Forecasting

??
 Learning outcomes

At the end of the session should be able to:

Explain the 3 time horizons and which models apply for
each
Identify the various forecasting methods available to
customer demand planning systems
– Explain when to use each of the 4 qualitative models
– Apply the naïve, moving average, exponential
smoothing and trend methods
Compute 3 measures of forecast accuracy
Conduct a linear regression and correlation analysis
Develop seasonal indexes
 What is forecasting?

Art and science of predicting a future event

Technique for using past experiences to project
experiences for the future
Underlying basis of all business decisions
Actual or expected demand - starting point for virtually all
planning systems
Production normally begins with expected demand, not
actual demand - a forecast of the demand
Good forecasts are an essential part of efficient service and
manufacturing operations
 Forecasting vs. prediction

Forecasting Prediction
Objective Subjective
Scientific Intuitive
Reproducible Non-reproducible
Free from bias Individual bias
Error analysis Error analysis limited
possible
 Introduction to forecasting
In virtually every decision they make, executives today
consider some kind of forecast
Forecasts are vital to every organization for every
significant management decision
Forecast of demand is an essential input for
planning
 Strategic importance of forecasting
Forecast is the only estimate of demand until actual
demand becomes known
Forecast of demand drives decisions in many areas:
Human resources
Capacity
Supply chain management
Inventory control/production planning
 Forecasting time horizons
Forecast classified by future time horizon it will cover
Short-range forecast
– Up to 1 year, generally less than 3 months
– Used for purchasing, job scheduling, workforce
levels, job assignments, production levels
Medium-range forecast
– 3 months to 3 years
– Used for sales and production planning, budgeting
Long-range forecast
– 3+ years
– Used for new product planning, facility location,
research and development, overall capacity needs
 Distinguishing differences
Medium term and long-run forecasts are distinguished
by 3 features:
Medium and long-range deal with more comprehensive
issues
Short term forecasts employ different methodologies
Short term forecasts tend to be more accurate
 Types of forecasts
Economic forecasts
– Address business cycle – predict inflation rate,
money supply, housing starts, etc.
Technological forecasts
– Predict rate of technological progress
– Impacts development of new products
– Predicts new product sales
Demand forecasts
– Predict demand for company’s existing
products and services
– Also called sales forecasts
Economic and technological fall outside the role of
Ops managers so not covered in course
 Hazards of forecasting

Underestimation
Overestimation
An underestimation of
An overestimation of
demand can result in
demand can result in:
various opportunity costs
Excess resources including:
Excess inventories Lost sales
The need to have Loss of goodwill
discount sales
Damage of company
image
Loss of market share
Loss of revenue presently
and in the future
 7 steps in forecasting
Determine the use of the forecast
– E.g. Disney World uses park attendance forecasts to
drive staffing, opening times, ride availability and food
supplies
Select the items to be forecasted
– Individual products or groups
Determine the time horizon of the forecast
Select the forecasting model(s)
Gather the data needed to make the forecast
Make the forecast
Validate and implement results
– Done by reviewing the forecast to make sure the
model, assumptions and data are valid
 The realities!
Forecasts are seldom perfect – in fact they are almost
always wrong
Most techniques assume an underlying stability in the
system
Product family and aggregated forecasts are more
accurate than individual product forecasts
Forecasts are more accurate for shorter time periods
Forecasts are no substitute for calculated demand
 Forecasting approaches/techniques
2 general approaches: Qualitative and Quantitative
Qualitative
– Generated from information that does not have a well-defined
analytical structure
– Used when situation is vague and little data exist
• New products (so no sales history)
• New technology
– Based on personal judgment, involves intuition, experience,
tends to be subjective e.g., forecasting sales on Internet
– Advantage – allows for some fairly quick results
Quantitative
– Used when situation is ‘stable’ and historical data exist
• Existing products
• Current technology
– Employ one or more mathematical models that rely on historical
data and/or causal variables to forecast demand e.g.,
forecasting sales of color televisions
 Forecasting Models

