This function takes probabilistic flu hospitalization forecast input and converts the forecasted values for each location to a categorical "change" indicator. The criteria for each level ("large decrease", "decrease", "stable", "increase", "large increase") was defined by the CDC (see link in references). The algorithm evaluates absolute changes in counts and rates (per 100k individuals) for the most recently observed week and a 2 week ahead forecasted horizon. This procedure runs independently for each location, and results in a formatted tabular output that includes each possible level and its corresponding probability of being observed (calculated from probabilistic quantiles) for every location.
Usage
forecast_categorical(
.forecast,
.observed,
method = "density",
format = "hubverse",
horizon = 4
)Arguments
- .forecast
A tibble with "submission-ready" probabilistic flu hospitalization forecast data (i.e., tibble contained in list element returned from format_for_submission)
- .observed
A tibble with observed flu admission data (i.e., tibble output from prep_hdgov_hosp)
- method
The categorical forecasting method to use; must be one of
"density"or"interpolation"; default is"density"- format
The submission format to be used; must be one of
"hubverse"or"legacy"; default is"hubverse"- horizon
The number of horizons ahead to forecast; must be one of
4or5; default is4
Value
A tibble with formatted categorical forecasts.
If format is "hubverse" the tibble will have the following columns:
reference_date: Date of reference for forecast submission
horizon: Horizon for the given forecast
target: Name of forecasted target
target_end_date: Last date of the forecasted target (e.g., Saturday of the given epidemiological week)
location: Name or geographic identifier (e.g., FIPS code) for location for the given forecast
output_type: Type of forecasted value (e.g., "quantile")
output_type_id: The quantile for the forecasted value if output_type is "quantile"
value: The forecasted value
If format is "legacy" the tibble will have the following columns:
forecast_date: Date of forecast
target: Horizon and name of forecasted target
target_end_date: Last date of the forecasted target (e.g., Saturday of the given epidemiological week)
location: Name or geographic identifier (e.g., FIPS code) for location for the given forecast
type: One of either "point" or "quantile" for the forecasted value
quantile: The quantile for the forecasted value;
NAif "type" is"point"value: The forecasted value
Examples
if (FALSE) { # \dontrun{
# Retrieve hospitalization data
h_raw <- get_hdgov_hosp(limitcols=TRUE)
# Prepare and summarize hospitalization data to weekly resolution
prepped_hosp <- prep_hdgov_hosp(h_raw)
# Create a keyed time series tibble with only locations of interest
prepped_tsibble <- make_tsibble(prepped_hosp,
epiyear = epiyear,
epiweek=epiweek,
key=location)
# Run with default constrained ARIMA, nonseasonal ETS, no NNETAR
hosp_fitfor <- ts_fit_forecast(prepped_tsibble,
horizon=4L,
outcome="flu.admits",
covariates=TRUE)
# Prepare forecast for quantile submission format
forc <- format_for_submission(hosp_fitfor$tsfor, method = "ts", format = "legacy")
# Run categorical summary of quantiles for the time series ensemble
forecast_categorical(forc$ensemble, prepped_hosp, method = "interpolation", format = "legacy")
} # }