Why null_model() perform better than any other regression methods in R tidymodels?

Question

I am testing several regression models using Tidyverse's parsnip. Initially the best performing one is rand_forest(), but after I add null_model(), it is the latter one that is best in terms of RMSE.

All are done after parameter tuning and cross-validated resampling.

Here is the result of null_model():

> show_best(null_grid_results, metric = "rmse")
# A tibble: 1 × 6
  .metric .estimator  mean     n std_err .config             
  <chr>   <chr>      <dbl> <int>   <dbl> <chr>               
1 rmse    standard   0.421    10  0.0701 Preprocessor1_Model1
> collect_metrics(null_grid_results) %>%
+   filter(.metric == "rmse") %>%
+   pull(mean) %>% mean()
[1] 0.4209793

And this is the random forest:

> show_best(random_forest_grid_results, metric = "rmse")
# A tibble: 5 × 8
   mtry min_n .metric .estimator  mean     n std_err .config              
  <int> <int> <chr>   <chr>      <dbl> <int>   <dbl> <chr>                
1  2971    28 rmse    standard   0.420    10  0.0700 Preprocessor1_Model15
2   945    21 rmse    standard   0.420    10  0.0703 Preprocessor1_Model16
3  1090    40 rmse    standard   0.420    10  0.0701 Preprocessor1_Model25
4  2074    32 rmse    standard   0.420    10  0.0702 Preprocessor1_Model13
5  1650    27 rmse    standard   0.420    10  0.0698 Preprocessor1_Model10
> collect_metrics(random_forest_grid_results) %>%
+   filter(.metric == "rmse") %>%
+   pull(mean) %>% mean()
[1] 0.4369285

The code snippet I used for performing null_model() is this:

library(tidyverse)
library(tidymodels)
library(rules)
library(baguette)
tidymodels_prefer()
library(doParallel)

# Skip showing steps for getting:
# prolif_feat_outcome_dat_train
# prolif_feat_outcome_dat_folds

null_model_spec <-    null_model() %>% 
  set_engine("parsnip") %>% 
  set_mode("regression") %>%  
  translate()

null_model_feature_preproc_rec <- recipe(prolif_outcome ~ ., data = prolif_feat_outcome_dat_train) %>% 
  step_zv(all_predictors()) 

null_model_wflow <- workflow() %>%
  add_model(null_model_spec) %>%
  add_recipe(null_model_feature_preproc_rec )

null_model_set <- extract_parameter_set_dials(null_model_wflow) 

grid_ctrl <- control_grid(
  verbose = TRUE,
  save_pred = TRUE,
  parallel_over = "everything",
  save_workflow = TRUE
)
nof_grid <- 25
ptm <- proc.time()
cls <- makePSOCKcluster(parallel::detectCores(logical = FALSE))
registerDoParallel(cls)

set.seed(999)
null_model_grid_results <- null_model_wflow %>%
  tune_grid(
    param_info = null_model_set,
    resamples = prolif_feat_outcome_dat_folds,
    grid = nof_grid,
    control = grid_ctrl
  )


stopCluster(cls)
proc.time() - ptm

show_best(null_model_grid_results, metric = "rmse")
collect_metrics(null_model_grid_results) %>%
  filter(.metric == "rmse") %>%
  pull(mean) %>% mean()

And this is by rand_forest():

random_forest_spec <-   rand_forest(mtry = tune(), min_n = tune(), trees = 1000) %>% 
  set_engine("ranger") %>% 
  set_mode("regression") %>% 
  translate()

random_forest_feature_preproc_rec <- recipe(prolif_outcome ~ ., data = prolif_feat_outcome_dat_train) %>% 
  step_zv(all_predictors()) 

random_forest_wflow <- workflow() %>%
  add_model(random_forest_spec) %>%
  add_recipe(random_forest_feature_preproc_rec )

random_forest_set <- extract_parameter_set_dials(random_forest_wflow) 

grid_ctrl <- control_grid(
  verbose = TRUE,
  save_pred = TRUE,
  parallel_over = "everything",
  save_workflow = TRUE
)
nof_grid <- 25
ptm <- proc.time()
cls <- makePSOCKcluster(parallel::detectCores(logical = FALSE))
registerDoParallel(cls)

set.seed(999)
random_forest_grid_results <- random_forest_wflow %>%
  tune_grid(
    param_info = random_forest_set,
    resamples = prolif_feat_outcome_dat_folds,
    grid = nof_grid,
    control = grid_ctrl
  )


stopCluster(cls)
proc.time() - ptm
saveRDS(random_forest_grid_results, file = paste0("/home/ubuntu/storage1/find_best_model_for_prolif_predictions_tidymodels/data/", wanted_dose, ".random_forest_grid_results.rds" ))

show_best(random_forest_grid_results, metric = "rmse")
collect_metrics(random_forest_grid_results) %>%
  filter(.metric == "rmse") %>%
  pull(mean) %>% mean()

I expected null_model() perform way worse than rand_forest()? My question is why null_model() perform best? Is my approach correct? If not what is the correct way to implement it?