In JSzitas/soothsayer: Tidy Time Series Forecast Model Selection

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

soothsayer

soothsayer provides tools for automation of time series model selection.

This is achieved by fitting your own version of Feature-based Forecast Model Selection (FFORMS), as described in . The package does this in a way that tries to leverage the tidy framework.

Workflow

• compute time series features (e.g. Box-Pierce statistic, time series entropy, ...)
• run a set of rules, or an oracle model for fitting models
• generate forecast combinations

The forecast rules are implemented directly in the model, i.e. the following is perfectly valid and runnable code:

library(fabletools)
library(soothsayer)

ex_data <- tsibbledata::aus_livestock %>%
  as.data.frame() %>%
  dplyr::group_by(Month, State) %>%
  dplyr::summarise(count = sum(Count)) %>%
  dplyr::ungroup() %>%
  tsibble::as_tsibble(index = "Month", key= "State")

train <- dplyr::filter(ex_data, Month <= tsibble::yearmonth("2017 Jan"))
test <- dplyr::filter( ex_data, Month > tsibble::yearmonth("2017 Jan") )

mdls <- fabletools::model(
  train,
  ar = fable::ARIMA( count ),
  arima = fable::ARIMA(count),
  soothsayer = soothsayer(count ~ 
                            # only fit arima if there is over 12 observations
                            rules(arima ~ .length > 12,
                                  # always fit an AR model 
                                  ar ~ TRUE,
                                  # only fit ets if length is over 15
                                  ets ~ .length > 15))
)

Note that this works seamlessly with fitting any model in the fable package and models which work in the fabletools framework. This includes, among others fable.tbats, the wrapper for the TBATS model.

This then works just like any other model in fable, and generating forecasts is seamless:

forecast(mdls, test) %>% 
  # only a subset of forecasts plotted for visibility
  dplyr::filter(State == "Western Australia") %>% 
  ggplot2::autoplot(test)

Combinations

soothsayer also supports forecast combinations. Following the example above, by default, soothsayer creates an equally weighed combination of all models which are matched:

# this is identical to the above 
mdls <- fabletools::model(
  train,
  ar = fable::ARIMA( count ),
  arima = fable::ARIMA(count),
  soothsayer = soothsayer(count ~ 
                            rules(arima ~ .length > 12,
                                  ar ~ TRUE,
                                  ets ~ .length > 15) + 
                            # note the explicit combiner, though
                            combiner(combiner_equal))
)

However, multiple combiners are implemented within the package - a good example being the 'greedy stacking' one (See [1]):

# this is identical to the above 
mdls <- fabletools::model(
  train,
  ar = fable::ARIMA( count ),
  arima = fable::ARIMA(count),
  soothsayer = soothsayer(count ~ 
                            rules(arima ~ .length > 12,
                                  ar ~ TRUE,
                                  ets ~ .length > 15) + 
                            # note the explicit combiner, though
                            combiner(combiner_greedy_stacking))
)

These allow you to implement specific types of weighed combinations, including preset weights (combiner_custom). Further, note that the combiner argument takes a function - any function with the following signature:

new_combiner <- function(.models, ...) {
  # your code goes here
}

this is expected to return a vector of weights in the same order as .models, where .models is the fitted model mable* object.

Adding soothsayers to existing mables

TODO: Document

References

• [1] Kurz, Maier & Rink; 'A greedy stacking algorithm for model ensembling and domain weighting', BMC Research Notes volume 13, Article number: 70 (2020) (See full text at https://bmcresnotes.biomedcentral.com/articles/10.1186/s13104-020-4931-7)

JSzitas/soothsayer documentation built on April 18, 2023, 12:59 a.m.