There was an interesting technical issue at a client using SAP SCM with substantial investment in demand planning using SCM tools. After some googling, I landed on this blog site written by a consultant who obviously makes a living on SAP. But strangely enough, this consultant has created a complete web of blog entries filled with complaints about SAP and APO particularly demand planning. The sole purpose of his writing seems to be scathing attacks on SAP. Unfortunately there was not any content that could solve problems or answer questions. Surely I did not find an answer to my question on his blog.
He does have pretty good complaints – very well articulated. These complaints are not new about SAP APO. Most users know these complaints and they live through this every day to their frustration. I have articulated them before – some of the problems are tool specific that SAP has not bothered to correct. It is a shame the company will not listen to the users or read the thousands of OSS notes written every day. The problems that SAP has to be blamed for include issues such as forecast error calculations, incorrect alert logic, manual effort required in life cycle planning among others.
However there are many problems that are created by the Integrators/Consultants including our “esteemed” blog writer due to their lack of knowledge of both demand forecasting as well as an understanding of the tool. If they don’t understand the tool, they should not be implementing it or training the users. The baggage they leave behind creates a mess that makes the software worthless and unusable. These problems include:
1. Implement APO to be used as a typing tool and tell the users Statistics are terrible so not worth using it.
2. Disable statistical modeling under the pretext of security. In reality, the consultants are worried about fielding questions from the users on statistics that they do not understand themselves.
3. Enabling options and parameters in the tool with out any idea of what they do to the resulting forecasts.
4. Deciding on important things such as forecast aggregations, forecasting levels and exception management without any process discovery or user input.
5. Finally not project planning the budget to include training to the users particularly on the Stats and Demand Planning.
Perhaps companies should divide the implementation plan into two parts – process design that includes the provision for model tuning and training and the second part that involves system design and making the system to work. The system design should follow the process design. The model tuning should be done by the same consultant that is responsible for process design. And definitely that person should be an Expert not only in the tool but also in best practice demand planning along with expert skills in training the planners.
Given how many implementations are currently plagued with tool problems aggravated by Consultant inefficiency and incompetence, perhaps more companies should think about re-implementations of APO DP. Just throw away the old concepts and practices and start thinking about how to fix the problems and make the tool more usable.
Finally SAP also needs to wake up and start fixing the problems in their most popular SCM module namely APO Demand Planning. It needs to fix the error calculations and the alert logic. More on that in a separate blog entry.
To set the record straight with this “esteemed” writer so he does not pollute the waters and mislead many planners, I would say the following:
The blog writer concludes that SAP APO calculates the MAPE incorrectly. I agree with him to a certain extent. SAP purports to compute MAPE using the academic definition of averaging percentages but it does not do this either. It goes into a hole when the actual demand is zero and makes the MAPE metric unusable. However, the other metrics namely MAD and RMSE are correct.
I strongly disagree with this Consultant/writer when he concludes that best fit models are erroneous because the error calculations are defective. If you poke around the underlying mechanics which is well documented in the APO online manual, you will know that the optimization is done using the Mean Absolute Deviation, which is superior to the MAPE; Mape is a percentage and has some awkward properties.
The Automodel selection 1 uses the MAD and picks only smoothing models while the Automodel selection 2 also claims to include linear regression models. But in practice Automodel selection 2 produces inferior results and expects the user to baby sit the modeling by feeding manual parameters!! In general Automodel is not for the faint of heart as many settings have to be correctly configured. Given the fact most configurations are done by junior consultants from big 5 consulting firms, you can guarantee that this is an unrealistic assumption.
Even with other forecasting packages we do not recommend best fit models or expert selection as the final model. They are good starting points, but the planner has to do more in getting to the right model and the demand forecast. They don’t have to be expert statisticians but they need to understand their business and have a preliminary understanding of what various models do.
Yes the Statistical models are straight forward in APO, in fact, they are basic. There is no complexity in them. They are not claiming to do Box-Jenkins or Transfer functions or ARMAX or any other models with esoteric names. However, I have found people use the MLR models very cleverly combined with forecast attributes. So it is all in implementation, model tuning and finally imparting that much elusive knowledge to the planners and find a way to sustain that knowledge.
My two cents = Do what you can and understand what you cannot. Some intellectual honesty will also go a long way!
Happy Labor Day weekend!