states_weighted_forc: States weighted forecast

Description

states_weighted_forc takes two or more forecasts, a data frame, matrix, or array of matching variables, an optional vector of time data associated with the matching variables, a matching window size, a matching function, and an error function. For each forecast period, matching_vars are standardized and the current state of the world is set as the the past matching_window periods of the matching variables. The current state is compared to all past periods of the matching variables using the matching function. The current state is matched to the past state that minimizes the matching function. The forecast error function is then used to compute the accuracy of each forecast over the matched past state. Forecast weights are computed based on this forecast accuracy, and the current period forecast is subsequently computed based on the forecast weights. Produces a weighted average of multiple forecasts based on how each forecast performed during the past state that is most similar to the current state of the world.

Usage

states_weighted_forc(
  ...,
  matching_vars,
  time_vec = NULL,
  matching_window,
  matching = "euclidean",
  errors = "mse",
  return_weights = FALSE
)

Value

Forecast object that contains the state weighted forecast.

Arguments

...: Two or more forecasts of class Forecast.
matching_vars: data frame, array, or matrix of variables used to match the current state of the world to a past state.
time_vec: Vector of any class that is equal in length to the data in matching_vars.
matching_window: Integer representing the window size over which the current state of the world is matched to a past state. Forecasts are also weighted based on their accuracy over matching_window periods.
matching: Character, "euclidean", "mse", or "rmse". Selects the function used to match the current state of the world to a past state.
errors: Character, either "mse", "rmse", "mae", or "mape". Selects what forecast accuracy function is used to evaluate forecast errors.
return_weights: Boolean, selects whether the weights used to weight forecasts in each period are returned. If TRUE, a data frame of weights and matched periods is returned to the Global Environment.

Details

Forecasts are weighted in each period with the function below. The error function used is MSE or RMSE depending on user selection. This example shows MSE errors. $$weight = (1 / MSE(forecast)) / (1 / sum(MSE(forecasts)))$$

Examples

Run this code


date <- as.Date(c("2010-03-31", "2010-06-30", "2010-09-30", "2010-12-31",
                  "2011-03-31", "2011-06-30", "2011-09-30", "2011-12-31",
                  "2012-03-31", "2012-06-30"))

future <- as.Date(c("2011-03-31", "2011-06-30", "2011-09-30", "2011-12-31",
                    "2012-03-31", "2012-06-30", "2012-09-30", "2012-12-31",
                    "2013-03-31", "2013-06-30"))

y  <- c(1.09, 1.71, 1.09, 2.46, 1.78, 1.35, 2.89, 2.11, 2.97, 0.99)
x1 <- c(4.22, 3.86, 4.27, 5.60, 5.11, 4.31, 4.92, 5.80, 6.30, 4.17)
x2 <- c(10.03, 10.49, 10.85, 10.47, 9.09, 10.91, 8.68, 9.91, 7.87, 6.63)

data <- data.frame(date, y, x1, x2)
matching_vars <- data[, c("x1", "x2")]

y1_forecast <- Forecast(
  origin = date,
  future = future,
  forecast = c(1.33, 1.36, 1.38, 1.68, 1.60, 1.55, 1.32, 1.22, 1.08, 0.88),
  realized = c(1.78, 1.35, 2.89, 2.11, 2.97, 0.99, 1.31, 1.41, 1.02, 1.05),
  h_ahead = 4L
)

y2_forecast <- Forecast(
  origin = date,
  future = future,
  forecast = c(0.70, 0.88, 1.03, 1.05, 1.01, 0.82, 0.95, 1.09, 1.07, 1.06),
  realized = c(1.78, 1.35, 2.89, 2.11, 2.97, 0.99, 1.31, 1.41, 1.02, 1.05),
  h_ahead = 4L
)

states_weighted_forc(
  y1_forecast, y2_forecast,
  matching_vars = matching_vars,
  time_vec = data$date,
  matching_window = 2L,
  matching = "euclidean",
  errors = "mse",
  return_weights = FALSE
)

states_weighted_forc(
  y1_forecast, y2_forecast,
  matching_vars = matching_vars,
  time_vec = data$date,
  matching_window = 3L,
  matching = "rmse",
  errors = "rmse"
)