Demand Forcasts - best way

[OliverA41]

Hello,

I have a general question about the topic "demand planning" and "forecast".

What is your strategie to find the best forecast model. In some situation, IBP find a"best" profile, but if I checked it, it is really not the best profile.

Do you have key figure, wich help you. Except MAPE.

How do you start in the detail - phase to find the best fcst modell.

Regards

Oliver

[lev_degtyarov · Accepted solution]

Hi Oliver,

My recommendations are:

1st Read methodology & basic techniques for time-series forecasting. I can recommend to start from sites:

1. https://towardsdatascience.com/tagged/time-series-analysis
2. https://medium.com/search?q=time%20series
3. https://www.analyticsvidhya.com/

As for example check these articles. Skip any code (R, python,etc.) if it mentioned and try to understand the concept.

1. Almost Everything You Need to Know About Time Series
2. Everything you need to know about Time Series
3. Basic Principles to Create a Time Series Forecast

2nd: After you understand the concepts then you can go with SAP IBP Demand configuration. IBP provides you most of the tools are usually used in TS forecasting, so you next step is to build your "best of breed" model from available IBP Demand blocks. Here you may need to do the research and some of calculations\modeling in IBP upfront. To mentioned just a few:

• what is appropriate approach: bottom-up , top-down , middle-out forecasts?
• what is the best level for demand forecasting?
• when do you need outliers cleansing and handling of missing data?
• If you need to clean outliers, how do you then plan to do it for seasonal sales not cutting the seasonal peak volumes?
• what forecasting technique to choose: can the sales be represented as stationary series and you use SARIMA\ARIMA? Or may be you need to use regression models - MLR, Boosting?
• etc...

3rd: IBP for demand is not a "red button". Do not blindly rely on provided algorithms in SAP IBP Demand, understand the purpose why you decided to use particular algorithm (read sites above & IBP help).

4th: I do not recommend to use MAPE as the best-fit metric for multiple model selections. This is because MAPE tends to select underestimating forecasts model. Think about the case when you have:

• sales = 100 unit
• forecast #1 = 80 units (MAPE#1 = |80 - 100| / 100 = 80%)
• forecast #2 = 120 units (MAPE #2= |120-100| / 100 = 120%)

Best-fit model with MAPE will choose forecast algo #1.

Best practice is to use RMSE or MAD (also known as MAE) in model selections. MAD provides a protection against outliers whereas RMSE provides the assurance to get an unbiased forecast. You can easily google for details or watch related videos on YouTube "MAE vs RMSE".

But when to use MAPE then? MAPE is easy to understand by planners and especially by "top managers", in this case you will calculate MAPE.

I hope this help you to deep dive.

Best Regards,

Lev Degtyarov

(former demand planner)

[lingaiahvanam]

Hi,

Demand planning provides you with several tools to generate forecasts for various scenarios or for specific parts of your business. Using historical data and scientifically based statistical algorithms, it allows you to improve the accuracy of revenue forecasts, align inventory levels with predictable changes in demand, and enhance profitability for a given channel or product.

SAP has provided the following tools to do demand planning and forecast planning.

• Statistical forecasting
• Demand sensing
• Promotion integration
• ABC/XYZ segmentation
• New product introduction

Demand planning purely depends on the company strategy how they want to generate the forecast for product groups, channels, and divisions.

IBP demand planning process suggested using the "Best Fit" model to derive the better forecast and again it depends on the input values wherever applicable in forecasting methods.

After the forecast run planner can able to review the best fil method selection based on historical data and forecast error methods.

Lat point to think about the historical data cleansing if historical data able to align properly to feed the data for statistical forecasting, overall forecast accuracy will also improve.

MPE - Mean Percentage Error

MAPE - Mean Absolute Percentage Error

MSE - Mean Squared Error

RMSE - Root-Mean-Square Error

MAD - Mean Absolute Deviation

MASE - Mean Absolute Scaled Error

WMPE - Weighted Mean Absolute Percentage Error

TE - Total Error

TEA - Total Absolute Error

Please note that this approach is at a high level again it depends on business requirements demand planning and forecasting methods selections may vary.

Best Regards,

Lingaiah

[OliverA41]

Hi Lingaiah,

thanks for the response.

Do you mentoined the different Forecast Errors. Wat is the best fcst error and why is it the best for you ?

thanks in advance

[riyazahmed_ca2]

Hi Oliver,

Not every Forecast Model need to be the best for every Sales Pattern. IBP Demand module comes with the concept called Time Series Analyses with which we can understand the sales pattern and decide the forecasting model accordingly. Normally Sales Pattern would be Constant, Trendy, Seasonal, Seasonal Trendy, Sporadic. For each of the pattern resulted from the Time Series Analyses, we can employ the different forecast model meant for the pattern to result the best fit results based on minimal forecast error.

If you dont have IBP Demand License, Time Series Analyses would be a manual activity, You will have limited forecast model and only RMSE or MSE is available for best fit calculation.

If you have IBP Demand License, You can go ahead Time Series Analyses run and you have multiple forecast models and multiple forecast errors based on which best fit selection can be done.

Refer the forecasting algorithm section in the below link

https://help.sap.com/viewer/6b0a6820ebf94ff4a15d68af6db7745b/2002/en-US

Regards,

Riyaz