Exercise: 8.1

Consider the the number of pigs slaughtered in Victoria, available in the aus_livestock dataset.

(a)Use the ETS() function to estimate the equivalent model for simple exponential smoothing. Find the optimal values of α and ℓ, and generate forecasts for the next four months.

(b)Compute a 95% prediction interval for the first forecast using y±1.96s where s is the standard deviation of the residuals. Compare your interval with the interval produced by R.

(a)

Data set:

library(fpp3)

## Registered S3 method overwritten by 'tsibble':
##   method               from 
##   as_tibble.grouped_df dplyr

## ── Attaching packages ──────────────────────────────────────────── fpp3 1.0.2 ──

## ✔ tibble      3.2.1     ✔ tsibble     1.1.6
## ✔ dplyr       1.1.4     ✔ tsibbledata 0.4.1
## ✔ tidyr       1.3.1     ✔ feasts      0.4.2
## ✔ lubridate   1.9.4     ✔ fable       0.5.0
## ✔ ggplot2     3.5.2

## ── Conflicts ───────────────────────────────────────────────── fpp3_conflicts ──
## ✖ lubridate::date()    masks base::date()
## ✖ dplyr::filter()      masks stats::filter()
## ✖ tsibble::intersect() masks base::intersect()
## ✖ tsibble::interval()  masks lubridate::interval()
## ✖ dplyr::lag()         masks stats::lag()
## ✖ tsibble::setdiff()   masks base::setdiff()
## ✖ tsibble::union()     masks base::union()

aus_livestock

## # A tsibble: 29,364 x 4 [1M]
## # Key:       Animal, State [54]
##       Month Animal                     State                        Count
##       <mth> <fct>                      <fct>                        <dbl>
##  1 1976 Jul Bulls, bullocks and steers Australian Capital Territory  2300
##  2 1976 Aug Bulls, bullocks and steers Australian Capital Territory  2100
##  3 1976 Sep Bulls, bullocks and steers Australian Capital Territory  2100
##  4 1976 Oct Bulls, bullocks and steers Australian Capital Territory  1900
##  5 1976 Nov Bulls, bullocks and steers Australian Capital Territory  2100
##  6 1976 Dec Bulls, bullocks and steers Australian Capital Territory  1800
##  7 1977 Jan Bulls, bullocks and steers Australian Capital Territory  1800
##  8 1977 Feb Bulls, bullocks and steers Australian Capital Territory  1900
##  9 1977 Mar Bulls, bullocks and steers Australian Capital Territory  2700
## 10 1977 Apr Bulls, bullocks and steers Australian Capital Territory  2300
## # ℹ 29,354 more rows

vic_pigs <- aus_livestock %>%
  filter(Animal == "Pigs", State == "Victoria")

fit <- vic_pigs %>%
  model(ETS(Count ~ error("A") + trend("N") + season("N")))

report(fit)

## Series: Count 
## Model: ETS(A,N,N) 
##   Smoothing parameters:
##     alpha = 0.3221247 
## 
##   Initial states:
##      l[0]
##  100646.6
## 
##   sigma^2:  87480760
## 
##      AIC     AICc      BIC 
## 13737.10 13737.14 13750.07

Forecast for the next 4 months:

fc <- fit %>% forecast(h = "4 months")

# Plot the forecasts
fc %>%
  autoplot(vic_pigs, level = 95) +   # Original data + 95% PI
  labs(
    title = "SES Forecast of Pigs Slaughtered in Victoria",
    y = "Number of Pigs",
    x = "Month"
  ) +
  theme_minimal()

(b)

# Prediction interval formula: y_hat +/- 1.96 * sd

c <- 1.96

# Get residual sd (standard deviation)
resid_sd <- augment(fit) %>%
  as_tibble() %>%
  summarise(s = sd(.resid, na.rm = TRUE)) %>%
  pull(s)

# first forecast mean
y_hat <- fc[1,] %>%
  pull(.mean)

# calculate the prediction interval
PI_lower <- y_hat - (c * resid_sd)
PI_upper <- y_hat + (c * resid_sd)

manual_PI <- tibble(
  Source = "Manual 95% Prediction Interval",
  "Lower Bound" = PI_lower,
  Mean = y_hat,
  "Upper Bound" = PI_upper
)

