Objectives should be able to:

chapter, you
studying this time-series and|its
stochasticcomponent
models s
After Understandthe conceptsof
few standard
" implement
Analyzeand
analysis time-series
Analyze and
implement forecasting using
Regression
Analyze and
implement time-series
Sup ort
forecasting using Artifical
Vety
Networks
analysis
Overview of time-series

6.1 Overview of time-series analysis

of time series arises as a particular phenomenon that varioe
Theconcept time-series
The frequency of observations in modeling may varybased onper secand
orper minute or even on
an hourly, daily, weekly, monthly, quarterly, or yeaty
basis. Figure 6.1illustrates a simple example of time-seriesanalysis which indc
an overall increasing trend in the phernomena with some seasonal variations. Sms
other common examples include quarterly sales earning trends, the study of nse
global temperature, speech recording, and stock exchange analysis.
The details that are recorded for time-series analysis can either be a set of continuous
observations or a set of discrete observations. Some physical quantities sut3
can be consder!
temperature and pressure vary continuously with time and hencediscrete-time sernies
as continuous observations. On the other hand, in the Case of intervaskr
the measurement is not noted continuously, but rather at discrete timee
a yearisa caR
instance, the profit made by a company atat the end of each quarter of
of adiscrete-time interval.
invariouspractia
Time series modeling and forecasting have vital significance in donedailyin
analysis domains. A good amount of real real research work is being
seriesanalysiss
this research area for
several years. The collected data for time
 observations based on which the
dTeads
the
pastVarious fitting models
trends have forecasting is done for
been developed for timeseries
s u i t s wellfor a
tuure particular area of study such as business, finance.
zingthat
angieeting
arndeconomics,
onastng

Year >

6.1: example
An ofa timeseriesanalysis
Figure
analysis involvethe following::
uses of time-series regarding the variationsin the values
Fewofthe comparisonmade
realisticin-depth different times.
" A variables at is predicted based on
ofdifferent phenomenon that
The future behavior of a particular
current situation with the
experience. comparing the actual
planning done by
Business
expected one. phenomenon underconsideration.
the past behaviorofthe time series
A studymade on stationarity. A
is its This indicates
characteristics ofthetime series over time.
One of theeimportant behavior does not change the variability is constant
considered stationary ifitsaboutthe samelevel,thatis, time seriestime-series
suchasthe
that the observed values vary propertiesofthe stationary
ofthe
over time. In turn, the statistical However, most donot
Constant over time. Non-stationarytimeseries
upwardor
mean and variance also remain non-stationary. drifting
either
problems that are encountered are and follows a trend by
have a constant mnean or variarnce
backward.
of time-series timeseriesnamely,
6.2 Components Components of
These
components
important Components.
four
Inthis we will study the
trend, section, seasonal, and irregular
cyclical, as shown in. Figure 6.2.
are distributed
or random
356 Data Sci

Components of
Time Series

Long-term
Short-term
Movement
irregular or Random
Movement or Trend Movement

Seasonal Variations Cyclic Variations

Figure 6.2: Components of time-series

overallinclination of data towards an
Trend: The trend showsthe
or decreasing drift during a long period of time. It is an average, long-term
tendency that may not always be increasing or decreasing in the same
increasing
direction throughout the given period of fime. That is, the tendencies mau k
stable, in increasing order, or decreasing order in different sections of time.
The overall trend should, however, signify whether it is upward, downward.
or stable.

Analysis of trends is profoundly done in many real-life cases such as in

share markets, birth and death rates, and agricultural production. The
trend analysis is mainly done by plotting the time series value on agraph in
accordance with time.
Seasonal: Seasonal pattern in time series refers to the rhythmic torces uia
occur pernodically in a regular manner throughout the considered time.Sua
Seasonal variations in data may be noted hourly, daily, weekly, que
or monthly and are often
conventions. caused due to natural forces or man-made
The best example of natural seasonal variation is found inthe
production of crops that depend on the four seasons found in ayear. Als,
sales of coffee
cups, raincoats, electric particular season
Compared
seasonal
to other
seasons of the year.
farns shoot up in a ofman-made
other variation can be the sale of Again, one example o rsome

important occasions.
cannot be taken as an An upward sweetsor during festive
downward
season
seasonal variation
Cyci c al: The
cydle. In cyclical indicator of
of interpreting business conditions.
variaistiolike
phases ofthis case, a cycle n of aatime series operatesin a
recognizable
the

in expansiison,termed
cycic period that encompasse
shownmoOvement
Figure 6.3. recessiason,the compl ete
deprBusiessinesson, Cycie
four-phase
and recovery. This
or the Economic
Grcleas
 357
Peak
uossa
Expansio,
Business
Variation

