Prep for Fall forecasting season

[dsweber2]

Essential

• [ ] Make sure we have a method that works for low/no data regime, such as
  • [x] cmu-delphi/exploration-tooling#126
  • [ ] 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
• [ ] Once the data from nhsn is live, do a mock-run of the production forecast to work out any kinks it might have
• [x] have a mathematical write-up
• [ ] decide on whether and how to do Peak timing, intensity, and sample trajectories

Further steps requiring backtesting

• [ ] #131
  • [ ] optum, either
  • [ ] outpatient https://cmu-delphi.github.io/delphi-epidata/api/covidcast-signals/doctor-visits.html
  • [ ] inpatient https://cmu-delphi.github.io/delphi-epidata/api/covidcast-signals/hospital-admissions.html
  • [ ] google search trends https://cmu-delphi.github.io/delphi-epidata/api/covidcast-signals/google-symptoms.html
  • [ ] NSSP
  • [ ] NWSS wastewater
• [ ] trailing window sizes
  • [ ] trailing windows that only use the comparable segment of last year
• [ ] add method to fit multiple models in the case that some geos are missing (e.g. mo for NSSP)
• [ ] cmu-delphi/exploration-tooling#127
• [ ] differencing
• [ ] flusion-style data augmentation
  • [ ] cmu-delphi/exploration-tooling#125
  • [ ] add NSSP as a fused data source (this works for COVID in addition to Flu)
• [ ] cmu-delphi/epipredict#415

@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 13
• NHSN details (probably linked previously)

[dshemetov]

Dmitry's tasks 2024-10-22:

• Calibration
  • Adapt Dan’s implementation into an exploration + production script and tune parameters
  • Learning rate - .1 in paper, probably somewhere between 0.05 and 0.5
  • Window in which to build the bounds - log, 3 or 4 values, 10, 25, 50, all
  • Grid across these. Maybe at some point do these independently for each quantile later.
  • There’s a for loop that probably just needs to be epix_slide
  • Just get best forecaster from David’s experiments and calibrate that
  • For now, can just use last year’s forecaster and forecasts to get intuition for tuning parameters
  • Lengthy Slack thread
• Exogenous data sources exploration (slack link)