Describe the product before you try to sell it. Break the frame into what it is made of: shape, rim type, material and finish, width, bridge, lens, price band, collection, channel, region. Every one of those has sold thousands of times before. A new frame is a new combination, not a new phenomenon. Pooling history across matching attributes, instead of copying one chosen twin, works: one published study reports error reductions of 20–60% across product groups using this approach. The capability is not the hard part. Most planning platforms now sell attribute-based forecasting, and there is a granted US patent whose whole argument is that vague, free-text attribute descriptions produce unreliable forecasts and must be replaced with measurable ones.
Read that patent as a warning rather than a feature list. The model takes weeks to build. The attribute definitions take years to govern. If “cobalt” becomes “electric blue” next season, the pooling quietly stops working and nobody gets an error message.
One more warning, because it is the mistake most likely to embarrass the project. The history you pool is incomplete. The winners sold out, so their recorded sales understate what they would have sold. Train on raw sales and the model will under-forecast exactly the frames you most need it to spot. That is precisely what happens to the Havana in Figure B.
Take the shape from a library, not from a pen. Level and shape are two different problems, so solve them separately. The cleanest published approach uses one model to set the level and a second to pick the launch curve. Group past launches into a handful of standard shapes, then assign the new frame to one of them based on its attributes. Every planning system already has the vocabulary for this called phase-in profiles, lifecycle profiles, launch profiles. What is missing is accountability: today the curve is chosen once, by one person, with no name on it and no accuracy measured against it afterwards. And the shape people assume is often simply wrong. In a study of 170 product life cycles, more than a fifth had no steady period at all. The S-curve everybody draws on the whiteboard is the wrong shape for a fifth of launches.
Treat the spread as an output, not a footnote. This is the part that turns the industry's favourite excuse into a tool. Everyone says eyewear cannot be forecast because it is about perception and five experienced people in a room will give five different numbers. The skiwear study found that this disagreement is the most useful signal available. When buyers forecast independently, instead of being pushed towards one agreed number, the spread between their individual answers predicted the accuracy of the final forecast almost perfectly. Where they agreed, the number held. Where they disagreed, it did not. So stop forcing agreement. Collect the spread, and use it to size the option rather than the argument.
Then buy the option. Commit the floor and reserve capacity for the rest. The encouraging finding from the same research is that you do not need to be a fast-fashion retailer to benefit: holding reactive capacity worth about 30% of the season captured close to half of the total possible saving.
Eyewear makes this harder than clothing. You cannot simply reorder faster because the tooling is a four- to eight-week commitment and suppliers publish minimums of three to five hundred pieces per style and colour. So the lever is not speed. It is deciding later. Commit the tooling and the front. Hold the colour split. Reserve a sheet of material and a slot at the plant instead of finished units. That is a negotiation with a supplier, not a setting in a system, and it is the part no platform will do for you.
And listen outside your own sales data. A study across three large clothing and footwear businesses found that adding social and search signals cut out-of-sample forecast error by 24% to 57%, and that colour and fit could be read months before the season started. In a category where a head of state wearing a pair of aviators at a January summit moved the maker's share price by around 65% in two days and where that maker, having sold a thousand pairs online, could not deliver them until May.
What That Looks Like
Figure A was the plan as it is published today: one borrowed curve, used four times, committed in full. Figure B is the same style planned the new way using a level from attributes, a shape from the library, a range instead of a single point, and a buy split into a committed floor and a reserved option released after an early read.
Put the two side by side and four things change.
- The weighted error at SKU level falls by more than half. That is the level at which money is committed, so that is the number that matters. The colourway that was going to sell gets a bigger opening commitment; the one that was not gets a smaller one. Nothing clever has happened. The model has simply stopped pretending that four different frames will sell the same quantity.
- Style-level error gets worse. This is the uncomfortable part, and it needs saying out loud. A model trained on incomplete history under-forecasts the winner, because the winner sold out and its recorded sales understate what it would have done. So the total for the style comes out lower than reality. If your forecast KPI is style-level accuracy, this will look like a failure in its first season. Change the KPI before you start, or do not start.
- Stock left over at season end drops, and so does unserved demand. Both at once, which is the whole point. Under the old plan you were wrong in both directions at the same time, short on one colour, long on three. Splitting the commitment lets you be less wrong on both.
- Cash committed at the tooling freeze falls. Part of the buy is now an option rather than an order. That is the real financial change: less money locked up before anyone has seen a customer.
Three cautions before anyone treats this as a proposal.
- A better forecast is still a bad forecast. The error does not disappear. It comes down far enough that an option sized from the spread can absorb what is left. The improvement is in how the buy is structured, not in the accuracy of the prediction and that is a more durable place for it to sit.
- The option is not free. Reserving material and a slot at the plant costs something, and some optioned units are never built at all. There is a price above which the trade stops paying, and that price comes out of a supplier negotiation rather than a model. Work it out before you promise anyone a saving.
- Not every style deserves an option. Most styles are not the one that sells out. The value of the spread is precisely that it tells you which ones are worth spending the option budget on and, just as usefully, which ones to leave alone.
