Detect and remove null models

Fixes #114

Adds remove_null_models arg to forecast function (default TRUE)
Captures models removed in artefacts

Download/extract this preprocessed dataset saved from last week to demonstrate/test

pht.rds.zip

library(tidyverse)
library(fiphde)

prepped_tsibble <- readRDS("pht.rds")

# Fit a model
hosp_fitfor_all <- ts_fit_forecast(prepped_tsibble,
                                   horizon=4L,
                                   outcome="flu.admits",
                                   covariates=c("hosp_rank", "ili_rank"),
                                   stepwise=FALSE,
                                   approximation=FALSE)

hosp_fitfor_all

> hosp_fitfor_all
$tsfit
# A mable: 1 x 4
# Key:     location [1]
  location                       arima          ets      ensemble
  <chr>                        <model>      <model>       <model>
1 US       <LM w/ ARIMA(2,1,0) errors> <ETS(M,A,N)> <COMBINATION>

$tsfor
# A fable: 12 x 10 [1W]
# Key:     location, .model [3]
   location .model      yweek       flu.admits .mean epiweek ili_mean ili_rank hosp_mean hosp_rank
   <chr>    <chr>      <week>           <dist> <dbl>   <dbl>    <dbl>    <dbl>     <dbl>     <dbl>
 1 US       arima    2022 W12   N(3298, 24748) 3298.      12     2.32       37      1.76        17
 2 US       arima    2022 W13   N(3685, 77730) 3685.      13     2.15       35      1.49        16
 3 US       arima    2022 W14  N(3937, 181881) 3937.      14     1.90       33      1.21        14
 4 US       arima    2022 W15  N(4118, 336191) 4118.      15     1.68       30      0.89        12
 5 US       ets      2022 W12  N(3148, 253769) 3148.      12     2.32       37      1.76        17
 6 US       ets      2022 W13  N(3382, 658575) 3382.      13     2.15       35      1.49        16
 7 US       ets      2022 W14 N(3616, 1259088) 3616.      14     1.90       33      1.21        14
 8 US       ets      2022 W15 N(3850, 2107647) 3850.      15     1.68       30      0.89        12
 9 US       ensemble 2022 W12   N(3223, 97251) 3223.      12     2.32       37      1.76        17
10 US       ensemble 2022 W13  N(3533, 262937) 3533.      13     2.15       35      1.49        16
11 US       ensemble 2022 W14  N(3777, 527037) 3777.      14     1.90       33      1.21        14
12 US       ensemble 2022 W15   N(3984, 9e+05) 3984.      15     1.68       30      0.89        12

$formulas
$formulas$arima
flu.admits ~ PDQ(0, 0, 0) + pdq(1:2, 0:2, 0) + hosp_rank + ili_rank
<environment: 0x7f87281907b0>

$formulas$ets
flu.admits ~ season(method = "N")
<environment: 0x7f87281907b0>

$nullmodels
# A tibble: 1 × 2
  location model
  <chr>    <chr>
1 05       arima

In theory, multiple models for the same location could be null, which would return multiple rows in the $nullmodels element of the returned list. The function removes entire rows, i.e., the entire location, if any models are null.

signaturescience / fiphde

Detect and remove null models #115