Walmart — Demand Forecasting — Product Case Questions
Role context: Senior Data Scientist, Demand Forecasting (US Stores & eCommerce) · Est. study time: 60 min · 5 questions
How to approach product cases here
Every case is one move in the same chain: understand the goal, break it into a data problem, choose the metric or method, name the bias and the trade-off, land on a decision and an action. Walk it out loud and survive the follow-ups.
Five things make Walmart forecasting cases distinctive.
1. A forecast is not a prediction, it is an order. Every number becomes physical inventory that either sells or gets binned. That is why the business metrics are waste dollars and in-stock, and why accuracy is explicitly the secondary metric in Walmart's own published account.
2. You observe sales, not demand. A stockout produces sales of zero with demand above zero. Train on that and you teach the model that out-of-stock items are unpopular, so you order less, so it stocks out again. The model's own output contaminates its future training data. This is the single most important data problem in retail forecasting and most candidates have never considered it.
3. The right forecast is a quantile, not a mean. Over-forecast a banana and you throw it away; under-forecast and you lose a sale and maybe the whole trip. The costs are unequal and item-specific, so the optimal order is the newsvendor quantile , not the conditional mean. MSE gives you the mean, which is right for nothing.
4. Waste and in-stock are the same dial. Order more and both rise. Any proposal to improve one without paying the other is either changing forecast quality (the real answer) or hiding the cost in someone else's P&L.
5. Scale forces the architecture. 500 million store-item forecasts across 11,000+ stores. You cannot fit and monitor 500M individual models, which is why global models exist here, and why cold start is a constant condition rather than an edge case.
The traps: optimizing accuracy and never mentioning waste, reading a sales drop as a demand drop, using MAPE on a grid that is mostly zeros, and believing a promo's gross lift.
This role is a forecasting specialist, so the five below weight diagnosis, measurement design, a model-launch decision, causal impact without a clean experiment, and cold-start forecasting. If you only have time for two, do bc1 and bc4.