Project -TSF

Time Series Forecasting

1
Problem Statement
For this assignment it is requested to download the Forecast package in R. The package contains
methods and tools for displaying and analyzing univariate time series forecasts including
exponential smoothing via state space models and automatic ARIMA modelling. Explore
the gas (Australian monthly gas production) dataset in Forecast package to do the following:

 Read the data as a time series object in R. Plot the data (5 marks)
 What do you observe? Which components of the time series are present in this dataset? (5
marks)
 What is the periodicity of dataset? (5 marks)
 Is the time series Stationary? Inspect visually as well as conduct an ADF test? Write down the
null and alternate hypothesis for the stationarity test? De-seasonalise the series if seasonality is
present? (20 marks)
 Develop an ARIMA Model to forecast for next 12 periods. Use both manual and auto.arima
(Show & explain all the steps) (20 marks)
 Report the accuracy of the model (5 marks)

2
Approach
In the TSF project first we have to down load the required packages in R and see the
gas data and convert the gas data into time series object. perform EDA and plot the
time series data to understand the insights in data . check the periodicity of the data.
Decompose the series to know about the trend ,seasonality and random error and plot
the time series to identify trend ,seasonality and error. From the plots confirm what
kind of components are there like upward trend or down ward trend ,additive
seasonality or multiplicative seasonality. check for missing values if any and treat
them. Inspect visually weather series Stationary or non stationary and conduct an
ADF test by writing down the null and alternate hypothesis for the stationary test and
De-seasonalise the series if seasonality is present. Split the data into test and train to
make accurate predictions. Perform ACF and PACF analysis to obtain p and q values
then apply ARIMA model to predict the data as per requirement and also plot the
data and check for the accuracy of the model . Apply AUTO ARIMA model to
predict the data for next 12 months and check the accuracy of the model .Compare all
the models and suggest the best model for the decision making and conclude .

3
CONTENTS

1. Load the data in R as time series object ----------------------------------5

2. Exploratory Analysis of the Data ------------------------------------------5
3. Decomposition of the time series -----------------------------------------10
4. periodicity of dataset -------------------------------------------------------13
5. Missing values check-------------------------------------------------------13
6. Stationarity of the time series---------------------------------------------14
7. ADF test ---------------------------------------------------------------------15
8. De- seasonalise--------------------------------------------------------------15
9. Splitting the data into test and train---------------------------------------16
10. ARIMA ----------------------------------------------------------------------17
11. AUTO ARIMA--------------------------------------------------------------21
12. Forecasting-------------------------------------------------------------------22
13. Accuracy of the model ARIMA-------------------------------------------23
14. Accuracy of the model AUTO ARIMA----------------------------------23
15. Model Comparison----------------------------------------------------------24
16. Conclusion -------------------------------------------------------------------25

4
1.Load Data in R as a time series object
The function ts is used to create time-series objects. These are vector or matrices with class
of "ts" (and additional attributes) which represent data which has been sampled at equispaced
points in time. In the matrix case, each column of the matrix data is assumed to contain a single
(univariate) time series. Time series must have at least one observation, and although they need not
be numeric there is very limited support for non-numeric series
> ##convert to time series object
> gasts_ts= ts(gas, frequency = 12,start = c(1956),end = c(1995,8))

in the above function gas is the original data set and ts function used to convert the data into time
series object and frequency = 12 given as the data is monthly data and also mentioned start and end
date

2 .Exploratory Analysis of the Data (EDA)

exploratory data analysis (EDA) is an approach to analyze the data sets to summarize their
main characteristics, often with visual methods.

2.1 structure of the data

> str(gasts_ts)
Time-Series [1:476] from 1956 to 1996: 1709 1646 1794 1878 2173 ...
structure of the data is time series

2.2 view the data

Invoke a spreadsheet-style data viewer on a matrix-like R object.
> View(gasts_ts)
2.3 start date
to get the starting date start function is used
> start(gasts_ts)
[1] 1956 1 is the start date

2.4 end date

to get the end date end function is used

> end(gasts_ts)
[1] 1995 8 is the end date

2.5 Frequency of the data

Basically means how much time elapses between two data points
> frequency(gasts_ts)
[1] 12 is the frequency that means data is monthly

5
2.6 class of the data
Many R objects have a class attribute, a character vector giving the names of the classes from whic
h the object inherits. here it is time series.
> class(gasts_ts)
[1] "ts

2.7 cycle of the data

cycle gives the positions in the cycle of each observation.

