resample datetimeindex via prod() function changes NaN to 1

Question

I am working with a rather large dataset. After applying the resample command in combination with the conversion method "prod" (multiplication), I realized that my NaN values were changed to 1, which is not what I intended. To give an example what happened:

# build random dataframe with one column containing NaN
import pandas as pd
import numpy as np

index = pd.date_range('1/1/2000', periods=7, freq='d')
df = pd.DataFrame(index = index, columns = ["Score 1", "Score 2", "Score 3"])

df["Score 1"] = np.random.randint(1,20,size=7)
df["Score 2"] = np.random.randint(1,20,size=7)
df["Score 3"] = [1, 2, 3, np.NaN, np.NaN, np.NaN, np.NaN]
print(df)

            Score 1     Score 2     Score 3
2000-01-01  6            7          1.0
2000-01-02  2            15         2.0
2000-01-03  8            19         3.0
2000-01-04  14           19         NaN
2000-01-05  17           8          NaN
2000-01-06  15           6          NaN
2000-01-07  12           18         NaN

Now lets say I want to resample my Dataframe from a daily to a 3-day Frequency with using the "prod" conversion method. I do so by:

df.resample("3d").agg("prod")
print(df)

            Score 1     Score 2     Score 3
2000-01-01  96          1995        6.0
2000-01-04  3570        2052        1.0
2000-01-07  12            18        1.0

Looking at the column "Score 3", my NaN values suddenly changed to 1, which is a surprise for me. This means that when multiplying NaN with each other, I would get =1. Does anyone why exactly a multiplication of NaN's equals one and what I could do to keep the NaN value in case it is multiplicated with itself?

Thanks in advance, any help is highly appreciated

Answer 1

pandas.DataFrame.prod function ( docs ) by default sets NaN to 1:

pd.Series([np.NaN, np.NaN]).prod()
# 1.0

You can circumvent this by setting the according keyword:

pd.Series([np.NaN, np.NaN]).prod(skipna=False)
# nan

In your case, you could apply that as

print(df)
            Score 1  Score 2  Score 3
2000-01-01       18       19      1.0
2000-01-02        9       18      2.0
2000-01-03       10        4      3.0
2000-01-04        4       15      4.0
2000-01-05       12        1      NaN
2000-01-06        1        3      NaN
2000-01-07        8        9      NaN

print(df.resample("3d").agg(pd.DataFrame.prod, skipna=False))
            Score 1  Score 2  Score 3
2000-01-01     1620     1368      6.0
2000-01-04       48       45      NaN
2000-01-07        8        9      NaN

Note that this will set all resampled time windows to NaN if the window contains at least one NaN value - I changed the example df slightly to show that. You can apply a lambda instead, checking if at least one element is not NaN :

print(df.resample("3d").apply(lambda x: x.prod() if any(x.notnull()) else np.nan))
            Score 1  Score 2  Score 3
2000-01-01     1620     1368      6.0
2000-01-04       48       45      4.0
2000-01-07        8        9      NaN