Chapter 1

Introduction

The following abbreviations are used in these notes:

TS Time Series
ma moving average filter
rv random variable
iid independently identically distributed
cdf cumulative distribution function
pdf probability density function
µ expected value of a random variable
σ 2 variance of a random variable
N(µ, σ 2 ) normal standard distribution of a random variable with expected value
µ and with variance σ 2
AR(p) Autoregressive model of order p
MA(q) Moving Average model of order q
ARMA(p,q) Autoregressive Moving Average model of order (p, q)
ARIMA(p,d,q) Autoregressive Integrated Moving Average model of order (p, q)
with the integration parameter d.

Also, CAPITAL Latin letters are used to denote random variables, small Latin
letters are used to denote realization of the random variables.

3
4 CHAPTER 1. INTRODUCTION

1.1 Introductory Definitions and Examples

A Time Series (TS) is a collection of observations made sequentially, usually in
time, but the observations may be collected in other domain as well, in another
kind of distance.
We denote a TS of a variable X as follows

X = (X1 , X2 , . . . , Xt , . . .)

or
X = {Xt }t=1,2,...
TS theory finds applications in a variety of fields. For example in
Economics: X = Unemployment, Consumption;
Meteorology: X = Changes in global temperature, Summer monsoon rainfall in
India;
Sciences: X = Chemical process temperature;
Sport: X = Olympic gold medal performance in track and field events.
TS has following features
1. It can be continuous or discrete,
• continuous TS is when observations are made continuously,
• discrete TS is when observations are made at specific times, usually
equispaced (every second, every day etc); TS is then discrete even if
the measured variable is continuous, like river level measured every
day.
2. Successive observations are usually correlated, so future values may be pre-
dicted from past values.
3. TS can be deterministic or stochastic,
• it is deterministic if the future observations may be predicted exactly,
• it is stochastic when such exact prediction is impossible due to ran-
domness of the observations. Then future can only be partly deter-
mined by past.
In this course we will consider discrete stochastic time series.

How do we analyze a TS? We can do it by:

1.2. SIMPLE DESCRIPTIVE TECHNIQUES 5

420
UK.Consumption

320

220

120
1960 1970 1980 1990 2000 Year

Figure 1.1: Yearly values of consumption in the UK, 1960-

2000, Billions of pounds sterling, 1990 prices. Source:
http://www.fgn.unisg.ch/eumacro/macrodata/macroeconomic-time-series.html

Description: Plot of the observations in time gives a general view of the phe-
nomenon variable X represents. It shows what kind of variation occurs in
time, is there any trend or seasonality, are there any unusual observations
(outliers), or turning points.

Explanation: Construction of a mathematical model which explains the observed

variability in the data.

Prediction: Based on the model we can predict, with some confidence, future
observations.

Control: If the predicted future values are ‘off target’ then it may be possible to
change the factors which influence the observed variable and so to control
the outcome.

1.2 Simple Descriptive Techniques

1.2.1 The Time Plot
A good picture is better than 1000 words.
6 CHAPTER 1. INTRODUCTION

12

10
Unemployment

0
1960 1970 1980 1990 2000 Year

Figure 1.2: Yearly values of unemployment percentage in the UK, 1960-

1999. Source: http://www.fgn.unisg.ch/eumacro/macrodata/macroeconomic-
time-series.html

0.4
Changes in global temperature

0.0

-0.4

-0.8

1880 1900 1920 1940 1960 1980 Year

Figure 1.3: Surface air ‘temperature change’ for the globe, 1880-1985. Degrees
Celsius. ‘Temperature change’ means temperature against an arbitrary zero point.
Source: J.Hansen and S.Lebedeff (1987).
1.2. SIMPLE DESCRIPTIVE TECHNIQUES 7

90

85
HighJump

80

75

70

1900 1920 1940 1960 1980

Year

Figure 1.4: The gold medal performance in the men’s high jump (mea-
sured in inches) for the modern Olympic series, 1896-1984. Source:
http://lib.stat.cmu.edu/DASL

1000

900
IndiaMonsoon

800

700

600

1813 1833 1853 1873 1893 1913 1933 1953 1973 1993
Year

Figure 1.5: Summer monsoon rainfall data from India. Units mm. Jun-Sep 1813
to Jun-Sep 1998.
8 CHAPTER 1. INTRODUCTION

26

24
Temperature

22

20

18

0 50 100 150 200 Time

Figure 1.6: Chemical process temperature readings, every minute. Degrees Cel-
sius. Source: Box and Jenkins (1976).
Sales

800

600

400

200

Jan 65 Jan 66 Jan 67 Jan 68 Jan 69 Jan 70 Jan 71 Year

Figure 1.7: Sales of an industrial heater; monthly data starting from January 1965
till December 1971. Source: Chatfield (2004). (The last value is added for com-
pleteness).
1.2. SIMPLE DESCRIPTIVE TECHNIQUES 9
UK consumption [Billions of £]

400

300

200

100
1960 1970 1980 1990 2000 Year

Figure 1.8: TS and a linear trend for the UK consumption data, compare Fig. 1.1

A plot representing the observations against time gives an initial analysis of the
data. A good TS plot should

• give a clear and self-explanatory title,

• state units of measurements,

• have carefully chosen scales, including the size of intercept,

• have clearly labelled axes,

• have appropriate plotting symbol and line (if a line is included).

A TS plot may reveal various features of the data, such as

• Trend, which indicates a long term change in the mean level. For example,
UK consumption data indicate a steady growth in time, a straight line would
well describe the trend of growth.

• Periodicity, which shows a pattern repeating in time. Sales of an industrial

heater data have both trend and periodicity (see Figure 1.7).

• Unusual features, such as a sudden single peak or a turning point. For ex-
ample a sudden drop of the temperature of the chemical process at the end
of measurement time (see Figure 1.6).
10 CHAPTER 1. INTRODUCTION

1.2.2 Deterministic Trend and Seasonality

A deterministic trend model with a seasonal effect can take either an additive
form,
Xt = mt + st + Yt , t = 0, 1, . . . , n, (1.1)
or a multiplicative form, such as

Xt = mt st Yt , t = 0, 1, . . . , n, (1.2)

or a mixed form,

Xt = mt st + Yt , t = 0, 1, . . . , n, (1.3)

where mt = m(t) is a (usually slowly changing) function of time, so called ‘trend

component’, st = s(t) is a periodical function of time and Yt is a random noise
component. Model (1.2) can be easily transformed to the additive form by taking
a logarithm of both sides. Model (1.3) is often referred to as a multiplicative one.

The most common trend function is a polynomial of a degree k ≥ 1, i.e.,

m(t) = β0 + β1 t + β2 t2 + . . . + βk tk (1.4)

A linear trend m(t) = β0 + β1 t is a special case of (1.4).