Adopt a Lean Forecasting Model for

More Accurate Projections

Thomas Brassøe
Stavanger October 22nd 2015
What We’ll Cover

Introduction to demand planning – Why is it so difficult?

Why advanced algorithms are not superior to simple models

Concrete tools and tips for keeping things simple

Can I only use APO to support simple forecasting

Wrap-up

2
Common Issue Relating Demand Planning

Many companies have a hard time dealing with the demand planning
and forecasting process
Things we often hear:

Forecast accuracy is too The forecast is a black box

low

We don’t have proper We don’t have time, too

competences many products, too little
time

A cumbersome and
complex process

3
One Fundamental Source to the Forecasting Challenges

1
• Complexity attracts! It seems intuitively right that more
complexity is more precise.
• Forecasting is a field of numbers that attracts
statisticians who are all about the number juggling
• Forecast consultancies and forecast software often
focuses on how much complexity they master

“Any intelligent fool can make things bigger and more

complex ... It takes a touch of genius – and a lot of
courage to move in the opposite direction”
– E. F. Schumacher

4
Complexity Attracts

Software providers Research papers

5
What We’ll Cover

Introduction to demand planning – Why is it so difficult?

Why advanced algorithms are not superior to simple models

Concrete tools and tips for keeping things simple

Can I only use APO to support simple forecasting

Wrap-up

6
Do We Really Need Advanced Algorithms, Parameter Optimization,
and Auto-Models?

“Keep it as simple as possible,

but not simpler, but no simpler”
A. Einstein

The Lean philosophy has taught us to ask the question:

Does it bring value?

7
What Is Advanced Algorithms Forecasting
When has something become complex?
• Well it pretty much depends on the eyes the beholder
• However, a few generic points will be used as reference in the presentation

Complex Simple
• Auto model selection, based on “ex-post”/“best-fit” • Manual chosen model
• Parameter optimization • Set parameters
• Several input variables • Few input variables
• Formulas only readable for mathematicians/ • Simple formulas
statisticians • Understandable calculation and outcome
• Black box solution

8
First Example

There is a lot of variance in the demand, meaning that:

• Step change (listings)

• Intermittence
Future History • Tender changes
• NPI/NPD
• Campaign heavy

Does it make sense to use advanced statistical algorithms to

create a future forecast?

Of course not – It doesn’t bring value

? • Simple statistical models, to set the level. Focus on

sales, marketing and category input.
• No statistical model at all

9
Second Example

There is a only little variance in the demand, meaning that:

• Stable products
• Clear seasonal
Future History
patterns
• Most high runners

Does it make sense to use advanced statistical algorithms to

create a future forecast?

?
Well, why should we? – It won’t create a better
accuracy – It doesn’t bring value
• Simple statistical models will do just fine – we don’t
need extra complexity

10
Do We Really Need Advanced Algorithms, Parameter Optimization
and Auto-Models?

• We think not
– The value-creation, being forecast accuracy, is found
other places
– In dynamic markets and changing conditions, patterns,
etc. We need to be proactive – statistical forecast can’t
do that.
• We need market intelligence
– However, statistical forecast can be really useful, if it is
managed in a simple lean way

If not:
• Black Box
• Forecast variation
• Frustration
• Lack of trust
• More work

11
What We’ll Cover

Introduction to demand planning – Why is it so difficult?

Why advanced algorithms are not superior to simple models

Concrete tools and tips for keeping things simple

Can I only use APO to support simple forecasting

Wrap-up

12
All We Need to Control the Statistical Forecast. Not More, Not Less.

Improve data quality by cleaning

Create a constant forecast for establishing a “level”

Apply seasonal indices to catch the seasonal variations

Introduce step change and trends for more control

13
Proper Data Quality Is a Prerequisite for Accurate Forecasts

Volume Volume
Promotions

Uncleansed
History Baseline
forecast
Time Time

The sales history, should be cleaned for all noise, Promotions and event should be added to the
which are not representative basic forecast
• Statistical models will not be able to forecast • Promotions and other events needs to be
promotions and outliers added on top
• Promotions will not have the same timing and • As with tenders, or other planning scenarios
volume year after year
• Bad data in  bad data out

14
SAP APO DP Provides Automatic Outlier Detection
But, it Quickly Gets Complicated!
• SAP APO DP provides automatic outlier detection It is based on standard deviations and medians
– But, it quickly gets complicated! • Outlier correction with ex-post method
• Outlier correction with Median method
• What should the Sigma be

Demand Upper Threshold Lower Threshold

Outliers?