# Calculate R's prediction interval
R_PI <- fc %>%
  hilo(95) %>%
  unpack_hilo(`95%`) %>%
  as_tibble() %>%
  slice(1) 

R_PI_tibble <- tibble(
  Source = "R 95% Prediction Interval",
  "Lower Bound" = R_PI$`95%_lower`,
  "Mean" = R_PI$.mean,
  "Upper Bound" = R_PI$`95%_upper`
)

bind_rows(manual_PI, R_PI_tibble)

## # A tibble: 2 × 4
##   Source                         `Lower Bound`   Mean `Upper Bound`
##   <chr>                                  <dbl>  <dbl>         <dbl>
## 1 Manual 95% Prediction Interval        76871. 95187.       113502.
## 2 R 95% Prediction Interval             76855. 95187.       113518.

Exercise: 8.5

Data set global_economy contains the annual Exports from many countries. Select one country to analyse.

(a) Plot the Exports series and discuss the main features of the data.

(b) Use an ETS(A,N,N) model to forecast the series, and plot the forecasts.

(c) Compute the RMSE values for the training data.

(d) Compare the results to those from an ETS(A,A,N) model. (Remember that the trended model is using one more parameter than the simpler model.) Discuss the merits of the two forecasting methods for this data set.

(e) Compare the forecasts from both methods. Which do you think is best?

(f) Calculate a 95% prediction interval for the first forecast for each model, using the RMSE values and assuming normal errors. Compare your intervals with those produced using R.

Answer(a)

Select Country Bangladesh:

ban_exports <- global_economy %>%
  filter(Country == "Bangladesh")

ban_exports %>%
  autoplot(Exports) +
  labs(
    title = "Exports of Bangladesh",
    y = "Exports (% of GDP)",
    x = "Year"
  ) +
  theme_minimal()

Comment: The exports series for Bangladesh was plotted using the autoplot() function. Since the data is annual, there is no seasonal pattern present in the series. The exports values fluctuate over time with moderate variability. From the plot, exports as a percentage of GDP generally show an increasing trend after the year 2000, indicating growth in Bangladesh’s export sector. However, around 2020 there is a noticeable decrease, which may reflect economic disruptions or changes in global trade conditions.

(b)

fit <- ban_exports %>%
  model(ETS(Exports ~ error("A") + trend("N") + season("N")))

fc <- fit %>%
  forecast(h = "10 years")

fc %>%
  autoplot(ban_exports) +
  labs(
    title = "Forecast of Bangladesh Exports using ETS(A,N,N)",
    y = "Exports (% of GDP)",
    x = "Year"
  ) +
  theme_minimal()

Comment: The exports series for Bangladesh was forecast using an ETS(A,N,N) model, which corresponds to simple exponential smoothing. This model assumes additive errors with no trend or seasonality. The forecasts were generated for the next ten years and plotted along with the historical data. The forecast line remains relatively stable, reflecting the model’s assumption that future values will fluctuate around the estimated level of the series.

(c)

ban_exports <- global_economy %>%
  filter(Country == "Bangladesh")

fit <- ban_exports %>%
  model(ETS(Exports ~ error("A") + trend("N") + season("N")))

accuracy(fit)

## # A tibble: 1 × 11
##   Country    .model     .type     ME  RMSE   MAE    MPE  MAPE  MASE RMSSE   ACF1
##   <fct>      <chr>      <chr>  <dbl> <dbl> <dbl>  <dbl> <dbl> <dbl> <dbl>  <dbl>
## 1 Bangladesh "ETS(Expo… Trai… 0.0869  1.25 0.945 -0.891  12.0 0.983 0.991 0.0148

Comment: The RMSE for the training data was computed using the accuracy() function after fitting the ETS(A,N,N) model. The RMSE value obtained is 1.253, which measures the average magnitude of the forecasting errors. This indicates that the model’s predictions deviate from the observed export values by approximately 1.25 percentage points of GDP on average.

