What Walmart, Amazon, and Target Actually Measured When They Rebuilt Demand Forecasting

Will Drewes

Founder, Fern Strategy · 3 min read

The numbers are out. Here's what they say.

Retail demand forecasting is one of the cleaner places to measure AI impact. Inventory carrying costs are real dollars. Stockouts and markdowns show up in the P&L. You don't have to argue about whether the outcome matters.

Three large retailers have now published or disclosed measurable results. The figures are worth looking at directly.

Walmart: $1.2 billion and 28.9% better accuracy

Walmart deployed machine learning across forecasting and inventory planning and reported 28.9% higher forecast accuracy alongside a $1.2 billion annual reduction in inventory carrying costs.

The carrying-cost number is the one to hold onto. Forecast accuracy is an operational metric. A billion-plus in carrying costs is a CFO metric. When those two move together, it confirms the model is doing real work at scale, not just fitting historical data cleanly.

Amazon: 96.7% on-time delivery, 17.3% less excess inventory

Amazon's deep-learning forecasting system evaluates over 100 million product-location combinations daily. The reported outcomes: 96.7% on-time delivery during peak seasons and 17.3% reduction in excess inventory.

The scale of the inference problem here matters for context. At 100 million combinations, no human planning team is touching individual SKU-location decisions. The system is making calls that would otherwise either be automated badly (static reorder points) or not made at all. The on-time delivery figure holding at 96.7% during peak is the stress test that validates the approach.

Target: 72% fewer stockouts, $2.3 million in markdown savings

Target's AI-driven inventory optimization system produced a 72% reduction in stockouts, a 31% reduction in excess inventory, and $2.3 million in annual markdown-loss reduction.

Stockouts and excess inventory moving in the same direction is the result most planning teams struggle to achieve manually. Cutting one usually means tolerating more of the other. The fact that both dropped 31-72% suggests the system is capturing demand signal that static safety-stock calculations miss entirely. The markdown savings are a downstream consequence: less excess means fewer units that have to be discounted to clear.

What this means for retail and supply chain leaders

These are not proofs of concept. Walmart, Amazon, and Target are running these systems at production scale, across full assortments, through peak periods.

The governance question is what most mid-market retailers haven't answered yet. When a model is making 100 million daily inventory decisions, or driving markdown strategy across thousands of SKUs, the audit trail matters. Which inputs drove which decisions? When the model's demand signal diverges from what the merchant team expected, who reviews it and on what basis?

Black-box forecasting at scale creates a specific kind of operational risk: the system performs well until it doesn't, and by the time the signal shows up in inventory or service levels, the decisions that caused it are weeks old and unreviable.

The retailers above have the engineering resources to build that infrastructure themselves. Most don't. The build-vs-buy question for demand forecasting has a reasonably clear answer on the model side. The harder question is whether the forecasting layer you're buying or building has the auditability your planning, finance, and ops teams can actually work with.

Start by asking what your current system can explain about its last 30 days of replenishment decisions. That answer tells you more than any benchmark.