Statistical Forecast Optimization

Problem Statement

The client utilized SAP IBP Demand Planning for forecasting product demands. However, market disruptions—including natural gas shortages, price volatility, increased demand for heat pumps, and evolving German legislation—resulted in severe instability and poor forecast accuracy. When we joined the company most product time series were seriously affected by these external shocks. The forecast setup as configured at that time was not robust enough to absorb these shocks.

Additionally, accessory products, typically sold alongside core products, were forecasted independently, disregarding their correlated sales patterns.

Solution Provided

• Implemented ensemble forecasting and optimized SAP IBP configuration to enhance forecast stability and accuracy for core product and accessories.
• Developed a specialized Python software package to predict accessory demands based on core product forecasts and product lifecycle management (PLM) data and preparation of integration with the client’s IT infrastructure.

Implementation & Execution

Initially, the client used the Best Fit forecasting algorithm of SAP IBP, selecting the single best-performing forecasting model (e.g., ARIMA, Exponential Smoothing, Gradient Boosting) based on historical backtesting. Our team replaced this with an ensemble approach, combining multiple models into a weighted forecast. In this, the same set of forecasting models is fitted to the time series and the actual forecast consists not only of a single model, but their weighted sum. Optimal model weights were determined from simulations of the SAP IBP forecasts with a certain holdout period and a Python-based grid search. The ensemble weights were periodically reoptimized for core products, accessories, and specific product subsets with notable seasonal demand patterns.

A “Core Product Governed Accessory Forecasting” (CPGAF) software was developed to leverage correlations between core products and their accessories, capitalizing on improved forecastability of core items. This solution incorporated historical correlation analysis and PLM insights to predict accessory demand effectively.

Moreover, our team provided detailed data analytics to identify key contributors to the forecasting accuracy across, e.g., individual product groups and countries. Further insights were gained in the analysis of different forecast levels, varying the product hierarchy and forecast interval.

Results & Impact

• Achieved a significant improvement in forecast stability and accuracy, boosting performance metrics by 15% to 30% across different countries.
• Successfully implemented the Python-based CPGAF package, enabling more accurate accessory forecasting driven by core product predictions.