Demand forecasting at small scale is not a modeling problem, it is a decision-support problem: three predictions, made honestly, cover every inventory decision a small seller takes, and the honest version of each includes its own uncertainty. The methods catalog covers the math; this is the operator’s view of what to predict and what to feed it.
The three predictions worth making
1. Next-period velocity per SKU. The workhorse: units per day over the coming reorder window, feeding reorder points and cover math. A weighted recent average with a trend nudge does the job; the operator’s edit is knowing WHEN to override it, a promotion planned, a competitor stocking out, a season turning.
2. Seasonal shape. The multiplier map of your year, December at 2.1x, February at 0.7x, built from last year per category, driving the Q4 buy and the post-season exits. One honest year of data makes a usable shape; category-level shapes cover young SKUs.
3. New-product analogies. No history means borrowing: the closest existing SKU’s launch curve, scaled by gut and capped by a deliberate first-order size. The forecast’s job here is bounding the bet, not blessing it.
Signals beyond your own sales history
Small sellers sit on leading indicators most forecasts ignore:
- Engagement precedes purchase: eBay watchers climbing, store cart adds, marketplace session counts on a listing, demand forming in public, days before it lands in the sales column.
- Channel mix shifts are demand news: a SKU’s Amazon share jumping means its combined velocity is about to, the mix data is a forecast input, not just a report.
- Supplier reality bounds everything: a forecast window is only as useful as measured lead times make it, forecasting demand precisely while guessing supply timing is precision theater.
- The calendar you already know: your own promotions, marketplace events, and category moments belong in the forecast as explicit overlays, tagged so they do not poison the baseline.
Handling error like an operator
Every forecast is wrong; the operational question is which direction hurts more per SKU:
- Understocking an A-item costs the stockout stack, margin, momentum, scramble premiums. Bias those forecasts up, or equivalently, hold more buffer.
- Overstocking the tail costs holding and dead-stock risk. Bias down, order small, let demand prove itself.
- High-variance SKUs get honesty, not precision: when the history whipsaws, the right forecast output is “wide range, buffer accordingly”, a number with error bars beats a confident lie, and buffer sizing IS the error bar in operational form.
Track one meta-metric: forecast versus actual per SKU, monthly. Ten minutes of review teaches you which SKUs your method handles and which need the operator override, the KPI habit applied to the predictions themselves.
Where automation fits
The continuous parts, velocity computation across channels, run-out projection, low-stock alerts, belong to software: live cross-channel velocity is exactly what a synced pool produces as a byproduct, and Unifystock ships the forecast layer (days of cover, stockout dates, risk flags) computed from it. The judgment parts, promotion overlays, new-product bets, direction-of-error calls, stay yours, the honest split between what machines watch and what operators decide.
Common questions
How much history before forecasts are trustworthy?
Thirty days for a workable velocity, a year for seasonal shape. Before that, forecast humbly and buffer deliberately.
Should I forecast revenue or units?
Units, always, for inventory. Revenue forecasts mix price effects into quantity questions and muddy both.
What about external data - trends, search volumes?
Useful directionally for category bets, rarely worth wiring into per-SKU math at small scale. Your own leading indicators (watchers, carts, mix) are closer to your demand.
When is a forecasting tool worth paying for separately?
Usually never at small scale - the forecast layer should come WITH your inventory platform, computed from the same live data, not sold as an add-on.
Predictions attached to decisions
Live counts across every channel from $49/month with unlimited orders, and per-SKU velocity, run-out dates, and risk flags, with the judgment calls left where they belong, on the plans above. See pricing.