calculate_intervals: Calculate bootstrap prediction intervals for forecasts
In nredell/forecastML: Time Series Forecasting with Machine Learning Methods

Description Usage Arguments Value Examples

The residuals from model training/fit are sampled i.i.d. for (a) each direct forecast horizon for a single time series and (b) each combination of direct forecast horizon and group for multiple time series.

calculate_intervals(
  forecasts,
  residuals,
  index = NULL,
  outcome = NULL,
  keys = NULL,
  levels = c(0.95),
  times = 100L,
  weights = NULL,
  keep_samples = FALSE
)

`forecasts`	A data.frame of forecasts.
`residuals`	A data.frame of residuals (e.g., `residuals(data_fit)`)
`index`	Optional for forecasts from `combine_forecasts()`. A string giving the name of the date column in `forecasts`.
`outcome`	Optional for forecasts from `combine_forecasts()`. A string giving the name of the forecast column in `forecasts`.
`keys`	Optional. For grouped time series, a character vector giving the column name(s) of the group columns. The key identifies unique time series of residuals for bootstrap sampling. For direct forecasting, a single time series will have one group per direct forecast horizon.
`levels`	A numeric vector with 1 or more forecast prediction intervals. A level of .95, for example, will return the 0.25 and .975 quantiles of the bootstrapped forecast distribution at each forecast horizon.
`times`	Integer. The number of bootstrap samples.
`weights`	Not implemented.
`keep_samples`	Boolean. If `TRUE`, a data.frame of `times` bootstrapped forecasts is returned in addition to the calculated forecast prediction intervals. The samples are in the list slot named 'samples'.

If forecasts is an object of class 'forecast_results', a forecast_results object with a new column for each lower- and upper-bound forecast in levels. If forecasts is a data.frame, the function return will be the same but without forecastML attributes. If, keep_samples is TRUE, a named list of length 2 is returned with 'forecasts' and 'samples'.

## Not run: 
  data("data_seatbelts", package = "forecastML")

  data_train <- create_lagged_df(data_seatbelts, type = "train", method = "direct",
                                 outcome_col = 1, lookback = 1:15,
                                 horizons = c(1, 6, 12))

  windows <- create_windows(data_train, window_length = 0)

  model_fn <- function(data) {
    model <- lm(DriversKilled ~ ., data)
  }

  model_results <- train_model(data_train, windows, model_name = "OLS",
                               model_function = model_fn)

  predict_fn <- function(model, data) {
    data_pred <- as.data.frame(predict(model, data))
  }

  data_fit <- predict(model_results, prediction_function = list(predict_fn), data = data_train)

  residuals <- residuals(data_fit)

  data_forecast <- create_lagged_df(data_seatbelts, type = "forecast",
                                    method = "direct", outcome_col = 1,
                                    lookback = 1:15, horizons = c(1, 6, 12))

  data_forecasts <- predict(model_results, prediction_function = list(predict_fn),
                            data = data_forecast)

  data_forecasts <- combine_forecasts(data_forecasts)

  data_forecasts <- calculate_intervals(data_forecasts, residuals, times = 30)

  plot(data_forecasts)

## End(Not run)