Refining sales forecasts for short-lived products, driving innovation

APPROACH

The TRENDY project leverages the combined expertise of the Fraunhofer Innovation Platform at the University of Twente and Innico to address the challenges of sparse time-series sales forecasting. The project begins with a comprehensive analysis of existing forecasting solutions, identifying their limitations when applied to products with short market lifespans. A critical focus will be placed on uncovering meaningful product features and incorporating these into advanced time-series forecasting models. By researching sub-problems such as integrating salesperson-driven insights and use-case-specific forecasting methods, the project aims to tailor solutions to the unique characteristics of sparse datasets.

The developed model will be rigorously tested, benchmarked against state-of-the-art alternatives, and integrated into Innico’s ERP systems. Explainability is a key focus, with intuitive visualisations and metrics ensuring transparency and usability. Dissemination activities, including reports, workshops, and technical documentation, will facilitate knowledge-sharing and ensure the model’s applicability across industries, improving sales forecasting for fast-moving, volatile markets.

OUTCOME

The TRENDY project successfully evaluated different forecasting models for different use-cases. To facilitate model comparison, an automated evaluation pipeline was developed and integrated into Innico’s web application. A total of 10 different forecasting models were implemented and benchmarked against 3 state-of-the-art methods. The models were assessed using five key performance metrics, and their impact was tested in a warehouse simulation environment to analyse real-world applicability.

• XGBoost outperformed the state-of-the-art but tended to underestimate, whereas the state-of-the-art overestimated
• Traditional models performed moderately, offering no significant improvements
• Zero-baseline model ranked second-best for lowest demand product categories, highlighting the research gap
• Warehouse simulations showed XGBoost and industry standard performed similarly overall
• Forecasting error reduced 25% for sparse items, stable for high-volume

The findings provide insights into the importance of integrating forecasting with stock strategies rather than relying solely on model accuracy. Future work will focus on item life cycles, transfer learning, and synthetic data to improve forecasting, especially for new products. Enhancing warehouse simulations with backorder management and dynamic ordering will further optimise inventory decisions.

This project was made possible through the Regio Deal supported by the Province of Overijssel and the Dutch State.

FOR MORE INFORMATION ABOUT TRENDY, PLEASE FEEL FREE TO REACH OUT TO