Forecasting
Techniques

Qualitative Time-Series Causal

Models Methods Methods

Delphi Moving Regression

Methods Average Analysis

Jury of Executive Exponential Multiple

Opinion Smoothing Regression

Sales Force Trend

Composite Projections

Consumer
Market Survey Decomposition
Qualitative approaches to forecasting
A Jury of executive opinion
– Pool opinions of high-level experts, sometimes
augment by statistical models
Delphi method
– Panel of experts, queried iteratively
Sales force composite
– Estimates from individual salespersons are
reviewed for reasonableness, then aggregated
Consumer market survey
– Ask the customer
 1. Jury of executive opinion
Or panel consensus
Involves small group of high-level experts and managers
– Opinions, experience, and technical knowledge of one or
more managers are summarized to arrive at a single
forecast
Group estimates demand by working together
• ‘2 heads are better than 1’ – discussion by group will
provide better forecast than 1 individual
Combines managerial experience with statistical models
Relatively quick
‘Group-think’ disadvantage
Also lower level staff may feel intimated by higher levels of
management
 2. Delphi method
Iterative group process, continues until consensus is
reached
Reduces ‘group-think’
Difference between jury of executive opinion is that in the
latter the experts are brought together in a meeting format
for the discussion
Delphi – individual forecasts made with each expert – each
expert develops own forecast; collective set then shared
with all experts, allowing each to then modify original
forecast. Continues until consensus is reached
 3. Sales force composite

Also called grass roots

Each salesperson projects his or her sales
Combined at district and national levels
Sales reps know customers’ wants
Tends to be overly optimistic – individual bias can
impair accuracy of forecast
 4. Consumer market survey
Ask customers about purchasing plans
Generally structured questionnaires submitted to
potential customers in the market
– Systematic approach used to determine external
customer interest through data-gathering surveys
What consumers say, and what they actually do are
often different
Sometimes difficult to answer e.g. how many hours will
you use the Internet next week?
 Quantitative approaches to
forecasting
2 categories of quantitative forecasting models
Time series analysis and projection - assume future is a
function of the past so use historical data to make a forecast
– Focuses entirely on patterns and pattern changes
– Relies entirely on historical data

Causal (or associative) models - incorporate variables or

factors that might influence the quantity being forecast
– Use highly refined and specific information about
relationships between elements and take special events
formally into account
– Past is important
 Time-series analysis
Most commonly used - predicts on assumption that the future is a
function of the past
Used when several years data for product or product line are
available
One common assumption - past demand follows some pattern,
which if analysed, can be used to develop projections for
future demand (assuming pattern continues in more or less the
same pattern)
Implies that the only real independent variable is time
Time-series - set of evenly spaced numerical data or
chronologically ordered points of raw data
Year 2004 2005 2006 2007
Sales 5,000 6,500 5,500 7,000
Forecast based only on past values, no other variables important
– Assumes that factors influencing past and present will continue
influence in future
 Time-series patterns
The repeated observations of demand for a service or a
product in their order of occurrence form a pattern known
as time-series
Five basic patterns of most demand time series:
– Horizontal – Data cluster about a horizontal line
– Trend - is the level going up or down
• Can be affected by economic conditions, fads,
demographics
– Seasonal - data pattern that repeats itself after a
period of days, weeks, months, or quarters
– Cyclical - data that occur every several years
– Random (noise) - random elements or ‘blips’ in data
caused by chance and unusual situations
 Demand patterns

Horizontal Trend
Data consistently increase or decrease
Data cluster about a horizontal line

Seasonal Cyclical
Data reveal gradual increases and decreases
Data consistently show peaks and valleys over extended periods
 1. Trend pattern
Gradual overall upward or downward movement of
the data over time
– Changes due to population, technology, age,
culture, etc.
Typically several years duration

Linear decreasing Linear increasing

trend trend
 2. Cyclical pattern (cycles)
Repeating up and down movements (that is, some rising
and falling demand) that occur every several years
Affected by business cycle, political, and economic factors
Multiple years duration (usually 2-10 years)
Often causal or associative relationships

0 5 10 15 20
 3. Seasonal pattern
Seasonality is a special type of cyclical pattern
Regular pattern of up and down fluctuations that repeat
after a period of days, weeks, months or quarters
Due to weather, customs, etc.
Occurs within a single year
6 common seasonality patterns
Number of
Period Length ‘seasons’ in pattern
Week Day 7
Month Week 4-4.5
Month Day 28-31
Year Quarter 4
Year Month 12
Year Week 52
 4. Random pattern
Erratic, unsystematic, fluctuations, ‘blips’
Due to random variation or unforeseen events e.g.
strike, tornado
Short duration and non-repeating
Unforecastable – arise from chance and cannot be
predicted

M T W T F
 Quantitative models

1. A Naive approach Time-Series Models

2. Moving averages
Assume future is a function of
3. Exponential the past so use historical data to
smoothing make a forecast
4. Trend projection

5. Linear regression Causal /Associative

Models
Incorporate variables or factors
that might influence the quantity
being forecast
Time-series models
 1. Naive approach
Assumes demand in next period is the same as
demand in most recent period
– e.g. If January sales were 68, then February
sales will be 68
Sometimes cost effective and efficient
Can be good starting point
Works best when horizontal, trend or seasonal
patterns are stable and random variation is small

