Sales data answers a narrower question than people think
Historical sell-through is genuinely the best predictor of future demand for a product you already sell, to a customer you already have. It is a bad predictor of almost everything else: a new category, a new market, a product you have not launched, or a customer who wanted something and could not find it, so there is no transaction to analyse in the first place. The absence of a sale is not the same as the absence of demand.
This is the gap most demand-forecasting advice glosses over. It treats "we have no data on this" as a reason to guess, when it is actually a reason to look somewhere other than your own sales system.
Step 1: Name the audience, not the category
"Streetwear buyers" is not an audience, it is a category with millions of different people and reasons inside it. A workable audience definition is specific enough to have a shared vocabulary: 22 to 30 year old women in a named city, a specific subculture, people who follow a specific set of creators. Demand language only makes sense once you know whose language you are reading.
Step 2: Read demand language, not just sales data
Three kinds of language matter more than volume: direct requests ("does anyone make this in a heavier weight"), comparisons ("I switched from X to Y because"), and complaints about what already exists. All three precede a purchase, which is exactly why they are useful for prediction and invisible to a sales report. This is the layer that shows demand for products that do not exist in anyone's catalogue yet, which no amount of your own sales analysis can surface.
Step 3: Separate volume from intensity
A thousand generic "cute" comments on a product photo are weaker signal than twenty specific, repeated requests for a feature nobody offers. Volume alone rewards whatever already got a lot of attention, which is usually already reflected in a sales number somewhere. Intensity, specificity and repetition are what catch demand before it is obvious.
Step 4: Check it against what already exists
Once you have a specific, repeated, unmet request, check whether it is genuinely unmet. This is where assortment and pricing data earns its place, cross-referencing a demand signal against your own catalogue and named competitors' to confirm the gap is real and not something you already carry under a different name.
Step 5: Test the smallest version of the buy
None of the above replaces committing real inventory. Treat a strong demand signal as a reason to run a limited buy or a pre-order, not a reason to bet the season on it. The actual sell-through on that small test becomes the highest-confidence data point of all, and it feeds back into step 1 for the next cycle.
What this looks like as a question, not a process
In practice this whole method compresses into a single question asked of the right source: "what are [named audience] asking for in [category] that nobody in our price tier currently sells." That is the shape of question Hugo, AI for fashion brands, is built to answer, reading live consumer conversation and returning the actual language, segmented and sourced, instead of a five-step framework you have to run manually every time.
See it applied to a specific category and audience in Hugo for fashion, or read what your bestsellers cannot tell you for the flip side of this same argument.
The sales report tells you what you already know how to sell. The prediction problem lives entirely outside it.