Evaluate the sensitivity of forecasts to varying exogenous values

This function evaluates the sensitivity of forecasts to varying exogenous values by modifying the exogenous values in the forecast horizon and comparing the resulting forecasts to the original forecasts. The function allows for both percentage-based and unit-based modifications of the exogenous values, and can also calculate impulse response functions if desired.

Usage

forecast_sensitivity(
  model,
  size = 0.5,
  size_type = "pct",
  grepl_variables = NULL,
  quiet = FALSE,
  impulse_response = TRUE,
  exog_fill_method = "AR",
  exclude_zero_change = TRUE,
  include_uncertainty = TRUE
)

Arguments

model: A model object of class 'osem'.
size: Numeric. If 'size_type' = 'pct' then 'size' must be larger than -1 as it represents the percentage value that the exogenous values will be multiplied with. Default is 0.5 (i.e. 50%). If 'size_type' = 'unit', then 'size' represents the unit value that will be added to the exogenous values (e.g. to run a one-unit IRF, use 'size = 1' and size_type = 'unit').
size_type: Character. Either "pct" or "unit". If "pct", the exogenous values will be multiplied with (1 + size). If "unit", the exogenous values will be increased by size. Default is "pct".
grepl_variables: Regular Expression Character. Can be used to select variables to be included in the sensitivity procedure.
quiet: Logical. Should messages about the forecast procedure be suppressed?
impulse_response: Logical. Should an impulse response function be calculated. Main reason to deactivate this might be improving the speed of this function.
exog_fill_method: Character, either 'AR', 'auto', or 'last'. When no exogenous values have been provided, these must be inferred. When option 'exog_fill_method = "AR"' then an autoregressive model is used to further forecast the exogenous values. With 'last', simply the last available value is used. 'auto' is an auto.arima model.
exclude_zero_change: Logical. Should cases where the change in forecasts is exactly zero be excluded from the output (i.e. set to NA). This can be useful if there are many variables and many of them have no effect on the forecasts, which can make the plot easier to read. This can happen if e.g. the modified variable is not part of the specification of the module or has been removed due to model selection.
include_uncertainty: Logical. Should the uncertainty of the forecasts be included in the plot. If TRUE, the function will calculate the difference in all forecast runs (not just the central estimate) between the modified and original forecasts and show the 5-95% and 25-75% quantiles of these differences in the plot as shaded areas.

Value

A list that contains a tibble with the original forecasts and the difference to the modified forecasts as well as plot(s).