DEMAND

FORECASTING
Basic Time Series Models
 Time Series Forecasting
2
Models
 Time series models predict the future
using past data
 Forecasts are estimated by
extrapolating past data into the future
 Require less data than causal models
 Usually not extraordinarily complex

 Most commonly used techniques

 Time Series Forecasting
3
Models
 Past trends may be a good indication of the
future, but forecasters must be wary
 External factors can have a significant impact on
the variable being forecast
 If something is going to happen in the future that
has not happened in the past, time series
analysis cannot respond to it

 Much like driving a car with the windshield

blocked out, only being able to see through
the rearview mirror
 Developing a time series forecast:
4

1. Identify time series data components

2. Choose appropriate time series methods
 Depends on the types of data
components present in the data
3. Evaluate different methods:
a) Calculate forecasts using historical (past) data
b) Evaluate forecasting errors for each forecast
method
c) Choose method which performs best
 The assumption is -- if it works well on past
data, it will work well in the future
4. Implement method of choice
 1. Identify time series data
components.
5

Demand
a) Random
(Stationary) Time

b) Trend (or
Demand

no trend)
Time
Demand

c) Seasonal
0 1 2 3 4 5 6 7 8 9 10 11 12
Time

e) Cyclical
Demand

Time
 2. Choose appropriate time series
methods.
6

Guide to Selecting Appropriate Time Series Methods:

Time Series Components
Trend and
Stationary Trend Seasonal Seasonality

Simple moving average


Weighted moving average

Simple exponential smoothing

Linear trend

Holt's exponential smoothing

Seasonal decomposition

Winter's exponential smoothing

Time series decomposition

Holt-Winter's exponential smoothing

MODELS FOR
RANDOM PATTERNS
 Example
8

 Candy store owner

 1 customer; rolls dice each day to decide how much to
buy
 Owner has no idea;Candy
wants to forecast demand
Sales
6
5
Demand

4
3
2
1
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Day
Demand

 Knowing what you know, what’s best possible forecast?

 Underlying Level
 Ignore the randomness
 What type of model could produce this forecast for us?
9 Running Average
 Running Average
10

 Forecast for a given period is simply the

average of demand from all previous
periods
 Ex: Forecast for period 10 is the average of
demand in periods 1-9
Candy Sales
6
5
Demand

4
3
2
1
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Day

Demand Simple Average Forecast

 This model quickly filters out the

 Running Average
11

 Problem: What happens if kid changes

rule?
 Now rolls 2 dice instead of 1
Candy Sales
12
11
10
9
8
Demand

7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

Day
Demand

 Level has shifted

 Running Average
12

 Level has shifted, but Running Average

forecast takes a long time to detect and
“catch up” Candy Sales
12
11
10
9
8
Demand

7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

Day
Demand Simple Average Forecast
13 Simple Moving Average
 Simple Moving Average Forecasting
Model
14

 This method works well when the demand is

fairly stable over time
 Random component only
 If trend is present, this model will lag demand
because of averaging effect

 Moving averages attempt estimate the level

while filtering out the random component

 Simple average places the same weight (or

emphasis) on all past demand periods
 Simple Moving Average Forecasting
Model
15

 The n-period moving average forecast is

given by,
t

A i
Ft 1  i t  n 1
n
where:
Ft+1 = forecast for Period t+1
n = number of periods used to
calculate moving
average, and
 Example 5.1: Simple Moving Average
16

Period Demand Forecast

1 1,600 n/a
2 2,200 n/a
3 2,000 n/a
4 1,600 n/a
5 2,500 1,850
6 3,500
7 3,300
8 3,200
9 3,900
10 4,700
11 4,300
12 4,400

Forecast (Period 6) = (2,200 + 2,000 + 1,600 + 2,500)/4 = 2,075

 Example: Candy Store
17

Candy Sales
12
11
10
9
8
Demand

7
6
5
4
3
2
1
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59

Day
Demand Simple Average Forecast 3 Period MA

 Moving Average picks up level shift quickly, but is quite

sensitive to randomness, (i.e. responsive to noise).
 Increasing the number of periods averaged in the forecast,
creates a less responsive forecast
 Example: Candy Store
18

 3 vs. 10 Period MA

Candy Sales
12
11
10
9
8
Demand

7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

Day
Demand Simple Average Forecast 3 Period MA 10 Period MA
 Example of Lag Effect
19

Exchange Rate With Japan Three-Quarter Moving Average Forecast

300.00

250.00

200.00
Yen per US Dollar

150.00

100.00

50.00

0.00
D 5

D 6

D 7

D 8

D 9

D 0

D 1

D 2

D 3

D 4

D 5

D 6

D 7

D 8

D 9

D 0

D 1

D 2

D 3
4
-8

-8

-8

-8

-8

-9

-9

-9

-9

-9

-9

-9

-9

-9

-9

-0

-0

-0

-0

-0
ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec

ec
D

Actual Rate Three-Quarter Moving Average Forecast

This is a problem with any of these “randomness only” models

20 Weighted Moving Average
 Weighted Moving Average
21
Forecasting Model
 The weighted moving average allows the forecaster to
assign weights to past periods
 “Regular” moving average model spreads them out evenly