1.600 1.600
1.400 1.400
1.200 1.200
1.000 1.000
800 800
600 600
400 400
200 200
0 0

15
Cleaning Doesn’t Need to Be Complex, Cumbersome and Time
Consuming – Introducing “Exclude Weeks”
Put some simplicity and knowledge in to the cleaning Promotion, and Manually set
outlier Exclude week “tick”

Automatically Exclude week “tick”

Orders
Non-ideal forecast Ideal forecast Basic Forecast
1.400 1.400
1.200 1.200
1.000 1.000
800 800
600 600
400 400
200 200
0 0

16
All We Need to Control the Statistical Forecast. Not More, Not Less.

Improve data quality by cleaning

Create a constant forecast for establishing a “level”

Apply seasonal indices to catch the seasonal variations

Introduce step change and trends for more control

17
Simple Exponential Smoothing (SES) Can Get You Pretty Far … (1/2)
Demand Forecast

• The SES has smoothing factors (α) and is an Alpha = 0,2

excellent forecast model 1.200
– Creates a constant future forecast
1.000
– Really easy to understand
800
– Alpha value controls the “reactiveness” of the
model 600
400
200

Alpha = 0,6
1.200
1.000
800
600
400
200

18
Simple Exponential Smoothing (SES) Can Get You Pretty Far … (2/2)

• We don’t want to create variance and uncertainty Alpha = 0,2

Demand Forecast
into our forecast
1000
– Reactiveness has its downside
– High alpha-values creates variance in our forecast 800
– Variance leads to uncertainty and bull-whip
600
– Not value adding
400

Alpha = 0,6
Demand Forecast
1000

800

600

400

19
So Is There a Right Answer at All?
• Yes, we think there is!
– A low alpha creates stability and hold variance down
– The only reasoning for having higher alpha values, is to “catch” step change
» We can do this smarter – by being proactive. I’ll illustrate this later.

Simple exponential smoothing, with Alpha 0,2 is what we need for level forecast

For NPIs, it can make sense to have a

slight higher alpha value initially (0,4),
in order to “find” proper level faster

20
All We Need to Control the Statistical Forecast. Not More, Not Less.

Improve data quality by cleaning

Create a constant forecast for establishing a “level”

Apply seasonal indices to catch the seasonal variations

Introduce step change and trends for more control

21
Seasonality Cannot Be Ignored
• Obviously, a level forecast (SES) is not sufficient if seasonality is present
– However, many companies have challenges with controlling the seasonal forecast
Typical issues

Products needs to have at least 1 Seasonality patterns can be Parameter tuning via a seasonal
year of history. This challenges difficult to see at the level of smoothing variable = Complexity
NPIs. forecast execution (e.g., KU/
customer/plant)

Holt Winters model:

Forecast
Brand new
Level

Slope

Seasonality

22
Group Seasonality Does the Job!
By accumulating the sales history by groups, we can actually solve at lot of seasonality challenges – simply

Group seasonal index

No. Description seasonal index
1500
320123 Chocolate 400g
1000
103235 Chocolate 100g
500
Products 321503 Caramel 300g

432503 Caramel 350g 0

Aug
Sep

Apr
Nov
Okt

Dec

Maj
Jul

Jan
feb
Mar

Jun
532592 Licorice 232gr

532942 Licorice 134gr

• “Law of large numbers” – accumulating
variance, for more valid indices
Product • Can be used for new products
Sweets
Group • Very easy to understand,
no complex formulas
Determine a range of products with same seasonal
pattern, or use product hierarchy if it makes sense

23
Example of How to Create Group Seasonality

• In SAP SCM 7 EHP2, a

similar functionality was
introduced. This
functionality works in a
similar way.
• This solution is, however,
a little bit more flexible

Selection options
• Product range/Product group
• Planning version
• Rolling periods vs. full-years

• Seasonality index number and

version
• Additional comments

24
Example of How to Create Group Seasonality (cont.)
Separate cockpit for index overview and maintenance makes the work much easier than considering complex
formulas

Seasonal index information

• Profile number
• Created by
• Created on date
• Comment

• Indices per week per year

• Possible to maintain, in
case of moving seasonality,
e.g., Easter

• Loaded in to a key figure in

the planning book, and
multiplied with the constant
forecast

25
All We Need to Control the Statistical Forecast. Not More, Not Less.

Improve data quality by cleaning

Create a constant forecast for establishing a “level”

Apply seasonal indices to catch the seasonal variations

Introduce step change and trends for more control

26
Be Very Careful When Using Trend Models …
What are the actual trends? Well, it depends on the parameter tuning you do, being Beta (β).

