Forecast Checks

Note: As of February 20, 2023 we are no longer collecting data or analyzing COVID-19 cases and as of March 6, 2023 we are no longer collecting data or analyzing COVID-19 deaths.

• header must minimarlly include location, target, type, quantile, value (required for zoltpy) and forecast_date, target_end_date

• each row must have the same number of columns as header

• location must be in "locations" column of locations.csv

• target must be in

  paste(1:20,  "wk ahead inc death")
  paste(1:20,  "wk ahead cum death")
  paste(0:130, "day ahead inc hosp")
  paste(1:8, "wk ahead inc case")

  county locations should have only "case" targets

• forecast_date and target_end_date must be in YYYY-MM-DD format. Additionally, forecast_date should be within ±1 day of the date mentioned in the forecast filename. E.g. - A file in data-processed/model/2021-04-12-model.csv should have forecast_date within 2021-04-11 - 2021-04-13.

• the set of quantiles for targets other than cases must include this entire set of quantiles

  c(0.01, 0.025, seq(0.05, 0.95, by = 0.05), 0.975, 0.99)

• the set of quantiles for "case" targets must include this entire set of quantiles

  c(0.025, 0.100, 0.250, 0.500, 0.750, 0.900, 0.975)

• checks quantile must be an int or float in [0, 1]

• checks value must be an int or float and non-negative, except for retractions as detailed below

  • Forecast retractions: If you want to retract some existing forecast rows in a file, you can do so by specifying NULL (no quote marks), not NA, None, or anything else. More details are mentioned here.

• validates date alignment as documented in the issue add additional validations

• validates quantiles and values (i.e., at the prediction level):

  • checks that entries in value must be non-decreasing as quantiles increase
  • checks that elements in the quantile are unique

• validates quantiles as a group:

  • there must be zero or one point prediction for each location/target pair

• Validates if the prediction value for a location is at least less than that location's population.

  • this check is run for all forecast submissions for all targets (in/cum deaths/cases).
  • the population truth data is present in the locations.csv file.
  • To check which predictions are violating, check the logs in the Github Actions build of your PR and the invalid predictions should be printed there.