I have data on frequencies (N), for combinations of [from, to, subset], and the month. Importantly, when N=0, the row is missing.

                   N      from      to      subset
month                                             
1996-01-01  8.956799         1       2           0
1996-02-01  2.068997         1       2           0
1996-03-01  1.086952         1       2           0
1996-05-01  7.103955         1       2           0

I want to compute a rolling average of N over these, considering the zeros.

df.groupby(['from', 'to', 'subset.rolling(3, center=True).mean()

This works, but does not treat missing rows as zeros. So, I first have to fill the misisng rows.

df.reset_index().set_index(['month', 'from', 'to', 'subset']).resample('1M', level=0).fillna(0)

This does not work, and gives me ValueError: Upsampling from level= or on= selection is not supported, use .set_index(...) to explicitly set index to datetime-like.

Can I not upsample for a multi-index? Should I try a different approach altogether?