• What would the ideal Trend forecast based on the parameter tuning
forecast look like?
• What should the trend Demand
be based on? 1800
– Few periods 1600
– Several periods 1400
– Lots of periods 1200
1000
800
600
400
200

Should the statistical forecast choose what is right?

• We think not

27
Step Changes Can Be/Are Nightmares for Demand Planners
Actual changes in the demand pattern (not noise) disrupts the statistical forecast

• Is this an ok forecast? Simple constant forecast

Demand Forecast
• Probably not … 1.200
– Constant forecast (SES)
1.000
takes time to “catch” up to
proper level 800
– In the eyes of an auto- 600 Ideal forecast
model selection, a trend is level
400
recognized
200

Forecast with trend recognition

Demand Forecast
1.200
1.000
800
These step changes can be
a huge challenge for 600 Ideal forecast
demand planners. Luckily level
400
there is a simple way.
200

28
Incorporate Intelligence into the Statistical Forecast and Increase
the Quality (1/2)
By allowing intelligence to be incorporated into the logic forecast, we can actually overcome the constrains of
the statistical forecast

Introducing a step change

functionality in SAP APO DP

Simple constant forecast – with

Simple constant forecast with step
step
change
change
Demand Forecast Demand
Demand Forecast
Forecast
1.200 1.400
1.200

1.000 1.200
1.000
1.000
800 800
800
600 600
600
400 400
400
200 200

29
Incorporate Intelligence into the Statistical Forecast and Increase
the Quality (2/2)
The step change functionality is a strong tool for controlling trends,
which is the expected steady increase/decrease of demand

Manually controlling trends

Simple constant forecast with step change • Should we let history determine future
Demand Forecast
trends?
1.400 • Or should we leave that decision to those
who know the market, that being sales?
1.200
1.000
800
600
400
200

30
Additional Tip for Phasing Out Products
The step change functionality can also be used for easy delisting or phasing out

Constant forecast with step change

Demand Forecast
1.200
1.000 Input index “0”, will
800 delist product

600
400
200
0

31
Statistical Forecast Made Simple …

Constant Seasonality Trend/step change

Constant forecast, for setting the Apply generic group seasonality if Add intelligence to the forecast, in
level. Only two models required. seasonality is expected terms of step changes

32
Statistical Forecast Made Simple …

That’s it – we really don’t need more than this

to control the statistical forecast, and create
the most value

“Simplicity is the ultimate

sophistication”
Leonardo da Vinci

33
What We’ll Cover

Introduction to demand planning – Why is it so difficult?

Why advanced algorithms are not superior to simple models

Concrete tools and tips for keeping things simple

Can I only use APO to support simple forecasting

Wrap-up

34
SAP APO is not the only tool for implementing a simple and more lean
Demand Planning process – introducing SAP IBP….
IBP for Demand 5.0
• Demand Sensing algorithms (short term
forecasting)
• Statistical Methods (mid-/ long-term
forecasting)
– Pre-Processing algorithms
– Time series algorithms
– Regression based methods
• Demand specific analytics, e.g.
– Demand Sensing Issues
• Forecast model assignment
• Integration with ERP and APO (HCI templates)
• Exception management
• Excel as a planning front-end

35
The Excel front-end for IBP also supports the Statistical Forecast
models you need for keeping it simple

36
Using the Process Modeling within IBP can support a lean Demand
Planning process

37
What We’ll Cover

Introduction to demand planning – Why is it so difficult?

Why advanced algorithms are not superior to simple models

Concrete tools and tips for keeping things simple

Can I only use APO to support simple forecasting

Wrap-up

38
Your Turn!

How to contact me:

Thomas Brassøe
Email: [email protected]

39