Business Objective
Our client is a global Fortune 500 quick restaurant chain operating multiple brands in 150 countries.
The client depended on a simple forecasting process using recent historical data and a pre-set adjustment applied for corporate events. The forecast engine was a black box with limited visibility into demand drivers. The objective was to help in developing forecasting models at the menu item/sales/transactions x store x daypart (15 min) level for the upcoming 4-week period factoring in potential drivers of demand. Subsequently, the model was to be used to convert item forecasts into inventory orders and also for labor planning.
Challenges
- Variety of data sources – data from historical sales, internal and external demand drivers
- Understanding the client’s existing black-box model
- Errors in backend migration resulted in inconsistent inventory data for one full year
- Revolving menu with limited period offers that kept changing all the time
- Arriving at a forecasting granularity of 15-minute interval for labor planning use case
Solution Methodology
- Developed a robust forecasting model for all sales, transactions and menu items at store x daypart level for the next 4-week period using Amazon DeepState
- Used the item level demand forecasts for:
- Inventory order forecasting by aggregating demand forecasts and factoring on-hand and previous order deliveries
- Labor optimization first required defining the various labor roles, their key drivers. Built forecast models for each driver type at appropriate granularity (example – service, order and delivery, required transaction-level forecast at 15-minute interval)
- Incorporated labor demand drivers into the model, eliminating the need for store managers to adjust forecasts manually
Business Impact
- Average inventory holding cost across the test stores dropped by ~50% for daily counted items and by ~8% for weekly counted items considering orders based on the new algorithm
- Developed a robust forecasting model with an accuracy that is 3-5% better than existing models. Improved labor forecasts for all role types led to an efficient roster that could be built three weeks in advance