Depression

Trough

Time

Figureé6.3: Cyclical variation in the Business cycle

cycles occurin a time-series as long-term oscillations that varies over

The expansion marks an
While increase in the cycle, recession leads to
Contraction.The troughindicatestheend of the depression. The measurement
time.

expansion phase is done by considering the trough value of the

of the of the current cycle.
previouscycle the peak
to
of time-series, as the name suggests, is
Iregular: The irregular component
regular variationandis: random. This variation may arise due to several
nota
uncontrollable or unpredictable situations such as natural disasters or wars.
time series can be realized by the short-term fluctuations
that
Iregularity in
series that are usually unsystematic and unpredictable.
orur in the
for a phenomenon or variable v is given as v, =ft). Here,
ime-series equationphenomenon
A or variable at time t. In such a case, the
time
the
-is the value of
has a corresponding recorded values v,, v,, D,, ...
for
eries valuest, t, ty ... ,, under study. Now, considering V, as the value of the
phenomenon or variable stated as:
the
t, the additive modelof the time series can be
variable at time ..(6.1)
,=T,+5,+C+R
of the trend, seasonal,cyclic, and random
Bere, T, S, C,and R, representsthet. values additive
the four
model indicates thatanalyzed
Auctuations respectively at time This independently and are neededto be
Omponents ofthe time seriesperform
parately. that
considers
components
thefour expressed
Again, themultiplicative model of the time:series
other. This model canbe
inthe timeesseries function proportionatelytoeach
ythe equation: .(6.2)
represented
v,=I, xS, xC, xR, data can be
As can time-series cyclical, and
be seen from Equations (6.1) and (6.2), oftrend, seasonality,
ether in teterms of addition or nultiplication
the time-sei

Time-series forecasting models

6.3 a standard forecasting
model is fitted to a
forecasting, given
In time-series the forecasting or estimation of future values is made based ontimthe.e
series data and data. Based on--the nature and pattern
previously provided
time-series
to be chosern that can generate an optimum foreractETs time-series,
asuitable model hasseveral time-series models have been proposed by rese
T the nast decades,
upgradation of appropriate time series forecasting mo
for the advancement and
time-series models are the stochastic modele &.
Three of the broad categories of models, and the support vector machines basei
artificial neural network-based common standard modeling techniques used
of the
model. Figure 6.6 lists out some broad classes of time-series models
that
time-series. The three
in forecasting of chapter are - the stochastic models, the
Support Vector
have been discussed in this the Artificial Neural Network (ANN) based
Machines (SVM) based models, and
models.
 Auto-Regression(AR)
NovingAverage(MAS
RegressiveMoving
Auto (ARMAJ
loAverage
Auto RegressiveIntegrated
ovingAverage(ARIBJAI
"Feedforward
Network Neural
"Timeagged
"Seasonal Neural NetWork
Network Artificial Neural
Linear
Fourier Kernel
Kernel
\nalysis361
SeasonalARIBIA "Probabilistic Neural Network
" Generalised Regression Radial (rbf)
Kernel Basis Function
"Autoregressive Conditíonal
Heteroskedastícity(ARCH) Neural Network "Pair Hidden
oBox-Jenkins
Time Series Analysis
Sub
sequencetMtearkchniov qModel
ue
TimeSeriesAnalysis using Artificiat
usingStochastic Neural tletwork Time Series Analysis
Models based NModels using Support Vector
MachiModels
nes based

Figure 6.6: Various Standard

time-series
Time-SSeries Models
each of the cases of models, the time-series is
In possible trend and/or
examine
sharp the
changesfound
found in the seasonality, the fluctuation of initially plotted to
the
fluctuation, and any other
seasonal components are then often untoward
data with time, any
The trend
and
Components from the series. An removed random
from the changes.
detect stationary series to
thenapplied forecasting
for data for the future. appropriate time-series model is

631 Time-series forecasting using

stochasticmodels
in the stochastic process,thechanges in values occur randomly over time.This
process contrasts the deterministicprocess which evolves systematically over time.
Stochastic modelsdo not give identical results if the modelis run several times. While
stochastic models are more realistic, deterministic modelsyare easier to interpret and
WOK upon. This is so as the probability of each step in the deterministic model is
aways 1(fully certain) whereas in the stochastic process it isn't so.
of random
atiematically, astochastic process can be expressed as a family
variables:
{Y(): teT} variable Y)is
The equation indicates that at every time t found in set T, arandomwhereasif Tis
witnessed. Jf the eset Tis finite, the equation isa discrete-timeprocess,
inAinite, the equation is acontinuous-time process. market where the stock market
The
gaphutaiurneplise example
eof the stochasticmodel is the stockparticular instance of time. the
Tactors at a values which make
generated based on various previous the
stochastic
yalue
is not Solely determined from the Another example of selections
nodepredlictioisn the
random
of future values a challenging task. which generates
Monte Carlo model
simulation
 Fundanl
Science
Data situations
362 complex where many random
moel exhaustive
samplesto uses
fromthe This method complexityof
this statistical sampling methods
method arises due to twO
computation.The simulated hundreds or thousands
mainhat
required.
input
model
heavyresultgenerated is
isaprobability distribution of possible of
re resqounise ha
a
The
generates outcomesim.eSotand
random numbers using the
method
The generator which
number stochastic c
time-series models
on ceep
r t
tendsto repeat after a certain

