Prep for Fall forecasting season

[dsweber2]

Essential

• Make sure we have a method that works for low/no data regime, such as
  • pure week-number or
  • an alternate data source, such as NSSP and/or wastewater
• adapt to purely weekly forecasts- unclear if that is the data regime we're actually in, but we need to be ready for it if it is. See here for up to date discussion
• Do a mock-run of the production forecast to work out any kinks it might have

Further steps requiring backtesting

• flusion-style data augmentation
• Exogenous data sources #131
• seasonal methods (note that this should be a small handful of features, aligned to the target_date)
  • PCA as a yearly component #132
  • indicators for pre-peak, peak, post-peak
  • sine/cosine terms (currently only in no_recent_data, also should be tested augmenting simple AR)
• quasi-geo indicator that creates a different intercept for states with similar errors or some sort of geo information #159
• ensemble of sufficiently different models (likely to be recurring)
• predicting differenced output rather than raw output (this is a difference with the flusion model) #160
• Calibration method #134
• case weights

@dajmcdon @dshemetov trying to transfer our disparate notes on what to do to an actionable gh issue. Any ideas I missed?

Timelines:
Forecasting hub websites and timing for the link to the meeting at the perfect time!):
All of these are due Wednesday night at 11PM Eastern.
RSV:

• website, github likely starting Oct 14 with RSV-NET, moving over to nhsn when available

COVID-19:

• github, epiweek totals. They have optional targets of trajectories (doesn't make much sense for us tbh). Likely starting Nov 13

Flu:

• no linked website/github yet, similar to last year/ COIVD-19. additional optional targets of peak week and peak intensity, which we may want to consider. I didn't quite catch which of the non-states they were forecasting; I think DC and maybe Puerto Rico? Likely starting Nov 20
• NHSN details (probably linked previously)