What is size curve optimisation and how does AI help buy the right sizes?

KEY TAKEAWAYS Summary by the editors

1. A size curve is the percentage split of a buy quantity across sizes; size curve optimisation estimates that split from demand rather than habit.
2. Raw sales are a biased basis for size curves, because sizes that sell out early stop recording sales while customers keep looking for them.
3. Research on a footwear retailer found that nearly 25 percent of demand unmet by stockouts shifted to adjacent sizes, and that accounting for this could raise profit by up to 28 percent in low-demand settings.
4. Size curves should differ by product type, fit, price point and location, but small stores rarely generate enough data alone, so models pool similar stores and products.
5. Because each size has its own GTIN, size-level sales and stock data can be captured in every channel, including from wholesale partners that share sell-out reports.

Size curve optimisation is the process of deciding how a buy quantity should be split across sizes, based on an estimate of true demand for each size rather than on a fixed ratio repeated every season. AI helps by learning size demand from sales and stock history, correcting for sizes that sold out, and producing different curves for different products, fits and locations.

What is a size curve in fashion buying?

A size curve, sometimes called a size profile or size ratio, is the share of each size in a buy or allocation. A women's knit might be bought 10 percent XS, 25 percent S, 30 percent M, 25 percent L and 10 percent XL, for example (an illustrative split, not a benchmark). Many businesses use one curve per category, adjusted occasionally. The problem is that size demand varies by fit, fabric stretch, price point, region and store, and a single curve guarantees that some locations run out of core sizes while others carry fringe sizes into the sale.

Why are sales a poor guide to size demand?

Sales only record demand when stock is available. If size M sells out in week three, the system shows no M sales for the rest of the season, and next year's curve allocates less M. This feedback loop, often called censored demand, quietly pushes curves away from the best-selling sizes.

Research quantifies part of the effect. Akchen and Caro, in a paper forthcoming in Manufacturing & Service Operations Management, studied women's boots at a footwear retailer and found that nearly 25 percent of demand left unmet by stockouts spilled over to adjacent sizes, roughly half to the next size up and half to the next size down. Recorded sales for neighbouring sizes therefore include customers who wanted another size. The authors found that in high-volume settings such as e-commerce, planning sizes without modelling substitution is close to optimal, while in low-demand settings such as seasonal physical retail, a model that accounts for substitution could improve profits by up to 28 percent.

Read also

Using AI in assortment planning without losing the brand

How does AI estimate the right size curve?

1. Collect sales and stock by size and day for each location, so the model knows when each size was actually available.
2. Estimate demand during available periods only, then infer what demand would have been for periods when a size was sold out.
3. Pool data across similar items and stores, because a single store and style rarely sells enough units to estimate a reliable curve on its own.
4. Separate by relevant attributes, such as fit (slim, regular, oversized), category, price band and region.
5. Convert curves into buy and pack decisions, respecting supplier pack ratios and minimum quantities.

The pooling step deserves attention. An early study published on arXiv, on a fashion discounter with many small branches, argued that estimating size demand directly from each branch's sales is statistically problematic because volumes are too small. The authors proposed a measure of which sizes were consistently scarcest and most plentiful per branch, used it to adapt pre-pack size distributions, and reported a gross yield almost one percentage point higher in test branches than in control branches.

What data is needed for AI size curves?

The essential input is sales and stock at size level for every location and channel. GS1 specifies that each product variation, meaning each size, each colour and each combination, needs its own GTIN, so size-level identification is standard wherever barcodes are scanned. Wholesale brands depend on partners for this: retail sales reports exchanged in standard formats, such as the EDIFACT SLSRPT message profiled by GS1 Austria for the fashion sector, carry sales quantities by GTIN, outlet and period.

• Size-level sales and stock history, ideally daily, by location.
• Consistent size coding across seasons and suppliers (an M in one fit block should mean the same in the next).
• Product attributes that affect sizing: fit, stretch, category, target customer.
• Returns data for online channels, since returns by size reveal fit problems rather than demand.

What are the limits of AI size curve optimisation?

Size curves can only redistribute a fixed buy; they cannot fix a wrong total quantity. They are also constrained by supplier pack ratios, which may force a standard mix per carton. Very small stores and niche sizes remain hard to forecast, and pooling assumptions can hide genuine local differences. Finally, extended size ranges are often under-represented in history because they were under-stocked, so models trained on that history can perpetuate the gap unless buyers deliberately test deeper size runs.

Read also

The AI vocabulary every fashion buyer should know

How should a buying team start?

Pick one category with stable fits and good data, compute stockout-adjusted curves alongside current ratios, and apply them to a test group of stores or a part of the buy. Measure availability of core sizes during the season and the share of fringe sizes left at the end. Because the change is limited to how a quantity is split, size curves are one of the lower-risk entry points for AI in buying.

Agree the success measures before the test starts. A curve that improves availability of medium sizes but leaves more extra-large stock at the end of the season may still be the right trade-off, but only if buying, allocation and finance have agreed in advance how to weigh lost sales against markdown cost.

Frequently asked questions

What is a size curve in retail?

A size curve is the percentage split of a buy or allocation across sizes. It determines how many units of each size are ordered or sent to each store.

Why do retailers run out of core sizes?

Often because size curves are based on past sales, which undercount sizes that sold out early. Each season the best-selling sizes receive a slightly smaller share, and stockouts repeat.

How much data do you need to optimise size curves?

More than a single store usually has for a single style. Models pool data across similar products and locations and need size-level sales and stock history, ideally daily, to separate demand from availability.

Do online and store size curves differ?

They can. Research suggests that in high-volume channels like e-commerce, simple size-aggregation planning works well, while low-volume store settings benefit more from modelling how customers substitute sizes.

GuideThe complete guide to AI in fashion merchandising and buyingRead the complete guide