The Challenge

• Non-AI based algorithm used to generate forecast for the base (fallback) region. Uses factors to calculate forecast for other regions: Lack of data (duration as well Covid patterns)
• Unclean attribute-data for products
• Inability to data for all regions: Multi-region forecasts (original processing allowed for processing only for one region)

Our Solution

A suite of artificial neural network (ANN)-based models was implemented to supplant a legacy algorithmic approach for demand-potential prediction across 12 regions encompassing over 290,000 stock-keeping units (SKUs) distributed across diverse product lines.

Our hybrid solution selects predictions from different models (FB Prophet, ANN, TFT) across different product groups, combines forecasts from different models, and further optimizes and tunes models for different product groups.