandas.DatetimeIndex(freq='10T', start='2012-01-01 00:00', periods=data.shape[0])
df1 = pandas.DataFrame(data=data, index=index_10T, columns=['A'])
print(df1)
A
2012-01-01 00:00:00 0
2012-01-01 00:10:00 8
2012-01-01 00:20:00 16
2012-01-01 00:30:00 24
2012-01-01 00:40:00 32
2012-01-01 00:50:00 40
2012-01-01 01:00:00 48
2012-01-01 01:10:00 56
2012-01-01 01:20:00 64
2012-01-01 01:30:00 72
2012-01-01 01:40:00 80
2012-01-01 01:50:00 88
2012-01-01 02:00:00 96
So then build a new 5-minute index and reindex the original dataframe
index_05T = pandas.DatetimeIndex(freq='05T', start=index_10T[0], end=index_10T[-1])
df2 = df.reindex(index=index_05T)
print(df2)
A
2012-01-01 00:00:00 0
2012-01-01 00:05:00 NaN
2012-01-01 00:10:00 8
2012-01-01 00:15:00 NaN
2012-01-01 00:20:00 16
2012-01-01 00:25:00 NaN
2012-01-01 00:30:00 24
2012-01-01 00:35:00 NaN
2012-01-01 00:40:00 32
2012-01-01 00:45:00 NaN
2012-01-01 00:50:00 40
2012-01-01 00:55:00 NaN
2012-01-01 01:00:00 48
2012-01-01 01:05:00 NaN
2012-01-01 01:10:00 56
2012-01-01 01:15:00 NaN
2012-01-01 01:20:00 64
2012-01-01 01:25:00 NaN
2012-01-01 01:30:00 72
2012-01-01 01:35:00 NaN
2012-01-01 01:40:00 80
2012-01-01 01:45:00 NaN
2012-01-01 01:50:00 88
2012-01-01 01:55:00 NaN
2012-01-01 02:00:00 96
and then linearly interpolate
print(df2.interpolate())
A
2012-01-01 00:00:00 0
2012-01-01 00:05:00 4
2012-01-01 00:10:00 8
2012-01-01 00:15:00 12
2012-01-01 00:20:00 16
2012-01-01 00:25:00 20
2012-01-01 00:30:00 24
2012-01-01 00:35:00 28
2012-01-01 00:40:00 32
2012-01-01 00:45:00 36
2012-01-01 00:50:00 40
2012-01-01 00:55:00 44
2012-01-01 01:00:00 48
2012-01-01 01:05:00 52
2012-01-01 01:10:00 56
2012-01-01 01:15:00 60
2012-01-01 01:20:00 64
2012-01-01 01:25:00 68
2012-01-01 01:30:00 72
2012-01-01 01:35:00 76
2012-01-01 01:40:00 80
2012-01-01 01:45:00 84
2012-01-01 01:50:00 88
2012-01-01 01:55:00 92
2012-01-01 02:00:00 96
build a 15-minute index and use that to pull out data:
index_15T = pandas.DatetimeIndex(freq='15T', start=index_10T[0], end=index_10T[-1])
print(df2.interpolate().loc[index_15T])
A
2012-01-01 00:00:00 0
2012-01-01 00:15:00 12
2012-01-01 00:30:00 24
2012-01-01 00:45:00 36
2012-01-01 01:00:00 48
2012-01-01 01:15:00 60
2012-01-01 01:30:00 72
2012-01-01 01:45:00 84
2012-01-01 02:00:00 96
In case someone is working with data without regularity at all, here is an adapted solution from the one provided by Paul H above.
If you don’t want to interpolate throughout the time-series, but only in those places where resample is meaningful, you may keep the interpolated column side by side and finish with a resample and dropna.
import numpy as np
import pandas
data = np.arange(0, 101, 3)
index_setup = pandas.date_range(freq='01T', start='2022-01-01 00:00', periods=data.shape[0])
df1 = pandas.DataFrame(data=data, index=index_setup, columns=['A'])
df1 = df1.sample(frac=0.2).sort_index()
print(df1)
A
2022-01-01 00:03:00 9
2022-01-01 00:06:00 18
2022-01-01 00:08:00 24
2022-01-01 00:18:00 54
2022-01-01 00:25:00 75
2022-01-01 00:27:00 81
2022-01-01 00:30:00 90
Notice resampling this DF without any regularity forces values to the floor index, without interpolating.
print(df1.resample('05T').mean())
A
2022-01-01 00:00:00 9.0
2022-01-01 00:05:00 24.0
2022-01-01 00:10:00 39.0
2022-01-01 00:15:00 51.0
2022-01-01 00:20:00 NaN
2022-01-01 00:25:00 79.5
A better solution can be achieved by interpolating in a small enough interval and then resampling. The result DF now has too much, but a dropna() brings it close to its original shape.
index_1min = pandas.date_range(freq='01T', start='2022-01-01 00:00', end='2022-01-01 23:59')
df2 = df1.reindex(index=index_1min)
df2['A_interp'] = df2['A'].interpolate(limit_direction='both')
print(df2.resample('05T').first().dropna())
A A_interp
2022-01-01 00:00:00 9.0 9.0
2022-01-01 00:05:00 21.0 15.0
2022-01-01 00:10:00 39.0 30.0
2022-01-01 00:15:00 51.0 45.0
2022-01-01 00:25:00 75.0 75.0