Operations Management

Previous Class (#7): Safety Stock and Newsvendor Model

» Why is safety stock important? How much of it to have?
» How does one perform marginal analysis for a product with one
selling period and uncertain (random) demand.
» Newsvendor model (minimize costs, maximize profit) and single-leg
revenue management (maximize revenue, maximize profit).
» Key Concepts: Overage cost, underage cost, critical fractile,
marginal analysis, newsvendor, single-leg revenue management.

Today’s Class (#8): Demand Forecasting

» Perform simple time series analysis: Simple moving average
(SMA), simple exponential smoothing (SES), linear trend line,
simple regression, and multiple regression
» Key Concepts: Forecast, trends, time series analysis, mean
absolute deviation, panel consensus, Delphi method.

1
Housekeeping Issues
 Final exam conflict declaration deadline
– Thursday, Dec 10th at 11:59pm by email

 Assignments
– Assignment #3 is due Nov. 25th, 11:59pm (submit at
https://utormat.io/#/login?r)

 Questions from last class?

2
Learning Objectives
 Strategic Role of Forecasting in Supply Chain
Management

 Components of Forecasting Demand

 Time Series Methods

 Regression Methods

 Forecast Accuracy
12-3
Forecasting
 Predicting the future

 Qualitative forecast methods

– Subjective

 Quantitative forecast methods

– Objective: based on mathematical formulas

12-4
Forecasting principles
1. Forecasts are always wrong and should
always include some measure of error
2. The longer the horizon, the larger the error
3. Method should be chosen based on need and
context
Value

5
Now Time
The Effect of Inaccurate
Forecasting in Supply Chains

12-6
Forecasting
 Quality Management
– Accurately forecasting customer demand is a key to
providing good quality service (timely service)

 Strategic Planning
– Successful strategic planning requires accurate
forecasts of future products and markets

12-7
Types of Forecasting Methods
 Time frame
» Short-range (daily, weekly, monthly, semi-annually)
• Taco Bell: Every 15 minutes
» Mid-range (annually, bi-annually)
 Hewlett-Packard 12-18 months
» Long-range (up to 20 years)
 Fiat auto company: 5 years
 Bombardier: 20 years

 Demand behavior

 Causes of behavior
12-8
Demand Behavior
 Trend
– a gradual, long-term up or down movement of
demand

 Seasonal pattern
– an up-and-down repetitive movement in demand
occurring periodically

 Cycle
– an up-and-down repetitive movement in demand

 Random variations
– movements in demand that do not follow a pattern
12-9
Forms of Forecast Movement

Demand
Demand

Random
movement

Time Time
(a) Trend (b) Cycle

Demand
Demand

Time Time
(c) Seasonal pattern (d) Trend with seasonal pattern
12-10
Forecasting Methods
 Qualitative
– use management judgment, expertise, and opinion to
predict future demand

 Time series
– statistical techniques that use historical demand data
to predict future demand

 Regression methods
– attempt to develop a mathematical relationship
between demand and factors that cause its behavior 12-11
Qualitative Methods
 Management, marketing, purchasing, and engineering are
sources for internal qualitative forecasts
 Delphi method
1. Choose experts to participate representing a variety of points of view
2. Obtain forecasts (and reasoning) from all participants
3. Summarize the results and redistribute them to the participants
along with appropriate new questions
4. Summarize again, refining forecasts and conditions, and again
develop new questions to distribute to all participants
5. Repeat Step 4 as necessary and distribute the final results

12-12
 Time Series Methods
 Assume that what has occurred in the past will
continue to occur in the future
 Relate the forecast to only one factor: time
 Time series techniques include
– Moving average
– Weighted moving average
– Exponential smoothing
– Linear trend line
 Linear regression: predictors can be time and others
» Simple predictor
» Multiple predictors

12-13
Moving Average
 Naive forecast
– demand in current period is used as next period’s
forecast

