8/7/2019

Data Analytics in Business

Operations Management

Bob Myers
Lecturer
Scheller College of Business

Forecasting

Learning Objectives
At the end of this lesson, you
should be able to:
• Discuss forecasting in the context
of operations management
• Discuss patterns of demand
• Identify qualitative and quantitative
forecasting methods

1
 8/7/2019

What is Forecasting?
Forecasting – prediction of future events used for planning purposes
Used for:
• Strategic planning (long term capacity decisions)
• Finance and Accounting (budgeting and cost control)
• Marketing (future sales trends and new product introduction)
• Production and Operations (staffing and supplier relations)

General Characteristics of Forecasts

• Forecasts are almost always wrong!
• Forecasts are more accurate from groups or families of items
• Forecasts are more accurate for shorter periods of time
• Every forecast should include an error estimate
• Forecasts are no substitute for actual demand

2
 8/7/2019

Patterns of Demand

• Trends

• Seasonality

• Cyclical elements

• Autocorrelation

• Random variation

Data can Exhibit Multiple Patterns

Seasonal variation

xx Linear
x
x x
x x Trend
x x
Sales

x x x
xx x x
x xx x x
x
x
x x x x x x
x x xxx x
x x x
x xxxxx
x
x x

1 2 3 4
Year

3
 8/7/2019

Time Series Components of Demand

. . . Randomness
Demand

Time

Time Series with…

…Randomness & Trend
Demand

Time

4
 8/7/2019

Time Series with…

…Randomness, Trend, and Seasonality
Demand

Dec. Dec. Dec. Dec.

Time

Some Important Questions

• What is the purpose of the forecast?
• Which systems will use the forecast?
• How important is the past in predicting the future?

Answers will help determine the time horizons, techniques, and level of detail in
the forecast

5
 8/7/2019

Type of Forecasting Methods

Qualitative
• Rely on subjective opinions from one or more experts

Quantitative
• Relay on data and analytical techniques

Quantitative Forecasting Methods

Time Series: models that predict future demand based on past history trends

Casual Relationships: models that use statistical techniques to establish

relationships between various items and demand (Ex: Linear Regression)

Simulation: models that can incorporate some randomness and non-linear

effects

6
 8/7/2019

Summary
1. Forecasting is trying to predict
future events for planning
purposes.
2. In Operations Management
forecasting Demand is key.
3. Demand can exhibit multiple
patterns
4. We will focus on Time Series
techniques

Data Analytics in Business

Operations Management

Bob Myers
Lecturer
Scheller College of Business

Exponential Smoothing

7
 8/7/2019

Learning Objectives
At the end of this lesson, you should be
able to:
• Explain exponential smoothing

Time Series: Exponential Smoothing

The Prediction of the future depends mostly on the most recent observation and
on the error for the latest forecast

Denotes the
Smoothing
importance of
constant alpha α
the past error

8
 8/7/2019

Why Exponential Smoothing?

• Uses less storage space for data
(although not a problem these days)

• Extremely accurate

• Easy to understand

• Little calculation complexity

Exponential Smoothing (ES)

Assume that we are currently in period t. We calculated the forecast for the last
period (Ft-1) and we know the actual demand last period (At-1):

𝐹𝑡 = 𝐹𝑡-1 +∝ (𝐴𝑡-1 − 𝐹𝑡-1ሻ

𝑤ℎ𝑒𝑟𝑒 0 ≤∝≤ 1
The smoothing constant a expresses how much our forecast will react to
observed differences:
• If a is low, there is little reaction to difference
• If a is high, there is a lot of reaction to differences

9
 8/7/2019

Plotting Exponential Smoothing Forecasts

Small a– Stable
Large a - Responsive

Example
Week Demand Forecast
1 820 820
2 775 820
3 680 811
4 655 785
5 750 759
6 802 757
7 798 766
8 689 772
9 775 756
10 760

10
 8/7/2019

How do you get started?

Week Demand Forecast
1 820 820 You must pick an initial forecast.
2 775 820 • Can set it equal to demand
• Can use another methods solution
3 680 811
• Note the initial forecast is not
4 655 785 calculated from the exponential
5 750 759 smoothing formula. Don’t use it in
evaluating the method
6 802 757
7 798 766
8 689 772
9 775 756
10 760

Summary
1. Exponential Smoothing is a common
method used to forecast random
behavior in demand.
2. It uses the prior forecast and error to
predict the next period.
3. The smoothing constant a
determines how much the error alters
the next prediction (forecast).

