r package framework

[vpnagraj]

moving towards our first submission (and beyond) we should try to get our code into a stable, distributable format ... ie. an R package

i'll set up the skeleton to conform to the usual standards (R code in R/, roxygen, buildignore stuff, etc)

[stephenturner]

keeping it all in this focustools repo?

[vpnagraj]

@stephenturner good question.

yeah that was my thought. no need to nuke all the scratch/* code to do that either ... we can just buildignore any directories that arent part of the pkg until we smooth out the internals

[stephenturner]

I'm going to suggest we clean up some of the nonstandard dirs laying around here. suggestions:

• move stuff in utils to scratch and rmdir utils
• move scratch dir into misc
• move internal dir into misc
• rename assets to data-raw

[vpnagraj]

yeah thats more or less what i was thinking

in progress now. will comment here when i push up to another branch / do a PR

[vpnagraj]

working on an outline for the package at https://github.com/signaturescience/focustools/tree/pkg

my todo before i open a PR and merge:

• [x] finish first pass at function docs
• [x] first pass at readme with demo of ts analysis
• [x] make sure devtools::check (pretty much) works
• [x] actually implement data-raw correctly and generate_sysdata.R for locations.csv
• [x] write ts_fit() function
• [x] move some of the misc exploration code (like what is in re_estimation.R script) to its own functions

[vpnagraj]

this is done and implemented as of #13

fully documented. no ERRORS, WARNINGS, NOTES per devtools::check()

there are a ton of new features with the R package API / directory structure:

• most importantly ... we no longer need to rely on sourcing helper functions scattered in the scratch dir
• the scratch dir still exists (and is ignored in Rbuildignore) though i've started moving some of those scripts to a nested scratch/archive. short term, we can hold onto those scripts ... long term we can nuke them.
• the ts_forecast() function now has a consistent API for all outcomes (icases, ideaths, cdeaths) thanks to a new, unexported helper (ts_cumulative_forecast())
• the new ts_fit() function wraps fable::model() so we can iterate over multiple outcomes/models simulatenously (note: there are some important caveats with that one, and you don't necessarily need to use to use ts_forecast() )
• ts_accuracy() is also new, and basically wraps ts_fit() and runs the fabletools::accuracy() measures on an arbitrary horizon (say you want to hold out horizon = 4 weeks for test, train on the rest)
• a few misc helpers (this_monday(), ts_futurecases()) to limit the code we have to write for generating submissions

the analysis in scratch/fable-submission-mockup-allmetrics.R is reimplemented using the new API in the README

we can talk about maintenance / pkg dev stuff on our next call. for the time being, if you're curious about anything here i'd recommend starting with https://r-pkgs.org/

heads up i'm going to punt some of the work to capture relevant exploration code as functions to its own issue(s)

closing this issue.

cc @stephenturner @chulmelowe