 The n-period weighted moving average forecast is given by,

t
Ft 1  wA
i t  n 1
i i

where:
Ft+1 = forecast for Period t+1
n = number of periods used to calculate moving average
Ai = Actual demand in Period i
wi = weight assigned to Period i (with Σwi = 1)
 Example 5.2: Weighted Moving
Average
22

weights of 0.4, 0.3, 0.2, and 0.1

Forecast???

F5 = 0.4( 1,600 ) + 0.3( 2,000 ) + 0.2(2,200 ) + 0.1(1,600 )= 1840

 Example: Candy Store
23

Weights: 0.4, 0.3, 0.2, 0.1

Candy Sales
12
11
10
9
8
Demand

7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

Day
Demand 4 Period WMA

 Weighted Moving Average also picks up level shift quickly,

but can sensitive to randomness, (i.e. responsive to noise).
 Increasing the number of periods in forecast and/or decreasing
the size of the weights, creates a less more responsive forecast
24
Simple Exponential
Smoothing
 Simple Exponential
25
Smoothing (SES)
 Like the previous time series models, the SES
model is suitable for data that show no trend
patterns

 Due to its simplicity, exponential smoothing is a

very popular forecasting technique

 Exponential smoothing forecasting is essentially

another weighted moving average forecast model
 AKA: Exponential Weighted Moving Average (EWMA)
 Simple Exponential
Smoothing (SES)
 Forecaster doesn’t need to choose weights
or number of periods to include!!
 Only one parameter to set – the smoothing
•
constant
The exponential smoothing equation places exponentially larger weights on more recent data

• Weight is determined by setting smoothing constant (a) between 0 and 1

Weighting
Time Period Structure a = 0.7 a = 0.2
Most rec ent time period a (1-a )0 0.700 0.200
One time period bac k a (1-a )1 0.210 0.160
T w o time periods bac k a (1-a )2 0.063 0.128
T hree time periods bac k a (1-a )3 0.019 0.102
Four time periods bac k a (1-a )4 0.006 0.082
Five time periods bac k a (1-a )5 0.002 0.066
Six time periods bac k a (1-a )6 0.001 0.052
Seven time periods bac k a (1-a )7 0.000 0.042
Eight time periods bac k a (1-a )8 0.000 0.034
N ine time periods bac k a (1-a )9 0.000 0.027
T en time periods bac k a (1-a )10 0.000 0.021
 Simple Exponential
27
Smoothing (SES)
 The next period’s forecast is simply the current period’s forecast adjusted by a
fraction of the difference between the current period’s actual demand and its
forecast.

Ft 1  Ft   At  Ft 
where:
Ft+1 = forecast for Period t +1
Ft = forecast for Period t
At= Actual demand in Period t, and
a = Smoothing constant

 Each new forecast (Ft+1) implicitly is built upon many past actual demands, each of which
receives less and less weight. This is because each past forecast (Ft) is itself a function of
prior demands
 Simple Exponential
Slide 28
Smoothing (SES)
 Choosing 
 If data has a lot of random variation – select low value of
alpha (close to 0)
 Puts more weight on historical observations
 If you want the forecast to respond greatly to recent
changes or shifts in the data – select a large alpha (close to
1.0)
 Puts more weight on recent observations
 Will chase randomness more however
 The larger the smoothing constant (a), the “less
smooth” the forecast, (or the more responsive it is
to randomness)
 Notes:
 In practice, small values of alpha tend to work best when
SES is the appropriate model to be using
 As a general rule, alpha should be kept between .01
and .30
 Example: Candy Store
Slide 29

Candy Sales
12
11
10
9
Demand

8
7

a = 0.1 6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

Demand SES

Day

Candy Sales
12
11
10
9
Demand

a = 0.6 7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960

Demand SES

Day
 Simple Exponential
30
Smoothing (SES)
 When using the Exponential smoothing method, your
are required to provide an initial forecast for the first
period of data.

 This can be done through numerous methods:

 Qualitative methods, such as the Delphi forecast
 Using an average of some kind
 Simply setting the initial forecast equal to the
demand of that period (naïve forecast).

There is little evidence to suggest that any particular

method of initializing the forecast is the best
 Example: Simple Exponential
Smoothing
Slide 31

Ft +1 = Ft + (At  Ft)
Example using:  = 0.3
Period Demand Forecast
1. 1600 You must initialize
1600
2. 2200
the first forecast!
1600
3. 2000 1780 Set: F1 = A1
4. 1600 1846