11
 8/7/2019

Data Analytics in Business

Operations Management

Bob Myers
Lecturer
Scheller College of Business

Exponential Smoothing with a

Trend

Learning Objectives
At the end of this lesson, you should
be able to:
• Explain exponential smoothing with a
trend adjustment

12
 8/7/2019

Recall: Exponential Smoothing (ES)

Assume that we are currently in period t. We calculated the forecast for the
last period (Ft-1) and we know the actual demand last period (At-1):
𝐹𝑡 = 𝐹𝑡-1 +∝ (𝐴𝑡-1 − 𝐹𝑡-1ሻ
𝑤ℎ𝑒𝑟𝑒 0 ≤∝≤ 1

The smoothing constant a expresses how much our forecast will react to
observed differences:
• If a is low, there is little reaction to difference
• If a is high, there is a lot of reaction to differences

What Happens if Demand is Trending?

What do you think will happen to the exponential
smoothing model when there is a trend in the data?

ES will lag behind trends

13
 8/7/2019

Exponential Smoothing With a Trend?

• How can we account for a trend in the data?

Sales Actual
Data
Forecast
Ft  Ft 1  a ( At 1  Ft 1 )

Forecast Including Trend

FITt  Ft  Tt

Ft  FITt1  α(At1  FITt1 ) FIT - Forecast Including Trend

δ - trend smoothing constant
Tt  Tt1  δ(Ft  FITt1 )
The idea is that the two effects are decoupled,
(F is the forecast without trend and T is the trend component)

14
 8/7/2019

Example: Kroger Canned Soup Sales

Ft  FITt1  α(At1  FITt1 ) Tt  Tt 1  δ(Ft  FITt 1 )

Month Actual Forecast Trend FIT

Jan 1325 1370 0 1370
Feb 1353 1334 -18.0 1316
FITt  Ft  Tt
Mar 1305 1346 -3.2 1342
Apr 1275 1312 -18.2 1294
May 1210 1279 -25.9 1253
Jun 1195 1219 -43.1 1176 α = 0.8, δ = 0.5
Jul ? 1191 -35.3 1156

Plotting ES with and without a Trend

ES: a=0.2

Actual
ES:
a=0.8

FIT: a=0.8,d=0.5

15
 8/7/2019

More on ES with and without a Trend…

• ES and FIT models require initial estimates for forecast and trend
(can use prior period as baseline to start)
• Values for a and d are found by trial and error (or using a solver to
optimize).
• If initial estimates are grossly incorrect, it may take a while for the
model to stabilize (higher a and d increase convergence speed if
initial estimates incorrect, but may tend to “overreact” to the noise in
the data)
• ES and FIT only use the prior period’s data for estimation of the
current period’s demand; this makes calculation and storage of data
easy

Summary
1. To account for a trend in data, the set of
three Forecast Including Trend
equations are much more accurate than
a simple Exponential Smoothing
equation.
2. The initial trend and forecast values
used can cause the model to take a few
periods to stabilize.

16
 8/7/2019

Data Analytics in Business

Operations Management

Bob Myers
Lecturer
Scheller College of Business

Exponential Smoothing with

Seasonality

Learning Objectives
At the end of this lesson, you should be
able to:
• Explain how to model seasonality with
Exponential Smoothing

17
 8/7/2019

Plotting Exponential Smoothing Forecasts

Small a – Stable
Large a - Responsive

Plotting the FIT ES Forecasts

ES: a=0.2

Actual
ES: a=0.8

FIT: a=0.8,d=0.5

18
 8/7/2019

What If My Data Has Seasonality?

Seasonality Calculation
• Measures the seasonal variation in demand
• Relates the average demand in a particular period to
the average demand for all periods

period average demand

The Seasonal Index 
average demand for all periods

19
 8/7/2019

Calculation of Seasonal Index

Monthly Sales of Ice Cream
January 10 July 600
February 10 August 700
March 10 September 350
April 50 October 100
May 150 November 10
June 400 December 10
Total annual sales 2400 cases
Average monthly sales 200 cases
June’s sales are 400 cases

Seasonal Index for June = month’s sales/avg sales = 400/200 = 2.0

Calculation of Seasonal Index

Monthly Sales of Ice Cream
SI (calc.) SI
Jan 10 = 10 / 200 0.05
Feb 10 = 10 / 200 0.05
Mar 10 = 10 / 200 0.05
Apr 50 = 50 / 200 0.25
May 150 = 150 / 200 0.75
Jun 400 = 400 / 200 2
Jul 600 = 600 / 200 3
Aug 700 = 700 / 200 3.5
Sep 350 = 350 / 200 1.75
Oct 100 = 100 / 200 0.5
Nov 10 = 10 / 200 0.05
Dec 10 = 10 / 200 0.05

TOTAL 2400
AVERAGE 200

20
 8/7/2019

Seasonal Series Indexing Sample Data

Seasonal
Month Year 1Year 2Year 3 Total Index Monthly Total (MT)
(SI) =
Average Month (AM)
Jan 10 12 11 33 0.33
Feb 13 13 11 37 0.37 Where:
Mar 33 38 29 100 1.00 1200
AM = = 100
Apr 45 54 47 146 1.46
12
May 53 56 55 164 1.64
Jun 57 56 55 168 1.68