Ft = A t – 1
 2. Moving averages
Used if little or no trend, seasonal or cyclical patterns
Horizontal pattern in time series is based on the mean
(average) of the demands
MA involves calculating the average demand for the n most
recent time periods and using it as a forecast for the next
time period
For the next time period, after the demand is known, the
oldest demand from the previous average is replaced with
the most recent demand and the average is recalculated
In this way, the n most recent demands are used, the
average ‘moves’ from period to period
 i. Simple moving average method
Used if little or no trend, seasonal or cyclical patterns
Simple moving average is a series of arithmetic averages
Uses a number of historical actual data values to generate
forecast
Nothing more than the mathematical average of the last
several periods of actual average – the MA serves as an
estimate of the next period’s demand

∑ demand in previous n periods

Moving average = n
where n= number of periods in the moving average
So for a 5-period moving average, n=5
Example: Simple moving average
Data for 100-room occupancy. We want to forecast
Saturday occupancy only. 3-period moving average

Sat Periods Occupancy

Aug 1 1 79
8 2 84
15 3 83
22 4 81
29 5 98
Sep 5 6 100
12 7 ?
 Example: Simple moving average
Each three-period moving-average forecast thus involves simply adding
the three most recent occupancy values and dividing by 3.

Actual 3-Period
Period Occupancy Moving Average

1 79
2 84
3 83
4 81 (79 + 84 + 83)/3 = 82
5 98 (84 + 83 + 81)/3 = 83
6 100 (83 + 81 + 98)/3 = 83
7 ? (81 + 98 + 100)/3 = 93
 Exercise 1: Simple moving average
Estimate simple moving average using the following
customer-arrival data

Month Customer arrival

1 800
2 740
3 810
4 790

Use a three-month moving average to forecast customer

arrivals for month 5
If the actual number of arrivals in month 5 is 805, what is
the forecast for month 6?
 ii. Weighted moving average
Used when a detectable pattern is present, weights can be
used to place more emphasis on recent values
– Weights used to place more emphasis on recent values
– Older data usually less important
SMA - each demand has same weight – 1/n
More responsive to changes in the underlying averages
because more recent periods are more heavily weighted
Choice of weights arbitrary
– Weights based on experience and intuition
– Deciding which weights to use requires some experience

∑ (weight for period n)

Weighted x (demand in period n)
=
moving average ∑ weights
Example: Weighted moving average
Data for 100-room occupancy. We want to forecast Saturday
occupancy only.

Sat Periods Occupanc

y
Aug 1 1 79
8 2 84
15 3 83
22 4 81
29 5 98
Sep 5 6 100
12 7 ?
Weights Applied Period
3 Last month
2 Two months ago
1 Three months ago
6 Sum of weights
 Example: Weighted moving average

Actual 3-Period Weighted

Period Occupancy Moving Average

1 79
2 84
3 83
4 81 [(3 x 83) + (2 x 84) + (79)]/6 = 83
5 98 [(3 x 81) + (2 x 83) + (84)]/6 = 82
6 100 [(3 x 98) + (2 x 81) + (83)]/6 = 90
7 ? [(3 x 100) + (2 x 98) + (81)]/6 =
96

In this example, more heavily weighting latest month – more accurate

projection
 Exercise 2: wMA
What is the forecast for March, April and May, based on a
weighted moving average applied to the following past
demand data and using the weights: 4, 3, 2 (largest weight
is for most recent data)?

Nov. Dec. Jan. Feb. Mar. April

37 36 40 42 47 43
 Possible problems with moving
averages
3 problems:
Increasing n smooths the forecast but makes it less
sensitive to changes
– Forecast line is smoother than demand line, showing
impact of taking an average
– More periods used in computing moving average,
smoother effect (graph on previous page)
Do not forecast trends well
– Will always lag behind any actual demand
– Both MA and WMA lag the demand from April
onwards (previous graph)
Require extensive historical data
 Impact of period lengths (n)
Longer moving average periods – more random elements or fluctuations
are smoothed (may be desirable in many cases)

