Inventory Forecasting for eCommerce: How to Predict Demand by SKU, Channel and Warehouse

SECTION 01Inventory forecasting in one paragraph

In brief: Inventory forecasting is predicting how many units of each SKU you will sell over a coming period, by channel and location, so you can decide what to buy, how much and where to place it. It combines sales history, seasonality, planned events and supplier lead times, and its output is a reorder decision, not just a number.

A forecast is only useful if it changes what you do. For an eCommerce brand, that means three decisions: how much to buy from suppliers, when to place the order so it lands in time, and which warehouse, marketplace fulfillment center or dark store should hold it. Get the forecast right and stock is where orders are. Get it wrong one way and you pay for stockouts, lost sales and marketplace ranking drops. Get it wrong the other way and cash sits in dead stock.

Inventory forecasting is the practical half of demand forecasting: demand forecasting predicts what customers will buy, and inventory forecasting turns that into stock. It is also the foundation of everything else in eCommerce inventory management: reorder points, safety stock, festive build-ups and warehouse transfers all start from a view of future demand.

SECTION 02Where inventory forecasting fits

In brief: Demand forecasting predicts how many units customers will order. Inventory forecasting takes that prediction and works out how much stock you need, where and by when, after allowing for supplier lead times, safety stock and stock already on hand or on order. This guide covers the second step; the demand side has its own guide.

The two work as a pair. A demand forecast on its own does not tell a buyer what to order this week; an inventory forecast without an honest demand view just reorders yesterday's mistakes. For the demand side, including the types of demand forecasting, the full range of methods and how to read demand signals, see our demand forecasting guide. The rest of this guide is about stock.

SECTION 03The data you need

In brief: You need daily or weekly sales by SKU, channel and location, ideally two years of it; a record of stockout days and promotions; supplier lead times and how much they vary; and current stock on hand, in transit and on order. Clean the history first: a forecast built on distorted sales repeats the distortion.

• Sales history by SKU, channel and location. At least 12 months to see one seasonal cycle, ideally 24 to see whether the pattern repeats. Weekly buckets are usually enough; daily helps for fast movers and quick commerce.
• Stockout days. When a SKU was out of stock, you sold what you had, not what customers wanted. Flag those days and exclude or adjust them, or your forecast will learn to under-stock.
• Promotions and events. Record every sale, discount and campaign with its dates. Promotional spikes should be modeled separately, then added back for the next event, not averaged into the baseline.
• Lead times. Supplier, production and transfer lead times, and how much they swing. Variable lead times need more safety stock than long but reliable ones.
• Stock position. On hand, reserved, in transit and on order, by location. A forecast without the current position cannot produce a reorder.
• Returns. In categories with high returns, net demand differs from gross orders. Forecast on what customers keep, or plan returned stock back in.

Three rules of thumb that practitioners repeat: forecast units, not revenue, because price changes distort revenue history; never trust a forecast built on stockout periods; and keep a written log of why the history looks odd in any given week, because nobody remembers six months later.

SECTION 04Choosing a method for each SKU

In brief: Use moving averages for stable SKUs, exponential smoothing when recent weeks matter more, seasonal methods when demand follows the calendar, regression when price, promotions or ads drive sales, and machine learning only with long, clean history across many SKUs. Most brands need two or three methods, chosen by SKU.

The table below compares the methods most used for SKU-level inventory forecasting. Our demand forecasting guide covers the wider set, including qualitative methods for markets with no history.

Which forecasting method should an eCommerce brand use?

Methods work best when matched to the SKU, so segment first. Group SKUs by value (ABC: the top sellers that make most of revenue, then the middle, then the long tail) and by how predictable they are. Steady A items can run on exponential smoothing or seasonal methods and deserve weekly review. Erratic long-tail SKUs rarely forecast well with any method; keep them on simple rules and small safety stock. Use regression for SKUs you promote heavily. Bring in machine learning when your history is long and clean enough to feed it, and keep a simple benchmark running alongside so you can tell whether it is actually better.

