Pandas rolling apply returning NaN in unexpected ways

I have a dataframe where one of my columns contains NaNs and I'm attempting to find the rolling average of the valid numbers in the column. I'm attempting to do this with rolling apply, using a custom function and passing the rolling window raw. Alternatives I've seen are to replace the NaNs with 0, but as I understand it this would artificially deflate my mean (more entries, same sum, smaller average).

import pandas as pd
import numpy as np

mySpecialNumber = 50

def nanMean(arr):
    return arr[~np.isnan(arr)].mean()

df = pd.read_csv('myfile.csv')
df['rollingAVG'] = df['nanColumn'].rolling(mySpecialNumber).apply(nanMean, raw=True)

The above fills df['rollingAVG'] with NaN values, only NaNs. Whereas I can use:

singleAVG = nanMean(df['nanColumn'][0:mySpecialNumber].values)

And get a totally reasonable average that accurately reflects the by-hand calculation.

Hope this isn't a duplicate; I couldn't find anything else like it around, so I suspect I must be goofing. I'm also open to alternatives that still produce the results I'm looking for, but this seemed like the most straightforward way I could find.

I'm on macOS 13.4.1, running python 3.11.4, pandas 2.0.3, and numpy 1.25.0.

Thanks in advance.

EDIT: As suggested I have previously tried to use

df['rollingAVG'] = df['nanColumn'].rolling(mySpecialNumber).mean()

as well as

df['rollingAVG'] = df['nanColumn'].rolling(mySpecialNumber).apply(np.mean, raw=True)

but the problem persists; df['rollingAVG'] is still full of NaNs.

I can also guarantee the array is full of NaNs; I'm not just looking at the first or last 50 entries, I have set np.set_printoptions(threshold=sys.maxsize) and can verify that df['rollingAVG'].values is a ~1300 entry array full of NaNs.

You should be providing a value to min_periods. Please check pandas documentation for this argument in .rolling() function:

> min_periods: int, default None Minimum number of observations in
> window required to have a value; otherwise, result is np.nan.
> For a window that is specified by an offset, min_periods will default
> to 1. For a window that is specified by an integer, min_periods
> will default to the size of the window.

In your case, min_periods defaults to window size, which is the value of mySpecialNumber, which results in output NaN if any of your 50 values in the window is NaN. Check this example:

df = pd.DataFrame({'a': [None, 2, 3]})

In [12]: df.rolling(3)['a'].mean()
Out[12]: 
0   NaN
1   NaN
2   NaN
Name: a, dtype: float64

If you pass min_periods, you'll start to see outputs:

In [13]: df.rolling(3, min_periods=1)['a'].mean()
Out[13]: 
0    NaN
1    2.0
2    2.5
Name: a, dtype: float64

I would recommend setting min_periods to something that make sense for your dataset, depending on how many values could be missing.