Decathlon Is Using AWS AI to Forecast Demand: Can Better Predictions Reduce Stockouts and Overstock?

Decathlon's Chronos-2 deployment shows that better AI demand forecasting can reduce both stockouts and excess inventory, but only when forecasts are connected to replenishment, allocation and exception rules. For Sydney and NSW retailers, distributors and multi-site operators, the practical opportunity is not simply a more accurate model. It is a forecast-to-order process that combines demand signals, promotions, lead times and stock constraints to make better weekly inventory decisions.

Retail forecasting has traditionally been discussed as a data-science problem. A model predicts what customers are likely to buy, the business compares the estimate with actual sales and the most accurate model wins.

Decathlon's recent work with Amazon Web Services makes the commercial question more interesting.

According to AWS: How Decathlon runs demand forecasting at scale with Chronos-2, the sporting-goods retailer is using the time-series foundation model to produce weekly forecasts across large product ranges and multiple supply zones. The system supports a 12-week replenishment horizon for purchasing decisions and a 52-week horizon for longer-term stock and capacity planning.

That turns the Decathlon Chronos-2 case into something more useful than another story about a large company adopting artificial intelligence.

It provides a way to examine the chain between forecast accuracy, inventory held, products actually available to customers and the sales that follow.

The Important Result Is Not Simply That The Forecast Became More Accurate

Decathlon benchmarked time-series foundation models against its existing forecasting approach using a large historical evaluation covering approximately 25,000 products per cutoff and 101 rolling forecast cutoffs.

AWS reports that the fine-tuned Chronos-2 system materially reduced weighted absolute percentage error, or WAPE, in the South East Asia and Latin America supply zones.

South East Asia — 12 weeks
Previous WAPE: 39%
Chronos-2 WAPE: 28%
Reported improvement: 11 percentage points
Latin America — 12 weeks
Previous WAPE: 53%
Chronos-2 WAPE: 38%
Reported improvement: 15 percentage points
South East Asia — 52 weeks
Previous WAPE: 44%
Chronos-2 WAPE: 38%
Reported improvement: 6 percentage points
Latin America — 52 weeks
Previous WAPE: 55%
Chronos-2 WAPE: 46%
Reported improvement: 9 percentage points

Those figures matter, but the stronger business evidence is what Decathlon says happened downstream.

AWS reports that, on average across Decathlon's supply zones, every percentage point of improvement in 12-week WAPE has historically corresponded with approximately 0.3 fewer days of inventory, a 0.3-point improvement in product availability and roughly 0.12 points of sales improvement.

Those relationships should not be treated as universal ratios for another retailer. They are Decathlon's reported operating results. They do, however, illustrate the measurement discipline that other businesses should adopt.

Forecast accuracy has commercial value when management can trace it through to stock, availability, working capital and sales.

A Forecast Cannot Reduce A Stockout By Itself

A prediction is information. A replenishment decision changes inventory.

The distinction matters because a retailer can build an excellent AI demand forecasting model and still have empty shelves.

Consider a forecast that correctly predicts demand for 420 units during the next replenishment cycle. The result still has to pass through several operational decisions:

How much stock is already available?
How much has already been ordered?
What stock is committed to other stores or customers?
What is the supplier lead time?
Is there a minimum order quantity or carton quantity?
How much safety stock is required?
Is warehouse capacity available?
Should stock be purchased, transferred between locations or held back?
What happens if the forecast changes materially next week?

If those decisions remain disconnected from the forecast, improved prediction accuracy may produce an impressive dashboard without materially changing product availability.

The Forecast-To-Replenishment Chain Is Where The Value Sits

The practical operating model can be viewed as a sequence rather than a standalone AI application.

