Detect accumulation frequency

[sbfnk]

          > I might also be in favour of some attempt to detect the accumulation in the current data to avoid having to respecify it

We could do that if all the gaps are equal. Same with the initial_accumulate argument.

Originally posted by @seabbs and @sbfnk in #867 (comment)

[jamesmbaazam]

I would rather be in favour of letting the user control this as has been suggested in #867 (comment) and #867 (comment). As we've discussed in the linked issue, this only works well if the accumulation pattern is regular. If it's not, what do we do?

I wonder if we want to do the forecast accumulation in the modelling step or as a post-processing step?

For example, we could generate daily forecasts and have a post-processing function that accumulates the data in any pattern using a setup similar to what fill_missing() does. There might be some computational issues but might be worth exploring.

[seabbs]

I like the idea of telling people if it has happened and how to override it

[seabbs]

For example, we could generate daily forecasts and have a post-processing function that accumulates the data in any pattern using a setup similar to what fill_missing() does.

I like this idea a lot but it would require a lot of work to reduce the pipeline like nature of the output or to get people used to more steps at least

[sbfnk]

If it's not, what do we do?

Nothing (i.e. the current default)?

[sbfnk]

For example, we could generate daily forecasts and have a post-processing function that accumulates the data in any pattern using a setup similar to what fill_missing() does.

I like this idea a lot but it would require a lot of work to reduce the pipeline like nature of the output or to get people used to more steps at least

I agree it's a nice idea - in addition to this there's also the issue that currently in the model observation noise is applied after accumulation.

[seabbs]

in addition to this there's also the issue that currently in the model observation noise is applied after accumulation.

As in this is a model fault or makes this hard? I am not sure its a model fault but a choice?

[sbfnk]

I'm not sure we'll be able to fit otherwise as the observation noise defines the likelihood and thus can only apply when there are observations.

[seabbs]

I think you could do it but not in any clean or reasonable way

[jamesmbaazam]

I believe this has been addressed in #901 and should be closed?

[sbfnk]

I believe this has been addressed in #901 and should be closed?

It was done for forecasts but not yet for initial_accumulate