I've successfully built a pipeline for hierarchial time-series forecasting using the fable
package. For now I've used multicore
function from the future
pacakge (in Databricks). And, for ~1k time series the runtime isn't too bad. Below is the code I have for that.
plan(multicore, workers = 14)
model_multicore_log_exts <- data_for_modeling %>%
model(
prophetML_additive_lin = prophet(log(dependent_variable) ~ exogenous_var1 + exogenous_var2 + exogenous_var3 + growth("linear") + season("year", type = "additive")),
prophetML_multiplicative_lin = prophet(log(dependent_variable) ~ exogenous_var1 + exogenous_var2 + exogenous_var3 + growth("linear") + season("year", type = "multiplicative")),
nn = NNETAR(log(dependent_variable) ~ trend() + season() + exogenous_var1 + exogenous_var2 + exogenous_var3, n_networks = 50)
)
But, I have the need for increasing that number substantially.
So, I'm wondering if there is a way to distribute with sparkr. I cannot find anything in the fable
CRAN documentation nor anything with a google search, so I am asking here.
Any suggestions/advice appreciated.
Thank you, Brian