S A Papo Workshop Slides 1
S A Papo Workshop Slides 1
S A Papo Workshop Slides 1
Data Integration
External buy
Overall importance
New ways of
doing Business in of knowledge
Corporate
benefits
& processes
Global capacities
Flexibility / Agility
IT IT purchasing
benefits benefits
IT architecture
cost reduction
Current functionality
Target functionality
100% (or more) functionality is NOT the aim of standard ERP/APS systems!
Disclaimer
Objectives of SAP APO
SAP APO is more than just a forecasting tool
Holistic integrated planning suite
Communication with legacy systems
Hammer
Response
540 responses
260 incomplete (reminders send, to only speculative interest, unwilling
to give email address (although not mandatory), Atrophy (number of
repeated questions), unsuitable respondent (industry sector &
position)
15 complete responses discarded (Consultants/academics, rushed
surveys (10-15 mins), highly inconsistent answers, middle-clicking {
same answer for every question in groups)
APO and APO+ (incl. specialists FSS add-ons) dominate todays market!
Study Results
Systems use of Respondents
Group responses by a forecasting software plus average response!
Interpret general
forecasting
Breakdown of responses by sector practices
and use of systems?
Food & Beverage
Classification:
Industrial/other
Industry Sector
23 respondents
Majority (72%) of use APO DP plus
respondents use an SAP (non APO) an extra FSS
ERP/APS system 5%
Specialist system
(incl. SAP APO etc.)
ERP/APS14 respondents
34%
still use SAP with
no APO modues
(e.g. R/3 MM )
Heterogeneous market of ERP/APS vs. stand-alone FSS & combination
Study Results
Systems use of APO+ Respondents
Arkieva/Zemeter,
Demand SAP APO+ FSS
SAP APO
Excel only
17%
only Solutions, Oliver JDA
21%
Wight eSOP tool, 16%
FSS
Servigistics,
Specialist
Only
SO99+, SPSS,
11% SAP APO +
Other FSS
Steelwedge,
12% TOOLBELT,
SAP (non
Vanguard
APO)
5%
29%
ERP/APS
34%
Forecast X
16%
23 users apply
SAP APO DP
plus an additional
Mercia
Specialist FSS 7%
Finmatica
INFOR (2005)
Forecast Pro SAS
16% 16%
Excel only
SAP APO SPECIALIST FSS SYSTEMS
only
17%
21%
Forecast Pro
FSS 21%
Specialist
Only
11% SAP APO +
Other FSS
12%
SAP (non
APO)
5%
ERP/APS
34%
Apollo Data
Technologies,
Applix Tm1, SAS
Avercast, Demand 10%
Solutions, Eviews,
GMT, Infor,
Predicast/Aperia,
Project i, Remira
LogoMate, Retail Forecast X
Express AMP2, 7%
SAF, SPSS, Stata,
TORA, TXT
R
7%
Most companies
Smaller number predict 1000
of items likely to items across
drive use of systems
MS Excel Small companies Large companies
more likely to more likely to
use MS Excel
Number of use SAP APO DP
forecasters is
independent of
APO+ users with
use of system(!?)Large companies
more #items
more likely to
than APO users
use SAP APO DP
drives system?
APO DP normally
forecast at item
or account level,
never store / DC
APO DP+ FSS
forecast also at
Daily forecasts DC level
are still rare (!) (VMI functions?)
Users forecast at
multiple time
buckets at the
same time!
Most companies
& system users
face the same
relative impact
Specialist FSS
Excel users apply Excel users apply
users apply
mainly judgment mainly judgment
mainly Stats
with any Stats! (without
with any Stats!)
Stats!
(without Judg.!)
Typical SAP APO usage employs 25% only stats, 25% only judgment
and 50% a combination of Statistics & Judgment
Study Results (all industries)
Use of Forecasting Methods
APO DP: 78% (45% ETS + 22% Av+ 11% Nave) use simple methods!
APO drives use of simpler methods than average software packages
What algorithms do you use?
Study Results (all industries)
Satisfaction with Forecast Accuracy
using Specialist
Software slightly
increases user
satisfaction
All users are only moderately happy with their forecast accuracy
APO DP not worse off than SAP R/3, ERP or Specialist software
Study Results (all industries)
Satisfaction with Forecast Accuracy
Users seem satisfied that better systems will not improve accuracy!
BUT: will more customer data not require new system (fucntions)?
