Start from each SKU's recent sales history, project it forward with a moving or weighted average, then adjust for trend, seasonality, and any planned promotions. The forecast has to extend at least as far as your total lead time, because you are ordering today to cover demand that lands a full lead time from now.
What is the best inventory forecasting method for sellers?
For most FBA sellers, a weighted moving average that leans on recent weeks is the best balance of accuracy and maintainability. More advanced methods (exponential smoothing, regression, machine learning) help at scale or with strong seasonality, but a model you cannot maintain is worse than a simple one you actually update.
T. Brian Jones is co-founder and CTO of Inventory Hero. He leads the engineering behind its Amazon data pipeline, demand forecasting, and the AI platform that lets sellers talk to their live inventory, sales, and supplier data in plain language.
At least your total lead time plus your review cycle, and further for seasonal SKUs. If your lead time is 80 days, a forecast that only looks 30 days out cannot tell you what to order today. Seasonal products need a forecast that reaches across the whole season so you can build before the ramp.
Read article
Amazon inventory forecasting is predicting each SKU's future demand so you can buy ahead of it, and on FBA the forecast has to reach at least as far out as your full lead time. The short version: start from recent sales history, project it forward with a weighted average, adjust for trend, seasonality, and promotions, and track how wrong the forecast turns out so you can tighten it. Below are the methods, how to handle seasonality and new products, and why a forecast that does not reach past your lead time is useless.
Inventory forecasting is the process of estimating how many units of each SKU you will sell over a future period, so you can order enough to meet that demand without overbuying. It is the input that feeds every restock decision: your reorder point and reorder quantity are only as good as the demand forecast underneath them. For the restock system that consumes the forecast, see Amazon FBA restock planning.
The reason FBA forecasting deserves its own discipline is the lead time. A local shop can react to a sales bump next week. You are committing cash months ahead, so a forecast that is wrong by the time your container lands is an expensive mistake, in either direction.
You do not need a complex model. You need one you will actually maintain. In rough order of sophistication:
Moving average. Average the last N periods of sales. Simple, stable, and slow to react. Good for steady SKUs.
Weighted moving average. The same idea, but recent periods count more, so the forecast follows real changes faster. This is the best default for most FBA SKUs. See inventory forecasting methods for the comparison.
Exponential smoothing. A weighted average where the weights decay smoothly. More responsive, still simple to run.
Seasonal and trend models. Add explicit seasonal indices and a trend term for SKUs with clear cycles. See how to model seasonality.
Regression and machine learning. Useful at scale or with many demand drivers, but only worth it once the simpler methods are maxed out and maintained.
The honest rule: a weighted moving average you update every week beats a machine-learning model you set up once and never touch.
Forecasting predicts demand; demand planning decides what to do about it. The forecast says "this SKU will sell about 600 units next month"; demand planning turns that into POs, accounts for cash and capacity, and reconciles it with what marketing and sourcing are planning. See demand planning for sellers for the wider loop. A forecast that never becomes a purchase order changes nothing.
A plain average works until a SKU has structure. Three signals need explicit handling:
Trend. Demand drifting up or down. A trailing average always lags a trend, so a growing SKU is chronically under-forecast. Lean on recent periods or add a trend term.
Seasonality. Predictable peaks (Q4, summer, a back-to-school window). Model these with seasonal indices so you build before the ramp, not after. See how to forecast Q4 inventory.
Promotions and events. A deal, a coupon, or a Prime Day spike is demand you created, not baseline. Forecast it separately and do not let the spike pollute your everyday average. See Prime Day inventory planning.
Say a SKU sold 280, 350, and 420 units over the last three months, clearly trending up. A flat three-month average forecasts (280 + 350 + 420) / 3 = 350, which is already well below last month's 420. A weighted average that leans on recent sales lands higher and closer to the real trajectory, and the harder you weight the latest month the closer it gets. The lesson for the pillar level is simple: on a moving SKU, the method you pick changes the number you order by a meaningful margin, so it is worth choosing deliberately. The full side-by-side of moving average, weighted average, and exponential smoothing lives in inventory forecasting methods. Run your velocity inputs in the sell-through rate calculator and reorder point calculator.
