Inventory forecasting for a small seller has a narrower job than the textbooks suggest: predict each SKU’s demand well enough to answer two questions, when do I run out, and how much should I reorder. The methods below are ranked by effort against those outputs, and the honest headline is that data hygiene beats model choice at small scale, every time.
The methods, in effort order
Simple moving average. Average the last 30 days of daily sales; project forward. Free, stable, and blind to trend and season. Right for steady mid-tail SKUs, which is most of a typical catalog.
Weighted moving average. Same idea, recent days weighted heavier (say, the last week counting double). Catches demand shifts weeks earlier than the simple average without new complexity. This is the workhorse setting for active SKUs.
Trend-adjusted (exponential smoothing with trend). Adds an explicit trend term: if a SKU is growing 5 percent weekly, the projection grows too. Worth it for your movers, unnecessary for the tail.
Seasonal adjustment. Multiply the base forecast by seasonal indices built from last year’s shape (December = 2.1x baseline, February = 0.7x). Essential for gift, outdoor, and holiday-driven catalogs; use category-level indices when a SKU lacks its own history.
Demand sensing / ML models. Learn from many signals (velocity shifts, channel mix, promotions). Genuinely useful at scale or high seasonality; at 200 SKUs with clean data, the gain over weighted-plus-trend is usually smaller than one week of bad inputs costs, our take on the AI layer.
What matters more than the model
Cross-channel completeness. A forecast built on one channel’s sales while three channels drain the same pool is precisely wrong. Combined velocity per SKU is the input, the multichannel correction applies to every method above.
Freshness. Forecasts recomputed monthly from exports answer last month’s questions. Continuous recomputation from live sales, which is what days-of-cover forecasting is, keeps the run-out date current as demand moves.
Anomaly handling. One viral day should inform, not define, the forecast. Cap outliers or use medians for thin-history SKUs, and annotate promotions so the model does not read your own discount as organic demand.
New products. No history means no forecast, borrow from the closest analog SKU, hold a deliberate buffer, and replace guesses with data at first review.
From forecast to decisions
The outputs that matter, per SKU:
- Run-out date: on-hand divided by forecast velocity, sorted soonest-first, that list IS your replenishment agenda. Unifystock computes it continuously across every connected channel as days-of-cover and stockout forecasts.
- Reorder quantity: enough to restore target cover at forecast velocity, bounded by holding-cost sanity.
- Confidence: wide-variance SKUs get bigger buffers, not braver forecasts.
Common questions
How much history do I need to forecast?
Thirty days gives a workable average; a year unlocks seasonality. Do not wait for perfect history, start simple and refine as data accrues.
Should I forecast weekly or continuously?
Decisions are made on your cadence, but the numbers should refresh continuously, a spike on Tuesday should move Friday’s purchase order.
How do promotions fit into forecasts?
Tag them. Promo-driven sales inform promo planning, not baseline demand. The classic failure is reordering to a discount spike and holding the excess for months.
Is forecasting worth it for the long tail?
The tail earns simple methods: a plain average plus a small buffer, reviewed rarely. Spend your forecasting attention where the money moves.
Fresh inputs, automatic outputs
Low-stock alerts from live cross-channel sales on every plan, from $49/month with unlimited orders, and per-SKU forecasts with run-out dates and reorder signals that maintain themselves on the plans above. See pricing.