This function allows the user to fit time series models and forecast values out to a specified horizon. Starting from a tsibble object (see make_tsibble), the function fits the models specified as a list in the "models" argument. The "Details" section provides more information on how to parameterize the models used. Note that if the input tsibble is "keyed" (e.g., grouped by location) then the procedure will fit and forecast independently for each grouping.
Usage
ts_fit_forecast(
prepped_tsibble,
outcome = "flu.admits",
horizon = 4L,
trim_date = "2021-01-01",
models = list(arima = "PDQ(0, 0, 0) + pdq(1:2, 0:2, 0)", ets =
"season(method=\"N\")", nnetar = NULL),
covariates = TRUE,
ensemble = TRUE
)Arguments
- prepped_tsibble
A
tsibblewith data formatted via make_tsibble- outcome
The outcome variable to model; default is
"flu.admits"- horizon
Number of weeks ahead to forecast
- trim_date
The date (YYYY-MM-DD) at which time series models should start fitting; default
"2021-01-01"; if set toNULLthe input data will not be trimmed (i.e., all data will be used to fit time series models)- models
A list of right hand side formula contents for models you want to run; default is
list(arima='PDQ(0, 0, 0) + pdq(1:2, 0:2, 0)', ets='season(method="N")', nnetar=NULL)which runs a constrained ARIMA, non-seasonal ETS, and ignores the NNETAR model; see "Details" for more information- covariates
Logical. Should flu hospitalization-specific covariates that should be modeled with the time series? If so, historical hospitalization and ILI rank for each epidemiological week, brought in with prep_hdgov_hosp, is added to the ARIMA model.
- ensemble
Logical as to whether or not the models should be ensembled (using mean); default
TRUE
Value
A list of the time series fit, time series forecast, and model formulas.
tsfit: A
mdl_dfclass "mable" with one row for each location, columns for arima and ets models.tsfor: A
fbl_tsclass "fable" with one row per location-model-timepoint up tohorizonnumber of time points.formulas: A list of ARIMA, ETS, and/or NNETAR formulas
Details
When fitting time series models, the set of models used (and their parameters) can be defined via a named list passed to the "models" argument. The list should contain elements that define the right-hand side of model formulas. The function internally uses the fable::fable package, and any models provided must be part of the fable ecosystem of time series models. The models passed must be named as "arima", "ets", and "nnetar". To skip any one of these models set the named argument for the given model to NULL. The "models" argument defaults to list(arima = "PDQ(0, 0, 0) + pdq(1:2, 0:2, 0)", ets = "season(method='N')", nnetar = NULL). To run an unconstrained ARIMA: list(arima='PDQ() + pdq()') (see fable::ARIMA). To run a seasonal exponential smoothing: list(ets='season(method=c("A", "M", "N"), period="3 months")') (see fable::ETS). To run an autoregressive neural net with P=1: list(nnetar="AR(P=1)") (see fable::NNETAR).
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) %>%
dplyr::filter(location %in% c("US", "51"))
# Run with default constrained ARIMA, nonseasonal ETS, no NNETAR
hospfor1 <- ts_fit_forecast(prepped_tsibble,
horizon=4L,
outcome="flu.admits",
covariates=TRUE)
# Run an unconstrained ARIMA, seasonal ETS, no NNETAR
hospfor2 <- ts_fit_forecast(prepped_tsibble,
horizon=4L,
outcome="flu.admits",
covariates=TRUE,
models=list(arima='PDQ() + pdq()',
ets='season(method=c("A", "M", "N"), period="3 months")',
nnetar=NULL))
# Run an unconstrained ARIMA, seasonal ETS, NNETAR
hospfor3 <- ts_fit_forecast(prepped_tsibble,
horizon=4L,
outcome="flu.admits",
covariates=TRUE,
models=list(arima='PDQ() + pdq()',
ets='season(method=c("A", "M", "N"), period="3 months")',
nnetar="AR(P=1)"))
} # }