if(!require("pacman")) install.packages("pacman")
pacman::p_load(tidyverse, ggplot2, forecast, caret, fpp2)

?visitors

## starting httpd help server ... done

This gives information about the data set, indicating it is a monthly australian short term overseas visitors time series data over May 1985 - April 2005. It has been taken from “Forecasting with exponential smoothing: the state space approach” By Hyndman, R.J., Koehler, A.B., Ord, J.K., and Snyder, R.D.

It also indicated that the scale is in thousands.

data(visitors)
force(visitors)

##        Jan   Feb   Mar   Apr   May   Jun   Jul   Aug   Sep   Oct   Nov
## 1985                          75.7  75.4  83.1  82.9  77.3 105.7 121.9
## 1986  98.0 118.0 129.5 110.6  91.7  94.8 109.5 105.1  95.0 130.3 156.7
## 1987 139.7 147.8 145.2 132.7 120.7 116.5 142.0 140.4 128.0 165.7 183.1
## 1988 161.3 180.4 185.2 160.5 157.1 163.8 203.3 196.9 179.6 207.3 208.0
## 1989 168.9 191.1 180.0 160.1 136.6 142.7 175.4 161.4 149.9 174.1 192.7
## 1990 176.2 192.8 189.1 181.1 149.9 157.3 185.3 178.2 162.7 190.6 198.6
## 1991 177.4 190.6 189.2 168.0 161.4 172.2 208.3 199.3 197.4 216.0 223.9
## 1992 196.1 238.2 217.8 203.8 175.2 176.9 219.3 199.1 190.0 229.3 255.0
## 1993 242.8 245.5 257.9 226.3 213.4 204.6 244.6 239.9 224.0 267.2 285.9
## 1994 250.5 304.3 307.4 255.1 214.9 230.9 282.5 265.4 254.0 301.6 311.0
## 1995 303.8 319.1 313.5 294.2 244.8 261.4 329.7 304.9 268.6 320.7 342.9
## 1996 317.2 392.7 365.6 333.2 261.5 306.9 358.2 329.2 309.2 350.4 375.6
## 1997 342.9 408.0 390.9 325.9 289.1 308.2 397.4 330.4 330.9 366.5 379.5
## 1998 346.2 353.6 338.6 341.1 283.4 304.2 372.3 323.7 323.9 354.8 367.9
## 1999 351.0 398.6 389.0 334.1 298.1 317.1 388.5 355.6 353.1 397.0 416.7
## 2000 360.8 434.6 411.9 405.6 319.3 347.9 429.0 372.9 403.0 426.5 459.9
## 2001 416.6 429.2 428.7 405.4 330.2 370.0 446.9 384.6 366.3 378.5 376.2
## 2002 379.3 437.2 446.5 360.3 329.9 339.4 418.2 371.9 358.6 428.9 437.0
## 2003 396.6 427.5 392.5 321.5 260.9 308.3 415.5 362.2 385.6 435.3 473.3
## 2004 420.2 454.8 432.3 402.8 341.3 367.3 472.0 405.8 395.6 449.9 479.9
## 2005 462.4 501.6 504.7 409.5                                          
##        Dec
## 1985 150.0
## 1986 190.1
## 1987 222.8
## 1988 245.8
## 1989 247.4
## 1990 253.1
## 1991 266.8
## 1992 302.4
## 1993 344.0
## 1994 384.0
## 1995 422.3
## 1996 465.2
## 1997 448.3
## 1998 457.6
## 1999 460.8
## 2000 559.9
## 2001 523.2
## 2002 534.0
## 2003 566.6
## 2004 593.1
## 2005

Checking if time series

is.ts(visitors)

## [1] TRUE

1. Analysing the properties of the times series data

#ts.plot(visitors) (base R style of plotting)
autoplot(visitors)+
  ggtitle("Overseas visitors to Australia")+
  xlab("Months")+
  ylab("Thousands of people")

ggseasonplot(visitors, polar = TRUE, )+
  ggtitle("Polar Season Plot: Visitors")

ggsubseriesplot(visitors)+
  ggtitle("Sub-series Plot: Overseas visitors to Australia")+
  xlab("Months")+
  ylab("Thousands of people")

ggAcf(visitors)+
  ggtitle("Autocorrelation Plot: Overseas visitors to Australia")

There seems to be a great fall in 2003, sometime in May. There is a strong positive trend in the data this can be seen in the time series and seasonal plots. There is also a seasonality in the data over months which can be verified in subseries and autocorrelation plots.

2. Spliting into train and test data sets The total number of years is 2005. So test is from 2004 and train is till 2004 or till the end of 2003

v_train <- window(visitors, end = c(2003, 4))
v_test <- window(visitors, start = c(2003, 5))

Holt-Winters’ multiplicative method

hwfc <- hw(v_train, seasonal="multiplicative")
autoplot(v_train) +
  autolayer(hwfc, series="HW multiplicative forecasts", PI=FALSE)+
  autolayer(v_test)+
  ggtitle("Holt-Winter's forecast: Overseas visitors to Australia")+
  xlab("Months")+
  ylab("Thousands of people")+
  guides(colour=guide_legend(title="Forecast vs Actual"))

3. I believe of the two Holt Winters’ Forecasts, multipliactive forecast is necessary here as the seasonality is increasing with time (year).

4. Forecasting using an ETS model, an additive ETS model applied to a Box-Cox transformed series, and a seasonal naïve method.

1. ETS model

v_ets <- v_train %>% ets() %>% forecast(h = length(v_test)) 
autoplot(v_ets)+
  ggtitle("ETS: Overseas visitors to Australia")+
  xlab("Months")+
  ylab("Thousands of people")

2. additive ETS model applied to a Box-Cox transformed series

v_ets.box <- ets(v_train, lambda = BoxCox.lambda(v_train), additive.only = T) %>% forecast(h = length(v_test)) 

autoplot(v_ets.box)+
  ggtitle("Box-Cox transformed: Overseas visitors to Australia")+
  xlab("Months")+
  ylab("Thousands of people")

3. seasonal naïve method

v_sn <- snaive(v_train, h = length(v_test))

autoplot(v_sn) + ggtitle(" Seasonal Naive Model: Overseas visitors to Australia")+
  xlab("Months")+
  ylab("Thousands of people")

5. Accuracies

accuracy(v_ets, v_test)[,"RMSE"]

## Training set     Test set 
##     14.53480     80.23124

accuracy(v_ets.box, v_test)[,"RMSE"]

## Training set     Test set 
##     14.97096     78.61032

accuracy(v_sn, v_test)[,"RMSE"]

## Training set     Test set 
##     31.15613     50.30097

Seasonal Naive is the best model for having lowest Test RMSE.

Residuals Test for Seasonal Naive

checkresiduals(v_sn)

## 
##  Ljung-Box test
## 
## data:  Residuals from Seasonal naive method
## Q* = 295.02, df = 24, p-value < 2.2e-16
## 
## Model df: 0.   Total lags used: 24

As per the test, it is visible that, the p-value is close to 0. This means that we can reject the hypothesis that it is white noise and conclude that the residuals are clearly distinguishable from a white noise series.