Study Results (all industries)
Satisfaction with Forecast Software
Annual conferences
The International Symposium on Forecasting (ISF) Peer Review
Professional conferences (Forecasting Summit, commercial
IBF, iE, SAS, Terra etc.) are typically non-peer review
Monthly
Maintain Analyse & Decision on Analyze
statistical
master data understand automated time series forecasting
& link with assortment forecasting Identify
& clean data
predecessors (ABC / XYZ) suitable
models &
parameters
Evaluate
forecast
quality
Apply
statistical
/ manual
Adjust & forecasting
Finalization release
of forecast forecast
The Forecasting Process
Statistical Modelling: Model Application
Initialization via overall analysis
Monthly
Maintain Analyse & Decision on Analyze
statistical
master data understand automated time series forecasting
& link with assortment forecasting Identify
& clean data
predecessors (ABC / XYZ) suitable
models &
parameters
Evaluate
forecast
quality
Apply
statistical
/ manual
Adjust & forecasting
Finalization release
of forecast forecast
The Forecasting Process
Statistical Modelling: Model Application
Initialization via overall analysis
Monthly
Maintain Analyse & Decision on Analyze
statistical
master data understand automated time series forecasting
& link with assortment forecasting Identify
& clean data
predecessors (ABC / XYZ) suitable
models &
parameters
Evaluate
forecast
quality
Apply
statistical
/ manual
Adjust & forecasting
Finalization release
of forecast forecast
Analysing Assortments
ABC-XYZ Analysis
41.4% of
profit
comes
from 20%
of
products
59.8% of
forecasting
error
comes
from 20%
of
products
A-Narts
Important for the business
Interactive MoSel for decision support Fewer NARTs (20%)
Cannot use Neural Nets as
Exponential Smoothing & SLR it cant be explained in e.g.
ISF
Automatic Automatic
X-Narts
Simple to forecast
Exponential Exponential Smoothing & SLR incl. Neural
BY, BZ, CY, & CZ -Narts
NARTs Smoothing & SLR Networks Less important
Avoid Neural Nets Many time series (50%)
similar high Neural Nets can be used
accuracy use the
simplest method
4.9 40
4.8
4.7 20
4.6
seasonal
0
4.5 0 0.1 0.2 0.3 0.4 0.5
2 4 6 8 10 12
frequency (Hz)
4.8
Create simple
seasonal diagram as
planning book
Program interactive
graphics container
with multiple
diagrams for APO
600
1500000
550
Units
GDP
500 1000000
450
500000
400
350
Show multiple
0
20 40
step60ahead80
Month
1948 1958 1968 1978 1988 1998 2008 2018
Year
UK forecasts
UK Hourly Electricity Demand
Demand UK Hourly Electricity Demand
UK
110000 110000
100000 100000
90000 90000
Demand
Demand
80000 80000
Demand
Demand
70000 70000
60000 60000
50000 50000
40000 40000
10/26/08 10/27/0810/28/0810/29/08
10/26/0810/27/08 10/28/08 10/29/08 10/30/08 10/31/08
10/30/0810/31/08 10/26/08 10/26/0810/27/08
10/26/08 10/27/0810/28/0810/29/08 10/30/0810/31/08
10/28/08 10/29/08 10/26/08
10/30/08 10/31/08
Day Day
Forecast Pro
Supports mainly
a manual Model
2 options for Selection by Forecaster
afully automaticAlpha [0,,1] , Beta [0,,1], Gamma [0,,1]
Parameters:
Model Selection Choice of 100x100x100 parameter combinations!
Solutions exist
Train demand planners in model selection
Use bolt-on systems for model selection APO-DP
Monthly
Maintain Analyse & Decision on Analyze
statistical
master data understand automated time series forecasting
& link with assortment forecasting Identify
& clean data
predecessors (ABC / XYZ) suitable
models &
parameters
SAPAPO DP
Evaluate
includes some
forecast
qualitymodels which
dont work Apply
statistical
SAPAPO DP / manual
excludes
Adjust some
& forecasting
Finalization release
of forecast
important
forecast
models
Baseline Model Application
Models in APO not functional
Forecasting Methods
Statistical Judgemental
Double ETS
Naive Methods Expert Opinion
(2nd order ETS)
Linear Trend ETS Moving Average
Median Method
Baseline Model Application
Models in APO not functional
Out-of-sample Out-of-sample
140 140
130
130
120
120
GDP
GDP
110
110
100
100
90
Data Data
80 APO 90 APO
External External
70 80
5 10 15 20 25 30 35 40 45 5 10 15 20 25 30 35 40 45
Month Month
3 years in-sample not enough to forget bad initialisation, requires higher smoothing
parameters Filter noise adequately?
Initialisation problem for ETS trend models is significant better avoid them!
