Skip to content

Latest commit

 

History

History
27 lines (16 loc) · 2.31 KB

UPDATING.md

File metadata and controls

27 lines (16 loc) · 2.31 KB
Raw

Updating the data and forecasts

The create-data, modelrunner, and dashboard folders contain the core components of the project. They are designed to be self-contained, and outputs from one that serve as inputs in others have to be manually copied over. See the respective README's for more details. I also tried to mark all places that require manual updates with "UPDATE:".

The general process is to:

  1. Update the external data (this part is shared with PART) in andybega/ds-external-data.
  2. Update the config settings in config.yml.
  3. Update the merged data using create-data, including the external data sources that feed into the final merged data. Copy the updated states.rds data to modelrunner/input/.
  4. Run the forecast model using modelrunner/R/rf.R. Copy modelrunner/output/fcasts-rf.csv to dashboard/data/.
  5. Update the dashboard by rebuilding the data and manually updating the text in the UI where needed.

Things to improve in the next cycle

Ideas for what to improve in the next update cycle.

  • In 2022, I ran tuning experiments to pick fixed HPs. However, I did that before the ERT-lite change, i.e. with "v12" not "v12.1" data. So those could probably be re-run.
  • I started a skeleton R package during the last update. There are two basic directions to go in with this: (1) throw everything--data, models, forecasts, dashboard--into a big monolithic demspaces package, or (2) do a demspaces-verse with more lighweight packages, e.g. for the dashboard. Right now I'm thinking do a lightweight package with the current forecasts and dashboard, and then leave everyhing else in a big dev package. Hmm.

Items from the 2021 update that I didn't spend time on (I think):

  • Clean up the differentiation between primary outputs and ancillary trackers for git. Like put the latter in tracker folders or something like that.
  • Makefile automation. Although the data updating requires manual, interactive checking, most of the other stuff down the road lends itself to full automation. To that end I started writing both versioned and un-versioned file names already, but this can probably be taken further. Also, make life is easier if output file names correspond to R script names; then I could use pattern rules instead of the current ugly mess of targets and recipes.