Wow, this looks really interesting. Was there any particular reason you decided on this approach?

Building up complicated time series transformations by composing simple functions helps me be sure I'm doing what I think I'm doing. Since the transformations are tacit expressions that don't define specific parameters they are then very easy to re-use and combine. And I have some combinators that can apply the functions in pretty flexible ways. Also, since its all lazy-evaluated on a per-column level I can work with huge tables, but only end up operating on the subset I need.

How is that more interesting than simple .modify_column(params)?

You definitely could do a similar column-level functional approach that way, but I think the simple syntax of the stack-oriented approach makes it easier to read and catch errors. The code would be a lot longer that way.