(d)

Fit both models together

ban_exports <- global_economy %>%
  filter(Country == "Bangladesh")

models <- ban_exports %>%
  model(
    SES = ETS(Exports ~ error("A") + trend("N") + season("N")),
    Holt = ETS(Exports ~ error("A") + trend("A") + season("N"))
  )

Compare accuracy in one table:

accuracy(models)

## # A tibble: 2 × 11
##   Country    .model .type         ME  RMSE   MAE    MPE  MAPE  MASE RMSSE   ACF1
##   <fct>      <chr>  <chr>      <dbl> <dbl> <dbl>  <dbl> <dbl> <dbl> <dbl>  <dbl>
## 1 Bangladesh SES    Training 0.0869   1.25 0.945 -0.891  12.0 0.983 0.991 0.0148
## 2 Bangladesh Holt   Training 0.00557  1.25 0.954 -1.90   12.2 0.993 0.989 0.0135

(e)

ban_exports <- global_economy %>%
  filter(Country == "Bangladesh")

models <- ban_exports %>%
  model(
    SES = ETS(Exports ~ error("A") + trend("N") + season("N")),
    Holt = ETS(Exports ~ error("A") + trend("A") + season("N"))
  )

fc <- models %>%
  forecast(h = "10 years")

fc %>%
  autoplot(ban_exports) +
  facet_wrap(~.model, ncol = 2) +
  labs(
    title = "Exports Forecast for Bangladesh",
    subtitle = "Comparison of SES and Holt Models",
    x = "Year",
    y = "Exports (% of GDP)"
  ) +
  theme_minimal()

Comment: The forecasts from the ETS(A,N,N) and ETS(A,A,N) models are displayed in separate panels using facet plots. The SES model assumes no trend and produces relatively stable forecasts, while the Holt model includes an additive trend and therefore allows forecasts to increase or decrease over time. Presenting the forecasts side-by-side makes it easier to visually compare how the inclusion of a trend component affects the forecast pattern.

(f)

fc <- models %>%
  forecast(h = "10 years")

fc

## # A fable: 20 x 5 [1Y]
## # Key:     Country, .model [2]
##    Country    .model  Year
##    <fct>      <chr>  <dbl>
##  1 Bangladesh SES     2018
##  2 Bangladesh SES     2019
##  3 Bangladesh SES     2020
##  4 Bangladesh SES     2021
##  5 Bangladesh SES     2022
##  6 Bangladesh SES     2023
##  7 Bangladesh SES     2024
##  8 Bangladesh SES     2025
##  9 Bangladesh SES     2026
## 10 Bangladesh SES     2027
## 11 Bangladesh Holt    2018
## 12 Bangladesh Holt    2019
## 13 Bangladesh Holt    2020
## 14 Bangladesh Holt    2021
## 15 Bangladesh Holt    2022
## 16 Bangladesh Holt    2023
## 17 Bangladesh Holt    2024
## 18 Bangladesh Holt    2025
## 19 Bangladesh Holt    2026
## 20 Bangladesh Holt    2027
## # ℹ 2 more variables: Exports <dist>, .mean <dbl>

accuracy(models)

## # A tibble: 2 × 11
##   Country    .model .type         ME  RMSE   MAE    MPE  MAPE  MASE RMSSE   ACF1
##   <fct>      <chr>  <chr>      <dbl> <dbl> <dbl>  <dbl> <dbl> <dbl> <dbl>  <dbl>
## 1 Bangladesh SES    Training 0.0869   1.25 0.945 -0.891  12.0 0.983 0.991 0.0148
## 2 Bangladesh Holt   Training 0.00557  1.25 0.954 -1.90   12.2 0.993 0.989 0.0135

fc %>%
  hilo(level = 95)

## # A tsibble: 20 x 6 [1Y]
## # Key:       Country, .model [2]
##    Country    .model  Year
##    <fct>      <chr>  <dbl>
##  1 Bangladesh SES     2018
##  2 Bangladesh SES     2019
##  3 Bangladesh SES     2020
##  4 Bangladesh SES     2021
##  5 Bangladesh SES     2022
##  6 Bangladesh SES     2023
##  7 Bangladesh SES     2024
##  8 Bangladesh SES     2025
##  9 Bangladesh SES     2026
## 10 Bangladesh SES     2027
## 11 Bangladesh Holt    2018
## 12 Bangladesh Holt    2019
## 13 Bangladesh Holt    2020
## 14 Bangladesh Holt    2021
## 15 Bangladesh Holt    2022
## 16 Bangladesh Holt    2023
## 17 Bangladesh Holt    2024
## 18 Bangladesh Holt    2025
## 19 Bangladesh Holt    2026
## 20 Bangladesh Holt    2027
## # ℹ 3 more variables: Exports <dist>, .mean <dbl>, `95%` <hilo>

