Skip to Content
Product Information
Author's profile photo Laura Tozzo

How to Take Advantage of Machine Learning Forecasting Algorithms in SAP Integrated Business Planning for Demand

Hi there!

I am a Product Manager for SAP Integrated Business Planning for demand at SAP and I am here to share some insights on how to leverage SAP IBP machine learning capabilities.

Let’s assume you are a demand planner who knows that the forecast of your product(s) does not depend only on the sales history, but it is influenced by other external factors. Let’s say you are interested in knowing which are the factors that influence your forecast, what is their importance and what is their impact.
Then you are in the right place! You can take advantage of the Machine Learning capabilities of SAP IBP for demand (available now!), based on the forecasting algorithms included in the Predictive Analytics Library (PAL) of SAP.
In particular, the algorithm we are interested in is Hybrid Gradient Boosting of Decision Trees (HGBDT).

Let’s have a short deep dive in the algorithm.

Gradient Boosting of Decision Trees – a short explanation

Gradient Boosting of Decision Trees is an ensemble machine learning technique for regression and classification problems.

Let me shortly explain what “Decision Trees” and “Gradient Boosting” stand for.

decision tree is a tree-like graph in which each internal node represents a “test” on an attribute (e.g. whether a coin flip comes up heads or tails), each branch represents the outcome of the test, and each leaf node represents the actual output/decision taken (or class label).

Decision trees belong to a class of supervised machine learning algorithms, which are used in both classification (predicts discrete outcome) and regression (predicts continuous numeric outcomes) predictive modeling.

The goal of the algorithm is to predict a target variable from a set of input variables and their attributes. The approach builds a tree structure through a series of binary splits (yes/no) from the root node via branches passing several decision nodes (internal nodes), until we come to leaf nodes. Source

Decision trees are called “weak learner”, because the predictions made by a single decision tree are better than a random method but still not very accurate.

Therefore the need for “Gradient Boosting”.

In statistical terms boosting is a method of additive modelling designed to provide a better predictive performance by combining multiple simple models together. In this case, Gradient Boosting takes the results of multiple decision trees, and looks to reduce the errors from each iteration sequentially to create a “stronger”, more complex predictor. A robust predictive model is therefore built by the linear addition of multiple decision trees, i.e. “Boosted”. Source

If you are already an expert on Gradient Boosting and would like more details on how it is implemented in SAP’s PAL see the official Help page.

Variable Impact Analysis – how it works

When using HGBDT to forecast a product in SAP IBP for demand, multiple things happen.
The Gradient Boosting algorithm first learns how the independent variables affected the sales in the past (even considering the combination of variables).

Then the forecast will be generated based on these learnings.
The forecast generation is independent from the variable impact analysis.

If the variable impact analysis is required, then the forecast (let’s call it total forecast from now on) will be split into a baseline forecast (including trend and seasonality) and impact key figures (for those of you who are not familiar with SAP IBP, consider key figure to a be a synonym of time series).

Baseline and impact key figures can also be defined for Ex-Post forecast.
You can choose the same key figure for more than one variable if you wish (then the impact of the variables will be summed up).
If no impact key figure is defined for a variable, the impact calculated for it is added to the baseline forecast or baseline ex-post forecast.
Small technical comment: if you choose the same baseline key figure for the forecast and the ex-post forecast, you should also choose the same impact key figure for them.

In the following example I am going to use the independent variables also mentioned in the image above: Marketing Budget, Product Price and Calendar Events (for which I use a categorical key figure), and corresponding impact key figures Marketing Budget Impact, Product Price Impact and Event Impact.

How to set up a forecast model with HGBDT in SAP IBP for demand

In the following I will assume you are familiar with the SAP Integrated Business Planning solution and therefore you already know how to set up a planning area, planning objects, key figures, application jobs, application job templates etc.
If you are not familiar with this, you may understand the content only partly, but this may not bring a lot of value for you (so I suggest to stop reading here!).

So let’s get started….

Moving forward I assume the following conditions have already been met:

  • Historical data (sales history) has been gathered
  • Historical data (sales history) has been cleansed.
  • Independent (external) variable data is already present in SAP IBP (as already mentioned before I will be using marketing budget, product price and a categorical key figure for calendar events).
  • Impact key figures have been defined (in the following example they are Marketing Budget Impact, Product Price Impact, event Impact)

Go to “Manage Forecast Models” and create a new model.

In general section decide a name and the granularity of the forecast together with the number of weeks that should be considered in the past and in the future.

Add any preprocessing steps you might need. Please consider that with Gradient Boosting correcting outliers might be counterproductive, as they may be due to the impact of an independent variable and removing them could hinder the learning process of the algorithm.
Note: if you use preprocessing step Outlier Correction with Outlier Correction Method “No correction” (i.e. the system identifies the outliers but does nothing about them), then IBP can generate a feature called “Periods with outliers” that can be consumed by HGBT.

In the “Forecasting Steps” tab select input key figures (sales history) and output key figures (total forecast).

Then select “Gradient Boosting of Decision Trees” as an algorithm.

At this point you already need to know which external variables are of potential interest in the computation of the forecast. You can add them in the “Independent Variables” section.

