How Instacart's item availability evolved over the pandemic

Viswa Mani Kiran Peddinti17 Jul 2023

Problem

Instacart sells items from more than 80,000 physical stores, each with its own inventory system, and shoppers pick the items by hand. It cannot know in real time whether an item is on the shelf, and an ordered item is not reserved. During COVID, demand surged and store inventory changed quickly. The availability model could not keep up, so the found rate (the share of ordered items the shopper finds) fell and out-of-stock cancellations rose.

Example from the article: Smaller spikes in out-of-stock cancellations still appear around holidays, floods in New York and baby formula shortages.

Solution in brief

Availability is a trade-off: show many items and the found rate drops, show few and available items are hidden. Thresholds on top of the model decide what is shown and can change without retraining, per retailer, location and category. The team also measured how found rate affects retention and added low-stock badges.

Questions this case answers
  • How to choose a display cutoff on a probabilistic availability score so that the size of the shown catalog and the fraction of fulfilled requests stay balanced
  • How to keep a prediction-based decision rule near its optimum during abrupt demand or supply shifts, when retraining the model is too slow
  • How to quantify the effect of order quality on long-term customer retention when the quality level cannot be assigned at random
  • How to decide which items to show when availability predictions are uncertain and shoppers may not find them?

Tools and methods

  • Availability probability model — estimates the chance that a shopper finds an item in a given store
  • Dynamic thresholds over the model — decide which items count as available; tuned per retailer, location or category
  • Found rate vs. retention analysis — experiments and analytical models that set the target found rate
  • Low-stock and "many in stock" badges, "Request" button — tell customers about availability risk
Continue reading — it's free

Google or a one-time email code. No password.