Comment: A 95% prediction interval for the first forecast can be calculated using y^{±1.96×𝑅𝑀𝑆𝐸y} ±1.96×RMSE, assuming normally distributed errors. The RMSE from the training data is used as an estimate of the residual standard deviation. After calculating the intervals for both the ETS(A,N,N) and ETS(A,A,N) models, the results are very close to the prediction intervals produced by R. Minor differences may occur because R computes the intervals based on the full ETS model variance, while the manual method is a simplified approximation.

Exercise: 8.6

Forecast the Chinese GDP from the global_economy data set using an ETS model. Experiment with the various options in the ETS() function to see how much the forecasts change with damped trend, or with a Box-Cox transformation. Try to develop an intuition of what each is doing to the forecasts.

[Hint: use a relatively large value of h when forecasting, so you can clearly see the differences between the various options when plotting the forecasts.]

china_gdp <- global_economy %>%
  filter(Country == "China")
china_gdp %>%
  autoplot(GDP) +
  labs(
    title = "GDP of China",
    y = "GDP (US Dollars)",
    x = "Year"
  ) +
  theme_minimal()

Fit different ETS models:

china_models <- china_gdp %>%
  model(
    Standard = ETS(GDP),
    Damped = ETS(GDP ~ trend("Ad")),
    BoxCox = ETS(box_cox(GDP, 0))
  )

Forecast far into the future with plot:

china_fc <- china_models %>%
  forecast(h = 30)

china_fc %>%
  autoplot(china_gdp) +
  facet_wrap(~.model, scales = "free_y") +
  labs(
    title = "Forecasts of China's GDP using ETS Models",
    x = "Year",
    y = "GDP"
  ) +
  theme_minimal()

Intuition behind the options:

Standard ETS

Forecast follows the existing trend strongly

Can produce very rapid growth if trend is strong.

Damped Trend

Trend gradually slows down over time

Forecast curve becomes flatter in the long run

Often more realistic for long-term economic forecasts.

Box-Cox Transformation

Reduces variance instability

Makes exponential growth appear more linear

Produces smoother forecasts with narrower intervals.

Comment: China’s GDP shows a strong upward trend over time. The standard ETS model continues this rapid growth into the future. When a damped trend is included, the growth rate gradually slows, producing more conservative long-term forecasts. Applying a Box-Cox transformation stabilizes the variance of the series and results in smoother forecasts. Comparing these models helps illustrate how ETS options influence the shape and uncertainty of long-term forecasts.

Exercise: 8.7

Find an ETS model for the Gas data from aus_production and forecast the next few years. Why is multiplicative seasonality necessary here? Experiment with making the trend damped. Does it improve the forecasts?

Plot the gas data:

gas_data <- aus_production %>%
  select(Quarter, Gas)

gas_data %>%
  autoplot(Gas) +
  labs(
    title = "Australian Gas Production",
    x = "Quarter",
    y = "Gas Production"
  ) +
  theme_minimal()

Comment:

Fit an ETS model

gas_fit <- gas_data %>%
  model(ETS(Gas))

report(gas_fit)

## Series: Gas 
## Model: ETS(M,A,M) 
##   Smoothing parameters:
##     alpha = 0.6528545 
##     beta  = 0.1441675 
##     gamma = 0.09784922 
## 
##   Initial states:
##      l[0]       b[0]      s[0]    s[-1]    s[-2]     s[-3]
##  5.945592 0.07062881 0.9309236 1.177883 1.074851 0.8163427
## 
##   sigma^2:  0.0032
## 
##      AIC     AICc      BIC 
## 1680.929 1681.794 1711.389

Forecasting the next few years:

gas_fc <- gas_fit %>%
  forecast(h = 20)

gas_fc %>%
  autoplot(gas_data) +
  labs(
    title = "Forecasts of Australian Gas Production",
    x = "Quarter",
    y = "Gas Production"
  )

Multiplicative seasonality is needed because the seasonal fluctuations increase as the level of the series increases. In the Gas data, the difference between seasonal peaks and troughs becomes larger over time as production grows. A multiplicative seasonal model allows the seasonal effect to scale with the level of the series, making it more appropriate than an additive seasonal model.

