> ## Documentation Index > Fetch the complete documentation index at: https://docs.stockful.app/llms.txt > Use this file to discover all available pages before exploring further. # Forecast models > Self-tuning demand models, forecast accuracy, and hub locations ## Self-tuning forecast models A demand forecast with its confidence band Stockful doesn't force every product onto a straight-line forecast. For each variant at each location it backtests a few demand models on that SKU's own history - a flat baseline, a **trend** model (demand rising or falling), and a **day-of-week** model (weekly buying patterns) - and adopts a richer model only when it clearly beats the baseline on recent weeks. Slow or sparse sellers stay on the simple baseline, so a forecast never gets worse by guessing at a pattern that isn't there. On the inventory detail page, the forecast chart shows the chosen model's projection: a curve rather than a straight line for SKUs with a real trend or weekly rhythm. Incoming shipments with an expected arrival date appear as a **step up** on the projection, so you can see stock replenish before it runs out. ## Warehouse / hub locations Locations with the **Hub (warehouse)** role measure demand differently: a hub's velocity reflects the total demand of every location it supplies, and its reorder points are sized for your supplier lead time. See [hub locations](/user-guide/locations#warehouse--hub-locations) for how roles are detected and overridden. ## Forecast accuracy Each week Stockful records the forecast it made, then grades it against what actually sold. The inventory detail page shows a **Forecast accuracy** figure once there are at least a couple of weeks to compare. You can also ask the AI assistant "how accurate are my forecasts?" or "where did the forecast miss, and why?" for a plain-English breakdown - including whether a miss lined up with a sale or a switch in model. Weeks a SKU was out of stock are left out of the grade, since there was nothing to sell.