fc_jags_mv: Fit a Bayesian forecast model to a multivariate set of...

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

Description

This function takes fitted values from a log-transformed forecast models and uses those as predictors in a Negative Binomial regression to model the discrete outcome time series. Multivariate covariance dependencies are incorporated via multivariate normal intercepts

Usage

fc_jags_mv(
  y,
  model = "arima",
  horizon,
  frequency,
  use_nb = TRUE,
  n.chains = 2,
  n.adapt = 1000,
  n.burnin = 10000,
  n.iter = 10000,
  thin = 10,
  auto_update = TRUE
)

Arguments

y	xts matrix. The discrete outcome series to be modelled. NAs are allowed
model	character either 'TBATS', 'arima' or 'stlf'
horizon	integer. The horizon to forecast. Defaults to frequency
frequency	integer. The seasonal frequency in y
use_nb	logical If TRUR, use a Negative Binomial likelihood with estimated overdispersion parameter r;
n.chains	integer the number of parallel chains for the model
n.adapt	integer the number of iterations for adaptation. See adapt for details. If n.adapt = 0 then no adaptation takes place
n.iter	integer the number of iterations of the Markov chain to run
auto_update	logical If TRUE, the model is run for up to 3 additional sets of n.iter iterations, or until the lower 15th percentile of effective sample sizes reaches 200

Details

The discrete series are first log-transformed and interpolated to fit univariate forecast models. Fitted values from the forecast models are then used as predictors in a Negative Binomial Bayesian regression to model the raw discrete series, while inter-series dependencies are captured using a multivariate normal intercept

Value

A list containing the posterior forecast from the forecast model and the original fitted jags.model

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