Similarly: avoid 2nd order Exponential Smoothing
Baseline Model Application
Models in APO missing
APO
1000 cannot run 1000
additive seasonal
800 800
model
600 600
400 400
200 200
Seasonality
0 0
expands as level
0 10 20 30 40 50 60 0 10 20 30 40 50 60
increases
This restricts the valid usage of the seasonal trend model in APO DP
Baseline Model Application
Models in APO missing
Exponential Smoothing Models in APO DP
-Seasonal Models
Because APO has only multiplicative seasonal EXSM, any time series that has
values close to zero becomes impossible to forecast.
At
Multiplicative Seasonality
Out-of-sample
Lt (1 ) Lt 1
350 St s
300
At
250 St 1 St s
200 Lt
Level Component
150 1500
100
1000
50
0 10 20 30 40 50 60
Multiplicative Seasonality 500
Out-of-sample
2000
0
10 20
Seasonal 30
Component 40
1500
Effect of 2
zero
1000 1.5
1
500
0.5
0 0
0 10 20 30 40 50 60 10 20 30 40
Forecasting Methods
Statistical Judgemental
68%
100
50
0
10 20 30 40 50 60 70 80 90 100
Period
Promotion 1 0.5
0
10 20 30 40 50 60 70 80 90 100
Promotion 2 0.5
0
10 20 30 40 50 60 70 80 90 100
The ScanPRO model
n p rj 3 T K
kjt
Qkjt krt lrjDlkrt jtXt Zk kj e
r 1 pkrt l 1 t 1 k 1
Cross-effects between products Interactions & cannibalism
Cross-effects between promotions Interactions, support & cannibalism
Cross-effects between stores Spatial cannibalisation & competition
6000
5000
4000
Sales
3000
2000
1000
0
5 10 15 20 25 30 35 40 45 50 55 60
2000
Period
+ Effect 1000
+ Effect
0
15 15
10 10
5 5
May-06
Jan-05
Jan-06
Mar-06
Nov-05
Nov-06
Jul-05
Sep-05
Jul-06
Sep-06
8
8
6
6
12 months of data 4
4
2 2
0 0
1 2 3 4 5 1 2 3 4 5 6
Monthly
Maintain Analyse & Decision on Analyze
statistical
master data understand automated time series forecasting
& link with assortment forecasting Identify
& clean data
predecessors (ABC / XYZ) suitable
models &
parameters
Evaluate
forecast
quality
Apply
statistical
/ manual
Adjust & forecasting
Finalization release
of forecast forecast
Baseline Data Analysis & Exploration
Allows various Data Exploration in SAP APO DP
overrides for
promo planning,
history correction,
Rational
Human
Causes
e.g. biases Sub-optimal Side effects:
Training Lower user
from company sales forecast Decompose Judgment
acceptance
policy
accuracy from Dont adjust
Human Use Analogies
Irrational Develop systems
Human Causes
Judgment
FSS features to
limited support judgment
information FSS features to
processing How to avoid sys- constrain judgement
capacity tematic errors in
final forecast?
Memory
support
Monthly
Maintain Analyse & Decision on Analyze
statistical
master data understand automated time series forecasting
& link with assortment forecasting Identify
& clean data
predecessors (ABC / XYZ) suitable
models &
parameters
Evaluate
forecast
quality
Apply
statistical
/ manual
Adjust & forecasting
Finalization release
of forecast forecast
Model Evaluation
Misleading Error Measures
650
600
In-sample data Out-of-sample 1 forecast origin
data only 1 measurement
550
low confidence in the
Units
500
measurement accuracy
450
BUT: more data (24) is
400
Forecast origin available
350
10 20 30 40
40 50
50 60
60 70
70 80
80
Month
n i 1 of MAPE exist
1 n Ai Fi 1 n Ai Fi
MAPE sMAPE
n i 1 Ai n i 1 ( Ai Fi ) 2
1 n
MAE Ai Fi
n i 1
At Ft
REt
At FNaive use Information Criteria
1 Relative Errors
n Ai Fi n for model selection?
GMRAE missing
AIC 2k 2 ln( L)
i 1 Ai FNaive
BIC 2 ln( L) k ln(n)
1 n At Ft
MASE
n i 1 MAEin sample1 stepahead Naive
Visualisation of error
distributions to check for
errors and bias for
relevant lead time!
Baseline Planning
Promotions Planning
You may use these slides for internal purpose, so long as they are
clearly identified as being created and copyrighted by Sven F.
Crone, Lancaster Centre for Forecasting.
You may not use the text and images in a paper, tutorial or external
training without explicit prior permission from the centre.