Jul 33 27 34 94 0.94
Aug 20 18 19 57 0.57
Sep 19 22 20 61 0.61 33
SIJAN = = .33
100
Oct 18 18 15 51 0.51
Nov 46 50 55 151 1.41
Dec 48 53 47 148 1.48
Total 395 417 398 1210 12.00

Calculate ES with Seasonal Forecast

1. Inputs: Realized seasonal demand in the previous period
(~At), seasonal forecast for the previous period (~Ft), and
smoothing constant a
2. De-seasonalize demand and forecasts to obtain At and Ft
• At=~At /SIt and Ft=~Ft /SIt
3. Use the exponential smoothing formula to obtain the de-
seasonalized forecast Ft+1
• Ft+1= a At +(1- a )Ft
4. Re-seasonalize the forecast to obtain ~Ft+1.
• ~Ft+1 = Ft+1 x SIt+1

21
 8/7/2019

Example: ES with Seasonality

ES w/ Seasonality
a = 0.5
De-Seasonalized De-Seasonalized
Actual Forecast SI Actual Forecast
Jan 9 10 0.05 180.0 200.0
Feb 8 9.5 0.05 160.0 190.0
Mar 11 8.8 0.05 220.0 175.0
Apr 53 49.4 0.25 212.0 197.5
May 160 153.6 0.75 213.3 204.8
Jun 390 418.1 2 195.0 209.0
Jul 590 606.1 3 196.7 202.0
Aug 720 697.7 3.5 205.7 199.3
Sep 370 354.4 1.75 211.4 202.5
Oct 120 103.5 0.5 240.0 207.0
Nov 12 11.2 0.05 240.0 223.5
Dec 8 11.6 0.05 160.0 231.7

1. De-seasonalize Actual and Forecast at time t (demand / SI)

2. Apply ES calculation: Ft+1 = Ft + a (At – Ft)
3. Re-seasonalize Forecast for t+1 (Ft+1 x SI)

Example: ES With Seasonality

22
 8/7/2019

Summary
1. This method can capture more
than one seasonal pattern.
2. This method can be combined
with FIT to capture demand that
has random behavior with a trend
and seasonality.

Data Analytics in Business

Operations Management

Bob Myers
Lecturer
Scheller College of Business

Error Methods

23
 8/7/2019

Learning Objectives
At the end of this lesson, you should
be able to:
• Discuss different Error measurements
used to evaluate Forecasting methods

How can we Compare Different Models?

We need a metric that provides estimation of accuracy
Forecast error = the difference between the actual and forecasted value
(also known as residual)

Errors can be:

1. Biased (consistent)

Forecast Error 2. Random

24
 8/7/2019

Measures of Forecast Accuracy

Error = Actual demand - Forecast
or
et = At - Ft

• et can be positive of negative

• Positive et means that the forecast was too low
• Negative et means that the forecast was too high

Measures of Forecast Error

RSFE – Running Sum of Forecast Error

RSFE= (A i  Fi )  ei

MFE – Mean Forecast Error (Bias)

 (A i  Fi ) RSFE
MFE= n
=
n

MAD – Mean Absolute Deviation

 A i  Fi
MAD = n
TS – Tracking Signal
RSFE
TS = MAD

25
 8/7/2019

Measuring Accuracy: MFE

• MFE = Mean Forecast Error; also called Bias
• Average error in the observation
n n

(A  F ) e
t t t
RSFE
MFE  i 1
 i 1

n n n
• A more positive or negative MFE implies worse performance, the
forecast on average is biased from the actual demand

Measuring Accuracy: MAD

• MAD = Mean Absolute Deviation
• Average absolute error in the observation
n n

 A F t t e t
MAD  i 1
 i 1
n n
• Higher MAD implies worse performance

26
 8/7/2019

Dartboard Analogy to MFE and MAD

• MFE is a measure of the overall average accuracy of the forecast
(lower MFE is better)
• MAD is a measure of the overall variability of the error terms (lower
MAD is better)

Large MFE Small MFE Small MFE

Large MAD Small MAD Large MAD

MFE and MAD (by the numbers)

E= Error
F= A= Actual (Sales Absolute
PD Forecast Sales Forecast) Error
RSFE =
1 1,000 1,200 200 200
 (A i  Fi )  200 2 1,000 1,000 0 0
3 1,000 800 -200 200
4 1,000 900 -100 100
5 1,000 1,400 400 400
Bias = MFE = 6 1,000 1,200 200 200
7 1,000 1,100 100 100
 (A i  Fi )  200  20 8 1,000 700 -300 300
N 10
9 1,000 1,000 0 0
10 1,000 900 -100 100
Mean Absolute Deviation (MAD) = 10,000 10,200 200 1,600