 Simple moving average

– uses average demand for a fixed sequence of periods
– stable demand with no pronounced behavioral
patterns

 Weighted moving average 12-14

Moving Average: Naïve Approach
Month Orders per month Forecast
January 120 ---
February 90 120
March 100 90
April 75 100
May 150 75
June 50 110
July 75 50
August 130 75
September 110 130
October 90 110
November --- 90

12-15
 Simple Moving Average

𝑛
  𝐷𝑖
∑
𝑖=1
𝑀 𝐴 𝑛=
𝑛

12-16
 3-month Simple Moving Average
3
Month Orders per Forecast   ∑ 𝐷𝑖
month 𝑖=1 120 +90 +100
𝑀 𝐴 3= = = 103.3
3 3
Jan. 120 ---
Feb. 90 ---
March 100 ---
April 75 103.3
3

May 150 88.3   ∑ 𝐷𝑖

𝑖=1 90 +100 +75
𝑀 𝐴 3= = = 88.3
June 50 95.0 3 3

July 75 78.3
Aug. 130 78.3
Sep. 110 85.0
Oct. 90 105.0 3

  ∑ 𝐷𝑖
𝑖=1 130 +110+ 90
Nov. --- 110.0 𝑀 𝐴 3=
3
=
3
=110.0

12-17
 5-month Simple Moving Average
5
Month Orders per Forecast   ∑ 𝐷𝑖 120 +90 +100 +75 +150
month
𝑖=1
𝑀 𝐴 5 = = = 99.0
5 5

Jan. 120 ---

Feb. 90 ---
March 100 ---
April 75 ---
May 150 ---
June 50 99.0
July 75 85.0
Aug. 130 82.0
Sep. 110 88.0
Oct. 90 95.0
5

 
𝑀 𝐴 =
∑ 𝐷𝑖
𝑖=1
=
50 +75 +130 +110+ 90
=91.0
Nov. --- 91.0 5
5 5

12-18
Smoothing Effects
Demand MA-3 MA-5
140

120

100

80
ORDERS

60

40

20

0
Jan. Feb. M a r. A p r. May Jun. J u ly Aug. Sep. O c t. Nov.
MONTHS

12-19
Time Series Analysis:
How many periods should be used?
 Advantages of more periods:
– More data points give a better estimate
– The effect of randomness is reduced (smoothed) by
averaging together a number of observations (smoothing
effect)
» When there is no trend in the data, using more observations
results in a forecast with lower error

 Disadvantages of more periods:

– A large number of observations will cause the moving
average to respond slowly to permanent changes
» When there is a trend in the data, using more observations
results in a forecast with higher error

20
Strictly Increasing Time Series

Demand MA-3 MA-5

950

900

850

800
ORDERS

750

700

650

600

550
Ja n . F e b . Ma r. A p r. Ma y Ju n . Ju l y A u g . S e p . Oct. N o v.

MONTHS

The 3-week simple moving average method more closely

follows demand when there is a clear trend.

21
Weighted Moving Average
 Adjusts moving average method to more closely
reflect data fluctuations
 

 Advantage: Flexibility
 Disadvantage: Need to decide all the weights
 Remedy: Exponential smoothing

12-22
Weighted Moving Average

Month Weight Data

August 0.17 130
September 0.33 110
October 0.50 90

November 103.4

𝑊
  𝑀 𝐴 3=(0.50)(90)+(0.33)(110)+(0.17)(130)= 103.4  orders

12-23
Exponential Smoothing
 Averaging method

 Weights most recent data more strongly

 Reacts more to recent changes

 Widely used, accurate method

12-24
 Exponential Smoothing
 
𝑭𝒕 +𝟏= 𝜶 𝑫 𝒕 + ( 𝟏 − 𝜶 ) 𝑭 𝒕
 

 Exponential smoothing method will be the same

as naïve method when .

