Introduction to Time Series Models
Types of Forecasting Methods
There are a number of classifications of forecast
methods;
Qualitative vs Quantitative
Time Series vs Causal
The above classifications are not mutually
exclusive
The forecaster needs to be aware of the
appropriate method to match the forecast
situation
2
Quantitative vs Qualitative
Prediction
It should be noted that mostly forecasts will be of
variables that are measured quantitatively e.g.: sales,
costs, exchange rates
The distinction between quantitative and qualitative
is how the prediction is derived.
Quantitative – the prediction is derived using some
algorithm or mathematical technique based on
quantitative data
Qualitative – the prediction is based primarily on
judgment or opinion 3
Time Series vs Causal
Time Series – These are methods which rely on
the past measurements of the variable of
interest and no other variables, e.g.: moving
average, exponential smoothing, decomposition,
extrapolation
Causal methods – where the prediction of the
target series or variable is linked to other
variables or time series, e.g.: regression,
correlation and leading indicator methods.
4
Sources of Data
Predictions and forecasts are based on relevant
current and past data.
The data sources can be classified into internal
and external;
Internal – sources that come from within the
organisation- eg sales data, employment records,
customer profiles and spending
External – data that is sourced from outside the
organisation e.g.: ABS data, other govt. agencies,
internet, trade organisations, commercial data
agencies. 5
Types of Data
A useful classification of data for forecasting is;
Time Series: A sequence of measurements on a
variable taken over specified successive intervals of
time
e.g.: monthly interest rates, sales/week, tourist
arrivals per annum
Cross-Sectional: Measurements on a variable that
are at one point in time but spread across a
population
e.g.: tourism spend across age groups, production
across sectors of the economy 6
Time Series Example
Consider annual Sales
Sales ($m)
($m) data from 2007 to 2016 140
Year Sales ($m)
120
2007 34
2008 44 100
2009 57
80
2010 67
2011 81
60
2012 89
2013 99 40
2014 112
20
2015 117
2016 125
0
2006 2008 2010 2012 2014 2016 2018
7
Forecasts for 2017 and 2018?
Sales ($m)
140
120
100
80
60
40
20
0
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017
8
Forecasts for 2017 and 2018?
Sales ($m)
140
120
100
80
60
40
20
0
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017
9
Sales, Trend Line and
Forecasts
Sales, Trend and Forecasts
160
140
120
Forecasts
100
2017, 2018
80
60
40
20
0
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Sales ($m) Trend Forecasts
10
Analysing the 2017, 2018 Forecasts
The previous example highlighted a number of key
points;
1. Evaluation of the time series for historical
patterns
2. Matching observed pattern to a relevant
algorithm (in this case a trend line)
3. Projection of the algorithm into the future for
forecasts
11
Exploring Time Series
Patterns
There are various patterns that are typically
associated with time series
These patterns can usually be ascribed to various
components of time series
The systematic components are typically
due to explainable factors
The forecaster needs to understand the
components of the time series to match the
appropriate forecast method or algorithm
12
Components of a Time Series
The components of a time series are:
▪ Level
▪ Trend
▪ Seasonal
▪ Cyclical
▪ Random
The random component is the only non-
systematic component
13
Level
Indicates the underlying value Data: GBP/$A exchange rate
of the series on the vertical For 12 days in January 2017
axis for a given time period.
GBP/$A Exchange Rate
0.6200
The level of the time series
may be constant over time 0.6190
0.6188
0.6184
or may change with the 0.6180 0.6180
0.6178
influence of the other 0.6170
components. 0.6160
0.6162
0.6164
0.6156
0.6153 0.6152
0.6150 0.6149
If the level remains relatively 0.6144
0.6146
constant over the entire time 0.6140
series a horizontal data 0.6130
pattern is observed 0.6120 14
1 2 3 4 5 6 7 8 9 10 11 12
Trend
Tendency for the underlying Data: Number of Credit Card
level of the time series to Accounts (000s) monthly, Jan
systematically increase or 2015 – Oct 2016
decrease from period to CC Accounts (000's)
period 16800
16600
The trend need not be
consistent over the entire 16400
time series or linear.
16200
Trends are usually caused by 16000
population changes,
15800
technology changes,
market expansions etc. 15600 15
Jan-2015
Apr-2015
Jul-2015
Oct-2015
Jan-2016
Apr-2016
Jul-2016
Oct-2016
Jan-2017
Further Trend Example
Data: Passenger Vehicle Sales (Australia), monthly 000’s
Passenger Vehicle Sales (000's)
43500
43000
42500
42000
41500
41000
40500
40000
39500
39000
Dec-2014 Mar-2015 Jun-2015 Sep-2015 Dec-2015 Mar-2016 Jun-2016 Sep-2016 Dec-2016 16
Seasonality
Systematic and repeatable Data: Food Sales ($m), NSW quarterly
fluctuations in the time Jan-2010 to Dec-2016
series that usually occur within Food Sales (m)
a well defined time period 13000.0
(year, week).
12000.0
Fluctuations typically repeat 11000.0
themselves in future
iterations of the set time 10000.0
period 9000.0
Occurs due to weather or 8000.0
institutional reasons e.g.: 7000.0
holidays, special celebrations
or accounting periods 6000.0 17
Dec-2009 Dec-2011 Dec-2013 Dec-2015
Further Seasonal Example
Overseas Visitors
1000000
900000
800000
700000
600000
500000
400000
300000
200000
Dec-05 Dec-06 Dec-07 Dec-08 Dec-09 Dec-10 Dec-11 Dec-12 Dec-13 Dec-14 Dec-15 Dec-16
18
Cyclical
Similar to seasonal fluctuations Data: Non-residential value, Aust.
but the cycle period is not as quarterly Dec-81 to Dec 16
regular as seasonality
Building - Non-Residential
Value
This makes the cyclical 8000000
component difficult to 7000000
predict 6000000
5000000
It is usually subjectively
4000000
assessed
3000000
2000000
Generally the economic
cycle will influence 1000000
the cyclical behaviour of the 0
Dec-15
Dec-01
Dec-03
Dec-05
Dec-07
Dec-09
Dec-81
Dec-83
Dec-85
Dec-89
Dec-91
Dec-93
Dec-95
Dec-97
Dec-99
Dec-87
Dec-11
Dec-13
Dec-17
series. 19
Random
The random component is non-systematic and not able
to be predicted with any accuracy
Typically the random component incorporates effects on the
time series that cannot be explained by the variables
that influence the systematic components
Includes one-off effects such as introduction of GST,
cataclysmic events (e.g.: a tsunami) or difficult to observe
and quantify effects such as confidence and security
The extent of the random component will determine the
maximum level of forecast accuracy achievable
20