In this example I want to analyze the impact of the variables.
As explained in the previous section, this is not a must: you could also just use Gradient Boosting to compute your forecast using the independent variables without variable impact analysis. In this case the Gradient Boosting algorithm will still take the independent variables into consideration when generating the forecast, but you would not be able to see how they affected the generated forecast.

If you do decide to analyze the impact of the key figures, however, you have to check the box “Analyze Impact on Forecast” and define impact key figures for this purpose. For non-categorical variables you also need to define the Zero-Impact Setting.

To complete the settings of the model, add any additional post-processing steps you need.

Finally, save the model.

The Zero-Impact Setting

The Zero-Impact Setting will have a big effect on the variable impact analysis: this setting should represent the “usual” value of the variable you are considering.
In my example, the marketing budget is usually zero, and only from time to time some budget is allocated. Then the expectation is that whenever the budget is zero there is not impact on the sales, but when it is positive the sales will change.

For the product price (again, referring to my example) it is different: the price of a product is never expected to be zero. Rather, it will have an average value, and whenever the price varies from the average sales are affected.
For this reason in this case the Zero-Impact Setting is not zero but rather the mean value.

Run Statistical Forecasting Job from Fiori or Excel

You can run the Forecasting Job from Fiori or through the SAP IBP Excel Add-in.

In the log you can see the importance of the external variables:

And after running the job this is how your forecast could look like:

I would like to highlight once more that one of the strengths of HGBDT, compared to other multivariate algorithms like MLR or ARIMAX, is that it is able to estimate the combined impact of multiple independent variables.
For example, let’s assume you have two independent variables. Then HGBDT is able to find out that if any of the two variables, then sales increase as well, but if both of them increase at the same time, then sales drop.

So that was it! I hope this helped you gain some insights into the Machine Learning capabilities of SAP IBP for demand.

Go and try it out yourself!


Find out more:



Assigned Tags

      You must be Logged on to comment or reply to a post.
      Author's profile photo Irmhild Kuntze
      Irmhild Kuntze

      Really a very nice description and helpful links, thanks!

      Author's profile photo Lokesh Reddy
      Lokesh Reddy

      Very Helpful...Simplified explanation...Thank you

      Author's profile photo Dario Scaini
      Dario Scaini

      Great article Laura Tozzo , thanks!

      One question, please: what's the meaning in the log of the "_LEFTx"?



      Author's profile photo Laura Tozzo
      Laura Tozzo
      Blog Post Author

      Hi Dario,

      this is because of the offset settings. As you can see in the screenshot, Marketing Budget has a period offset of +1 and Product Price has an offset of +2.
      The log shows the impact of each key figure and then also of the key figure shifted of the corresponding period offset.
      So _LEFTx means the key figure had a period offset greater or equal to x.
      (It will be LEFT for +offsets and RIGHT for -offsets)

      Hope this answers your question.



      Author's profile photo Dario Scaini
      Dario Scaini

      Great, thanks!



      Author's profile photo Alejandro Serrano
      Alejandro Serrano

      Hi Laura,

      You've written a very nice article. I've also tested this functionality and i was surprised when I realized that the sum of baseline forecast and impact key figures is not equal to the total forecast. In your figure I understood that this sum should be equal to the total. 

      • Am I missing any setup in the model?
      • Might be related to the level where I launch the GB?
      • Is it mandatory to set an Impact Key Figure for Forecast for all the GB variables?



      Alex Serrano & Sergi Moros

      Author's profile photo Laura Tozzo
      Laura Tozzo
      Blog Post Author

      Hi Alex,

      the sum of baseline forecast and impact key figures is always equal to the total forecast, unless the sum is below zero and you have set the parameter MIN_FORECAST in the Parameter Group FORECAST equal to zero. In this case the Total Forecast will show 0 and it will not be equal to the sum.
      Is this your case?

      Answering your other question: no, it is not mandatory to set an impact key figure for all the variables. If you do not set a key figure than the impact will automatically be summed to the baseline.

      Hope this helps



      Author's profile photo Guillermo Torrens
      Guillermo Torrens

      Hi Laura,


      we are implementing SAP IBP. We have already implemented the Times series Forecasting as a Best Fit calculation.  Nos we want to introduce Machine learning. The question is how do we manage Times Series  +`Machine learning?  I mean, do we include the machine learning algotrithms in the models we are launching now with the rest of exponential smothing...?  Or there is a different better aproach?  We are a bit lost how to mix the times series with gradient bossting models.


      Many thanks in advance

      Author's profile photo Laura Tozzo
      Laura Tozzo
      Blog Post Author

      Hi Guillermo,

      you can have classical time series forecasting approached and Machine Learning in the same Best Fit model, and I would suggest to do so. The system will then pick the algorithm that performs best, independently on whether it is a classical forecasting approach or machine learning.

      Best regards


      Author's profile photo julio Andrades
      julio Andrades

      Hi Laura, I found this very article about Machine Learning, we implemented SAP IBP two year ago, and now need to include independent variables, for this we need to create a News Key Figure?, how can almacenate the información.

      Thanks a lot



      Author's profile photo Laura Tozzo
      Laura Tozzo
      Blog Post Author

      Hi Julio,


      yes independent variables are technically speaking just additional key figures that you add to your forecast model.

      Best regards