12-25
 Effect of Smoothing Constant

Naïve method
 Chose high values of  12-26
 Exponential Smoothing Method
Period Month Demand Assume: F1 = D1
1 Jan. 37
2 Feb. 40 𝑭 𝟐(𝐅𝐞𝐛.)=𝜶𝑫𝟏(𝐉𝐚𝐧.)+( 𝟏−𝜶 ) 𝑭𝟏(𝐉𝐚𝐧.)=𝟎.𝟑 ( 𝟑𝟕) +( 𝟏 −𝟎.𝟑) 𝟑𝟕=𝟑𝟕
3 March 41
4
5
April
May
37
45
𝑭 𝟑(𝐌𝐚𝐫.)=𝜶𝑫𝟐(𝐅𝐞𝐛.)+(𝟏 −𝜶 ) 𝑭𝟐(𝐅𝐞𝐛.)=𝟎.𝟑 (𝟒𝟎)+(𝟏 −𝟎.𝟑)𝟑𝟕=𝟑𝟕.𝟗
6 June 50
7 July 43
8 Aug. 47
9 Sep. 56
10 Oct. 52
11
12
Nov.
Dec.
55
54
𝑭 𝟏𝟑(𝐉𝐚𝐧.)=𝜶𝑫𝟏𝟐(𝐃𝐞𝐜.)+(𝟏 −𝜶 ) 𝑭𝟏𝟐(𝐃𝐞𝐜.)=𝟎.𝟑 (𝟓𝟒 )+ (𝟏 −𝟎.𝟑) 𝟓𝟎.𝟖𝟒=𝟓𝟏.𝟕𝟗
12-27
Exponential Smoothing
Forecast,
Period Month Demand
1 January 37 37.00 37.00
2
2 February
February 40
40 37.00
37.00 37.00
37.00
3
3 March
March 41
41 37.90
37.90 38.50
38.50
4
4 April
April 37
37 38.83
38.83 39.75
39.75
5
5 May
May 45
45 38.28
38.28 38.37
38.37
6 June 50 40.29 41.68
6 June 50 40.29 41.68
7 July 43 43.20 45.84
7 July 43 43.20 45.84
8 August 47 43.14 44.42
8 August 47 43.14 44.42
9 September 56 44.30 45.71
9 September 56 44.30 45.71
10 October 52 47.81 50.85
10 October 52 47.81 50.85
11 November 55 49.06 51.42
11 November 55 49.06 51.42
12 December 54 50.84 53.21
12
13 December
January 54
--- 50.84
51.79 53.21
53.61
13 January --- 51.79 53.61
12-28
Exponential Smoothing
Demand Alpha=0.3 Alpha=0.5
60

55

50
DEMAND

45

40

35

30
1 2 3 4 5 6 7 8 9 10 11 12

periods

12-29
Time Series Analysis:
Exponential Smoothing

30
Extrapolating Moving Averages
 As we are concerned about forecasting likely more
than one period into the future, what happens as we
extrapolate with these forecast methodologies?

 When we go more than one period into the future, we

do not have the actual values (Dt) to use and the
previous period’s forecast is used.
Simple Moving Average - Data
Week Muffins SMA
Jun 02 150 A
 The forecasts become
Jun 09 195 A quite flat.
Jun 16 156 A  Let’s look at the graph
Jun 23 168 A 167.00
Jun 30 165 A 173.00
Jul 07 180 A 163.00
Jul 14 171.00 F 171.00
Jul 21 172.00 F 172.00
Jul 28 174.33 F 174.33
Aug 04 172.44 F 172.44
Aug 11 172.93 F 172.93
Aug 18 173.23 F 173.23
Aug 25 172.87 F 172.87
Simple Moving Average - Graph
250

