inst/doc/vignette_02_hourly_fixed.R

In tscv: Functions and Utilities for Tidy Time Series Forecasting and Time Series Cross-Validation

## ----include = FALSE----------------------------------------------------------
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  dev = "png",
  # fig.path = "figures/tscv-",
  fig.height = 5,
  fig.width = 7
)

## ----packages, message = FALSE, warning = FALSE-------------------------------
# Load relevant packages
library(tscv)
library(tidyverse)
library(tsibble)
library(fable)
library(feasts)

## ----abbreviations, echo=FALSE, warning=FALSE, message=FALSE, results='hide'----
Sys.setlocale("LC_TIME", "C")

## ----clean_data, fig.alt = "Plot raw data"------------------------------------

series_id = "bidding_zone"
value_id = "value"
index_id = "time"

context <- list(
  series_id = series_id,
  value_id = value_id,
  index_id = index_id
)

# Prepare data set
main_frame <- elec_price %>%
  filter(bidding_zone %in% c("DE", "FR", "NO1", "SE1"))

main_frame

main_frame %>%
  plot_line(
    x = time,
    y = value,
    color = bidding_zone,
    facet_var = bidding_zone,
    title = "Day-ahead Electricity Spot Price",
    subtitle = "2019-01-01 to 2020-12-31",
    xlab = "Time",
    ylab = "[EUR/MWh]",
    caption = "Data: ENTSO-E Transparency"
    )

summarise_data(
  .data = main_frame,
  context = context
)

summarise_stats(
  .data = main_frame,
  context = context
)


## ----split_data---------------------------------------------------------------

# Setup for time series cross validation
type = "first"
value = 2400      # size for training window
n_ahead = 24      # size for testing window (= forecast horizon)
n_skip = 23       # skip 23 observations
n_lag = 0         # no lag
mode = "slide"    # fixed window approach
exceed = FALSE    # only pseudo out-of-sample forecast

split_frame <- make_split(
  main_frame = main_frame,
  context = context,
  type = type,
  value = value,
  n_ahead = n_ahead,
  n_skip = n_skip,
  n_lag = n_lag,
  mode = mode,
  exceed = exceed
)

# For illustration, only the first 50 splits are used
split_frame <- split_frame %>%
  filter(split %in% c(1:50))

split_frame


## ----train_models-------------------------------------------------------------

# Slice training data from main_frame according to split_frame
train_frame <- slice_train(
  main_frame = main_frame,
  split_frame = split_frame,
  context = context
  )

train_frame

# Convert tibble to tsibble
train_frame <- train_frame %>%
  as_tsibble(
    index = time,
    key = c(bidding_zone, split)
    )

train_frame

# Model training via fabletools::model()
model_frame <- train_frame %>%
  model(
    "SNAIVE" = SNAIVE(value ~ lag("week")),
    "STL-NAIVE" = decomposition_model(STL(value), NAIVE(season_adjust)),
    "SNAIVE2" = SNAIVE2(value),
    "SMEDIAN" = SMEDIAN(value ~ lag("week"))
    )

model_frame

# Forecasting via fabletools::forecast()
fable_frame <- model_frame %>%
  forecast(h = n_ahead)

fable_frame

# Convert fable_frame (fable) to future_frame (tibble)
future_frame <- make_future(
  fable = fable_frame,
  context = context
  )

future_frame



## ----accuracy_horizon, fig.alt = "Plot accuracy by horizon"-------------------

# Estimate accuracy metrics by forecast horizon
accuracy_horizon <- make_accuracy(
  future_frame = future_frame,
  main_frame = main_frame,
  context = context,
  dimension = "horizon"
)

accuracy_horizon

# Visualize results
accuracy_horizon %>%
  filter(metric == "MAE") %>%
  plot_line(
    x = n,
    y = value,
    facet_var = bidding_zone,
    facet_nrow = 1,
    color = model,
    title = "Evaluation of forecast accuracy by forecast horizon",
    subtitle = "Mean absolute error (MAE)",
    xlab = "Forecast horizon (n-step-ahead)",
    caption = "Data: ENTSO-E Transparency, own calculation"
    )


## ----accuracy_split, fig.alt = "Plot accuracy by split"-----------------------

# Estimate accuracy metrics by forecast horizon
accuracy_split <- make_accuracy(
  future_frame = future_frame,
  main_frame = main_frame,
  context = context,
  dimension = "split"
)

accuracy_split

# Visualize results
accuracy_split %>%
  filter(metric == "MAE") %>%
  plot_line(
    x = n,
    y = value,
    facet_var = bidding_zone,
    facet_nrow = 1,
    color = model,
    title = "Evaluation of forecast accuracy by split",
    subtitle = "Mean absolute error (MAE)",
    xlab = "Split",
    caption = "Data: ENTSO-E Transparency, own calculation"
    )