Business Objective
Our client is a leading US-based grocery retailer with 100+ categories and 10,000 + SKU’s. The current inventory planning process for promo and non-promo time periods relied heavily on business rules developed over time. There was a need to revisit those and develop predictive models in order to have a robust demand forecasting process that accurately accounts for promotional impact.
Challenges
- Building a full-blown forecasting solution at an Item x Store x Day level
- Forecasting for recently introduced products and brand new products
Solution MethodologyÂ
- Collected sales and promotional data. Blended them to have a uniform Item x Store x Day level granularity.
- Pre-processed to de-trend, de-seasonalize, and remove holiday effects
- Built models for baseline and promotional lift estimation, using a combination of time series and frequency-based approach (for sparse data).
- Developed Sub-Category/Category models for SKUs that didn’t meet range (# of stores) or time period criteria
- Used store level daily sales distribution indices to break down weekly to daily sales
- Identified similar products for new products by clustering for 10+ features (like sales, regional preference, complement, substitute, promoted group, etc.) and modeled based on them.
- Automated end-to-end process and enabled calibration for daily/weekly refresh.
Business Impact
- Solution tested against current approach: yielded item x store x day level forecast accuracy improvements ranging from 5% to 80%. Weighted average improvement was around 15%.
- Estimated business value due to stock-out avoidance and improved service levels was around 1.5% to 2% on sales.