Experiment with a damped trend:

# Fit another model with a damped trend.
gas_damped <- gas_data %>%
  model(ETS(Gas ~ trend("Ad")))

report(gas_damped)

## Series: Gas 
## Model: ETS(M,Ad,M) 
##   Smoothing parameters:
##     alpha = 0.6489044 
##     beta  = 0.1551275 
##     gamma = 0.09369372 
##     phi   = 0.98 
## 
##   Initial states:
##      l[0]       b[0]      s[0]    s[-1]   s[-2]     s[-3]
##  5.858941 0.09944006 0.9281912 1.177903 1.07678 0.8171255
## 
##   sigma^2:  0.0033
## 
##      AIC     AICc      BIC 
## 1684.028 1685.091 1717.873

Forecast:

gas_damped_fc <- gas_damped %>%
  forecast(h = 20)

gas_damped_fc %>%
  autoplot(gas_data)

Comparing forecast accuray:

accuracy(gas_fit)

## # A tibble: 1 × 10
##   .model   .type        ME  RMSE   MAE   MPE  MAPE  MASE RMSSE    ACF1
##   <chr>    <chr>     <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>   <dbl>
## 1 ETS(Gas) Training -0.115  4.60  3.02 0.199  4.08 0.542 0.606 -0.0131

accuracy(gas_damped)

## # A tibble: 1 × 10
##   .model              .type       ME  RMSE   MAE   MPE  MAPE  MASE RMSSE    ACF1
##   <chr>               <chr>    <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>   <dbl>
## 1 "ETS(Gas ~ trend(\… Trai… -0.00439  4.59  3.03 0.326  4.10 0.544 0.606 -0.0217

Comment: An ETS model was fitted to the Gas production data from the aus_production dataset. The selected model includes multiplicative seasonality because the seasonal fluctuations increase as the level of the series grows. Forecasts for the next few years show continued growth in gas production with strong seasonal patterns. When a damped trend model was tested, the long-term forecasts became flatter, but the improvement in forecast accuracy was small. Therefore, the standard ETS model with multiplicative seasonality remains appropriate for this data.

The damped trend slightly reduces long-term growth in the forecasts, but the accuracy measures are very similar to the non-damped model, indicating that it does not significantly improve the forecasting performance.

Expercise : 8.8

Recall your retail time series data (from Exercise 7 in Section 2.10).

(a) Why is multiplicative seasonality necessary for this series?

(b) Apply Holt-Winters’ multiplicative method to the data. Experiment with making the trend damped.

(c) Compare the RMSE of the one-step forecasts from the two methods. Which do you prefer?

(d) Check that the residuals from the best method look like white noise.

(e) Now find the test set RMSE, while training the model to the end of 2010. Can you beat the seasonal naïve approach from Exercise 7 in Section 5.11?

Time series data:

takeaway <- aus_retail %>%
  filter(Industry == "Takeaway food services") %>%
  summarise(Turnover = sum(Turnover))

takeaway %>%
  autoplot(Turnover) +
  labs(
    title = "Australian Takeaway Food Turnover (Total)",
    y = "Turnover (AUD)",
    x = "Month"
  )

(a)

Multiplicative seasonality is required because the size of the seasonal fluctuations increases as the level of the series increases. In the retail data, the seasonal peaks and troughs become larger over time as sales grow. This means the seasonal effect is proportional to the level of the series, making a multiplicative seasonal model more appropriate than an additive one.