Observe demand. Collect sales, stock, promotion, price, product and relevant external signals.
Generate the forecast. Estimate expected demand over the horizon that matters to purchasing and distribution.
Measure uncertainty. Distinguish a stable forecast from one with a wide range of possible outcomes.
Calculate the inventory position. Account for available stock, inbound orders, committed inventory and expected lead time.
Apply replenishment policy. Convert the forecast into proposed order, transfer or allocation quantities.
Review exceptions. Escalate unusual orders, major forecast movements, supply constraints and high-value decisions.
Execute. Create or approve the purchase order, transfer or allocation.
Measure the result. Compare the forecast with actual demand and track stockouts, excess inventory, availability, margin and working capital.

The model occupies only one part of that chain.

The Hidden Data Problem: Sales Are Not Always The Same As Demand

Retail systems contain a deceptively simple historical variable: units sold.

The problem is that units sold only reveal demand when the product was actually available.

Suppose a Sydney store sells 20 units of a product every Saturday until midday, then runs out. Its point-of-sale system records 20 units sold. Actual customer demand may have been 24, 30 or 40 units.

If a forecasting system repeatedly treats sold-out periods as evidence of lower demand, it can learn from the shortage and recommend another shortage.

A credible forecasting dataset therefore needs more than transaction history. Depending on the business, teams should consider:

Periods when products were unavailable.
Inventory on hand and inbound stock.
Promotional periods.
Price changes and markdowns.
Returns and cancellations.
Range changes and discontinued products.
New product launches.
Store openings, relocations and closures.
Supplier interruptions.
One-off bulk orders.
Seasonal events.
Known future events that could materially change demand.

This is one reason foundation models do not eliminate the operational work around forecasting.

Better modelling cannot repair a business process that has never distinguished genuine low demand from an inability to supply the customer.

What Chronos-2 Changes

Amazon Science: Introducing Chronos-2 — from univariate to universal forecasting describes Chronos-2 as a time-series foundation model capable of univariate, multivariate and covariate-informed forecasting.

In practical terms, this means the model can work not only with historical demand but with related signals that may help explain what happens next.

Known future information might include a planned promotion, a scheduled price or the number of stores expected to carry a product. Other businesses may have useful external signals such as weather, events or commercial calendars.

Decathlon also changed the operating burden around its forecasting stack. Its previous systems required much more frequent model retraining. AWS says Chronos-2 is now fine-tuned approximately every six months while weekly batch forecasting continues between those training runs.

The significance is not that every retailer should use the same model or training schedule.

It is that a foundation model can potentially reduce the amount of specialised model engineering required to introduce forecasting across additional categories or regions.

For Sydney Retailers, Geography Can Be Part Of The Forecasting Problem

A retailer operating across NSW rarely has one homogeneous demand curve.

Demand in a Sydney CBD location can differ materially from a suburban centre, a coastal market, Western Sydney, Newcastle or a regional NSW location. Sporting goods make the distinction particularly easy to understand, but the same issue appears in hardware, apparel, building materials, furniture, food distribution and specialist wholesale.

A business may need to account for different combinations of:

Local customer mix.
Seasonality.
School and public-holiday periods.
Local events.
Weather sensitivity.
Promotions.
Delivery lead times.
Store capacity.
E-commerce fulfilment demand.
Available stock elsewhere in the network.

The Australian Bureau of Statistics now publishes the Monthly Household Spending Indicator with state and territory data. The broader lesson for individual retailers is that demand conditions move over time and can differ geographically. Internal forecasting needs considerably more detail than a statewide economic indicator, but external data can provide useful context.

Better Forecasts Can Reduce Stockouts In Three Different Ways

1. Ordering Earlier

When demand is likely to exceed the normal run rate and supplier lead time is long, earlier visibility gives purchasing teams more time to act.

This is particularly important for imported products, seasonal ranges or suppliers with constrained production windows.

2. Ordering The Right Quantity

A business can avoid a stockout and still make a poor inventory decision by dramatically over-ordering.

Better demand forecasts can narrow the gap between the stock needed to reach the target service level and the unnecessary buffer held because management does not trust its numbers.

3. Moving Existing Stock Instead Of Buying More

Some stock problems are allocation problems rather than purchasing problems.

One location may be approaching a shortage while another holds several weeks of excess inventory. A network-wide forecast can support transfers before the business places another supplier order.

