The forecast is not a chatbot
Why inventory demand planning needs purpose-built ML, not an LLM, and why a forecast only becomes a plan when bottom-up meets top-down.
Every planning team has now had the same meeting. Someone opens a chat window, asks it to forecast demand for the top sellers, and a confident answer appears in seconds. So the question follows: why not just use ChatGPT, Claude or Gemini, and skip the forecasting platform?
This whitepaper answers it, specifically. A large language model is an extraordinary tool for language and the wrong tool for a demand forecast, because forecasting inventory demand is a different kind of problem than predicting the next word. It makes the case in four parts, then hands you a checklist to pressure-test any vendor, us included.
What's inside
- What a language model actually does, and the three properties (tokenization, no distribution, no grounding in your data) that make it the wrong tool for the numbers.
- What a real demand forecast requires: probability not a point, trained on your history at SKU by size by channel, cold-start for newness, and censored-demand correction.
- Why a forecast is still not a plan, and how bottom-up reconciles with top-down so finance and merchants work from one number.
- Where language models genuinely belong: the copilot that reads the forecast and drafts the move, not the model that computes it.
- A six-question checklist to pressure-test any forecasting vendor before you buy.