For most Amazon FBA sellers, a weighted moving average that leans on the most recent weeks is the best balance of accuracy and maintainability. Simple moving averages lag trends; advanced models help at scale but only if you keep them updated. The best method is the most responsive one you will actually maintain.
What is the difference between a moving average and a weighted moving average?
A simple moving average treats every period in the window equally. A weighted moving average gives recent periods more weight, so the forecast reacts faster to a change in demand. On a rising or falling SKU, the weighted version is consistently closer to reality.
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.
Exponential smoothing forecasts the next period as a blend of the last actual sales and the last forecast, controlled by a smoothing factor (alpha) between 0 and 1. A higher alpha reacts faster to change; a lower alpha is smoother and steadier. It is a responsive method that automates cleanly.
Read article
The main inventory forecasting methods are the simple moving average, the weighted moving average, and exponential smoothing, and for most Amazon FBA SKUs a weighted moving average is the right default. The short version: simple averages are stable but lag changes, weighted averages follow recent demand faster, and exponential smoothing gives you one tuning knob for responsiveness. Below is how each works, a side-by-side on the same SKU, and how to choose.
Average the last N periods of sales and use that as the forecast. With N = 3 and sales of 400, 460, and 520, the forecast is (400 + 460 + 520) / 3 = 460. It is stable and easy, but every period counts equally, so it lags a trend: here it forecasts 460 even though last month was already 520 and climbing. The window N is the dial: a short window (4 weeks) reacts faster but is noisier, a long one (8 to 12 weeks) is smoother but slower to notice a real change. Four to eight weeks suits most FBA SKUs. Best for mature, steady products where demand bounces around a flat line.
The same window, but recent periods get more weight. Using weights of 0.5 for last month, 0.3 for two months ago, and 0.2 for three months ago on the same data:
Month
Units
Weight
Contribution
Three months ago
400
0.2
80
Two months ago
460
0.3
138
Last month
520
0.5
260
Weighted forecast
478
The weighted method forecasts 478 versus the simple 460, closer to the rising reality. The weights are yours to set: lean harder on the recent period for trending SKUs, flatter for steady ones. This is the best default for most FBA SKUs because it reacts without being noisy.
Exponential smoothing forecasts the next period as forecast = alpha x last actual + (1 - alpha) x last forecast, where alpha is between 0 and 1. A high alpha (say 0.5) reacts fast to change; a low alpha (say 0.2) is smoother. It needs only the last forecast and the last actual, so it automates cleanly and never requires storing a long window. It is the natural next step up from a weighted average when you want a single responsiveness dial. Start alpha around 0.3 for a typical FBA SKU, then tune it by minimizing your past forecast error: raise it for fast-moving or trending products, lower it for steady ones.
The 0.5 / 0.3 / 0.2 weights in the example above are a reasonable starting point for a weighted average, and alpha around 0.3 is a safe start for exponential smoothing. Do not agonize over the exact values. Pick a sensible start, then track your forecast error per SKU over four to six weeks and adjust: if you are chronically under-forecasting a SKU, shift more weight to the most recent period (or raise alpha); if the forecast is jumpy, do the opposite. The tuning matters far less than actually running the method every week.
None of the three base methods handles a seasonal SKU on its own. The standard add-on is a seasonal index: divide each period's sales by the year's average to get a multiplier per period, so a December index of 1.8 means December typically runs 80 percent above the yearly average. You forecast the baseline with one of the methods above, then multiply by the period's index. See how to model seasonality for the full method. Whatever you pick, it still has to forecast at least one full lead time out, because that is how far ahead you are buying.
The data comes from the same place for every method: your Seller Central Business Reports (units ordered by ASIN, by date), with out-of-stock days excluded so the zeros do not drag the forecast down, the same correction you make for sales velocity. Once you have a forecast, demand planning turns it into purchase orders, and you can run the numbers through the reorder point calculator. For the full picture, see Amazon inventory forecasting.
Inventory forecasting methods trade responsiveness against stability: a simple moving average for steady SKUs, a weighted moving average as the default for anything trending, and exponential smoothing when you want one dial you can automate. Pick the most responsive method you will maintain, layer seasonality where it exists, and always forecast past your lead time.