flusight_hub_formatter: Format predictions for submission to FluSight forecast Hub

In cmu-delphi/epipredict: Basic epidemiology forecasting methods

Format predictions for submission to FluSight forecast Hub

Description

This function converts predictions from any of the included forecasters into a format (nearly) ready for submission to the 2023-24 FluSight-forecast-hub. See there for documentation of the required columns. Currently, only "quantile" forcasts are supported, but the intention is to support both "quantile" and "pmf". For this reason, adding the output_type column should be done via the ... argument. See the examples below. The specific required format for this forecast task is here.

Usage

flusight_hub_formatter(object, ..., .fcast_period = c("daily", "weekly"))

Arguments

object	a data.frame of predictions or an object of class canned_epipred as created by, e.g., arx_forecaster()
...	<dynamic-dots> Name = value pairs of constant columns (or mutations) to perform to the results. See examples.
.fcast_period	Control whether the horizon should represent days or weeks. Depending on whether the forecaster output has target dates from layer_add_target_date() or not, we may need to compute the horizon and/or the target_end_date from the other available columns in the predictions. When both ahead and target_date are available, this is ignored. If only ahead or aheads exists, then the target date may need to be multiplied if the ahead represents weekly forecasts. Alternatively, if only, the target_date is available, then the horizon will be in days, unless this argument is "weekly". Note that these can be adjusted later by the ... argument.

Value

A tibble::tibble. If ... is empty, the result will contain the columns reference_date, horizon, target_end_date, location, output_type_id, and value. The ... can perform mutations on any of these.

Examples

library(dplyr)
library(epiprocess)
weekly_deaths <- covid_case_death_rates %>%
  filter(
    time_value >= as.Date("2021-09-01"),
    geo_value %in% c("ca", "ny", "dc", "ga", "vt")
  ) %>%
  select(geo_value, time_value, death_rate) %>%
  left_join(state_census %>% select(pop, abbr), by = c("geo_value" = "abbr")) %>%
  mutate(deaths = pmax(death_rate / 1e5 * pop * 7, 0)) %>%
  select(-pop, -death_rate) %>%
  group_by(geo_value) %>%
  epi_slide(~ sum(.$deaths), .window_size = 7, .new_col_name = "deaths_7dsum") %>%
  ungroup() %>%
  filter(weekdays(time_value) == "Saturday")

cdc <- cdc_baseline_forecaster(weekly_deaths, "deaths_7dsum")
flusight_hub_formatter(cdc)
flusight_hub_formatter(cdc, target = "wk inc covid deaths")
flusight_hub_formatter(cdc, target = paste(horizon, "wk inc covid deaths"))
flusight_hub_formatter(cdc, target = "wk inc covid deaths", output_type = "quantile")

cmu-delphi/epipredict documentation built on March 5, 2025, 12:17 p.m.