> cycle(gasts_ts)
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
1956 1 2 3 4 5 6 7 8 9 10 11 12
1957 1 2 3 4 5 6 7 8 9 10 11 12
1958 1 2 3 4 5 6 7 8 9 10 11 12
....................................................
2.8 structureof the data

> str(gas)
Time-Series [1:476] from 1956 to 1996: 1709 1646 1794 1878 2173 ...
given data set structure is Time-Series [1:476]

data is time series

2.9.1 summary of the data

> summary(gas)
Min. 1st Qu. Median Mean 3rd Qu. Max.
1646 2675 16788 21415 38629 66600
mean is 21415
Median is 16788

2.9.2 head of the data

> head(gas)
Jan Feb Mar Apr May Jun
1956 1709 1646 1794 1878 2173 2321
displayed the first 6 records beginning from 1956

2.9.3tail of the data

> tail(gas)
Mar Apr May Jun Jul Aug
1995 46287 49013 56624 61739 66600 60054
displayed the few end 6 records 1995

6
2.9.4 Plotting method for objects inheriting from class "ts".

2.9.4.1 monthly plot

plot.ts(gasts_ts,main="Australian monthly gas production",xlab="year",ylab="ml")

2.9.4.1 .2 yearly plot

plot.ts(gasts_ts.yr,main="Australian yearly gas production",xlab="year",ylab="ml")

7
2.9.4.3 seasonplot

seasonplot(gasts_ts,year.lables=TRUE,year.lables.left=TRUE,col=1:40,pch=15,main="Australi
an monthly gas production",xlab="year",ylab="ml")
This is like a time plot except that the data are plotted against the seasons in separate years.

2.9.4.4 monthplot

These functions plot seasonal (or other) subseries of a time series. For each season (or other
category), a time series is plotted.

>monthplot(gasts_ts,main="Australian monthly gas production",xlab="year",ylab="ml")

8
In the above plot it is evident that the gas production is more in the months of May, June ,July,
August among all July is the month in which gas production is very high

2.9.4.5 Boxplot

Produce box-and-whisker plot(s) of the given (grouped) values.

> boxplot(gasts_ts ~ cycle(gasts_ts),main="Australian yearly gas production",xlab="year",ylab="m

l")

In the above plot it is evident that the gas production is more in the months of May, June ,July,
August among all July is the month in which gas production is very high

A box plot is a standardized way of displaying the distribution of data based on a five number
summary (“minimum”, first quartile (Q1), median, third quartile (Q3), and “maximum”). It can tell
you about your outliers and what their values are. It can also tell you if your data is symmetrical,
how tightly your data is grouped, and if and how your data is skewed.

Plot function is used to see the distribution of the data

Recall that the gas data set is a time series of monthly gas production in Australia from 1956 to
1995. The data in its orginal shape looks like the following:

2.9.4.6 plot
plot.ts(gasts_ts,main="Australian monthly gas production",xlab="year",ylab="ml")

Plot function is used to see the distribution of the data

9
in the above plot there is upward trend also evidence of some seasonality

What do you observe? Which components of the time series are present in this dataset? (5 marks)
What is the periodicity of dataset? (5 marks)

3. Decomposition of the time series

components of the time series are present in the dataset can be obtained by decomposition

Decomposition is the tool that can separate different components in a time series data so we can see
trend , seasonality and random error

3.1Decomposition Model-Additive
> decomp_gasts_ts = decompose(gasts_ts,type="additive")
> plot(decomp_gasts_ts)

3.2Decomposition Model-multiplicative
> decomp_gasts_ts1 = decompose(gasts_ts,type="multiplicative")
> plot(decomp_gasts_ts1)

from the above plots we an increasing trend and additive seasonality are evidenced .

