DSS25 - Improving Forecasts: Compositional Data in Nestlé's (...) - N. Szwagierczak, W. Stasiuk

Full Title: Improving Forecasts: Compositional Data in Nestlé's Demand Planning Within Nestle, thousands of markets - retail customer - product combinations need to be forecasted every day. Demand planners require SKU-level forecasts, but input data, including planned promotions and sales, is often available at a higher level, such as product groups. This creates a challenge, as splitting forecasts from higher to lower hierarchy can lead to losses of 10-30 percentage points in Demand Planning Accuracy (DPA), measured by 1 - Weighted Mean Absolute Percentage Error (WMAPE). This problem can be addressed by so-called Compositional Data Analysis (Aitchison, 1982). To tackle the issue, we tested four methods: hierarchical modelling using minimum trace, VARMA with center log ratio transformation, Dirichlet regression with covariates, and a grid search across all possible values. We developed an algorithm using Dirichlet Regression and estimation of forecast risk based on historical SKU share distributions. This method resulted in improved KPIs, with an increase in DPA and a double-digit percentage point reduction in bias on back-test data. Consequently, the innovations translate into substantial reduction in inventory (~1.5%) and shortages (~2.5%), based on common business heuristics.