The Same System Can Also Create Overstock If Decision Rules Are Poor

Forecasting systems can fail commercially even when their statistical performance looks respectable.

Promotion uplift is treated as permanent demand
Possible operational consequence: Excess purchasing after the campaign ends.
Control: Promotion flags and post-event normalisation.
Stockout periods are interpreted as weak demand
Possible operational consequence: Repeated under-ordering.
Control: Availability and lost-sales treatment.
New products inherit the wrong comparison history
Possible operational consequence: Initial allocation is too high or too low.
Control: Cold-start rules and comparable-product review.
Forecast rises but supplier lead time is ignored
Possible operational consequence: The order arrives after demand has passed.
Control: Lead-time-aware replenishment.
Every forecast change automatically becomes an order
Possible operational consequence: Order volatility and unnecessary stock.
Control: Thresholds, review bands and minimum change rules.
Store forecasts are considered independently
Possible operational consequence: One site over-orders while another needs stock.
Control: Network-level transfer and allocation logic.

This is where operational design becomes more important than choosing the most fashionable model.

Forecast Accuracy Should Not Be The Only KPI

A retailer can improve WAPE while failing to improve the economics of inventory.

Management should therefore measure the model and the operating outcome separately.

Forecast
Useful metric: WAPE, bias, RMSE or category-appropriate accuracy metric.
What it tells management: How closely predictions match observed demand.
Availability
Useful metric: In-stock rate or product availability.
What it tells management: Whether customers can actually buy the product.
Shortage
Useful metric: Stockout days or lost-sales estimate.
What it tells management: Where insufficient inventory is costing revenue.
Inventory
Useful metric: Days or weeks of supply.
What it tells management: How much working capital is tied up.
Excess
Useful metric: Aged stock and markdown exposure.
What it tells management: Whether purchasing is creating future write-down risk.
Commercial
Useful metric: Gross margin, sales and working-capital movement.
What it tells management: Whether improved forecasting is producing business value.

A strong pilot should establish these baselines before the AI system goes live.

A Mid-Sized Business Does Not Need Decathlon's Scale To Test The Idea

The lesson from Decathlon is not that a Sydney retailer needs tens of thousands of products before AI demand forecasting becomes worthwhile.

A smaller retailer or distributor can begin with one commercially meaningful inventory problem.

A disciplined pilot might look like this:

Select one category or distribution flow. Choose products where stockouts, excess inventory or purchasing volatility already create measurable cost.
Choose the decision horizon. A four-week forecast may be irrelevant if overseas replenishment takes 12 weeks.
Build the baseline. Measure existing forecast accuracy, stockouts, availability, inventory days and markdowns.
Repair the data. Identify missing transactions, stockout periods, discontinued products, promotions and abnormal orders.
Benchmark more than one forecasting method. Do not assume a foundation model will automatically outperform the current system on the business's own data.
Run in shadow mode. Compare AI recommendations with actual planner decisions before allowing automated purchasing.
Introduce controlled replenishment. Start with proposed quantities and exception review rather than unrestricted purchase-order creation.
Measure inventory outcomes. Continue only if better forecasting translates into better availability, lower excess stock or another agreed commercial result.

This is consistent with a broader principle in AI consulting and implementation planning: the business problem, workflow and measurement model should be defined before the technology stack is treated as the project.

Compute May Be The Cheap Part

One of the more striking details in AWS's Decathlon case study is the reported inference economics.

AWS says Decathlon can run weekly Chronos-2 inference on CPU infrastructure, with the specific workload costing approximately US$0.03 per weekly inference run. That figure represents model inference for the described architecture, not the total cost of implementing or operating an enterprise forecasting system.

For another business, the larger cost may sit elsewhere:

Extracting reliable stock and sales data.
Connecting ERP, POS, warehouse and e-commerce systems.
Designing purchasing rules.
Reconciling product identifiers.
Building exception workflows.
Testing forecasts across changing seasons.
Monitoring model performance.
Managing the organisational change required for planners to use the system.