so components of the time series are Trend and additive seasonality

10
3.3 plotting different components
plot(decomp_gasts_ts$trend)

there is upward trend in the above plot

3.4 plot(decomp_gasts_ts$seasonal)

constant seasonality can be seen in the above plot this is called additive

11
3.5 plot(decomp_gasts_ts$random)

some random error also seen in the above plot

Deseasonalising the data helps in understanding the relative effect of 12 month data

We can understand the long term movement of the series after seasonal effect is eliminated

plot(stl(gasts_ts,s.window = 'periodic'))

3.6 decomp_gasts_ts = stl(gasts_ts,s.window = 'periodic')

3.7 Comparison of actual and de seasonal

Deseason<-deseasonal_gasts$time.series[,2]+deseasonal_gasts$time.series[,3]

ts.plot(Deseason,gasts_ts,col=c("red","blue"),main="comparison of actual and de seasonal")

12
seasonality is strong but consistent

there was a strong growth from 1970

4. Periodicity of dataset
the periodicity of dataset :Estimate the periodicity of a time-series-like object by calculating the
median time between observations in days.

A simple wrapper to quickly estimate the periodicity of a given data. Returning an object of
type periodicity.

This calculates the median number of days between observations as a difftime object, the numerical
difference, the units of measurement, and the derived scale of the data as a string.

The time index currently must be of either Date or POSIX class, or coercible to such.

The only list item of note is the scale. This is an estimate of the periodicity of the data in common
terms - e.g. 7 day dialy data is best described as ‘weekly’, and would be returned as such.

Possible scale values are:

‘minute’,‘hourly’, ‘daily’,‘weekly’, ‘monthly’,‘quarterly’, and ‘yearly’.

An object containing a list containing the difftime object, frequency, units, and suitable scale

> periodicity(gasts_ts)

Monthly periodicity from Jan 1956 to Aug 1995

5.0 missing values check

> sum(is.na(gas))
[1] 0
NO missing values

13
Is the time series Stationary? Inspect visually as well as conduct an ADF test? Write down the null
and alternate hypothesis for the stationarity test? De-seasonalise the series if seasonality is
present? (20 marks)

6.0 Stationary means no trend and no seasonality (level data)

Stationary time series is one whose properties will not depend on time

series will not have any predictable pattern. Stationary time series another name white noise.

6.1 visualize time series by Simple Moving Average

> library(TTR)
> gasts_sma3 = SMA(gasts_ts, n=12)
> plot.ts(gasts_sma3)

> ts.plot(gasts_sma3, gasts_ts, gpars = list(col = c("black", "red")))

the time series is not Stationary because in the above plot we clearly evident that there is trend and
also seasonality present

14
7.0 Check for stationary - ADF test
ADF test : Augmented Dickey–Fuller Test
Computes the Augmented Dickey-Fuller test for the null that x has a unit root.

> adf.test(gasts_ts, alternative = "stationary")

Augmented Dickey-Fuller Test

data: gasts_ts
Dickey-Fuller = -2.7131, Lag order = 7, p-value = 0.2764
alternative hypothesis: stationary

as p-value = 0.3992is high null will fly and series is non-stationary

visually inspected and also conducted ADF test to check the series is stationary or non stationary

in both the cases it is evident that the series is non stationary

stationeries the series if the series is non stationary - it is done by taking the difference between
consecutive observations know as difference series of oreder1 that is we have to remove trend and
seasonality

8.0. De-seasonalising the series

it is possible to De-seasonalise the series by using seasadj function

seasadj function returns seasonally adjusted data constructed by removing the seasonal component.

