library(fpp3)

## Apple stock price

apple <- gafa_stock |> filter(Symbol == "AAPL")

apple |>
  autoplot(Close)

# Use trading day as index
apple <- apple |>
  mutate(trading_day = row_number()) |>
  update_tsibble(index = trading_day, regular = TRUE)

fit <- apple |>
  model(
    #Seasonal_naive = SNAIVE(Close),
    Naive = NAIVE(Close),
    Drift = RW(Close ~ drift()),
    Mean = MEAN(Close)
  )

fc <- fit |>
  forecast(h = 20)

fc |> autoplot(apple |> tail(50), level = NULL)


# Use date as index
apple <- apple |>
  update_tsibble(index = Date, regular = TRUE) |>
  fill_gaps()

fit <- apple |>
  model(
    Seasonal_naive = SNAIVE(Close),
    Naive = NAIVE(Close),
    Drift = RW(Close ~ drift()),
    Mean = MEAN(Close)
  )

fc <- fit |>
  forecast(h = 20)

fc |> autoplot(apple |> tail(50), level = NULL)

## ---- Holiday tourism by state --------------

holidays <- tourism |>
  as_tibble() |>
  filter(Purpose == "Holiday") |>
  summarise(Trips = sum(Trips), .by = c("State", "Quarter")) |>
  as_tsibble(index = Quarter, key = State)

holidays |> autoplot(Trips)

## Fit models
fit <- holidays |>
  model(
    Seasonal_naive = SNAIVE(Trips),
    Naive = NAIVE(Trips),
    Drift = RW(Trips ~ drift()),
    Mean = MEAN(Trips),
    tslm = TSLM(Trips ~ trend() + season())
  )

fit |> select(tslm) |> filter(State == "Victoria") |> report()

## Produce forecasts

holidays_fc <- fit |>
  forecast(h = "4 years")

holidays_fc |>
  autoplot(holidays, level = NULL)

holidays_fc |>
  filter(.model == "Seasonal_naive", State == "Victoria") |>
  autoplot(holidays, show_gap = FALSE)

holidays_fc |>
  hilo(level = 95) |>
  mutate(
    lower = `95%`$lower,
    upper = `95%`$upper
  )

# Piecewise linear trend
fit <- holidays |>
  filter(State == "Victoria") |>
  model(
    with_bend = TSLM(Trips ~ season() + trend(knots = yearquarter("2013 Q1"))),
    without_bend = TSLM(Trips ~ season() + trend())
  )
fc <- fit |>
  forecast(h = "4 years")
fc |>
  autoplot(holidays |> filter(State == "Victoria")) +
  geom_line(data = as_tibble(fitted(fit)), aes(y = .fitted, colour = .model))

# Transformations
airp <- as_tsibble(AirPassengers)
fit <- airp |>
  model(
    with_log = SNAIVE(log(value) ~ drift()),
    with_no_log = SNAIVE(value ~ drift())
  )
forecast(fit, h = "3 years") |>
  autoplot(airp)

# Decomposition model

stl_model <- decomposition_model(
  STL(Trips ~ season(window = "periodic") + trend(window = 15)),
  RW(season_adjust ~ drift()),
  SNAIVE(season_year)
)
holidays |>
  model(stl_model) |>
  forecast(h = "4 years") |>
  filter(State == "Victoria") |>
  autoplot(holidays, level = NULL)