Issue 110: Add epiweekly fits and scoring

[SamuelBrand1]

This PR closes #110 .

NB this PR uses forecasttools functions, in particular daily_to_epiweekly. Very handy.

The goal is to compare the epiweekly forecasts of baseline models (which are fitted to epiweekly data) to the epiweekly grouped forecasts from pyrenew_hew.

The daily analysis is unaffected; that is baseline models are still run for daily reporting as well.

Tasks to contribute

• Add generate epiweekly truth data from daily truth data step
• Add run baseline models on epiweekly
• Add extra pyrenew-hew postprocess to chunk the daily posterior preds into epiweekly
• check how much revision the scoring function needs (might be zero)

Testing

At each step the test mode pipeline is run to check for failure or breaking changes.

[SamuelBrand1]

Added a step to convert daily resolved data to epiweekly cf generate_epiweekly.R.

[SamuelBrand1]

Modified timeseries_forecasts.R to run baseline forecasters for both for daily and epiweekly models.

[SamuelBrand1]

Added scoring on epiweekly scale for pyrenew. The algorithm is that if epiweekly = TRUE then the pyrenew forecasts for counts and other counts are group summed by epiweek (which auto does by draw), pivoted out to create the correct weekly percentage (which is not the mean of the daily percentages), and then pivoted longer.

Dates for the week are supplied by more forecasttools helpers.

No other changes were need because the scoring is flexible enough to ingest this.

[SamuelBrand1]

This is ready for review now.

Note that the plotting is unchanged and the ultimate scoring end product is both:

• score_table.rds
• epiweekly_score_table.rds

The collate functions seem to be currently broken, but I'm interpreting that as a new issue.

[SamuelBrand1]

Also note that this PR has inherited the failing CI from #160

[damonbayer]

@SamuelBrand1 Since we are making even more model results and anticipate adding PyRenew-EW to the workflow soon, I would really like to push toward #101. Could the fable-related parts of that issue be accomplished in this PR (i.e. let's store all the results from these time series models in a fable directory)?

[damonbayer]

Nothing that would necessarily prevent a merge, but I had lots of comments and questions.

[SamuelBrand1]

I've closed the comments marked Outdated and we are onto our last few.

I think @damonbayer remains the reviewer here because this PR is a mix of me and @dylanhmorris

[damonbayer]

@dylanhmorris This PR has my tentative approval. Please respond to (or make changes based on) the remaining comments before merging.