> deseasonal_gasts=seasadj(decomp_gasts_ts)
> plot(deseasonal_gasts)

15
in the above plot we have only trend and error no seasonality

9.0 Split data into train and test sets

The fast and powerful methods that we rely on in machine learning, such as using train-test splits
and k-fold cross validation, The goal of time series forecasting is to make accurate predictions
about the future that can be easily done by splitting the series into test and train .

gasts.train <- window(gasts_ts, end = c(1994,8))

# msales.train <- window(msales_ts, start = c(2014,1),end = c(2016,12))

gasts.test<- window(gasts_ts, start = c(1994,9))

gasts.train

gasts.test

16
 Develop an ARIMA Model to forecast for next 12 periods
 Use both manual and auto.arima (Show & explain all the steps) (20 marks)

10. ARIMA : Auto Regression Integrated Moving Average

ARIMA Model is applied to non-stationary series

steps for analysis of ARIMA model

10.1 visualization (Check for stationary)

10.2 stationarization

10.3explore ACF and PACF to find p,q values

10.4 Build ARIMA model

identification of training and test periods

decision on model parameters

compare model using accuracy measures

10.5 prediction

10.1 visualization and ADF test(Check for stationary) are done and found the value of p and
decided that series is non stationary. non-stationary confirmation is important for applying
ARIMA model. next step is to stationerise the series

10.2 stationarizationt is done by taking the difference between consecutive observations know as
difference series of order 1 that is we have to remove trend and seasonality .

#Differencing the time series data

count_d1 = diff(deseasonal_gasts, differences = 1)

plot(count_d1)

17
adf.test(count_d1, alternative = "stationary")

Augmented Dickey-Fuller Test

data: count_d1
Dickey-Fuller = -18.14, Lag order = 7, p-value = 0.01
alternative hypothesis: stationary

Next step is to apply ARIMA for that we require p,d,q values .p,q values will be obtain from ACF
and PACF .

the following steps to be performed to get p,q values for ARIMA model

10.3 ACF and PACF plots

acf(gasts_ts)

acf(gasts_ts, lag.max = 24)

pacf(gasts_ts, lag.max = 24)

#There are significant autocorrelations with many lags in our gas series, as shown by the ACF plot.

#PACF plot shows that there could be monthly seasonality since the plot peaks at intervals of 12

18
## acf and pacf for differencing time series

Acf(count_d1, main='ACF for Differenced Series')

Pacf(count_d1, main='PACF for Differenced Series')

##From the ACF plot, there is a cut off after lag 0. This implies that q=0.

PACF cuts off after lag 0. Hence p=0.

##Splitting into training and test sets

gaststrain = window(deseasonal_gasts, end = c(1994,8))

gaststest= window(deseasonal_gasts, start = c(1994,9))

10.4 ARIMA MODEL

gastsARIMA = arima(gaststrain, order=c(0,1,0),season=list(order = c(2,0,2), period=12))

gastsARIMA

plot(forecast(gastsARIMA, h = 12), xlab="Time", ylab="Log Monthly Gas Production", main =

"Australian Monthly Gas Production")

tsdisplay(residuals(gastsARIMA), lag.max=15, main='Model Residuals')

> gastsARIMA

Call:
arima(x = gaststrain, order = c(0, 1, 0), seasonal = list(order = c(2, 0, 2),
period = 12))

Coefficients:
sar1 sar2 sma1 sma2

19
 0.6833 0.2338 -0.3135 -0.2470
s.e. 0.3257 0.3004 0.3197 0.1964

sigma^2 estimated as 2775067: log likelihood = -4095.86, aic = 8201.71

>

plot(forecast(gastsARIMA, h = 12), xlab="Time", ylab="Log Monthly Gas Production", main =

"Australian Monthly Gas Production")

10.5 PREDICTION done for the next 12 periods

There are no significant autocorrelations present. If the model is not correctly specified, that will
usually be reflected in residuals in the form of trends, skeweness, or any other patterns not captured
by the model. Ideally, residuals should look like white noise, meaning they are normally distributed.
Residuals plots show a smaller error range, more or less centered around 0.

tsdisplay(residuals(gastsARIMA), lag.max=15, main='Model Residuals')

20
11. Auto ARIMA
the auto.arima() function Returns best ARIMA model according to either AIC, AICc or BIC value.
The function conducts a search over possible model within the order constraints provided

> autoarima1<-auto.arima(gaststrain, seasonal=TRUE)

> autoarima1
Series: gaststrain
ARIMA(1,1,5)(0,1,1)[12]