(b)

# Apply Holt-Winters multiplicative method

takeaway_fit <- takeaway %>%
  model(
    HW = ETS(Turnover ~ error("M") + trend("A") + season("M")),
    HW_damped = ETS(Turnover ~ error("M") + trend("Ad") + season("M"))
  )

# Forecast the next few years

takeaway_fc <- takeaway_fit %>%
  forecast(h = "3 years")

takeaway_fc %>%
  autoplot(takeaway) +
  labs(
    title = "Takeaway Food Turnover Forecasts",
    y = "Turnover (AUD)",
    x = "Month"
  )

Compare accuracy:

accuracy(takeaway_fit)

## # A tibble: 2 × 10
##   .model    .type        ME  RMSE   MAE     MPE  MAPE  MASE RMSSE  ACF1
##   <chr>     <chr>     <dbl> <dbl> <dbl>   <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 HW        Training 0.0873  18.6  13.7 -0.0673  1.96 0.308 0.329 0.172
## 2 HW_damped Training 2.02    18.6  13.7  0.259   1.95 0.309 0.329 0.160

(c)

The RMSE for the Holt-Winters multiplicative model is 18.58, while the damped trend model has an RMSE of 18.57. Since the damped model has a slightly lower RMSE, it provides marginally better one-step-ahead forecasts. However, the difference is very small, so both models perform similarly.

(d)

takeaway_fit %>%
  select(HW_damped) %>%
  gg_tsresiduals()

## Warning: `gg_tsresiduals()` was deprecated in feasts 0.4.2.
## ℹ Please use `ggtime::gg_tsresiduals()` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.

Comment: The residual diagnostics for the damped Holt-Winters model show that the residuals fluctuate randomly around zero and exhibit no significant autocorrelation. Therefore, the residuals resemble white noise, indicating that the model provides an adequate fit to the data.

(e)

# Training and test data

train <- takeaway %>%
  filter_index(. ~ "2010 Dec")

test <- takeaway %>%
  filter_index("2011 Jan" ~ .)

# Fit the Holt-Winters model on training data

hw_model <- train %>%
  model(
    HW_damped = ETS(Turnover ~ error("M") + trend("Ad") + season("M"))
  )

# Forecast the test period

hw_fc <- hw_model %>%
  forecast(new_data = test)

# Compute test RMSE

accuracy(hw_fc, test)

## # A tibble: 1 × 10
##   .model    .type    ME  RMSE   MAE   MPE  MAPE  MASE RMSSE  ACF1
##   <chr>     <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 HW_damped Test   20.8  105.  87.8 0.714  6.30   NaN   NaN 0.916

# Seasonal naïve model

naive_model <- train %>%
  model(SNAIVE(Turnover))

naive_fc <- naive_model %>%
  forecast(new_data = test)

accuracy(naive_fc, test)

## # A tibble: 1 × 10
##   .model           .type    ME  RMSE   MAE   MPE  MAPE  MASE RMSSE  ACF1
##   <chr>            <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 SNAIVE(Turnover) Test   175.  228.  180.  11.9  12.4   NaN   NaN 0.954

Comment: The Holt-Winters damped model has a test RMSE of 105.1, which is much lower than the seasonal naïve RMSE of 228.0. Therefore, the Holt-Winters model provides more accurate forecasts and clearly beats the seasonal naïve approach.

Exercise: 8.9

For the same retail data, try an STL decomposition applied to the Box-Cox transformed series, followed by ETS on the seasonally adjusted data. How does that compare with your best previous forecasts on the test set?

# Train–test split

train <- takeaway %>%
  filter_index(. ~ "2010 Dec")

test <- takeaway %>%
  filter_index("2011 Jan" ~ .)

# STL decomposition with Box-Cox + ETS

stl_fit <- train %>%
  model(
    STL_ETS = decomposition_model(
      STL(box_cox(Turnover, 0)),
      ETS(season_adjust)
    )
  )

# Forecast the test period

stl_fc <- stl_fit %>%
  forecast(new_data = test)

# Compute test RMSE

accuracy(stl_fc, test)

## # A tibble: 1 × 10
##   .model  .type    ME  RMSE   MAE   MPE  MAPE  MASE RMSSE  ACF1
##   <chr>   <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 STL_ETS Test  -239.  260.  239. -17.2  17.2   NaN   NaN 0.913

Comment: The STL + ETS model produced a test RMSE of 259.91, which is higher than both the Holt-Winters damped model (105.1) and the seasonal naïve model (228.0). Therefore, the Holt-Winters damped model provides the best forecasts for this dataset.

Data 624: Home Work 5

Mohammad Zahid Chowdhury

2026-03-08

Exercise: 8.1

Consider the the number of pigs slaughtered in Victoria, available in the aus_livestock dataset.

