A leading consumer goods manufacturer has a direct to retail distribution model in a market – load products on vans from their warehouses and sell directly to retailers. The client had deployed a model to predict the necessary loading for all products for all sales agents across sales offices. However, it was observed that the quantum of unloads at the end of the sales run was high. Consequently, the client wanted to leverage machine learning to optimize loading and reduce loading/unloading time and manpower cost.
INXITE OUT APPROACH
Features were generated for the following class of demand impact factors:
- Historical purchase pattern of the retailer for each product
- Pricing variations due to permanent price changes and short terms promotions
- Holiday and festival linked demand variations
- Impact of weather on the retail demand
To predict the retail demand for each product for each sales agent on a particular day, a teacher-forcing approach was adopted, where a sequence of models is used, starting with a Purchase Intent classifier, followed by a Purchase Quantity regressor.
- Purchase Intent classifier: Classification model was implemented to predict whether a particular POS would buy a particular product on a visit date based on their historical purchase patterns, promotions on offer, festival, events and other factors.
- Purchase Quantity Regressor: The model is designed to forecast the quantity of each product that a retailer is likely to purchase, provided that the Purchase Intent classifier indicates that a purchase is probable. The model was implemented to forecast the daily demand of over 40 products at each point of sale (POS) within more than 30 direct sales offices. This model employs over 30 engineered features that characterize the outlet profile, past purchase patterns, events, promotions, weather, location, and other relevant factors.
Additionally, a dynamic buffer has been implemented to improve the accuracy of the predictions.
- Carton loading and unloading quantity was reduced by 20-30% across the sales offices.
- Optimized man-power cost by 20%+ due to reduced loading/unloading and reconciliation workload.
- Stock shortages measured as percentage of days where the van returned with no stock was maintained at levels amenable for business