Coefficients:
ar1 ma1 ma2 ma3 ma4 ma5 sma1
-0.7980 0.3613 -0.4045 -0.1756 -0.1505 0.0923 -0.5829
s.e. 0.0806 0.0885 0.0586 0.0567 0.0509 0.0526 0.0410

sigma^2 estimated as 2354968: log likelihood=-3947.93

AIC=7911.86 AICc=7912.19 BIC=7944.76
> tsdisplay(residuals(autoarima1), lag.max=45, main='Auto ARIMA Model Residuals')

11.1 Ljung box test

#H0: Residuals are independent
#Ha: Residuals are not independent

> library(stats)
> Box.test(gastsARIMA$residuals)

11.2 Box-Pierce test

data: gastsARIMA$residuals
X-squared = 38.813, df = 1, p-value = 4.663e-10

> Box.test(autoarima1$residuals)

21
11.3 Box-Pierce test
data: autoarima1$residuals
X-squared = 0.01426, df = 1, p-value = 0.9049

12.1Forecasting with the ARIMA model

> library(forecast)
> fcast <- forecast(gastsARIMA, h=12)
> plot(fcast)

12.2 Forecasting with the AUTO ARIMA model

> fcast1 <- forecast(autoarima1, h=12)
> plot(fcast1)

13.Accuracy of the forecast ARIMA

> f7=forecast(gastsARIMA)
> accuracy(f7, gaststest)
> f7=forecast(gastsARIMA)
> f7
Point Forecast Lo 80 Hi 80 Lo 95 Hi 95
Sep 1994 57676.22 55541.34 59811.09 54411.21 60941.23
Oct 1994 55815.81 52796.63 58834.98 51198.38 60433.23
Nov 1994 54799.69 51101.97 58497.40 49144.52 60454.85
Dec 1994 53929.36 49659.61 58199.12 47399.34 60459.39
Jan 1995 52593.87 47820.14 57367.60 45293.08 59894.66

22
Feb 1995 52649.52 47420.16 57878.88 44651.91 60647.14
Mar 1995 54063.55 48415.19 59711.90 45425.14 62701.96
Apr 1995 54928.42 48890.07 60966.76 45693.57 64163.26
May 1995 58899.32 52494.69 65303.95 49104.28 68694.36
Jun 1995 60658.64 53907.57 67409.71 50333.76 70983.51
Jul 1995 61935.92 54855.33 69016.50 51107.10 72764.74
Aug 1995 63111.41 55715.98 70506.84 51801.07 74421.74
Sep 1995 61121.29 53168.66 69073.93 48958.79 73283.80
Oct 1995 59129.30 50656.02 67602.58 46170.53 72088.06
Nov 1995 57914.47 48950.74 66878.20 44205.62 71623.32
Dec 1995 57067.51 47638.80 66496.22 42647.54 71487.48
Jan 1996 56000.07 46128.27 65871.88 40902.45 71097.70
Feb 1996 56078.12 45782.27 66373.98 40331.97 71824.27
Mar 1996 56406.89 45703.77 67110.00 40037.89 72775.89
Apr 1996 58001.59 46906.16 69097.03 41032.59 74970.60
May 1996 61590.77 50116.42 73065.12 44042.26 79139.28
Jun 1996 63411.39 51570.24 75252.54 45301.91 81520.86
Jul 1996 64267.28 52070.36 76464.19 45613.70 82920.85
Aug 1996 65046.41 52503.81 77589.01 45864.16 84228.67

In the above data we can see the predicted data for the next 12 periods

ME RMSE MAE MPE MAPE MASE

Training set 36.70681 1664.057 999.1565 -2.975644 16.351715 0.5474465
Test set -4824.55970 5738.348 4824.5597 -9.977174 9.977174 2.6434181
in the above train and test data MAPE difference is 6.37

