G'day Newb!

You are fortunate that a community member asked a similar question - the attached sample codes may be what you are looking for!

Hope this helps.

This example deals with nans (not a number). It happens when the fundamental value is absent for certain stocks.
Around 40% of the original calculated values are nan without a screen.
They can be screened on your input values being used (the Two-value screen section) and then shows that the built-in Fundamentals.roic still has some nans. That's on stocks where your two fundamental inputs were ok, they had populated values. So your route might be better than the built-in.
Then the complete screen section just adds removal of Fundamentals.roic nans, and I called that ROIC_f for short.
The ROIC() class is looking at a window of time and doing forward fill as a way of dealing with nans.
So, maybe that'll help some too. One might wonder why the minimum values (ROC vs ROIC_f) are so different, by the way.
Run that and look at the logging window, it also shows specific stock highs and lows for each column, I didn't paste them here.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  
No screen  
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  
        #screen  = NI.notnull() & IC.notnull(),  
        #screen  = NI.notnull() & IC.notnull() & ROIC_f.notnull(),

2019-01-03 05:45 log_data:146 INFO Stocks: 8688  
                      min                  mid               max  
    IC      -1848599000.0         4436131000.0               nan     NaNs 3247/8688  
    NI     -22769000000.0          123832000.0               nan     NaNs 3584/8688  
   ROC     -16295.4934498      0.0410566685425               nan     NaNs 3586/8688  
ROIC_f         -87.080808             0.035863               nan     NaNs 3473/8688  
 class     -16295.4934498     0.00941190521827     193.560874895

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  
Two-value screen  
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  
2019-01-03 05:45 log_data:146 INFO Stocks: 5102  
    pipe  = Pipeline(  
        screen  = NI.notnull() & IC.notnull(),  
        #screen  = NI.notnull() & IC.notnull() & ROIC_f.notnull(),

2019-01-03 05:45 log_data:146 INFO Stocks: 5102  
                      min                 mid                max  
    IC      -1848599000.0         758137000.0     565951000000.0  
    NI     -22769000000.0           5734000.0      18540000000.0  
   ROC     -16295.4934498     0.0107877060859       46.899068323  
ROIC_f         -87.080808            0.013287                nan     NaNs 230/5102  
 class     -16295.4934498     0.0107877060859       46.899068323

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  
Complete screen  
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -  
2019-01-03 05:45 log_data:146 INFO Stocks: 4872  
    pipe  = Pipeline(  
        #screen  = NI.notnull() & IC.notnull(),  
        screen  = NI.notnull() & IC.notnull() & ROIC_f.notnull(),

2019-01-03 05:45 log_data:146 INFO Stocks: 4872  
                      min                 mid                max  
    IC      -1848599000.0         759468000.0     565951000000.0  
    NI     -22769000000.0           5683000.0      18540000000.0  
   ROC     -16295.4934498     0.0105196594322       46.899068323  
ROIC_f         -87.080808            0.012086         165.954222  
 class     -16295.4934498     0.0105196594322       46.899068323  

Hi Gary Blue,

I routinely avoid the .fillna() or nanfill() methods in data series as the data fills have produced (for me anyway) misleading and unpredictable results. I prefer to .dropna() outright for reason of data hygiene - use only those data that fit the purpose to compute the factors. However, on occasions when I really must approximate the data gaps, I use the np.polyfit() method instead.