A brand or retailer sells a thousand pairs of a shoe. Good volume. Then the returns and the stranded inventory tell a different story: the size range they bought was nowhere close to what sold, and it bled the e-commerce business for a season. The product is right, the size curve was wrong.
You can have healthy weeks of supply at the top line and, two clicks down, a size mix that's broken. And for most teams, the curve behind it is still built by hand. Someone pulls last year's sales, eyeballs the split, and types in percentages. It's slow, it's manual, and it's one of the biggest reasons a well-bought assortment sells through badly.
Good size curve optimization comes down to a set of decisions. Most of which planners never make on purpose. Think of it as a decision stack: get the layers right and the curve becomes an asset. Skip them and you've automated a guess.
1. Pick the right grain, and know when to jump up a level
The first decision in size curve planning is the level you build from. A style-color curve is sharper than a class curve, which is sharper than a department curve. But granularity only helps when there's enough data underneath it. The real skill is the jump rule: build at the most granular level that has signal, and roll up to class, department, or supplier when the granular data is too thin. Done well, this is also how you get optimal size curves by store instead of one national average trying to fit every door.
2. Define "enough data" before you build
Set a confidence threshold per level: an explicit unit count below which you don't trust the granular curve. This is the difference between a number and a feeling. Build a supplier or style curve on too few units and its performance on the next buy falls apart, no matter how sharp the grain looks. When a partition falls under the threshold, that's your signal to jump a level, and the jump should be visible so a planner can see the curve rolled up rather than silently defaulting.
3. Trim both tails, not just the sparse one
Most planners worry about too little data. The error is letting the wrong data in. Filter by participation, not a raw count, so the curve isn't dragged around by the bottom 20% of products that behave erratically. Then cap the top too. A single viral item can distort a curve as badly as a sparse one, and it isn't how you want to size the next buy. Trimming the top few percent is often as valuable as trimming the bottom.
4. Normalize with size mapping, and watch for artificial sparsity
Map inconsistent size labels to a core set so one curve can serve many items, whether a supplier uses numeric or alpha sizes, or a simple small-medium-large run. But mapping has a trap worth knowing. A rare size type like "tall" can make a curve look sparse when it isn't, because the system now expects a size the item wasn’t going to carry. That false gap will push you to jump a level you didn't need to.
5. Have a deliberate plan for brand-new sizes
Introducing a size with no history is one of the hardest calls in size optimization, and it's where the manual habit is worst: planners pick the sizes and hand-key what percentage should go into each. For a genuinely new size, that guess is all you have, so make it a method instead. Jump a level and inherit the broader curve, borrow from a comparable partition, or map to an analog size that behaves like it. Anything but typing in a number and hoping.
6. Shape the curve off real demand, not what you happened to sell
This is the one that quietly poisons everything else. If you build a curve from raw sales, you're building it from constrained sales. The sizes that stocked out first look weaker than they were, so next season you buy fewer of them, and they stock out again. The curve becomes a self-fulfilling prophecy. Trustworthy size curve forecasting starts from unconstrained demand: strip out the weeks and locations where a size wasn't actually in stock, so you're measuring what customers wanted, not what the shelf allowed.
7. Weight recent history, and re-validate when you change it
Prefer a dynamic curve that updates as new sales land over a static one you refresh by hand, so you're not constantly moving the goalposts. From there, weighting recent periods more heavily than old ones often improves accuracy. One discipline is non-negotiable, though: any time you change the weighting, re-validate it against what sold. Which leads to the move almost everyone skips.
8. Close the loop with backtesting
Here's the habit in most size curve work. Teams build the curve, buy against it, and never check it. Weekly recaps compare what was expected to what actualized, and then that hindsight dies in a spreadsheet. Nothing feeds back into the curve. Backtesting closes the loop: measure last season's curve against what sold, quantify the miss, and feed the correction forward. It's the least glamorous step and the most valuable, because it's the only one that tells you whether the other eight are working. A curve you've never backtested is a hypothesis instead of a data-backed plan.
9. Differentiate by channel
One curve rarely holds across every channel. A DTC curve and a wholesale curve aren’t typically the same shape. In wholesale you might carry a narrow size break, offering a store just one or two sizes per category instead of the full size run. Retail buys size in bulk, while allocation sizes each store individually. Some fringe sizes are must-carry for presentation no matter what the math says, which calls for a minimum set independent of the curve rather than a percentage the curve would never justify on its own.
Nine decisions, one reliable curve
None of this requires a bigger team, but treating the size curve as a stack of decisions you make on purpose instead of a number you inherit. The brands and retailers who get retail size optimization right aren't the ones who know what level to trust it at, which data to keep out, and how to check themselves after the fact.
If you're evaluating assortment planning software to help with size curve optimization, that's the checklist to hold it against: does it pick the right grain, set a real data threshold, trim both tails, normalize cleanly, handle new sizes, build off unconstrained demand, weight and re-validate, close the loop, and split by channel? The mechanics are what separate a curve you can defend from a sizing curve you're quietly hoping about.