3-week MA 6-week MA
450 — forecast
forecast
430 —
Patient arrivals

410 —

390 —
Actual patient
370 — arrivals
| | | | | |
0 5 10 15 20 25 30
Week
 3. Exponential smoothing
Major drawback of moving averages – need to continually carry
large amount of historical data
As each new piece of data is added, oldest observation is
dropped and new forecast calculated
In most instances most recent occurrences are more indicative
of the future than those in distant past
If premise is valid – importance of data diminishes as past
becomes more distant
Exponential smoothing – most used of all forecasting techniques
Requires only 3 items of data
– Last period’s forecast (most recent forecast)
– Actual demand for last period
– Smoothing constant, alpha (α), where 0 ≤ α ≤ 1.0
So look at error in last forecast and take a fraction of that error
and adjust that forecast to get new forecast
 Exponential smoothing
Smoothing constant, alpha, (), chosen by the forecaster
– Ranges from 0 to 1; usually between 0.05 and 0.5 for
businesses
– Subjectively chosen
Fairly easy to use and accurate
Involves little record keeping of past data
Concept fairly simple – latest estimate of demand is equal to old
estimate adjusted by a fraction of the difference between last
period’s actual demand and the old estimate
So method assigns exponentially decreasing weights as data gets
older – recent data are given relatively more weight than older data
Called exponential smoothing because each increment in the past is
decreased by (1 - ) - decreasing weight as observations get older
(since future more dependent on recent past than on distant past)
 Exponential smoothing
New forecast = Last period’s forecast + a (Last period’s actual
demand – Last period’s forecast)

Ft = Ft – 1 + a(At – 1 - Ft – 1)

Ft = a At – 1 + (1– a) Ft – 1

where Ft = new forecast

Ft – 1 = previous forecast
a = smoothing (or weighting)
constant (0 ≤ a ≤ 1)
 Example: Exponential smoothing
 

Sat Periods Occupancy

Aug 1 1 79
8 2 84
15 3 83
22 4 81
29 5 98
Sep 5 6 100
12 7 ?
 Example: Exponential smoothing
 

Sat Periods Occupancy Forecast

Aug 1 1 79 79
8 2 84 79
15 3 83 81.5
22 4 81 82.25
29 5 98 81.63
Sep 5 6 100 89.81
12 7 ? 94.91
 Choosing appropriate value for 
Determines level of smoothing and the speed of reaction to
differences between forecasts and actual occurrences
Where  = 0, forecast is always the same number (F t = F t – 1)
Where  = 1, forecast is naïve forecast (F t = A t – 1)
 close to 1 (larger value) - gives greater weighting to recent
levels of demand
 closer to 0 (small value) – gives more weight to past data -
greater smoothing effect
– Little weight placed on recent demand and treat past
demand more uniformly
– Less responsive to recent changes
 Effects of smoothing constants
Weight Assigned to
Most 2nd Most 3rd Most 4th Most 5th Most
Recent Recent Recent Recent Recent
Smoothing Period Period Period Period Period
Constant (a) a(1 - a) a(1 - a)2 a(1 - a)3 a(1 - a)4

a = .1 .1 .09 .081 .073 .066

a = .5 .5 .25 .125 .063 .031

 =0.5 – new forecast based almost entirely on demand in

the last 3 or 4 periods (total = 0.938 for 4 period)
 =0.1 – little weight placed on recent demand (adding up
all these still not close to 1 so will need to take more older
periods into account)
 Exponential smoothing

0    1
weight
Decreasing weight given 
to older observations
(1  )
(1  ) 2
(1  ) 3

today
 Impact of different 
Higher  (0.5) - more responsive – more closely follows
actual demand
Lower  (0.1) – smoother (smooth out fluctuations, which
is desirable in some instances)

225 –
Actual a = .5
demand
200 –
Demand

175 –
a = .1
| | | | | | | | |
150 1– 2 3 4 5 6 7 8 9
Quarter
 Exponential smoothing and moving
average

3-week MA 6-week MA
450 –
forecast forecast
430 –

410 –
Patient arrivals

390 –

370 –
Exponential
smoothing
 = 0.10
| | | | | |
0 5 10 15 20 25 30
Week
 Exponential smoothing
Approach easy to use and has been successfully applied
to virtually every type of business
Simplicity also a disadvantage when underlying average
is changing – in case of demand time series with trend
Smoothing constant – difference between an accurate
and an inaccurate forecast
Higher  values may help reduce forecast error but if
large  values are required (e.g. >0.50) chances are
good that another model is needed because of a
significant trend or seasonal influence in demand series
 Exercise 3: Exponential smoothing

a. Using α = 0.10, calculate the exponential smoothing

forecast for week 4. Assuming forecast for week 3 is
398.

b. If the actual demand turned out to be 415, what is

the forecast for week 5?