14.Accuracy of the forecast AUTO ARIMA

> f8=forecast(autoarima1)
> f8
Point Forecast Lo 80 Hi 80 Lo 95 Hi 95
Sep 1994 55775.04 53808.39 57741.70 52767.30 58782.78
Oct 1994 52940.19 50682.99 55197.38 49488.10 56392.27
Nov 1994 50502.14 48034.33 52969.96 46727.95 54276.34
Dec 1994 50000.26 47423.86 52576.67 46059.99 53940.53
Jan 1995 48081.04 45424.00 50738.07 44017.45 52144.63
Feb 1995 48366.59 45560.18 51173.00 44074.56 52658.62
Mar 1995 48624.82 45732.22 51517.43 44200.97 53048.68
Apr 1995 51116.04 48097.60 54134.48 46499.73 55732.35
May 1995 56021.63 52915.34 59127.92 51270.97 60772.29
Jun 1995 58568.31 55351.39 61785.24 53648.45 63488.18
Jul 1995 59629.38 56325.20 62933.56 54576.08 64682.69
Aug 1995 60913.95 57509.58 64318.32 55707.42 66120.48
Sep 1995 57920.65 54164.24 61677.06 52175.71 63665.59

23
Oct 1995 55030.20 51064.50 58995.90 48965.18 61095.22
Nov 1995 53107.86 48966.88 57248.83 46774.78 59440.94
Dec 1995 52290.63 48003.15 56578.10 45733.50 58847.76
Jan 1996 50501.44 46086.73 54916.15 43749.72 57253.16
Feb 1996 50683.22 46114.81 55251.63 43696.44 57670.00
Mar 1996 51024.26 46330.61 55717.92 43845.94 58202.59
Apr 1996 53449.40 48615.67 58283.13 46056.85 60841.95
May 1996 58407.73 53452.00 63363.46 50828.59 65986.86
Jun 1996 60912.33 55826.58 65998.08 53134.35 68690.31
Jul 1996 62006.97 56803.01 67210.94 54048.20 69965.75
Aug 1996 63264.75 57938.53 68590.96 55119.01 71410.49

above data is the forecasted data as per the requirement

> accuracy(f8, gaststest)

ME RMSE MAE MPE MAPE MASE
Training set 29.745 1501.153 863.6272 2.904595 7.691125 0.4731888
Test set -1447.701 3118.433 2045.9007 -3.216644 4.275448 1.1209667
In the above data we can see the predicted data for the next 12 periods
and the MAPE difference is 3.4156 i.e not having much between train and test data

15.Comparisonof all models

SES ME RMSE MAE MPE MAPE MASE ACF

TRAIN 503.1702 4199.373 2673.83 0.9465959 13.8564 1.5242 0.77

44

TEST 3231.0385 8979.490 7750.868 14.93589 4.4184 0.61

3.6560301 62

SES DIFF

TRAIN -11.76 2811.701 1660.831 Inf Inf 1.6443 0.25

72

TEST 874.92598 5794.017 4768.963 112.6735 119.217 4.7217 0.28

04

HOLTS

TRAIN -12.97084 2808.726 1657.486 -0.08706183 8.35102 0.9448 0.25

671

24
TEST 13912.174 17131.712 14573.414 24.83965377 26.487448 8.3077 0.70
48 246

HW ADDITIVE

TRAIN 90.16106 1579.848 1046.256 0.4750107 12.261360 0.5964 0.20

29

TEST -123.9088 3561.418 2594.813 -0.6258161 5.458679 1.4792 -0.1

3 00

HW
MULTIPLICATI
VE
TRAIN 50.93662 1402.270 753.7654 0.3352298 3.542997 0.4296 -0.1
917

TEST -879.3179 2504.105 1910.9812 -2.0427764 3.965536 1.0893 -0.0

4 745

ARIMA TRAIN 41.36362 1677.981 1003.626 -1.85887 14.31550 0.5409 -0.3

876

6984.5573 7252.501 6984.557 13.21462 13.21462 3.7649 -0.0

TEST 1 084

AUTO ARIMA

TRAIN 18.40389 1551.733 876.413 2.861080 7.609896 0.4724 -0.0

157 3

TEST 1797.0596 2623.714 2420.378 3.410957 4.528056 1.3046 -0.4

3 641 1

16. Conclusion
The MAPE difference is 3.41567 between train and test data

SO AUTO ARIMA Model is the best model to predict the data for the next 12 periods for
Australian gas production

25
26