Challenge

Significant challenges in data quality and availability, with inconsistent and incomplete data for marketing spend, product features, and reseller promotions, and actuals data only available from April 2019. Ensuring forecast accuracy, particularly for end-of-life periods and capturing local promotions, added to the complexity, along with the need for model generalizability across regions like Germany, Great Britain, and France.

Approach

Conducted a thorough data audit, identified and documented gaps, and developed and validated models using Machine Learning, Statistical, and Heuristic approaches. They ensured the explainability of forecasts by analyzing the impact of product features, holidays, and past trends. Models were created at the Country-Product Family-Week level and disaggregated to the Reseller-SKU level, incorporating business feedback for finetuning and integrating into Kinaxis.

Outcome

This approach resulted in improved forecast accuracy, with over 75% monthly accuracy for 6 out of 9 Product Family x Country combinations and insightful forecasts highlighting the importance of holidays and product features on sales. Ultimately, detailed models were successfully integrated into Kinaxis, and the forecasting process was comprehensively documented and automated.