APO DP Forecasting

Forecasting
Dr. Stefan Theis

SAP
Use of statistical methods
Demand Planning
Forecasting
Statistical methods can support the planning process but they cannot solve basic planning problems
E.g.: Periodicity change from months to weeks
Powerful forecasting software can calculate millions of forecasts on the lowest level of detail but this is not always the appropriate planning level
Nobody can control / check millions of forecasts With millions of data sets everything happens (Murphys law) It is sometimes better to plan on higher (controllable) level and breakdown the results to detail using fixed rules
SAP AG 2001, SCM, Dr. Stefan Theis 2
Content
Data Preparation
2 3 4 5 6
Selection of Forecasting Method
Univariate Forecasting
Causal Analysis (MLR)
Composite Forecasts
Evaluation
Data preparation (outlier correction)
Statistical methods can only run on appropriate data Therefore: This step is the most important one ! Adaptations can be necessary for:
Start of real history Level changes Negative / zero values Missing values Special events (e.g. strike, promotions, earth quake) Causal effects
With increasing data amount, the data preparation gains more and more importance but causes tremendous effort
Input for forecasting

The result of the data preparation process should be representative for the ordinary development corrected by all unnormal influences (for univariate forecasting) all unnormal influences not covered by the explanatory variables (for MLR) If data shows no structure (visually identifiable), it will be a challenge to create a forecast (no matter which method will be applied)
Recommendation
The standard algorithm for outlier correction implemented in APO 3.0 seems to be not appropriate in many projects Therefore a macro is recommended. This has the following advantages: There is no unique definition for an outlier. A macro can be defined for any customer definition By working with different key figures, the original history remains unchanged and the corrected quantities can easily be saved Please note: If the sum of corrected quantities is not approximate 0, the forecast has to be adapted correspondingly !
Pirelli example
Improvement of forecast accuracy at Pirelli
4000
3500
3000
Forecast Error Index
2500
2000
1500
1000
500
Const
Trend
Seasonal
Seasonal Linear
Linear Regression
Trend - Season
Regression Forecast Method
Original history
Cleaned history
Results based on 44 SKUs
Content
Data Preparation
2 3 4 5 6
Composite Forecasts
Evaluation
Selection of forecasting methods I
It is not appropriate to use the same forecasting method for all items Basic classifications include:
Forecast / Planning - Horizon: short <-> medium <-> long Linear <-> Non-linear development of the trend Univariate forecast <-> Causal analysis (MLR) Product type: new, mature, sporadic (e.g. parts) Different parameter settings
Selection of forecasting methods II
The assignment can be based on

logical reasons
product
classification (e.g. part, standard)
planning
purpose / business requirements

of products of parameters
pilot study
grouping assignment
Ex-Post error measures can be used but should not be the only criteria
The assignment should be checked in regular intervals
Automatic model selection in APO
Model selection 1 (Parameters fixed)

Test for trend and season (50 and 53) Test for trend (51) Test for season (52) Seasonal model with test for trend (54) Trend model with test for season (55)
Model selection 2 (Parameter kombination tested)

Test for trend and season (56)
Never use model selection for regular updates !
Comments on selection methods implemented in APO
Advantages:
Easy to use Checks a multitude of different methods Fast
Disadvantages:
No real optimization Selection criteria is only MAD The results cannot be used directly, but selections have to be created manually and assigned to corresponding forecast profiles
Alternative process using macros

1) Define different forecast profiles 2) Define any measurement for accuracy comparison 3) Create activity with the following steps (per characteristic combination)
a) Calculate 1. forecast profile and accuracy measurement (macro) b) Calculate 2. forecast profile and accuracy measurement (macro) c) ... d) Compare accuracy measurements and assign forecast profile with lowest accuracy (macro)
4) Run batch job with above activity for all characteristic combinations 5) Result: Optimal forecast profile is assigned per characteristic combination
Macro: Measurement calculation
Very easy error measure not to be used in reality
Macro: Compare measurement and assign profile

