[robjhyndman]

Reported in https://github.com/orgs/OTexts/discussions/164.

Problem

Combining models with arithmetic gave forecasts on the wrong scale when one of the models was a transformed decomposition_model(), while mean() gave the right answer:

STLF <- decomposition_model(
  STL(log(Turnover) ~ season(window = Inf)),
  ETS(season_adjust ~ season("N"))
)
train |>
  model(ets = ETS(Turnover), stlf = STLF, arima = ARIMA(log(Turnover))) |>
  mutate(
    combination1 = (ets + stlf + arima) / 3,   # ~860, far too low
    combination2 = mean(ets, stlf, arima)      # ~1299, correct
  )

Cause

.cj_parts() unpacks any fit that inherits from model_combination into its components. A decomposition_model fit also inherits from model_combination, so + split STLF into ETS on season_adjust plus SNAIVE on season_year. Those components add up to log(Turnover), and the back-transformation is stored on the outer STLF model, which was dropped. combination1 ended up with four weights of 1/3 and averaged forecasts in dollars with forecasts in log-dollars. Without a transformation the split is harmless, which is why this went unnoticed.

Fix

.cj_parts() no longer unpacks decomposition models; each one is kept as a single component. Added a regression test and a NEWS entry.

Testing

• The example above now gives identical forecasts for combination1 and combination2.
• The new test in test-combination.R fails without the fix and passes with it.
• The combination and decomposition tests pass.

🤖 Generated with Claude Code