12-practical.qmd

---
title: "ETC3550/ETC5550 Applied forecasting"
author: "Ch12. Some practical issues"
institute: "OTexts.org/fpp3/"
pdf-engine: pdflatex
fig-width: 7.5
fig-height: 3.5
format:
  beamer:
    theme: monash
    aspectratio: 169
    fontsize: 14pt
    section-titles: false
    knitr:
      opts_chunk:
        dev: "cairo_pdf"
include-in-header: header.tex
execute:
  echo: false
  message: false
  warning: false
---


```{r setup, include=FALSE}
source("setup.R")
```

# Models for different frequencies

## Models for different frequencies

### Models for annual data

 * ETS, ARIMA, Dynamic regression
\pause

### Models for quarterly data

 * ETS, ARIMA/SARIMA, Dynamic regression, Dynamic harmonic regression, STL+ETS, STL+ARIMA
\pause

### Models for monthly data

 * ETS, ARIMA/SARIMA, Dynamic regression, Dynamic harmonic regression, STL+ETS, STL+ARIMA

## Models for different frequencies

### Models for weekly data

 * ARIMA/SARIMA, Dynamic regression, Dynamic harmonic regression, STL+ETS, STL+ARIMA, TBATS
\pause

### Models for daily, hourly and other sub-daily data

 * ARIMA/SARIMA, Dynamic regression, Dynamic harmonic regression, STL+ETS, STL+ARIMA, TBATS

# Ensuring forecasts stay within limits

## Positive forecasts
\fontsize{9}{9}\sf

```{r, fig.height=3.8}
recent_prices <- prices |> filter(!is.na(eggs))
recent_prices |>
  model(ETS(log(eggs) ~ error("A") + trend("A") + season("N"))) |>
  forecast(h = 50) |>
  autoplot(recent_prices)
```

# Forecast combinations

## Forecast combinations

### Clemen (1989)
"The results have been virtually unanimous: combining multiple forecasts leads to increased forecast accuracy. \dots In many cases one can make dramatic performance improvements by simply averaging the forecasts."

## Forecast combinations
\fontsize{9}{10}\sf\vspace*{-0.2cm}

```{r}
aus_cafe <- aus_retail |>
  filter(Industry == "Cafes, restaurants and catering services") |>
  summarise(Turnover = sum(Turnover))
fc <- aus_cafe |>
  filter(Month <= yearmonth("2013 Sep")) |>
  model(
    ETS = ETS(Turnover),
    ARIMA = ARIMA(Turnover)
  ) |>
  mutate(
    Combination = (ETS + ARIMA) / 2
  ) |>
  forecast(h = "5 years")
```

## Forecast combinations
\fontsize{10}{10}\sf\vspace*{-0.2cm}

```{r combineplot, echo=TRUE, fig.height=4.2}
fc |> autoplot(aus_cafe, level = NULL) +
  labs(
    x = "Year", y = "$ billion",
    title = "Australian monthly expenditure on eating out"
  )
```

## Forecast combinations
\fontsize{10}{15}\sf
```{r combineaccuracy, dependson="combine1"}
fc |> accuracy(aus_cafe)
```

# Missing values

## Missing values
\fontsize{12}{13}\sf

**Functions which can handle missing values**

 * `ARIMA()`
 * `TSLM()`
 * `NNETAR()`
 * `VAR()`
 * `FASSTER()`

**Models which cannot handle missing values**

 * `ETS()`
 * `STL()`
 * `TBATS()`

\pause

### What to do?
 1. Model section of data after last missing value.
 2. Estimate missing values with `interpolate()`.

## Missing values
\fontsize{12}{12}\sf
```{r}
gold <- as_tsibble(forecast::gold)
gold |> autoplot(value)
```

## Missing values
\fontsize{12}{12}\sf
```{r, fig.height=3}
gold_complete <- gold |>
  model(ARIMA(value)) |>
  interpolate(gold)
gold_complete |>
  autoplot(value, colour = "red") +
  autolayer(gold, value)
```

# Outliers

## Outliers
\fontsize{11}{13}\sf

```{r, fig.height=3.4}
fit <- gold |>
  model(ARIMA(value))
augment(fit) |>
  mutate(stdres = .resid / sd(.resid, na.rm = TRUE)) |>
  filter(abs(stdres) > 10)
```