SAP IBP for Demand


After completing this lesson, you will be able to:

  • Explain what's new for Demand

SAP IBP Demand Forecasting

  • Prediction intervals provide information on the expected minimum and maximum forecast for any confidence level set by the user
  • In case of this real-life example, the range calculated for 80% confidence covered 10 points out of 12 of the actual sales.
  • The range for 95% confidence only missed one point out of 12 of the actuals.
  • Maximum two confidence intervals can be calculated at a time.
  • Users can save either or both the lower and upper boundaries for any of the intervals.

Prediction intervals are now available for all Exponential Smoothing algorithms

  • If you add multiple algorithms to the forecast model, intervals are only calculated in Choose Best Forecast mode, and only for those planning objects where the winning algorithm supports it
  • There is no main output key figure for prediction intervals
  • Prediction intervals are calculated based on the distribution of the historical data and the predictable patterns like seasonality and trend

About prediction intervals in general  

  • Intervals are not calculated for test phase
  • Intervals are only calculated for the forecast horizon, not for the past (ex-post ranges are not calculated)
  • Global configuration parameter MIN_FORECAST is applied to prediction intervals, too
  • Phase-in and Phase-out curves are applied to prediction intervals, too

Some background - What is Gradient Boosting of Decision Trees?

Gradient boosting of decision trees is a machine learning technique that creates a prediction model by combining multiple weak decision trees models.

Each decision tree is not built independently, but rather each one depends on the previous since it aims to correct the errors of the total ensemble to that point. This model can be used for both regression (forecasting use case) and classification problems.

Extensive Testing was completed on several large customer data sets.

Basic Settings: 104 Weeks of history, only sales history, test setting of 13 weeks to determine best algorithm

****Individual results will change depending on settings and data used.

Log in to track your progress & complete quizzes