nu m b o t t a n d o nm
Let us now
of the
discuss a few models
over
have been rigorously used
time. Thesefor
forecasting and risk
th at
analysis. and have ben
ite r a t i o ns
developed
dealing with timeseries
data
ad a p t e d in
6.3.1.1TheAutoregressive Moving
(ARMA)
Average model
Autoregressive Moving Average the
model is strictly based
The time-series models,
namely AR (Autoregressive) model,
0n two
standard linear time-series models
Linear have and the
MA
(MovingAverage) model.
due to the
simpliity of the model and ease in implementation. gai
try to understand how the AR
ned populari
To understandty
us at first and MA
the ARMA model, let models
individuallywork.
time-series data available to predict future
The AR modeluses the past
similar to the concept of linear regression. However, by mainly relying on th
data for future prediction, there is almost no chance of getting 100% accuracy in
behavior,
predicted results. The AR model is denoted by AR(p), where p denotes the order of
the process. For example, AR(3) indicates the third-order autoregressive proceS In
general, the autoregressive model is defined as:

i=1

Here, C15 a constant, , ,, .., .are the model parameters (which cannot be
and e, is the stochastic whitee Zer)
are to be estimated using the noise that defines the randomness. These parameters
variables
usedfrom the past data are known autocorrelation function. The input valuecis
determined by considering theconsidered
process asmeantheHas
lag shown in the
variables. constant equation
Thefollowing

It should be
almost noted that if the
time-seri
icontmpossiainibnlge purely
as it is esisa becomes
process, forecasting discrete
white no1se
signal
varwhiiacnce.h areTo uncorrelated random. In such a case, the white noise
isa
finite
Imean and
bothsummar random variables that zero walk by
linear izine, the AR(p) model is an extensionhas
model
\nature. This of the random govemed
is linearly
 theprevious
term, along with the coefficients for each term.
-Series Analysis A 363
nmodelis an extension of the
AR
ARmodelmainly depends on the lagged
MA
The
model ofwith a major
values
resultsWhereasthe
af MA model
mainly the data dif erence in that
further forecasting. The MA depends on provided for the
model can be the errorsas: of the
for
forecasts

Y, =c+e,+)ج
defined forecasting
previous
i=1

theAR model,in
the MA
like model, cis a constant, ,, ,, ... ,
Here (which cannot be
parameters s zero) and e, E E,. are the model
define the randomness. The E,are the white
that
These parameters are to be estimated noise error
autocorrelation function. These
terms
error terms are using the
known
(iid) random numbers sampled from a
distrlbuted
MA(9), where g
independent
normal distribution. identically
The MA
is
model denoted by denotes the order of the process. For
third-order,moving average is denoted by MA(3) and can be example,
written as:
the
Y- c+e, +), Et, t , E3
Now,. combiningthe concepts ofthee AR model and the MA model, the ARMA model
developed. The ARMA(P,q) model combines the.
the AR(p) and MA(g) models
has been
asshownbelow:

Y,=c+¬, +2 Øy,+Ø,44 /=1

i=1

main notion of combining the concepts of the AR model and the MAmodel in
The requirements - the AR(p) model
the ARMA model is to equally consider both the based on its own lagged values, and
encompasses regressing the feature considered
model includes modeling theerror termstor several times in the past. By
the MA(q) give much
concepts of both the models, the ARMA is expected toARMA(1, 1)
Cmbinng the models. Now, if we consider the
individual two
erresults than the equation can be written as
Moael which considers asingle time value. the

canbe usedto predictfuture

This showst modelthat particular product
the basic calculation of the ARMA numbersfor a model
Tends. For many
if we want to find how current month(Y), the ARMA
instance,
shocld be manufactured by a Companyin requiredto be
this manufacturedinthe
product
be used to check products were predictingthe number of the
ast how predictionof
many in involvedinthe
month y) how much error was
quired for the last
involved
(¬,), and also the
error
TuTrber for the currentmonthmonth (e).