Forecasting functional basis coefficients (Internal)

For internal use: This function is primarily used internally by forecast.ffc_gam(). Most users should call forecast() directly on their ffc_gam object instead of using this function, which requires properly structured coefficient data and has specific format requirements.

Usage

# S3 method for class 'fts_ts'
forecast(
  object,
  model = "ARIMA",
  h = get_stan_param("h", "forecast"),
  n_samples = 25,
  stationary = FALSE,
  ...
)

Arguments

object

An object of class fts_ts containing time-varying basis function coefficients extracted from an ffc_gam object using fts_coefs()

model

A character string representing a valid univariate model definition from the fable package, ensemble methods, or one of the built-in Bayesian dynamic factor models. Note that if a fable model is used, the chosen method must have an associated generate() method in order to simulate forecast realisations. Valid models currently include: 'ENS', 'ARDF', 'GPDF', 'VARDF, 'ETS', 'ARIMA', 'AR', 'RW', 'NAIVE', and 'NNETAR'. The 'ENS' option combines ETS and Random Walk forecasts with equal weights, hedging bets between exponential smoothing and random walk assumptions to provide more robust predictions when model uncertainty is high.

h

A positive integer specifying the length of the forecast horizon

n_samples

A positive integer specifying the number of forecast realisation paths to simulate from the fitted forecast model

stationary

If TRUE, the fitted time series models are constrained to be stationary. Default is FALSE. This option only works when model == 'ARIMA'

...

Other arguments to pass to the Stan dynamic factor models (i.e. the (V)AR order lag = ... or the number of factors K = ...)

Value

A tsibble object containing the forecast prediction (.sim) for each replicate realisation (.rep) at each timestep in the forecast horizon h

Details

Basis function coefficient time series will be used as input to the specified model to train a forecasting model that will then be extrapolated h timesteps into the future. A total of times forecast realisations will be returned for each basis coefficient

See also

predict(), fts()

Author

Nicholas J Clark

Examples

# Extract coefficients and generate forecasts
mod <- ffc_gam(
  deaths ~ offset(log(population)) + sex +
    fts(age, k = 8, bs = "cr", time_k = 10),
  time = "year",
  data = qld_mortality,
  family = poisson(),
  engine = "bam"
)
coefs <- fts_coefs(mod, summary = FALSE, n_samples = 5)

# Generate ETS forecasts
forecast(coefs, model = "ETS", h = 3)
#> # A tsibble: 2,625 x 6 [1Y]
#> # Key:       .basis, .realisation, .model, .rep [875]
#>    .basis          .realisation .model  year .rep   .sim
#>    <chr>                  <int> <chr>  <dbl> <chr> <dbl>
#>  1 fts_bs_s_age__1            1 ETS     2021 1     -1.93
#>  2 fts_bs_s_age__1            1 ETS     2022 1     -1.90
#>  3 fts_bs_s_age__1            1 ETS     2023 1     -1.86
#>  4 fts_bs_s_age__1            1 ETS     2021 10    -1.93
#>  5 fts_bs_s_age__1            1 ETS     2022 10    -1.88
#>  6 fts_bs_s_age__1            1 ETS     2023 10    -1.83
#>  7 fts_bs_s_age__1            1 ETS     2021 11    -1.92
#>  8 fts_bs_s_age__1            1 ETS     2022 11    -1.88
#>  9 fts_bs_s_age__1            1 ETS     2023 11    -1.84
#> 10 fts_bs_s_age__1            1 ETS     2021 12    -1.93
#> # ℹ 2,615 more rows