Although a significant business benefit can be gained, we suggest that you use such advanced methods with care. The forecast can become a black box and then the planner will not trust the result at all and will instead overwrite these. Furthermore, it is important to mention that the Gradient Boosting algorithm is not good for trending items and the algorithm can potentially overfit the training data and thus correspond too closely with the sales history and give a poor forecast.
Machine learning is now applied in multiple ways to SAP IBP. First Demand Sensing was introduced as a short-term forecasting algorithm that uses pattern recognition to automatically calculate and update regression weights to estimate the short-term daily forecast. In release 1808, K-Means and DBSCAN were introduced for custom alerts. K-Means and DBSCAN are clustering methods used to avoid the static threshold definition of alerts, which is useful when patterns in the data change, and to avoid too many or too few alerts.
4. Excel front-end filter using key figure values enables more exception-based planning
With the 1811 upgrade, a new type of planning filter becomes available when you create planning views. The new value-based filters allow filtering on key figure values. The new filters add on to the classic attribute filters, and a user is now able to shrink the planning view even further and focus on the things that are relevant to see. For instance, you may use a filter on the product IDs where the forecast for the next quarter is below a certain threshold or filter out combinations where a certain key figure is empty.
The filter’s time level is independent from the one selected in the planning view – you might use a filter on historical or future values not necessarily visible in the view. However, as for this upgrade, you can only apply one value-based filter per planning view, and the key figure needs to be part of the view.