200

150
Actual
Sales

Forecast
100

50

0
0 2 4 6 8 10 12 14
Week
Weighted Moving Average - Data
Week Muffins Ft
Jun 02 150 A
 Again, the forecasts
Jun 09 195 A appear to level off.
Jun 16 156 A  The graph shows the
Jun 23 168 A 166.5
same result.
Jun 30 165 A 169.8
Jul 07 180 A 164.1
Jul 14 173.1 F 173.1
Jul 21 173.55 F 173.55
Jul 28 174.705 F 174.705
Aug 04 174.038 F 174.038
Aug 11 174.14 F 174.14
Aug 18 174.222 F 174.222
Wtd. Moving Average - Graph
250

200

150
Sales

100
Actual
Forecast
50

0
0 2 4 6 8 10 12 14
Week
 Exponential Smoothing ()
𝛼= 0.3
Month Dt Ft 𝛼Dt (1-𝛼)Ft Forecast
May 210 210.0 63.0 147.0 210.0 Jun
June 204 210.0 61.2 147.0 208.2 Jul
July 215 208.2 64.5 145.7 210.2 Aug
August 220 210.2 66.0 147.2 213.2 Sep
September 206 213.2 61.8 149.2 211.0 Oct
October 204 211.0 61.2 147.7 208.9 Nov
November 220 208.9 66.0 146.2 212.2 Dec
December 225 212.2 67.5 148.6 216.1 Jan
January 216.1 216.1 64.8 151.2 216.1 Feb
February 216.1 216.1 64.8 151.2 216.1 Mar
Regression Methods
To predict more than one period in future, we
use linear regression technique.

 Linear regression
– mathematical technique that relates a dependent variable to
an independent variable in the form of a linear equation
– Simple regression model uses the same formulae as in
linear trend line.

12-37
 Regression forecasting technique

  𝑦= 𝑎+𝑏 𝑥
 

     

10-38
Regression forecasting technique
(periods) (demand)
1 37 37 1
2 40 80 4
3 41 123 9
4 37 148 16
5 45 225 25
6 50 300 36
7 43 301 49
8 47 376 64
9 56 504 81
10 52 520 100
11 55 605 121
12 54 648 144
78 557 3867 650

12-39
Regression line
  

 Slope

 Intercept

 Regression line

10-40
 Linear trend line: y = 35.2 + 1.72x
Forecast for period 13: y = 35.2 + 1.72(13) = 57.56 units

Chart Title
60

55
f(x) = 1.72 x + 35.21
R² = 0.8
50
DEMAND

Actual demand
45

40

35
1 2 3 4 5 6 7 8 9 10 11 12

pERIODS

12-41
Linear regression allows you to obtain forecast values
for previous and future periods
(periods) (demand)

1 37 36.92
36.92   Plug in different values of x into:
2
2 40
40 38.64
38.64
3
3 41
41 40.36
40.36
4
4 37
37 42.08
42.08
5 45 43.8
5 45 43.8
6 50 45.52
6 50 45.52
7 43 47.24
7 43 47.24
8 47 48.96
8 47 48.96
9 56 50.68
9 56 50.68
10 52 52.4
10 52 52.4
11 55 54.12
11 55 54.12
12 54 55.84
12
13 54 55.84
57.56
13 57.56
10-42
Can linear regression be used to
forecast seasonal demand?
Demand (1000’s per season)
Year 1 2 3 4
2002 12.6 8.6 6.3 17.5
2003 14.1 10.3 7.5 18.2
2004 15.3 10.6 8.1 19.6
25

19.6
20
17.5 18.2
15.3
15 14.1
12.6
10.3 10.6
10 8.6 8.1
7.5
6.3
Is linear trend line effective in estimating demand?
5

0
0 1 2 3 4 5 6 7 8 9 10 11 12

10-43
Seasonal Adjustments
 Repetitive increase/ decrease in demand
 Use seasonal factor to adjust forecast

