DataFrame.pct_change should default fill_method=None instead of 'pad'

[selik]

Both DataFrame.pct_change and Series.pct_change should have default behavior that matches the "obvious" implementation:

lag = series.shift(1)
change = (series - lag) / lag

Unfortunately, the default behavior is set to fill_method='pad' which surprises many people (#6389). We should start a deprecation cycle to change the default to behave as expected, fill_method=None. I don't believe this will be disruptive, as most people think that's how it behaves currently, as far as I know.

[WillAyd]

Can you add a reproducible example? It's not quite clear from the code what you would be looking for.

FWIW #25291 has some new functionality which maybe covers what you are looking for, though that PR has gone stale. If it covers it would certainly welcome you picking that one up

[selik]

Looks like @TomAugspurger made a comment to this effect on that PR. I'll check it out... someday :-) .

[WillAyd]

OK - closing this as a duplicate of what the PR already solves then, but again if it's something different feel free to provide a reproducible code sample with expected values and can reopen

[selik]

Ah, no, I don't think that PR was meant to address this issue directly. There happened to be a discussion of this issue as related to PR #25291, in that there's a discussion of what the best default should be. That PR addresses #25006, which looks at the behavior of fill_method='pad'.

[selik]

I misspoke earlier. The referenced PR does not put the default behavior into a deprecation cycle, which is the goal of this issue.

[maartenb]

As this is not a duplicate, can the duplicate label be removed?

What needs to be done to get this issue resolved/implemented?

[jreback]

this is a duplicate of #25006

there is a closed PR that almost ready (referenced from the above issue)

[maartenb]

The approach @selik proposed is different and well defined. If there is a sequence of NaNs it might not make sense to do a pct_change over the gap. Also, it seems that PR for #25006 is not really in play anymore, possibly because it also needs a deprecation period?

[selik]

@jreback I looked at PR #25006 and don't think it's going to solve this issue. The discussion of that issue does not indicate changing the default behavior from fill_method='pad' to fill_method=None.