library(tidyverse)
library(tidymodels)
library(modeltime)
library(timetk)
library(tidyquant)

#Unemployment Rate (UNRATE)
df_unrate <- 
  tq_get("UNRATE", get = "economic.data") %>% 
  select(date, unrate = price)
  
#Federal Funds Effective Rate (FEDFUNDS)
df_fedfunds <- 
  tq_get("FEDFUNDS", get = "economic.data") %>% 
  select(date, fedfunds = price)

#Nasdaq 100
df_nasdaq <- 
  tq_get("^NDX") %>% 
  tq_transmute(select = close,
               mutate_fun = to.monthly,
               col_rename = "nasdaq") %>% 
  mutate(date = as.Date(date))


#Merging the datasets
df_merged <- 
  df_unrate %>% 
  left_join(df_fedfunds) %>% 
  left_join(df_nasdaq) %>% 
  drop_na() 


#Split Data 
splits <- 
  time_series_split(
  df_merged,
  assess     = "1 year",
  cumulative = TRUE
)

df_train <- training(splits)
df_test <- testing(splits)



#Recipe
recipe_ml <- 
  recipe(nasdaq ~ ., df_train) %>%
  step_date(date, features = "month", ordinal = FALSE) %>%
  step_dummy(all_nominal_predictors(), one_hot = TRUE) %>% 
  step_mutate(date_num = as.numeric(date)) %>%
  step_normalize(all_numeric_predictors()) %>%
  step_rm(date)

#Model spec
mod_spec <- 
  boost_tree(trees = tune(),
             min_n = tune(),
             tree_depth = tune(),
             learn_rate = tune()) %>%
  set_engine("xgboost") %>% 
  set_mode("regression")

#Hyperparameter Tuning
mod_param <- extract_parameter_set_dials(mod_spec)

set.seed(1234)
model_tbl <- 
  mod_param %>% 
  grid_random(size = 50) %>%
  create_model_grid(
    f_model_spec = boost_tree,
    engine_name  = "xgboost",
    mode         = "regression"
  )

#Extracting the model list
model_list <- model_tbl$.models

#Workflowsets
model_wfset <- 
  workflow_set(
  preproc = list(recipe_ml),
  models = model_list, 
  cross = TRUE
)


#Fitting Using Parallel Backend
model_parallel_tbl <- 
  model_wfset %>%
  modeltime_fit_workflowset(
    data    = df_train,
    control = control_fit_workflowset(
      verbose   = TRUE,
      allow_par = TRUE
    )
  )



#Accuracy 
model_parallel_tbl %>% 
  modeltime_calibrate(new_data = df_test) %>% 
  modeltime_accuracy() %>%
  table_modeltime_accuracy()



#Calibration to the test set for the best model
calibration_tbl <- 
  model_parallel_tbl %>%
  filter(.model_desc == "RECIPE_BOOST_TREE_27") %>% 
  modeltime_calibrate(df_test)
  


#Prediction Intervals
calibration_tbl %>% 
  modeltime_forecast(new_data = df_test, 
                     actual_data = df_merged %>%
                                   filter(date >= as.Date("2024-07-01"))) %>%
  plot_modeltime_forecast(.interactive = FALSE,
                          .legend_show = FALSE,
                          .line_size = 1.5,
                          .color_lab = "",
                          .title = "NASDAQ 100") +
  
  labs(subtitle = "<span style = 'color:dimgrey;'>Predictive Intervals</span><br><span style = 'color:red;'>ML Model</span>") + 
  scale_x_date(expand = expansion(mult = c(.1, .15)),
               labels = scales::label_date(format = "%b'%y")) +
  scale_y_continuous(labels = scales::label_currency()) +
  theme_minimal(base_family = "Roboto Slab", base_size = 20) +
  theme(legend.position = "none",
        plot.background = element_rect(fill = "azure", 
                                       color = "azure"),
        plot.title = element_text(face = "bold"),
        axis.text = element_text(face = "bold"),
        plot.subtitle = ggtext::element_markdown(face = "bold"))