I ask myself the same question regularly. The answer to this depends on what type of dataframe you are creating, that is, are you creating a time series, cross section, panel. My dataframes tend to become cross sectional that look like this usually:

date ticker var1 var2 var2

In this case, I would think it would be easier to do the computation before the dataframe is created, using the class within the pipeline. If you wanted to do this in the dataframe, you would need to add all those past ROAs to the dataframe. Instead, if you do it in the pipeline you can do that ahead of time and keep the dataframe simple.

The answer provided by @Niccola Tartaglia is good advice. Generally, do computations in pipeline when possible.

However, it's not always a clear cut answer. For simple arithmetic and boolean logic between factors, pipeline uses numexpr, which is very fast, faster than pandas or numpy manipulation of dataframe columns. However, if one is applying a very selective screen to your pipeline, there may be much less data to operate on after the pipeline is run. In this case, even though the computations may be slower, the total time may be faster (by operating on less data).

Logic inside custom factors is less clear cut. If one executes identical numpy calculations inside a custom factor vs after the pipeline is returned, the speed will probably be comparable. Again, it depends upon how much data is being operated on so pay attention to masks and screens.

Pipeline is typically more memory efficient too. If one is outputting a number of columns just to combine them later in pandas, you'll save memory by doing the work in pipeline. Additionally, it's easy to unintentionally end up with copies of the output dataframe when doing calculations after a pipeline is returned.

So, generally, if you can do something easily in pipeline, it's probably faster and/or more memory efficient to do it there.

Now, the second half to the original question was "how do I go about this? Simple if statements never seem to work" . True, in defining calculations for pipeline, 'if statements' are not defined. Often one doesn't need them however. Maybe take a look at this post for some ideas on how to implement Piotroski F score (https://www.quantopian.com/posts/piotroskis-f-score-algorithm )

Good luck.