 𝐒𝐞𝐚𝐬𝐨𝐧𝐚𝐥 𝐟𝐚𝐜𝐭𝐨𝐫 𝒊 : 𝑺𝒊 = 𝑫 𝒊
∑ 𝑫𝒊
𝒊

12-44
Seasonal adjustments for linear regression
Step 1 – Subtotals
Demand (1000’s per season)
Year 1 2 3 4 Total
2002 12.6 8.6 6.3 17.5 45.0
2003 14.1 10.3 7.5 18.2 50.1
2004 15.3 10.6 8.1 19.6 53.6
Total 42.0 29.5 21.9 55.3 148.7

Step 2 – Seasonal factors

  𝐷1 42.0   𝐷3 21.9
𝑆 1= = =0.28 𝑆 3= = =0.15
∑ 𝐷 148.7 ∑ 𝐷 148.7
  𝐷2 29.5   𝐷4 55.3
𝑆 2= = =0.20 𝑆4= = =0.37
∑ 𝐷 148.7 ∑ 𝐷 148.7
10-45
Can linear regression be used to
forecast seasonal demand?
1. We now estimate the linear regression line using the aggregate
annual demand of the previous three years: 45, 50.1, and 53.6.
 
2. The linear regression line is:

3. We plug in for 2005 (fourth year)

4. We now seasonalize 58.17 based on seasonal factors:

10-46
It seems we can use linear regression
effectively for seasonal data.
Demand
25

20

15

10

0
0 2 4 6 8 10 12 14 16 18

Note that I have included a different way of using

regression to forecast each season of 2005.
10-47
Linear Regression with Multiple
Predictors
x y
x*y x^2
(wins) (attendance* 1000)

4 36.3 145.2 16
6 40.1 240.6 36
6 41.2 247.2 36
8 53 424 64
6 44 264 36
7 45.6 319.2 49
5 39 195 25
7 47.5 332.5 49
49 346.9 2167.7 311

12-48
 Example 2: Linear Regression
  
 Slope

 Intercept

 Regression line (attendance forecast for (i.e., wins))

10-49
Example 2: Linear Regression
60000

50000
f(x) = 4060.92 x + 18464.37
Attendance, y

40000

30000

20000

10000
3 4 5 6 7 8 9

Wins, x

12-50
Correlation and Coefficient of
Determination
 Correlation,
  r
– Measures of strength of relationship between dependent
and independent variables
– Varies between -1.00 and +1.00

 Coefficient of determination,
– Percentage of variation in the dependent variable that
results from changes in the independent variable

12-51
 Computing Correlation and coefficient of
determination for Example 2

Coefficient of determination ():

12-52
Example 2: Multiple Regression

 Consider a case where two factors influence attendance:

Predict attendance based on these two factors

Varying values of and yield different attendance values

12-53
Multiple regression: More than one
independent variables influence demand
 Study
  the relationship between demand and two or
more independent variables:

10-54
Multiple regression: More than one
independent variables influence demand

 To use multiple regression, go to

– Excel, data, data analysis, regression

10-55
Multiple Regression Example
y
x1 (wins)
x2 ($ - promotion) (attendance* 1000)
4 29500 36300
6 55700 40100
6 71300 41200
8 87000 53000
6 75000 44000
7 72000 45600
5 55300 39000
7 81600   47500
49 527400 The lower
346700 P-value, the stronger the predictor.

SUMMARY OUTPUT
General rule: p-value should be less than 0.05
Regression Statistics
Multiple R 0.949
R Square 0.901
Adjusted R Square 0.861
=19,094.424
 
Coefficients P-value
Intercept 19094.424 0.006
x1 (wins) 3560.996 0.064
x2 ($ - promotion) 0.0369 0.731

 
 Multiple Regression Example
 Equation for the linear regression line:

Predict attendance for 7 wins ( and 60,000 promotions (

12-57
Takeaways From the Excel Output
  
– We can use to select the few independent variables that have more impact
on the dependent variable if we have many independent variables.
Forecasting Accuracy
Forecasting error
– difference between forecast and actual demand

