ensemble_base: Fit a series of univariate models and return a weighted...

In nicholasjclark/mvforecast: Tools to fit, interrogate and forecast multivariate statistical and machine learning timeseries models

Description

This function fits six simple univariate forecast models to the series and calculates weights that minimise the mean absolute scaled error of a weighted ensemble forecast

Usage

ensemble_base(
  y,
  frequency,
  lambda = NULL,
  k = 1,
  bottom_series = FALSE,
  discrete = FALSE
)

Arguments

y	A ts object containing the series to forecast
frequency	The seasonal frequency of y
lambda	numeric. The Box Cox power transformation parameter for all series. Must be between -1 and 2 inclusive. If y contains zeros, lambda will be set to max(c(0.7, lambda)) to ensure stability of forecasts
k	integer specifying the length of the forecast horizon in multiples of frequency
bottom_series	logical. If TRUE, the two auto.arima models will be ignored to ensure bottom level series forecasts are not overconfident
discrete	logical Is the series in y discrete? If TRUE, use a copula-based method relying on the Probability Integral Transform to map the series to an approximate Gaussian distribution prior to modelling. Forecasts are then back-transformed to the estimated discrete distribution that best fits y. Default is FALSE

Details

A total of seven simple univariate models are tested on the series. These include stlf with an AR model for the seasonally-adjusted series, ets, auto.arima, auto.arima with fourier terms K = 4 regressors, rwf with drift, naive and snaive. The mean absolute scaled error (MASE) of the in-sample series is calculated for each series, and a weighted mean is optimised to minimise in-sample MASE using the "L-BFGS-B" algorithm in optim.

Value

A list object containing the ensemble forecast and the ensemble residuals

nicholasjclark/mvforecast documentation built on Dec. 22, 2021, 2:11 a.m.