Migrate the delphiBackfillCorrections package from the covid-19 repo

[nmdefries]

Description

Move the covid-19/jingjing/backfill_correction branch and associated commit history to covidcast-indicators.

Used this approach to preserve the commit history. Additional Slack discussion here.

Additional package setup (e.g. tests on push) will be done in a separate PR.

Changelog

• Current version of delphiBackfillCorrections package.Slack discussion here

Fixes

Partially addresses #1700

[krivard]

some followups, but also @nmdefries something has gone a bit weird with the file organization here -- there are files that used to be in delphiBackfillCorrection that are now only found deep in the delphiBackfillCorrection.Rcheck directory. What is supposed to be here?

[krivard]

are there any constraints on the name, like does it have to fit a certain pattern? otherwise how does the code know which columns do what?

[nmdefries]

Whoops, I added local test output files when renaming the directory. Removed in fb38e07

[jingjtang]

are there any constraints on the name, like does it have to fit a certain pattern? otherwise how does the code know which columns do what?

@krivard There would be a constraint when the functions are used for tooling.

But here, for the daily production, the uncertainty is created by Quidel since Quidel has multiple age group signals. We don't want to add burden on the storage of backfill files. So, currently, for all the other indicators, num and denom are the two fixed columns, but we have some extra ones like num_age_0_4 etc. for Quidel.

[krivard]

I believe this is a complete pass, and I'm provisionally fine with merging this as-is or with the suggested changes.

I do want to flag a couple of things for later discussion and possible refactor:

• keeping track of models and their training end dates
• flow of training and test data through prediction and training modes

[krivard]

If you don't set it to 1, what should it be set to?

[jingjtang]

Since we run it daily, we can just set it to 1. If for some reasons we don't have the prediction result for several days and don't traverse the training date, then we can produce the prediction for those days at once by change the value of testing window

[krivard]

by change the value of testing window

change it to what?

[jingjtang]

To what ever number of days of prediction we missed.

[jingjtang]

keeping track of models and their training end dates

We have training_end_dates to be a prefix of the model names, so we do keep records of themodels

[krivard]

We have training_end_dates to be a prefix of the model names, so we do keep records of themodels

Yes -- the system of model filenames and additional metadata files is needlessly complicated. We can discuss a refactor in a future version.

[nmdefries]

@krivard This is ready to merge.

Additional package setup (e.g. tests on push) will be done in a separate PR. Expect some changes after the planning meeting.