Week Patient Arrivals

1 400
2 380
3 411
4 415
 Choosing a time-series method

Manager must consider several factors in selecting a

forecasting technique for a particular time-series
One important consideration is forecast performance –
determined by forecast errors
Managers need to know how to measure forecast errors
and how to detect when something is going wrong with
the forecasting system
 Measures of forecast error
Forecast error = Actual demand - Forecast value
=At–Ft
Popular measures of the forecast error: MAD, MSE

1. Mean Absolute Deviation (MAD)

∑ |Actual - Forecast|
MAD =
n

2. Mean Squared Error (MSE)

∑ (Forecast Errors)2
MSE =
n
 Example: Determining MAD & MSE
Using α = 0.10, find MAD and MSE

Week Patient Arrivals

1 400
2 380
3 411
4 415
 Solution: Determining MAD & MSE

Patient Error
Week Forecast Error |(A-F)|
Arrivals Squared
1 400 400 400-400 = 0 0
2 380 400 380-400 = 20 400
3 411 398 411-398=13 169
4 415 399.3 415-399.3=15.7 246.49
Total 48.7 815.49
MAD 48.7/4=12.175
MSE 815.49/4=203.87
 Exercise 4: MAD & MSE
Using α = 0.50, find MAD and MSE

Week Patient Arrivals

1 400
2 380
3 411
4 415
 Comparing forecasting models
Using α = 0.10 and exponential smoothing technique, find MAD

∑ |forecast errors| Error

WeekMADPatient
= Forecast Error |(A-F)|
Arrivals n Squared
1 400 400 0 0
2 For a =380
.10 400 20 400
3 411 390 21 441
= 12.175
4 415 400.5 14.5 210.25
Total
For a = .50 55.5 1051.25
MAD 13.875
= 13.875
MSE 262.81

Is a MAD of 12.175 good or bad?

Smoothing constant of 0.1 is preferred – smaller MAD
 MAPE
Mean absolute percent error (MAPE)
– Average of the absolute diff between the forecasted
and actual values, expressed as a percentage of
actual values
MAD and MSE – values depend on magnitude of item
being forecast
– If forecast item is measured in thousands, MAD and
MSE can be very large

3. Mean Absolute Percent Error

(MAPE)
n

MAPE =
∑ 100|Actuali - Forecasti|/Actuali
i=1
n
 Example: Determining MAPE
Using α = 0.50 and exponential smoothing technique,
find MAPE

Patient
Week Forecast Error |(A-F)| Abs % error
Arrivals
1 400 400 0 100*0 = 0
2 380 400 20 100* 20/380 = 5.26
3 411 390 21 100* 21/411 = 5.11
4 415 400.5 14.5 100* 14.5/415 = 3.49
Total 13.87
MAPE =13.87/4 = 3.47%
 Criteria for selecting time-series
methods

Forecast errors provide important information for choosing

best forecasting method for service/product
Also guide managers in selecting best values for parameters
needed for the method
– n for moving average method
– weights for weighted moving average
–  for exponential smoothing
2 criteria to use in making forecast method and parameter
choices:
– Minimising MAD, MSE and MAPE
– Minimising forecast error last period
 4. Trend projections
Time-series technique fits a trend line to a series of historical data
points and then projects line into the future for medium to long-
range planning
Several trend equations like quadratic but looking as linear here
Linear trends can be found using the least squares technique
Least squares line described by y-intercept (height at which it
intercepts the y-axis) and slope
Expressed using equation:

^ = a + bx
y
^ where y = computed value of the
variable to be predicted (dependent
variable)
a = y-axis intercept
b = slope of the regression line
x = the independent variable (in this
 Least squares method
Least-squares method – finds a best fitting straight line (line of best fit)
Asterisks are actual observations; then fit a trend line to these values
Values of Dependent Variable

Actual observation Deviation7

(y value)
Deviation5 Deviation6

Deviation3

Deviation4

Deviation1
Deviation2
Trend line, ^y = a + bx
(error)

Time period
 Least squares method

Actual observation
Values of Dependent Variable

Deviation7
(y value)

Deviation5 Deviation6

Deviation3
Least squares method minimises
the Deviation
sum of4 the squared errors
(deviations)
Deviation1
(error) Deviation2
Trend line, y^= a + bx

Time period
 Least squares method
Equations to calculate values for a and b

^ a = y-axis intercept
y = a + bx
b=slope of line
x = known value of independent variable (in
Sxy - nxy this case time)
b=
Sx2 - nx2 y= known value of dependent variable
(variable to be predicted)
n=no of observations
a = y - bx x bar = average of the value of the x’s
y bar = average of the value of the y’s
 Example: Least squares method
Time Electrical Power
Year Period (x) Demand
2001 1 74
2002 2 79
2003 3 80
2004 4 90
2005 5 105
2006 6 142
2007 7 122
2008 ?
2009 ?
 Example: Least squares method
Time Electrical Power
Year Period (x) Demand (y) x2 xy
2001 1 74 1 74
2002 2 79 4 158
2003 3 80 9 240
2004 4 90 16 360
2005 5 105 25 525
2006 6 142 36 852
2007 7 122 49 854
2008 ?
2009 ?
 Solution: Least squares method
Time Electrical Power
Year Period (x) Demand x2 xy
2001 1 74 1 74
2002 2 79 4 158
2003 3 80 9 240
2004 4 90 16 360
2005 5 105 25 525
2006 6 142 36 852
2007 7 122 49 854
∑x = 28 ∑y = 692 ∑x2 = 140 ∑xy = 3,063
x=4 y = 98.86
=28/7 =692/7