FORECAST_PROFILE_SET( Profilename ; Model to be used (S = Univariate, M = MLR, C = Composite) ; Version (Often using the command: KEYFS_VERSION( Row:... ) ;
Forecast Key Figure (can be omitted)

)
Activity
Batch job
Result: Assignment of forecast profiles
Content
Data Preparation
2 3 4 5 6
Composite Forecasts
Evaluation
Univariate forecasting methods I
Constant
Exponential smoothing 1st order (10/11) Exponential smoothing 1st order with a - adaptation (12) Moving average (13) Moving weighted average (14) Croston (80)
Exponential smoothing
Basic weighting principle:
Weight of current value: Parameter
Weight of last periods value: (1 Parameter)
Example of weighting (in %) of historical values

Parameter 0.1 0.3 0.5 0.7 Current period 10 30 50 70
First historical period

9 21 25 21
Second historical period

8 15 13 6
Third historical period

7 10 6 2
Example: Forecast of constant history
Constant Forecast Methods

1050
1025
1000
975
950 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Hist
FS10/11
Month FS12
FS13
FS14
Example: Forecast of trend-seasonal history
Constant Forecast Methods

1900
1500
1100
700
300 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Month
Hist
FS10/11
FS12
FS13
FS14
Univariate forecasting methods II
Trend
Exponential smoothing 1st order / Holt (20/21) Exponential smoothing 2nd order (22) Exponential smoothing 2nd order with a - adaptation (23) Linear Regression (94)
Season (without trend)

Exponential smoothing 1st order / Winters (30/31)
Trend - Season
Exponential smoothing 1st order (40/41) Seasonal linear regression (35) New in APO 3.0 !
Example: Forecast of trend history
Trend / Seasonal Forecast Methods

3000 2500 2000 1500 1000 500 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Month
Hist
FS20/21
FS22
FS23
FS30/31
FS40/41
Example: Forecast of seasonal history

3000 2500 2000
1500 1000 500 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Month Hist FS20/21 FS22 FS23 FS30/31 FS40/41
Example: Forecast of trend-seasonal history

3000 2500 2000 1500 1000 500 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
Month
Hist
FS20/21
FS22
FS23
FS30/31
FS40/41
Univariate forecasting methods III
Others
History (60) Manual Forecasting (70) No forecast (98) External Forecast / User Exit (99) New in APO 3.0 ! New in APO 3.0 !
Planning equipment (APO 3.0)

Outlier correction Ignore zero consumption Promotion planning Phase in / out profiles Like profiles
Trend dampening
There is only slight statistical methodology behind these (more or less) manual modelling procedures !
(Dis-) Aggregation with fixed or detailed proportional factors Macros
Content
Data Preparation
2 3 4 5 6
Composite Forecasts
Evaluation
Multiple linear regression (MLR)

MLR can assess how the development of one (dependent) variable can be explained by several (independent) variables (and a constant value) For a causal analysis MLR does the final calculation of the regression coefficients The input data for the MLR, i.e. the modelling of the causal effects is the key issue
Causal analysis: Typical variables

Trend (any function) Seasonality (dummy or Fourier swingings) Climatic conditions (e.g. temperature, precipitation) Economy (e.g. GDP, inflation, unemployment rate) Product specific (e.g. price/costs, new model/version, advertising, marketing activities, promotions) Demography (e.g. population in age classes) Others (e.g. life cycle, replacement demand, distribution)
Causal analysis: Typical business questions

What can we do in order to reach sales of X units ? What is the cost optimal solution ? How will the market react, if we (our competitor) increase the price by Y % ? How much of our sales are influenced by climatic conditions (e.g. ice cream, beverages) ? What is the risk of an economic weakness for our sales ? What is the target group for our advertising ? Which variables drive the long term development ?
Causal analysis challenges I
Data challenges
Historical data for all variables to be included in the model Comparison to competitors products seems often to be appropriate but data is not always available Forecasts are needed for all variables
Logical challenges
Which variables influence the development ? What are the effects of each variable ?
Causal analysis challenges II
Statistical challenges
Autocorrelation Multicollinearity
Modelling challenges
Outliers Trend Seasonality Effect of each variable
Time
dynamic Elasticity
Time
dependent weighting
Substancial experience is required for modelling causal effects !
Causal analysis advantages

