Walmart — Demand Forecasting — Product Sense & Metrics

Role context: Senior Data Scientist, Demand Forecasting (US Stores & eCommerce) · Est. study time: 45 min · 5 practice questions

How to prepare for this role

This is a deep forecasting role, not a generic product analytics one. The team works in global deep models (Temporal Fusion Transformers, N-BEATS, PatchTST), sequence modeling and attention-based architectures, and GPU training pipelines. Expect the interview to go deep on forecasting craft.

The day-to-day is building global forecasting models across thousands of retail series, applying generative AI to synthesize time-series data and build analytics assistants, and shipping neural networks for forecasting, anomaly detection, and causal inference through automated batch pipelines. The scale is what forces the architecture: Walmart runs roughly 500 million store-item forecasts across 11,000+ stores.

Where to spend your prep time

• Why accuracy is the secondary metric here, and what the primary ones are.
• Quantile forecasting and the pinball loss, and why the mean is the wrong answer.
• Censored demand: you observe sales, not demand, and the difference is the job.
• Intermittent demand, and why MAPE is undefined on most of the data.
• Backtesting without leakage, at the horizon that actually matters.

What Is Demand Forecasting at Walmart

Walmart is the largest US grocer. This team serves Walmart US stores and eCommerce, spanning groceries, entertainment, and sporting goods.

The framing that matters: a forecast here is not a prediction, it is an order. Every number the model emits becomes physical inventory sitting in a real store, and it is either sold or thrown away. That single fact separates this from every forecasting job in a purely digital product.

Three facts define the work.

Forecasting is an operational system, not an analysis. Walmart's platform manages about 500 million store-item forecasts, with a metrics space of roughly 75 billion historical and 25 billion future data points, drawing on ~100 TB from 10+ sources. It runs on containers so data scientists can ship Python, R, or C++, and it moved from on-premises to cloud in 2019. That scale is why the models here are global: you cannot fit, tune, diagnose, and monitor 500 million individual ARIMA models. Global models are an operational necessity here, not a research preference. That is the honest answer to "why deep learning for forecasting?"

The business goal is waste and availability, and Walmart says so in print. Their published primary goals are waste reduction, measured in dollars year over year, and in-stock percentage at the weekly item-store level. Forecast accuracy is explicitly listed as the secondary metric. The platform launched in June 2018 in Meat and Produce, chosen for high waste impact, reached all key US departments by July 2020, and expanded to Canada, Mexico, and the UK. Reported gains were about 300 basis points of accuracy improvement in the US over a 52-week backtest and about 500 basis points internationally.

The legacy failure explains the modern architecture. The old exponential smoothing replayed last year's seasonal pattern and missed calendar shifts (Easter moves, holidays land on different weekdays), macro and micro factors (payroll calendars, SNAP benefits, pricing, promotions, weather), and geographic variation. Walmart's own example is the one to remember: "Chayote squash is an extremely popular dish in New Orleans during Thanksgiving but has no spike in demand in other parts of the country." A model that cannot represent store-level context cannot see that, and a per-store model has too little data to learn it.

The product is a mature platform under active modernization. It has been live since 2018 and deployed globally. What is new is the current wave of work: transformer-based global models, generative AI for synthetic data, LLM analytics assistants, and neural networks for anomaly detection and causal inference.