A forecast is only as good as its inputs, all of which you already have:
Sales history comes from your Seller Central Business Reports (units ordered by ASIN, by date). Pull at least a year for seasonal SKUs so you can see the full cycle. Then exclude out-of-stock days, or those zeros read as low demand and quietly shrink the forecast. To find those days, cross-reference the Inventory Ledger or your FBA inventory history (which shows when sellable units hit zero) against the date range of your sales pull, and drop the zero-stock days from the denominator. It is the same correction you make when calculating sales velocity.
Trend and seasonality are read from that same history, but you have to look back far enough to see them: compare each period to the same period a year before, not just to last month, or you will mistake a seasonal dip for a declining SKU.
Lead time sets how far out the forecast must reach, and it is the full chain from PO placed to units sellable, not the supplier's quoted production time. A SKU made in 30 days but 50 more days in freight, customs, and check-in needs an 80-day forecast horizon. See inventory lead time.
Planned events (promos, deals, launches) come from your own calendar, not the history. Add their lift on top of the baseline, and keep the spike out of the baseline afterward so it does not inflate the next several forecasts.
Pulling and updating all of this per SKU every week is the real work, and it is exactly what a forecasting tool exists to automate.
A forecast frequently asks for more inventory than your cash can cover, which quietly turns forecasting into a prioritization problem. The forecast itself does not decide what to buy; it feeds the decision. When the total demand across SKUs exceeds your available cash, rank by the cost of being wrong: stockout cost first (a fast, high-margin SKU about to run dry loses the most for every day it is out of stock, including the rank you then rebuild), then by how fast the cash comes back (margin and sell-through). Fund the high-risk, high-return SKUs, and defer slow movers and anything with months of cover already on hand. The full mechanics of that reconciliation live in demand planning and the cash-constrained walkthrough in restock planning; the point at the forecasting stage is that a good forecast is what lets you make that call deliberately instead of reacting to whichever SKU stocks out first.
A brand-new SKU has no sales history, so you cannot project it forward. Forecast it from the closest thing you have: a comparable product in your catalog, the category's typical launch curve, and a deliberately conservative first order so you are not betting cash on an unproven guess. See new product forecasting and inventory for a product launch. The first reorder, once real sales arrive, matters more than the launch forecast.
A forecast you never grade never improves. Track forecast error per SKU, the gap between what you predicted and what actually sold, and tighten the method where the error is worst. Two numbers make this concrete:
MAPE (mean absolute percentage error) is the average size of your misses. For each period take the absolute error divided by actual sales, then average across periods. A SKU forecast at 500 that sold 600 missed by 100 / 600 = 17% that month; average that over several months and you have its MAPE. It tells you how noisy the SKU is.
Bias is the direction of the misses. If your forecast sits below actual most months, you are systematically under-forecasting, which on FBA shows up as recurring stockouts; consistently above means you are overbuying.
The fix depends on which number is bad. High MAPE with low bias is a genuinely noisy SKU that needs more safety stock, not a cleverer forecast. Consistent negative bias means a trend or seasonality you are not modeling yet. Forecast accuracy is a KPI you watch per SKU, not a one-time guess, and you spend your effort on the worst offenders.
Forecasting shorter than your lead time. The most common and most expensive error: a 30-day forecast cannot tell you what to order when your stock takes 80 days to arrive.
Letting a promo spike pollute the baseline. A Prime Day or coupon spike is demand you created, not your run rate. Leave it in the history and it inflates the next several forecasts and walks you into overbuying.
Ignoring out-of-stock days. Days the SKU could not sell read as low demand and quietly shrink the forecast, which sets your reorder point too low and causes the next stockout.
A model you do not maintain. A weighted average you update weekly beats a machine-learning model you configured once and abandoned. Match the method to the time you will actually give it.
One forecast for a whole variation family. Sizes and colors sell at different rates, so a blended parent forecast over-buys the slow children and starves the fast ones. Forecast at the level you actually place orders.
Amazon inventory forecasting is predicting per-SKU demand far enough ahead to buy against your full lead time, using a method simple enough to maintain and rich enough to capture trend, seasonality, and events. Start with a weighted moving average, handle the three signals a flat average misses, track your forecast error, and feed the result into your restock plan.