(a)Use the ETS() function to estimate the equivalent model for simple exponential smoothing. Find the optimal values of α and ℓ, and generate forecasts for the next four months.

(b)Compute a 95% prediction interval for the first forecast using y±1.96s where s is the standard deviation of the residuals. Compare your interval with the interval produced by R.

(a)

Data set:

Forecast for the next 4 months:

(b)

Exercise: 8.5

Data set global_economy contains the annual Exports from many countries. Select one country to analyse.

(a) Plot the Exports series and discuss the main features of the data.

(b) Use an ETS(A,N,N) model to forecast the series, and plot the forecasts.

(c) Compute the RMSE values for the training data.

(d) Compare the results to those from an ETS(A,A,N) model. (Remember that the trended model is using one more parameter than the simpler model.) Discuss the merits of the two forecasting methods for this data set.

(e) Compare the forecasts from both methods. Which do you think is best?

(f) Calculate a 95% prediction interval for the first forecast for each model, using the RMSE values and assuming normal errors. Compare your intervals with those produced using R.

Answer(a)

Select Country Bangladesh:

(b)

(c)

(d)

Fit both models together

Compare accuracy in one table:

(e)

(f)

Exercise: 8.6

Forecast the Chinese GDP from the global_economy data set using an ETS model. Experiment with the various options in the ETS() function to see how much the forecasts change with damped trend, or with a Box-Cox transformation. Try to develop an intuition of what each is doing to the forecasts.

[Hint: use a relatively large value of h when forecasting, so you can clearly see the differences between the various options when plotting the forecasts.]

Fit different ETS models:

Forecast far into the future with plot:

Intuition behind the options:

Standard ETS

Forecast follows the existing trend strongly

Can produce very rapid growth if trend is strong.

Damped Trend

Trend gradually slows down over time

Forecast curve becomes flatter in the long run

Often more realistic for long-term economic forecasts.

Box-Cox Transformation

Reduces variance instability

Makes exponential growth appear more linear

Produces smoother forecasts with narrower intervals.

Exercise: 8.7

Find an ETS model for the Gas data from aus_production and forecast the next few years. Why is multiplicative seasonality necessary here? Experiment with making the trend damped. Does it improve the forecasts?

Plot the gas data:

Comment:

Fit an ETS model

Forecasting the next few years:

Experiment with a damped trend:

Forecast:

Comparing forecast accuray:

The damped trend slightly reduces long-term growth in the forecasts, but the accuracy measures are very similar to the non-damped model, indicating that it does not significantly improve the forecasting performance.

Expercise : 8.8

Recall your retail time series data (from Exercise 7 in Section 2.10).

(a) Why is multiplicative seasonality necessary for this series?

(b) Apply Holt-Winters’ multiplicative method to the data. Experiment with making the trend damped.

(c) Compare the RMSE of the one-step forecasts from the two methods. Which do you prefer?

(d) Check that the residuals from the best method look like white noise.

(e) Now find the test set RMSE, while training the model to the end of 2010. Can you beat the seasonal naïve approach from Exercise 7 in Section 5.11?

Time series data:

(a)

(b)

Compare accuracy:

(c)

The RMSE for the Holt-Winters multiplicative model is 18.58, while the damped trend model has an RMSE of 18.57. Since the damped model has a slightly lower RMSE, it provides marginally better one-step-ahead forecasts. However, the difference is very small, so both models perform similarly.

(d)

Comment: The residual diagnostics for the damped Holt-Winters model show that the residuals fluctuate randomly around zero and exhibit no significant autocorrelation. Therefore, the residuals resemble white noise, indicating that the model provides an adequate fit to the data.

(e)

Comment: The Holt-Winters damped model has a test RMSE of 105.1, which is much lower than the seasonal naïve RMSE of 228.0. Therefore, the Holt-Winters model provides more accurate forecasts and clearly beats the seasonal naïve approach.

Exercise: 8.9

For the same retail data, try an STL decomposition applied to the Box-Cox transformed series, followed by ETS on the seasonally adjusted data. How does that compare with your best previous forecasts on the test set?

Comment: The STL + ETS model produced a test RMSE of 259.91, which is higher than both the Holt-Winters damped model (105.1) and the seasonal naïve model (228.0). Therefore, the Holt-Winters damped model provides the best forecasts for this dataset.