This is why businesses comparing an off-the-shelf forecasting platform with a custom implementation should assess the full workflow rather than model cost alone. Elyment's build-versus-buy AI framework applies the same principle to AI software decisions more broadly.

Do Not Let A Forecast Become An Unverified Customer Promise

Forecasting becomes a different risk when an internal prediction is exposed directly to customers.

A system may predict that replenishment will arrive next Thursday. That is not necessarily the same as confirmed stock availability or a guaranteed delivery date.

The Australian Competition and Consumer Commission: False or misleading claims guidance specifically notes that businesses should not mislead consumers about whether goods are in stock or when they may be supplied.

Retail systems should therefore distinguish between:

A forecast of future demand.
A supplier's estimated arrival.
A confirmed inbound shipment.
Stock physically received.
Stock available to promise.
Stock reserved for an identified order.

AI can improve prediction without changing the evidentiary standard required before a business makes a firm availability statement to a customer.

Governance Should Follow The Consequence Of The Decision

Private NSW retailers are not generally subject to the NSW Government's mandatory AI Assessment Framework simply because they use forecasting software.

The NSW AI Assessment Framework nevertheless illustrates a useful lifecycle principle: AI systems should be reassessed when their data, purpose, functionality or risk context materially changes.

That principle has practical value in retail.

A forecasting tool initially used to advise a planner presents a different operational risk from the same tool after it is connected directly to purchase orders, warehouse transfers, supplier commitments or customer-facing availability.

Authority should increase progressively.

A sensible sequence is:

Forecast only.
Forecast plus planner recommendation.
Automatic replenishment proposals.
Automatic execution within defined quantity and value thresholds.
Human approval for high-value, unusual or low-confidence exceptions.

The Real Business Case Is Better Inventory Decisions

Decathlon's adoption of Chronos-2 is significant because it connects a modern foundation model with one of retail's oldest problems.

Buy too little and the customer finds an empty shelf.

Buy too much and cash becomes inventory that may eventually need to be discounted.

Better predictions can narrow that gap. They cannot eliminate it.

Supplier reliability, lead times, minimum-order quantities, promotion strategy, warehouse capacity, merchandising decisions and human judgement still determine what happens after the forecast arrives.

Sydney and NSW retailers considering similar systems should therefore avoid beginning with the question, "Which forecasting model should we deploy?"

A more useful starting question is:

Which inventory decision are we trying to make better, what evidence should change that decision, and how will we prove that the result reduced stockouts or unnecessary stock?

Businesses that need to connect forecasting with inventory data, approval rules, ERP systems, reporting and operational workflows can review Elyment's AI systems and software development capability in Sydney. The implementation objective should be a measurable operating system, not an isolated forecasting demonstration.

Review The Forecast Before It Starts Driving Stock

AI FORECASTING · INVENTORY CONTROL · OPERATIONAL DELIVERY

Map demand data, replenishment rules, ERP integrations, purchasing thresholds, exception handling and performance measures before an AI forecast is connected to live inventory decisions.

Request An AI Operations Review

The Bottom Line

Decathlon's reported results suggest that modern AI demand forecasting can materially improve the quality of retail planning. Chronos-2 reduced forecast error across the supply zones reported by AWS while Decathlon linked those improvements with lower inventory, higher product availability and increased sales.

The more important lesson for other businesses is operational.

Forecast accuracy is not the finish line. The system must know what stock exists, what is already arriving, how long suppliers take, which constraints apply, when a planner should intervene and how the eventual decision affected availability and working capital.

When those elements are connected, AI demand forecasting can become part of a disciplined inventory operating model.

When they are not, a more sophisticated prediction can simply give an old stock problem a better-looking dashboard.

Sources and References

AI FORECASTING · INVENTORY CONTROL · OPERATIONAL DELIVERY

Review The Forecast Before It Starts Driving Stock

Map demand data, replenishment rules, ERP integrations, purchasing thresholds, exception handling and performance measures before an AI forecast is connected to live inventory decisions.

Review Your AI Operations