Full modelling flexibility Explanation of the historical influence of causal factors Enables what-if simulations Risks / opportunities can be estimated The MLR results are easy to understand and can directly be used to optimize the strategies
Content
Data Preparation
2 3 4 5 6
Composite Forecasts
Evaluation
Composite forecast
Enables the combination of different forecasts with a constant or time dependent weighting Can be used for the one number principle in combination with an internal or external collaboration process Different studies proved that composite forecasts deliver in average a higher forecast accuracy than every single method included in the composite forecast
Content
Data Preparation
2 3 4 5 6
Composite Forecasts
Evaluation
Evaluation
The planning / forecasting process has to be reviewed permanently / in regular intervals. This can include the analysis of:
KPIs (e.g. service level, out of stock) Financial data (e.g. turnover, profit) Promotions / advertising Special influences (e.g. strike) Causal effects Forecast accuracy (-> model selection)
The exclusive comparison of the forecast with real sales is not appropriate !
The results should be documented and archived

APO DP Forecasting

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

APO DP Forecasting

Diunggah oleh

Hak Cipta:

Format Tersedia

Forecasting

Dr. Stefan Theis

Use of statistical methods

SAP AG 2001, SCM, Dr. Stefan Theis 2

Selection of Forecasting Method

Causal Analysis (MLR)

Data preparation (outlier correction)

SAP AG 2001, SCM, Dr. Stefan Theis 4

Input for forecasting

SAP AG 2001, SCM, Dr. Stefan Theis 5

SAP AG 2001, SCM, Dr. Stefan Theis 6

Forecast Error Index

Regression Forecast Method

Selection of Forecasting Method

Causal Analysis (MLR)

Selection of forecasting methods I

SAP AG 2001, SCM, Dr. Stefan Theis 9

Selection of forecasting methods II

The assignment can be based on

classification (e.g. part, standard)

purpose / business requirements

SAP AG 2001, SCM, Dr. Stefan Theis 10

Automatic model selection in APO

Model selection 1 (Parameters fixed)

Model selection 2 (Parameter kombination tested)

Never use model selection for regular updates !

SAP AG 2001, SCM, Dr. Stefan Theis 11

Comments on selection methods implemented in APO

SAP AG 2001, SCM, Dr. Stefan Theis 12

Alternative process using macros

SAP AG 2001, SCM, Dr. Stefan Theis 13

Macro: Measurement calculation

Very easy error measure not to be used in reality

SAP AG 2001, SCM, Dr. Stefan Theis 14

Macro: Compare measurement and assign profile

Forecast Key Figure (can be omitted)

SAP AG 2001, SCM, Dr. Stefan Theis 15

SAP AG 2001, SCM, Dr. Stefan Theis 16

SAP AG 2001, SCM, Dr. Stefan Theis 17

Result: Assignment of forecast profiles

SAP AG 2001, SCM, Dr. Stefan Theis 18

Selection of Forecasting Method

Causal Analysis (MLR)

Univariate forecasting methods I

SAP AG 2001, SCM, Dr. Stefan Theis 20

Basic weighting principle:

Weight of current value: Parameter

Weight of last periods value: (1 Parameter)

Example of weighting (in %) of historical values

First historical period

Second historical period

Third historical period

SAP AG 2001, SCM, Dr. Stefan Theis 21

Example: Forecast of constant history

Constant Forecast Methods

SAP AG 2001, SCM, Dr. Stefan Theis 22

Example: Forecast of trend-seasonal history

Constant Forecast Methods

SAP AG 2001, SCM, Dr. Stefan Theis 23

Univariate forecasting methods II

Season (without trend)

SAP AG 2001, SCM, Dr. Stefan Theis 24

Example: Forecast of trend history

Trend / Seasonal Forecast Methods