 A i  Fi  1600  160
N 10

27
 8/7/2019

Measuring Accuracy: Tracking Signal

TS = Tracking Signal
• Measure of how often our estimations have been above or
below the actual value. It is used to decide when to re-
evaluate the model

RSFE
TS 
MAD
• Positive tracking signal – most of the time, the actual values
are above the forecasted values
• Negative tracking signal – most of the time, the actual values
are below the forecasted values
• If TS < -4 or TS > 4, investigate! ( |TS| >4 )

Let’s Compare Accuracy Measures

28
 8/7/2019

MAD and TS for ES with a = 0.2

Month Actual ES: a=0.2 error RSFE MFE Abs error MAD TS

Jan 1325 1370 -45 -45 -45.0 45 45.0 -1.0

Feb 1353 1361 -8 -53 -26.5 8 26.5 -2.0

Mar 1305 1359 -54 -107 -35.7 54 35.7 -3.0

Apr 1275 1349 -74 -181 -45.3 74 45.3 -4.0

May 1210 1334 -124 -305 -61.0 124 61.0 -5.0

Jun 1195 1309 -114 -419 -69.8 114 69.8 -6.0

Compare Accuracy Measures

TS < -4

29
 8/7/2019

MAD and TS for ES with a = 0.8

Month Actual ES: a=0.8 error RSFE MFE Abs error MAD TS

Jan 1325 1370 -45 -45 -45.0 45 45.0 -1.0

Feb 1353 1334 19 -26 -13.0 19 32.0 -0.8

Mar 1305 1349 -44 -70 -23.3 44 36.0 -1.9

Apr 1275 1314 -39 -109 -27.3 39 36.8 -3.0

May 1210 1283 -73 -182 -36.4 73 44.0 -4.1

Jun 1195 1225 -30 -212 -35.3 30 41.7 -5.1

Let’s Compare Accuracy Measures

TS < -4

30
 8/7/2019

MAD and TS for FIT with a = 0.8 and d = 0.5

FIT: Abs
Month Actual a=0.8,d=0.5 error RSFE MFE error MAD TS
Jan 1325 1370 -45 -45 -45.0 45 45.0 -1.0
Feb 1353 1316 37 -8 -4.0 37 41.0 -0.2
Mar 1305 1342 -37 -45 -15.0 37 39.7 -1.1
Apr 1275 1294 -19 -64 -16.0 19 34.5 -1.9
May 1210 1253 -43 -107 -21.4 43 36.2 -3.0
Jun 1195 1176 19 -88 -14.7 19 33.3 -2.6

Let’s Compare Accuracy Measures

31
 8/7/2019

Comparing the Models

MFE MAD TS

ES: a=0.2 -69.8 69.8 -6.0

ES: a=0.8 -35.3 41.7 -5.1

FIT: a=0.8,d=0.5 -14.7 33.3 -2.6

• The tracking signals for both ES models are below -4, thus indicating
that these models are not fitting the data well
• The FIT model yields the smallest MFE, MAD, and TS; this is the only
model with an acceptable tracking signal

Summary
1. Error measurements are
needed to evaluate different
forecast models
2. The tracking signal is useful
to alert us to changes in
patterns of demand

32
 8/7/2019

Data Analytics in Business

Operations Management

Bob Myers
Lecturer
Scheller College of Business

Forecasting Recap

Learning Objectives

• Discuss and recap lessons from this week

• Assess tie back to analytics

33
 8/7/2019

Which Forecasting Method Should You Use

 Gather historical data of what you want to
forecast
 Divide the data into initiation and evaluation sets
 Use the first data set to develop the models
 Use the second data set to evaluate the models
 Compare the MADs and MFEs of each model,
keeping an eye on the tracking signals
 Incorporate seasonality and trending aspects into
the model if the data displays those
characteristics

Recap
 Forecasting is based on the assumption that the
past predicts the future.
 Components of demand include the average,
trends, seasonal elements, cyclical elements,
random variation, and autocorrelation.
 Qualitative forecasting is used when hard data is
unavailable.
 Exponential Smoothing uses the most recent data
and forecast error.
 Forecast Including Trend (FIT) extends
Exponential Smoothing to applications with a clear
trend.
 Seasonality can be incorporated into forecasting
models

34
 8/7/2019

Recap
 Forecast error is the difference between
the actual and forecasted values
 Mean Forecast Error (Bias) is as
measure of the overall average of the
forecast to actual demand
 Mean Absolute Deviation is a measure of
variation in the error between the forecast
and actual demand
 Tracking Signals tell whether a forecast is
above or below actual and by how much

35