The Fifth Elephant 2019

Operational excellence is one of the key tenets in any retail business. Promotions are a core part of any price investment strategy in a high-low market. Promotions involve cost in providing discounts and other supports. Efficiency in utilizing the budget available for the most rewarding price investment strategy is what we are driving through this paper. The investment required for reduction of price is being optimized to provide maximum returns. The solution is developed at category level to assist category managers in investing their resources.
This is a modularized solution workflow with re-usable components: Candidate Item Selection, Volume Prediction and Optimization.
For Candidate Item Selection, after filtering the items that contribute to top 95% of sales in the category, a blend of three defining factors of items is being considered. A heuristic method to determine a score that defines the importance of an item in the category. An unsupervised clustering based approach based on the sales driving characteristics of items to exclude items that are unsuitable for promotions. An elasticity score derived from a price-volume regression model to determine the responsiveness of volume to the discounts. The items that match the criteria of heuristic score threshold and elasticity threshold are included after eliminating the cluster of items that unresponsive. Another factor considered into the mix is the flavor group concept. A Flavor group is a group of line items that need to be priced similarly to maintain the consistency rule of pricing. Hence, if one item of the flavor group is being considered a Key Item, then the complete flavor group is to be considered to be given the same discount.
With the candidate items selected, our novel forecasting technique provides weekly demand predictions for each item. First, we introduce sophisticated statistical corrections to estimate lost-sales adjusted demand as a function of daily observed sales and on-hand quantity at a store/item level. Thereafter, we employ optimal time series methods to predict future demands using historical demands along with external variables such as price, promotion, holiday index etc. Demand prediction for different discount depths for each item is done for different promotion durations. Applicable discount depths and durations are determined from historical patterns, with the provision of adding more inputs. This two step approach ensures robust forecasts of not just observable sales but in fact hidden demands of every item ensuring a holistic review of the past.
The outputs of the previous step is then fed into a dynamic programming based optimization technique to maximize revenue given the constraints of budget and same pricing for flavor group items. Revenue is calculated as the sales of predicted volume at the discount depth. Budget is considered as the cost involved in providing the difference of price due to the discount depth for the predicted volume of items. Different budgets are estimated based on constant discount depth scenarios for all candidate items. The algorithm is a customized approximated version of dynamic programming with forward and backward propagation, that allows for scaling up of the solution with lesser than O(n2) complexity.
The solution provides data-backed recommendations for investment into the most effective promotions. Recommendations are provided for each budget scenario and a budget with maximum returns is also recommended for future planning. The projected gain from the solution is 15-20% higher than a constant discounting practice.

Outline

Operational excellence is one of the key tenets in any retail business. Promotions are a core part of any price investment strategy in a high-low market. Promotions involve cost in providing discounts and other supports. Efficiency in utilizing the budget available for the most rewarding price investment strategy is what we are driving through this paper. The investment required for reduction of price is being optimized to provide maximum returns. The solution is developed at category level to assist category managers in investing their resources.

Speaker bio

Abhishek Mungoli ~ Data Scientist, WalmartLabs
Ashwini Chandrashekharaiah ~ Senior Data Scientist, WalmartLabs
Diptarka Saha ~ Data Analyst II, WalmartLabs