Uncovering Critical Products in Retail Baskets: A Predictive Modelling Approach to Increase Order Fulfilment

Order fulfilment is a key goal for retailers. It is impacted by item unavailability, poor item quality and delivery mishaps. Although the north star is to resolve these challenges completely, in practice, retailers would benefit by prioritizing perfect fulfilment of a few critical items for customers. This is based on the hypothesis that certain items are more important to the customer and their fulfilment impacts customer satisfaction to a large extent. This leads us to the problem of predicting what is critical for a customer. Since this can vary from one purchase to another, we propose a solution to identify critical items in a single purchase (‘Basket’) in real-time. These crucial items are termed as ‘Basket Breakers’ because failure to fulfil them results in low customer satisfaction with the purchase. In this work, we discuss a data science methodology to identify the set of k-critical items and evaluate how many of them are critical-in-reality. We use a combination of inventory, price, item-properties and historical feedback in our method.

The proposed solution is a part of Walmart’s Substitution program. This program was rolled out to explicitly seek customer’s substitution-preferences for each of the basket-items. Although retailers would want to know customer’s explicit substitute-preferences for every item, this is cumbersome and time-consuming for the customer. Hence, the customers are prompted to give their choice for a few critical items alone. When prompted for the right items, this serves well to compensate for imperfect inventory counts and sub-optimal substitutes for out-of-stock items. The proposed solution is highly relevant for Grocery and other fast-moving categories where an item may be available at the time of customer placing the order, but may go out-of-stock at the time of pick-up or delivery, leading to partial order fulfilment. Inventory prediction for these categories is challenging due to the lack of precise data as well as the dynamics in real-time purchase patterns. This makes the identification of a few ‘critical’ items even more impactful. Our offline evaluation results and online-production tests are encouraging. The results also show significant impact on business metrics involving order fulfilment and customer satisfaction.