blgt.multi.forecast function

Rlgt LSGT Gibbs run in parallel

Fit a list of series and produce forecast, then calculate the accuracy.

blgt.multi.forecast(
  train,
  future,
  n.samples = 20000,
  burnin = 10000,
  parallel = T,
  m = 1,
  homoscedastic = F
)

Arguments

• train: A list of training series.
• future: A list of corresponding future values of the series.
• n.samples: The number of samples to sample from the posterior (the default is 2e4).
• burnin: The number of burn in samples (the default is 1e4).
• parallel: Whether run in parallel or not (Boolean value only, default TRUE).
• m: The seasonality period, with default m=1, i.e., no seasonality specified.
• homoscedastic: Run with homoscedastic or heteroscedastic version of the Gibbs sampler version. By default it is set to FALSE, i.e., run a heteroscedastic model.

Returns

returns a forecast object compatible with the forecast package in R

Examples

## Not run: demo(exampleScript)
## Not run:

## Build data and test
library(Mcomp)
M3.data <- subset(M3,"yearly")

train.data = list()
future.data = list()
for (i in 1:645)
{
  train.data[[i]] = as.numeric(M3.data[[i]]$x)
  future.data[[i]] = as.numeric(M3.data[[i]]$xx)  
}

## Test -- change below to test more series
w.series = 1:20
# w.series = 1:645        # uncomment to test all series

# use 10,000 posterior samples; change n.samples to 20,000 to test that as well if you want
s = system.time({
  rv=blgt.multi.forecast(train.data[w.series], future.data[w.series], n.samples=1e4)
})

s                         # overall timing info
s[[3]] / length(w.series) # per series time

mean(rv$sMAPE)            # performance in terms of mean sMAPE
mean(rv$InCI)/6           # coverage of prediction intervals -- should be close to 95%
## End(Not run)