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
Content
Data Preparation
2 3 4 5 6
SAP AG 2001, SCM, Dr. Stefan Theis 3
Univariate Forecasting
Composite Forecasts
Evaluation
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
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
2500
2000
1500
1000
500
Const
Trend
Seasonal
Seasonal Linear
Linear Regression
Trend - Season
Original history
SAP AG 2001, SCM, Dr. Stefan Theis 7
Cleaned history
Results based on 44 SKUs
Content
Data Preparation
2 3 4 5 6
SAP AG 2001, SCM, Dr. Stefan Theis 8
Univariate Forecasting
Composite Forecasts
Evaluation
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
planning
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
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
4) Run batch job with above activity for all characteristic combinations 5) Result: Optimal forecast profile is assigned per characteristic combination
Activity
Batch job
Content
Data Preparation
2 3 4 5 6
SAP AG 2001, SCM, Dr. Stefan Theis 19
Univariate Forecasting
Composite Forecasts
Evaluation
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
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
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
Trend
Exponential smoothing 1st order / Holt (20/21) Exponential smoothing 2nd order (22) Exponential smoothing 2nd order with a - adaptation (23) Linear Regression (94)
Trend - Season
Exponential smoothing 1st order (40/41) Seasonal linear regression (35) New in APO 3.0 !
Month
Hist
FS20/21
FS22
FS23
FS30/31
FS40/41
Month
Hist
FS20/21
FS22
FS23
FS30/31
FS40/41
Others
History (60) Manual Forecasting (70) No forecast (98) External Forecast / User Exit (99) New in APO 3.0 ! New in APO 3.0 !
Trend dampening
There is only slight statistical methodology behind these (more or less) manual modelling procedures !
Content
Data Preparation
2 3 4 5 6
SAP AG 2001, SCM, Dr. Stefan Theis 30
Univariate Forecasting
Composite Forecasts
Evaluation
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 ?
Statistical challenges
Autocorrelation Multicollinearity
Modelling challenges
Outliers Trend Seasonality Effect of each variable
Time
dynamic Elasticity
Time
dependent weighting
Content
Data Preparation
2 3 4 5 6
SAP AG 2001, SCM, Dr. Stefan Theis 37
Univariate Forecasting
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
SAP AG 2001, SCM, Dr. Stefan Theis 39
Univariate Forecasting
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 !