∑xy - nxy 3,063 - (7)(4)(98.86)

b= = = 10.54
∑x – nx
2 2 140 - (7)(4 )
2

a = y - bx = 98.86 - 10.54(4) = 56.70

 Solution: Least squares method
Time Electrical Power
Year Period (x) Demand x2 xy
2001 1 74 1 74
2002 2 79 4 158
2003 3 80 9 240
2004 4 90 16 360
The trend line
2005 5 is 105 25 525
2005 6 142 36 852
2007 ^
y = 756.70 + 10.54x122 49 854
∑x = 28 ∑y = 692 ∑x2 = 140 ∑xy = 3,063
x=4 y = 98.86

∑xy - nxy 3,063 - (7)(4)(98.86)

b= = = 10.54
∑x - nx
2 2 140 - (7)(4 )
2

a = y - bx = 98.86 - 10.54(4) = 56.70

 Solution: Least squares method

Trend line,
y^ = 56.70 + 10.54x
160 –
150 –
140 –
Power demand

130 –
120 –
110 –
Demand for 2008 (estimate)
100 –
= 56.70 + 10.54 (8) = 141 MW
90 –
80 – Demand for 2009 (estimate)
70 – = 56.70 + 10.54 (9) = 152 MW
60 –
50 –
| | | | | | | | |
2001 2002 2003 2004 2005 2006 2007 2008 2009
Year
 Requirements for least squares
Must meet 3 requirements in order to use least squares:
Always plot the data to ensure a linear relationship
Do not predict time periods far beyond the database
– e.g. cannot take 5 years sales data and project 10
years into the future – world is too uncertain
Deviations around the least squares line are assumed
to be random
– Normally distributed – most observations close to line
and a smaller number farther out
 Associative forecasting
Time-series forecasting – time is the only independent variable
considered
Associative forecasting – considers several variables related to
quantity being predicted
Also called causal method/ causal relationship forecasting -
used when historical data are available and relationship between
factor to be forecasted and other factors can be identified
Used when changes in one or more independent variables can be
used to predict the changes in the dependent variable
– Most common associative forecasting model is linear
regression analysis
One problem with this method is that we need a forecast for x – the
independent variable – before estimating the dependent variable
and it can be difficult for some variables such as GDP,
unemployment
 Linear regression
One variable (dependent variable) is related to one or more
independent variables by a linear equation
Dependent variable (e.g. demand for doorknobs) – one
manager wants to forecast
Independent variables (such as advertising expenditure
and new housing data) – assumed to affect the dependent
variable and thereby ‘cause’ results observed in the past
 Linear regression analysis

Using same mathematical model as in least-squares

‘x’ is no longer only time
^
y = a + bx
^ where y = computed value of the
variable to be predicted (dependent
variable)
a = y-axis intercept
b = slope of the regression line
x = the independent variable thought to
predict the value of the dependent
variable
 Example: Linear regression
Nobel Construction Company renovates old homes in Arima,
Trinidad. Over time, the company has found that its dollar volume of
work is dependent on the Arima area payroll. Table below lists
Noble’s revenues and amount of money earned by Arima workers
during last 6 years. Management wants to establish a mathematical
relationship to help predict sales.

Sales Local
Payroll
($ millions), y ($
billions), x
2.0 1
3.0 3
2.5 4
2.0 2
2.0 1
3.5 7
 Solution: Linear regression
First: Determine whether there is a linear relationship between Arima
payroll (independent) and sales (dependent variable)
Then use least squares criterion

4.0 –

3.0 –
Sales

2.0 –

1.0 –
| | | | | | |
0 1 2 3 4 5 6 7
Area payroll
 Solution: Linear regression

Sales, y Payroll, x x2 xy
2.0 1 1 2.0
3.0 3 9 9.0
2.5 4 16 10.0
2.0 2 4 4.0
2.0 1 1 2.0
3.5 7 49 24.5
∑y = 15.0 ∑x = 18 ∑x2 = 80 ∑xy = 51.5

x = ∑x/6 = 18/6 = 3 y = ∑y/6 = 15/6 = 2.5

∑xy - nxy 51.5 - (6)(3)(2.5)

b= = = .25
∑x2 - nx2 80 - (6)(32)

a = y - bx = 2.5 - (.25)(3) = 1.75

 Solution: Linear regression
^
y = 1.75 + .25x Sales = 1.75 + .25(payroll)

