cross_validation_detailedexample_2_UPDATE.qmd

```{r}
library(fpp3)
```

```{r}
google_stock <- 
  gafa_stock %>%
  filter(Symbol %in% c("GOOG", "AAPL"), year(Date) >= 2015) %>%
  group_by(Symbol) %>% 
  arrange(Symbol, Date) %>% 
  mutate(day = row_number()) %>%
  ungroup() %>% 
  update_tsibble(index = day, regular = TRUE)

stocks_2015 <- google_stock %>% filter(year(Date) == 2015)
```

```{r}
stocks_2015_tr <- 
  stocks_2015 %>% 
    stretch_tsibble(.init = 20, .step = 1) %>% 
    select(Symbol, .id, everything())
```

```{r}
# How many training datasets do we have for each stock asset?
stocks_2015_tr %>% 
  as_tibble() %>% 
  group_by(Symbol) %>% 
  summarize(
    n_tr_sets = max(.id)
  )

# How many observations does each training dataset have?
stocks_2015_tr %>% 
  as_tibble() %>% 
  group_by(Symbol, .id) %>% 
  summarize(
    nobs = n()
  )
```

```{r}
stocks_2015_fit_cv <- 
  stocks_2015_tr %>%
          model(
              Drift = RW(Close ~ drift()),
              Mean = MEAN(Close),
              SES = ETS(Close ~ error("A") + trend("N") + season("N"))
          )
```

```{r}
# Number of forecasts:

# Number of tseries = 466
# Nmodels = 3
# Nforecasts per model = 8
466 * 3 * 8
```

```{r}
stocks_2015_fc <- 
stocks_2015_fit_cv %>% 
  forecast(h=8)

stocks_2015_fc
```
```{r}
stocks_2015
```

```{r}
cv_accuracy <- 
stocks_2015_fc %>% 
  accuracy(stocks_2015) %>% 
  arrange(Symbol, RMSE)

cv_accuracy
```

```{r}
stocks_2015
```

```{r}
# Step 1. Compute forecasts - Done
stocks_2015_fc

# Step 2 - Bring the test data information to the table containing the forecasts
# LEFT JOIN
stocks_2015_close <- 
  stocks_2015 %>% 
  select(Symbol, day, Close) %>% 
  rename(Close_data = Close)

errors_manual <- 
  stocks_2015_fc %>% 
    left_join(stocks_2015_close, by=c("Symbol", "day"))

# Step 3 - Compute the error at each point in time, for each training dataset
errors_manual <- 
  errors_manual %>% 
    mutate(
      fc_error = Close_data - .mean,
      fc_error_pc = fc_error / Close_data
    )

# Step 4 - Average depending on the desired output
cv_errors_manual <- 
  errors_manual %>% 
  as_tibble() %>% # Cast it as a normal dataframe to have group_by work normally
  group_by(Symbol, .model) %>% 
  summarize(
    MAE = mean(abs(fc_error), na.rm=TRUE),
    RMSE = sqrt(mean(fc_error^2, na.rm=TRUE)),
    MAPE = mean(abs(fc_error_pc), na.rm=TRUE)
  ) %>% 
  arrange(Symbol, RMSE)

all.equal(cv_errors_manual$RMSE, cv_accuracy$RMSE)
```

```{r}
# Create the column h and render back to a fable
stocks_2015_fc_fable <- 
  stocks_2015_fc %>% 
    group_by(.id, Symbol, .model) %>% 
    mutate(
      h = row_number()
    ) %>% 
  select(.id, Symbol, .model, h, day, everything()) %>% 
  ungroup() %>% 
  as_fable(response = "Close", distribution = Close)
```

```{r}
cv_accuracy_h <- 
stocks_2015_fc_fable %>% 
  accuracy(stocks_2015, by=c("Symbol", ".model", "h")) %>% 
  arrange(Symbol, h, RMSE)

cv_accuracy_h %>% 
  filter(h==4)
```

```{r}
# Step 1. Compute forecasts and add the column h
stocks_2015_fc_fable

# Step 2 - Bring the test data information to the table containing the forecasts
# LEFT JOIN
stocks_2015_close <- 
  stocks_2015 %>% 
  select(Symbol, day, Close) %>% 
  rename(Close_data = Close)

errors_manual <- 
  stocks_2015_fc_fable %>% 
    left_join(stocks_2015_close, by=c("Symbol", "day"))

# Step 3 - Compute the error at each point in time, for each training dataset
errors_manual <- 
  errors_manual %>% 
    mutate(
      fc_error = Close_data - .mean,
      fc_error_pc = ((Close_data - .mean) / Close_data)*100
    )

# Step 4 - Average depending on the desired output
cv_errors_manual_h <- 
  errors_manual %>% 
  as_tibble() %>% # Cast it as a normal dataframe to have group_by work normally
  group_by(Symbol, .model, h) %>% 
  summarize(
    MAE = mean(abs(fc_error), na.rm=TRUE),
    RMSE = sqrt(mean(fc_error^2, na.rm=TRUE)),
    MAPE = mean(abs(fc_error_pc), na.rm=TRUE)
  ) %>% 
  arrange(Symbol, h, RMSE)

all.equal(cv_errors_manual_h$MAPE, cv_accuracy_h$MAPE)
```