Quantity forecasting is a mature, mostly-solved problem
Given a product that already exists, with a season or more of sell-through history, statistical forecasting is genuinely good at predicting how many units to buy, in which sizes, for which locations. Time-series models, seasonality adjustments, and increasingly machine-learning approaches that factor in weather, local events and pricing all work reasonably well on this narrower question. This is not the part of the industry that is struggling.
The confusion comes from calling this the whole of demand forecasting, when it only answers the question for products you have already decided to make.
The harder problem has no name and gets the least investment
Before any quantity model can run, someone had to decide what the product is. That decision, in most organisations, is a mix of design instinct, a macro trend report everyone in the category also read, and whatever performed last season. It is rarely informed by direct evidence about what customers actually want, because that evidence has historically been expensive and slow to gather: a research agency, a survey, months of turnaround.
This is the part of forecasting that actually determines outcomes. A perfectly calculated quantity for the wrong product still fails. A slightly-off quantity for the right product still sells through with adjustments. Yet almost every "demand forecasting" tool and vendor in the market is built for the first, easier, already-mostly-solved problem.
Why statistical models cannot solve the product question
A time-series model needs history to project forward. A genuinely new product, category, market or season has none. This is not a limitation that more data of the same kind fixes, since there is no "same kind" of data available yet. What exists instead, before a single unit is made, is language: what people are already asking for, comparing, and complaining about in the relevant category. That is a fundamentally different kind of evidence, qualitative and conversational rather than numeric, and it needs a different method to read at scale.
A practical split
| Question | Right tool | Data type |
|---|---|---|
| How many units of this known product, in these sizes, for these stores | Statistical / ML forecasting on sell-through history | Quantitative, internal |
| What product should this even be, for a new category, market or season | Live consumer conversation analysis | Qualitative, external |
| How confident should we be in a colour or silhouette direction | Macro trend forecasting (WGSN, Stylus) or attribute tracking (Heuritech) | Editorial or image-based, external |
| Are we positioned well against named competitors | Assortment intelligence (EDITED) | Quantitative, external |
Most forecasting failures that get blamed on "the model was wrong" are actually failures one row up: the product decision was made without the qualitative half of this table, and no quantity model downstream could have corrected for that.
Where this changes an actual buying process
The fix is not replacing statistical forecasting, which works fine for what it does. It is adding the missing input earlier, before the product decision locks, rather than only tightening the quantity math after. In practice that means asking a demand-language question, for a named audience and category, at the same point in the calendar you would previously have only consulted a macro trend report or your own instinct.
Hugo, AI for fashion brands, is built for exactly that earlier step: a specific question, answered from live consumer conversation, with the source language attached, in time to actually affect what gets bought rather than only how much of it.
A perfectly forecasted quantity of the wrong product is still a miss. Most of the money is lost one decision earlier than the forecasting model ever runs.