TSENS fails with sharp increases

[stephenturner]

If there is a sharp increase in the number of events (e.g., hospitalizations), the TSENS model will fail. A few recent examples showing locations for influenza hospitalizations where TSENS failed (the green band is the CREG model used in the submission.R script at the project root, ignore these):

[stephenturner]

Editing the submission script to remove the ARIMA model restrictions:

hosp_fitfor <- ts_fit_forecast(prepped_hosp_tsibble,
                               horizon=4L,
                               outcome="flu.admits",
                               covariates=c("hosp_rank", "ili_rank"),
                               stepwise=FALSE, approximation=FALSE,
                               models=list(arima='PDQ() + pdq()',
                                           ets='season(method="N")',
                                           nnetar=NULL))

Results in no null model / failures to converge. Note the extremely and implausibly sharp increase in forecasted hospitalizations.

The above plots show the unrestricted ARIMA model. Below show the same locations (CA and US) with the ETS model ensembled.

We might consider removing the ARIMA restrictions (with retrospective evaluation). Modifying the ts_fit_forecast() function to conditionally remove restrictions for a specific location in the event of a model failure at one or more locations will be difficult. Conditionally removing restrictions and re-running in the event of a null model / failure is more tractable, but should necessitate some retrospective evaluation of performance against the restricted model.

[stephenturner]

Closing for now. Reopening if this persists, possible to consider relaxing constraints on ARIMA after retrospective evaluation.

[stephenturner]

Code in dd6728e to remove restrictions.