If payroll next year is

estimated to be $6 4.0 –
billion, then: 3.25

Sales
3.0 –

Sales = 1.75 + .25(6) 2.0 –

Sales = $3,250,000
1.0 –
| | | | | | |
0 1 2 3 4 5 6 7
Area payroll
 Measures of accuracy
Common measures reported:
– Correlation coefficient, r
– Coefficient of determination
 Correlation coefficient, r
How strong is the linear relationship between the variables?
Correlation does not necessarily imply causality – so regression lines are
not cause-effect relationships
Another way to evaluate relationship between 2 variables is to compute
the coefficient of correlation
Coefficient of correlation, r, measures degree or strength of linear
relationship
– Values range from -1 to +1
– r = +1: implies that changes in direction (increase or decrease) of
independent variable are always accompanied by changes in same
direction by dependent variable
– r = -1: means that decreases in independent variable are always
accompanied by increases in dependent variable, and vice versa

nSxy - SxSy
r=
[nSx2 - (Sx)2][nSy2 - (Sy)2]
 Correlation coefficient
y y

(a) Perfect x (b) Positive x

positive correlation:
correlation: 0<r<1
r = +1

y y

(c) No x (d) Perfect negative x

correlation: correlation:
r=0 r = -1
 Example: Correlation
For Nobel Construction

Sales, y Payroll, x x2 xy
2.0 1 1 2.0
3.0 3 9 9.0
2.5 4 16 10.0
2.0 2 4 4.0
2.0 1 1 2.0
3.5 7 49 24.5
∑y = 15.0 ∑x = 18 ∑x2 = 80 ∑xy = 51.5

(6)(51.5)  (18)(15) =39/43.3 = 0.901

r
2 2
[( 6)(80)  (18) ][( 6)(39.5)  (15)
r=0.91 appears to be a strong positive relationship confirm the
closeness of the relationship between the two variables meaning as
payroll went up Nobel sales went up
What would r = -0.91 tell you?
 Coefficient of determination
Other measures exist – such as coefficient of
determination, r2
– Square of the coefficient of correlation
Will always be positive and range from 0 to 1
Measures the percent of variation in the dependent
variable y that is explained by the regression equation
– Higher coefficient of determination – better the variance that the
dependent variable is explained by the independent variable
In Nobel Construction - r2 = 0.81
– 81% of the total variation in Nobel’s sales can be explained by the
linear relationship between sales and Arima’s payroll
– Measure is used by decision makers to indicate how well the linear
regression line fits the data points
– Better fit closer r2 will be to 1
 Exercise 5
The sales manager of a large apartment rental complex feels the
demand for apartments may be related to the number of newspaper ads
placed during the previous month. She has collected the data shown in
the accompanying table.
Ads purchased Apartments leased • Find a mathematical equation
by using the least squares
regression approach.
15  6 • If the number of ads is 30, we
 9  4 estimate the number of
40 16 apartments leased using the
regression equation
20  6 • Compute coefficients of
25 13 determination and correlation
25  9
15 10
35 16
 Multiple Regression Analysis
If more than one independent variable is to be used in the model,
linear regression can be extended to multiple regression to
accommodate several independent variables
– E.g. using Nobel Construction – say we want to include
annual interest rates in the model, in addition to payroll

^
y = a + b1x1 + b2x2 …

Computationally, this is quite complex and generally done on the

computer
 Seasonal variations in data
Seasonal variations are regular up and down movements in a
time series – relate to recurring events such as weather or
holidays – so pattern we see repeating during the year
– E.g. demand for suntan lotion – highest in summer
Seasonality may be applied to hourly, daily, weekly, monthly
recurring patterns
To identify seasonal component – compare same period year
to year
Understanding seasonal variations is important for capacity
planning – to handle peak loads
The multiplicative seasonal model can adjust trend data for
seasonal variations in demand
– Seasonal factors are multiplied by an estimate of average
demand to produce a seasonal forecast
 Seasonal variations in data
Seasonal factor/index is the amount of correction needed
in a time series to adjust for the season of the year
– Indicates how a particular season compares with an
average season
– An index of 1 indicates an average season
– An index > 1 indicates the season is higher than
average
– An index < 1 indicates a season lower than average
 Seasonal variations in data
Steps in the process
1. Find average historical demand for each season
2. Compute the average demand over all seasons
3. Compute a seasonal index for each season
4. Estimate next year’s total demand
5. Divide this estimate of total demand by the number
of seasons, then multiply it by the seasonal index for
that season. This provides the seasonal forecast.
 Example: Seasonal index
Past Demand
Month 2005 2006 2007
Jan 80 85 105
Feb 70 85 85
Mar 80 93 82
Apr 90 95 115
May 113 125 131
Jun 110 115 120
Jul 100 102 113
Aug 88 102 110
Sept 85 90 95
Oct 77 78 85
Nov 75 72 83
Dec 82 78 80
 Example: Seasonal index
Demand Seasonal average Average Seasonal
Month 2005 2006 2007 2005-2007 Monthly Index
Jan 80 85 105 90 94 Step 1: Average (Jan)
Feb 70 85 85 80 94 = (80+85+105)/3
Mar 80 93 82 85 94
Apr 90 95 115 100 94 Step 2: Average for
May 113 125 131 123 94 all months= total/12
Jun 110 115 120 115 94 = 1128/12=94
Jul 100 102 113 105 94
Aug 88 102 110 100 94
Sept 85 90 95 90 94
Oct 77 78 85 80 94
Nov 75 72 83 80 94
Dec 82 78 80 80 94

