Upsampling & interpolation with .resample()

Manipulating Time Series Data in Python

Stefan Jansen

Founder & Lead Data Scientist at Applied Artificial Intelligence

Frequency conversion & transformation methods

.resample(): similar to .groupby()
Groups data within resampling period and applies one or several methods to each group
New date determined by offset - start, end, etc
Upsampling: fill from existing or interpolate values
Downsampling: apply aggregation to existing data

Getting started: monthly unemployment rate

unrate = pd.read_csv('unrate.csv', parse_dates['Date'], index_col='Date')

unrate.info()

DatetimeIndex: 208 entries, 2000-01-01 to 2017-04-01
Data columns (total 1 columns):
UNRATE    208 non-null float64 # no frequency information
dtypes: float64(1)

unrate.head()

            UNRATE
DATE
2000-01-01     4.0
2000-02-01     4.1
2000-03-01     4.0
2000-04-01     3.8
2000-05-01     4.0

Reporting date: 1st day of month

Resampling Period & Frequency Offsets

Resample creates new date for frequency offset
Several alternatives to calendar month end

Frequency	Alias	Sample Date
Calendar Month End	M	2017-04-30
Calendar Month Start	MS	2017-04-01
Business Month End	BM	2017-04-28
Business Month Start	BMS	2017-04-03

Resampling logic

ch2_3_v2 - Upsampling & Interpolation.015.png

Resampling logic

ch2_3_v2 - Upsampling & Interpolation.016.png

Assign frequency with .resample()

unrate.asfreq('MS').info()

DatetimeIndex: 208 entries, 2000-01-01 to 2017-04-01
Freq: MS
Data columns (total 1 columns):
UNRATE    208 non-null float64
dtypes: float64(1)

unrate.resample('MS') # creates Resampler object

DatetimeIndexResampler [freq=<MonthBegin>, axis=0, closed=left, 
                        label=left, convention=start, base=0]

Assign frequency with .resample()

unrate.asfreq('MS').equals(unrate.resample('MS').asfreq())

True

.resample(): returns data only when calling another method

Quarterly real GDP growth

gdp = pd.read_csv('gdp.csv')

gdp.info()

DatetimeIndex: 69 entries, 2000-01-01 to 2017-01-01
Data columns (total 1 columns):
gpd    69 non-null float64 # no frequency info
dtypes: float64(1)

gdp.head(2)

            gpd
DATE
2000-01-01  1.2
2000-04-01  7.8

Interpolate monthly real GDP growth

gdp_1 = gdp.resample('MS').ffill().add_suffix('_ffill')

       gpd_ffill
DATE
2000-01-01  1.2
2000-02-01  1.2
2000-03-01  1.2
2000-04-01  7.8

Interpolate monthly real GDP growth

gdp_2 = gdp.resample('MS').interpolate().add_suffix('_inter')

            gpd_inter
DATE
2000-01-01  1.200000
2000-02-01  3.400000
2000-03-01  5.600000
2000-04-01  7.800000

.interpolate(): finds points on straight line between existing data

Concatenating two DataFrames

df1 = pd.DataFrame([1, 2, 3], columns=['df1'])

df2 = pd.DataFrame([4, 5, 6], columns=['df2'])

pd.concat([df1, df2])

   df1  df2
0  1.0  NaN
1  2.0  NaN
2  3.0  NaN
0  NaN  4.0
1  NaN  5.0
2  NaN  6.0

Concatenating two DataFrames

pd.concat([df1, df2], axis=1)

   df1  df2
0    1    4
1    2    5
2    3    6

axis=1: concatenate horizontally

Plot interpolated real GDP growth

pd.concat([gdp_1, gdp_2], axis=1).loc['2015':].plot()

ch2_3_v2 - Upsampling & Interpolation.032.png

Combine GDP growth & unemployment

pd.concat([unrate, gdp_inter], axis=1).plot();

ch2_3_v2 - Upsampling & Interpolation.034.png

Let's practice!

Manipulating Time Series Data in Python