Ask a planner how the size curve got set and the honest answer is usually some version of last year, plus a feeling. The category has a curve, the curve gets copied onto the new styles, the new styles ship, and the mediums sell out while the extra-larges sit. Everyone knows this happens. Almost nobody rebuilds the curve to stop it, because rebuilding it by hand is a week of work and the category number looks fine either way.
It is not fine. A size curve is a demand forecast wearing a boring name, and a blanket national curve is the coarsest possible version of that forecast. It answers the question how does this category split across sizes on average, when the question you actually need answered is how does this specific style split across sizes in this specific door. Those are different questions, and the gap between them is where the margin leaks.
A size curve that sells is built per-style and per-door, from demand, not copied from a category and pushed everywhere. That sounds like more work. Done properly it is less, because the machine builds it and the merchant checks it, instead of the merchant building it and hoping.
One national curve is wrong in every door at once
The average size profile does not exist in any single store.
Start with why the blanket curve fails. A national size curve is an average across every door, and an average describes a population that does not shop anywhere in particular. The flagship in the city center skews small. The suburban outlet skews large. The resort location swings seasonally. Blend them into one curve and you get a profile that matches none of them, then you print that mismatch into every store's allocation.
The result is the same two failures repeated in every door. The sizes that door actually sells run out early, so you lose full-price sales to a stockout the category number never shows. The sizes that door does not sell pile up, so you carry them to markdown. You did not buy too much or too little in total. You put the units in the wrong sizes in the wrong doors, which is a distribution mistake dressed up as a demand miss.
The reason this survives is that the failure is invisible at the level anyone looks. A regional manager sees the region sold through. A merchant sees the style sold through. Neither of them is looking at the size-by-door grid, because that grid has thousands of cells and no one reads it by hand. So the blanket curve produces a style-level number that looks healthy while quietly failing in most of the individual cells that make it up, and the mismatch persists precisely because it hides one level below where the reporting stops.
size-level availability leaders hold per door, versus the mid-eighties a blanket split typically delivers, without shipping a single extra unit.
The units are the same. The buy is the same. The only thing that changed is that each size landed where the demand for that size actually was. That is the whole game, and it is invisible at the category level because a category can post a healthy sell-through while quietly failing at the size level in half its doors.
Build it from demand, one style, one door, at a time
A worked example, because this is a mechanical process, not an art.
Here is how you actually build one. Take a women's ankle boot going into 40 doors. The wrong way: pull the footwear category curve, which says 5 percent size 5, 15 percent size 6, 25 percent size 7, 25 percent size 8, 18 percent size 9, 12 percent size 10, and push that identical split to all 40 doors. Every door gets the same boxes regardless of who shops there.
The right way starts by clustering the doors on how they actually trade. The 40 doors are not 40 snowflakes; they fall into a handful of archetypes. Say three: a small-size-skewing urban cluster, a mid-range mall cluster, and a large-size-skewing suburban cluster. Each cluster gets its own curve, derived from that cluster's own footwear sell-through, not the national blend. The urban cluster might peak at size 6 to 7, the suburban at size 8 to 9. Same total units across the 40 doors, allocated to how each cluster trades.
Then you go one level finer, to the style. This ankle boot is not the average footwear item. It is a mid-price, everyday silhouette, and boots like it historically skew half a size larger than the footwear average because customers size up for socks. So the style-level curve shifts right of the category curve, and that shift gets applied on top of the cluster shape. Category tells you the rough shape, cluster tells you the door, and style tells you the specific product. You stack all three.
For a brand-new style with no history, you do not guess. It inherits its curve from its nearest historical cluster, matched on category, price point and silhouette, so a first-season boot launches with the curve of the boots most like it rather than the footwear category average. That is the difference between a launch that ships with a real forecast and one that ships with a shrug.
Blanket national curve versus per-style, per-door demand
Same total units. The blanket curve stocks out the peak sizes in some doors and buries the ends in others. The demand-built curve puts each size where it sells.
Twelve points of size-level availability recovered, on the units you already bought, is the difference between a customer buying the boot in her size and a customer walking out. It never appears in the category sell-through, which is exactly why it goes unfixed for so long.
Let the model build the curve, the merchant approve it
The point is not to do more work. It is to stop copying last year and calling it a plan.
Nobody is going to hand-build a three-layered curve for every style across every door cluster on a spreadsheet. That is why it does not happen, and that is why the blanket curve survives. The answer is to make the curve a first-class part of the plan that the platform builds from demand: category shape, door cluster, style tilt, and cluster-inherited curves for newness, all derived automatically and presented to the merchant to check, not to construct.
Then it stays alive in-season. When a hero size starts running hot in one cluster and cold in another, the model catches the drift and reallocates before it shows up on a markdown report weeks later. The curve you shipped in month one is not the curve you are running in month three, because demand moved and the plan moved with it.
The buyer's job changes shape in the process, and for the better. Instead of spending the week assembling curves from last year's exports and gut feel, the buyer spends it reviewing curves the model already built and overriding the handful that need a human read: a store that is about to be remodeled, a size run a supplier cannot make, a launch the brand wants to seed deliberately. The mechanical work goes to the machine and the judgment stays with the merchant, which is the right division of labor. The curve gets built from demand every time, not just the seasons someone had a spare week to redo it, and the merchant's attention goes to the exceptions that actually need it rather than the arithmetic that never did.
A size curve is not a template you inherit. It is a forecast you build, per style, per door, from the demand that is right in front of you.
Get the curve right and the markdown report gets shorter, availability climbs, and you did it with the exact same units in the exact same buy. The only thing that changed is where they landed. That is the quiet power of the size curve as a lever: it does not ask for more open-to-buy, more suppliers or more risk. It asks you to stop copying last year and start putting the sizes where the demand for them lives, one style and one door at a time, which is a plan the machine can build and the merchant can trust.