Forecast models were degrading across 13 markets, with MAPE as high as 45%. Nine weeks of rebuilding brought overall error to 4%
Multiple forecasting models supporting commercial planning and supply chain were showing measurable performance degradation across the business. The Massive Forecast model carried a MAPE of up ot 45%, and the trend was worsening. Products with many shade variants, such lipstick and nail polish with 20 or more tones, showed the worst accuracy rates. BIAS was unstable, overfitting was prevalent, and the train/test split methodology had never been properly validated.
Profitmind addressed four distinct forecasting models across a nine-week engagement structured in four sprints. The work introduced a percentile-based forecasting methodology that predicts value ranges rather than point estimates, reducing the commercial impact of forecast misses. Proper train/test splits and cross-validation were implemented for the first time. Hyperparameter optimization was performed, COVID-era data was excluded from training sets, and new features were engineered to improve signal quality. The approach was validated across all 13 countries simultaneously.
The results were consistent across every market: One Latin American country improved from 40% MAPE to 2%; Another from 35% to 3%. Every country saw material improvement, confirming that the methodology generalizes across diverse market conditions rather than being tuned to a single geography. The overall weighted MAPE reduction of 90% was achieved across 20 cosmetic product types over a nine-week engagement. These improvements directly reduce supply chain costs, promotional waste, and stockout exposure across the full portfolio.