 MAD
– mean absolute deviation

 MAPD
– mean absolute percent deviation

 MSE
– mean squared deviation

12-59
Mean Absolute Deviation (MAD)
S|Dt - Ft |
MAD = n
where
t = period number
Dt = demand in period t
Ft = forecast for period t
n = total number of periods
|| = absolute value
Why not just mean deviation?
 What is the ideal MAD?
What does a high MAD indicate?
12-60
MAD Example
Diff Abs. Diff
Period Demand,
1 37 --- --- ---
2
2 40
40 37.00
37.00 3.00
3.00 3.00
3.00
3
3 41
41 37.90
37.90 3.10
3.10 3.10
3.10
4
4 37
37 38.83
38.83 -1.83
-1.83 1.83
1.83
5 45 38.28 6.72 6.72
5 45 38.28 6.72 6.72
6 50 40.29 9.69 9.69
6 50 40.29 9.69 9.69
7 43 43.20 - 0.20 0.20
7 43 43.20 - 0.20 0.20
8 47 43.14 3.86 3.86
8 47 43.14 3.86 3.86
9 56 44.30 11.70 11.70
9 56 44.30 11.70 11.70
10 52 47.81 4.19 4.19
10 52 47.81 4.19 4.19
11 55 49.06 5.94 5.94
11
12 55
54 49.06
50.84 5.94
3.15 5.94
3.15
12
13 54
557 50.84
51.79 3.15
49.31 3.15
53.39
13 557 51.79 49.31 53.39
12-61
MAD Calculation

  ¿ 53.39
𝑀𝐴𝐷=∑∨ 𝐷𝑡 − 𝐹 𝑡∨ = =4.85 ¿
𝑛 11
Number of periods for which we
have forecasted values.

12-62
Time Series Analysis:
Other ways to measure forecast error
 Mean Absolute Deviation (MAD)
MAD = ∑ |Dt – Ft| / n
where n = the number of time periods
 Mean Squared Error (MSE)
MSE = ∑ (Dt – Ft)2 / n
 Mean Absolute Percent Deviation (MAPD)
– Measure error as a percent of actual values

MAPD = 100[ ∑ |Dt – Ft| / Dt ] / n

63
First example: Comparison of
Forecasts

Forecast MAD MAPD

Exponential smoothing ( 4.85 9.6%

Exponential smoothing ( 4.04 8.5%
Linear regression: 2.29 4.9%

12-64
Housekeeping
 Assignment 3 is due on Nov. 25th, 11:59pm.

 Report for case study # 2 is due on Nov. 30th, 11:59

pm

 Presentation for case study # 2 is due on Dec. 1st.

11:59 pm.

 Next week – Linear Programming: Introduction to

Management Science
10-65
Alternative use of linear regression for
seasonal data
 The idea is to fit a linear line to the data of each season
across different years.

 We have the data for the first season of three different

years and we can therefore forecast the demand for each
season using the data of the same season in previous
years.
Season 1:
Linear regression for data of season 1
18

16
f(x) = 1.35 x + 11.3
14 R² = 1
12

10

0
0 1 2 3 4
Season 2:
Linear regression for data of season 2
12
f(x) = x + 7.83
10 R² = 0.86
8

0
0 1 2 3 4
Season 3:
Linear regression for data of season 3
9

8 f(x) = 0.9 x + 5.5

7
R² = 0.96
6

0
0.5 1 1.5 2 2.5 3 3.5
Season 4:
Linear regression for data of season 4
20

19.5
f(x) = 1.05 x + 16.33
R² = 0.96
19

18.5

18

17.5

17

16.5

16

15.5

15
0 1 2 3 4
Calculations
 Plug in x = 4 in the formulas of the previous four
slides to get the forecasted value for the intended
season of year 2005.

 Plugging x = 1, 2, and 3 in the formulas of the

previous four slides will provide the forecast for years
2002, 2003, and 2004.

 Note that this linear regression approach yields lower

MAD than the previous linear regression method
explained.