Total average demand = 1128

 Example: Seasonal index
Past Demand Average Average Seasonal
Month 2005 2006 2007 2005-2007 monthly demand Index
Jan 80 85 105 90 94 0.957
Feb 70 85 85 80 94
Mar 80 93 82 85 94
average 2005-2007 monthly demand
Apr Seasonal index =
90 95 115 100 94
average monthly demand
May 113 125 131 123 94
Jun 110 115 120 115
= 90/94 = .957 94
Jul Step
1003 102 113 105 94
Aug 88 102 110 100 94
Sept 85 90 95 90 94
Oct 77 78 85 80 94
Nov 75 72 83 80 94
Dec 82 78 80 80 94
 Example: Seasonal index
Demand Average Average Seasonal
Month 2005 2006 2007 2005-2007 Monthly Index
Jan 80 85 105 90 94 0.957
Feb 70 85 85 80 94 0.851
Mar 80 93 82 85 94 0.904
Apr 90 95 115 100 94 1.064
May 113 125 131 123 94 1.309
Jun 110 115 120 115 94 1.223
Jul 100 102 113 105 94 1.117
Aug 88 102 110 100 94 1.064
Sept 85 90 95 90 94 0.957
Oct 77 78 85 80 94 0.851
Nov 75 72 83 80 94 0.851
Dec 82 78 80 80 94 0.851
 Example: Seasonal index
Demand Average Average Seasonal
Month 2005 2006 2007 2005-2007 Monthly Index
Jan 80 85 105 90 94 0.957
Feb 70 85 85 80 94 0.851
Steps 4&5
Mar 80 93 82 85 94 0.904
Forecast for 2008
Apr 90 95 115 100 94 1.064
May Assuming
113 125 expected
131 annual demand123 = 1,200, 94 1.309
Jun use seasonality
110 115 120indices to forecast
115 monthly 94 1.223
Jul demands
100 102 113 105 94 1.117
Aug 88 102 110 1,200 100 94 1.064
Sept 85 Jan
90 95 12 x .957
90 = 96 94 0.957
Oct 77 78 85 80 94 0.851
Nov 75 72 83 1,200 80 94 0.851
Dec 82 Feb
78 80 12 x .851
80 = 85 94 0.851

Without seasonality average sales would be 94; with seasonality fluctuate

from 85 to 131 in May
 Example: Seasonal index
Demand Average Average Forecast
Month 2005 2006 2007 2005-2007 Monthly demand
Jan 80 85 105 90 94 96
Feb 70 85 85 80 94 85
Mar 80 93 82 85 94 90
Apr 90 95 115 100 94 106
May 113 125 131 123 94 131
Jun 110 115 120 115 94 122
Jul 100 102 113 105 94 112
Aug 88 102 110 100 94 106
Sept 85 90 95 90 94 96
Oct 77 78 85 80 94 85
Nov 75 72 83 80 94 85
Dec 82 78 80 80 94 85

So sales fluctuate from 85 to 131 – notice the huge difference if we

had just used average of 120
 Exercise 6: Multiplicative seasonal
method
The manager of the Stanley Steemer carpet cleaning company needs
a quarterly forecast of the number of customers expected next year.
The carpet cleaning business is seasonal, with a peak in the third
quarter and a trough in the first quarter. Following are the quarterly
demand data from the past 4 years:

Quarter Year 1 Year 2 Year 3 Year 4

1 45 70 100 100
2 335 370 585 725
3 520 590 830 1160
4 100 170 285 215

The manager wants to forecast customer demand for each quarter of

year 5, based on an estimate of total year 5 demand of 2,600
customers