SECTION 05How to forecast inventory, step by step

In brief: Clean the sales history, segment SKUs, pick a method per segment, forecast by SKU and channel, layer in planned events and launches, split the forecast by warehouse, convert it into reorder points and order quantities, then measure accuracy and repeat on a fixed cycle, weekly for fast movers.

1. Clean the history. Remove or adjust stockout days, separate promotion spikes, and correct one-off events such as a bulk B2B order.
2. Segment SKUs. Group by value and by how predictable demand is, so effort goes where the money is.
3. Choose a method per segment. Use the method-picker table above, and keep a "same as last period" forecast as the benchmark.
4. Forecast by SKU and channel. Marketplaces, quick commerce and your own site behave differently; forecast each, then add them up.
5. Add events and launches. Layer planned sales, campaigns and new SKUs on top of the baseline.
6. Split by warehouse. Allocate the forecast to the locations that will serve the demand, using each region's share of orders.
7. Set reorder points and quantities. Turn the forecast into when to order and how much, allowing for lead time and safety stock.
8. Measure and repeat. Compare forecast with actual, track accuracy and bias, and re-run on a fixed cycle.

Clean the history, forecast by SKU and channel, add events, split by warehouse, set reorder points, measure and repeat.

SECTION 06Forecasting by channel and location

In brief: Forecast at the level you make each decision. Buy from suppliers on total demand across channels, but forecast and allocate by channel and warehouse, because a marketplace sale, a quick-commerce dark store and your webstore draw on stock differently and on different clocks. One stock pool across channels is what makes this work.

This is where single-channel forecasting breaks. A brand selling on its webstore, two marketplaces and quick commerce does not have one demand curve; it has several, each with its own rhythm:

• Marketplaces spike around their own sale events and reward in-stock listings with ranking. A stockout costs the lost sale and the visibility that follows.
• Quick commerce runs on dark-store stock replenished every 24 to 48 hours, with fill-rate expectations on every purchase order. The forecast needs to work at city or dark-store level and a short horizon.
• Your webstore follows your own campaigns, email sends and ad spend, which you can plan precisely.
• B2B and wholesale orders arrive in lumps. Forecast them separately so one large order does not inflate the retail baseline.

Location adds a second layer. If 40% of orders ship to one region, that warehouse should hold roughly that share of stock, adjusted for the speed each channel needs. Forecasting by location also tells you when to move stock between warehouses rather than buy more, which our guide to multi-warehouse management covers in detail. For channel-specific planning, see our quick commerce guide.

The rule underneath all of this: forecast low, act high. Forecast at SKU by channel by location so the detail is honest, then add up for purchasing, where supplier minimums and lead times apply to the total.

SECTION 07Planning for sale events and new launches

In brief: For sale events, start from the uplift the last comparable event delivered, adjust for this year's discounts, channels and growth, and build stock early enough to cover lead times. For new launches, borrow the demand curve of a similar existing SKU, buy conservatively and re-forecast weekly once real sales arrive.

Sale seasons are where forecasts are made or broken, and they are changing shape. EasyEcom's Festive Report 2025 found that festive shopping in India now begins months in advance, with momentum building from August, and that by the final weeks daily order volumes were up 300 to 400%. Orders processed across brands roughly doubled, from 21.8 million to 43.4 million. A plan built around one peak week would have been short for weeks.

A worked example: forecasting a sale event

Suppose a SKU sells a baseline of 50 units a day. During last year's main sale week it sold 175 units a day, an uplift of 3.5 times. This year you are offering a deeper discount and adding a quick-commerce channel, so you plan for 4 times, or 200 units a day, for seven days: 1,400 units against a normal week's 350. If supplier lead time is 30 days, that extra stock has to be ordered about five weeks before the sale, not when the sale is announced. Then plan the other side: demand often dips right after a big event, and returns from the sale come back in the following weeks.

The same logic applies to Black Friday and Prime Day in the US, and to marketplace sale events everywhere: uplift from the last comparable event, adjusted for what changed, built ahead of lead time.

Forecasting a new launch

A new SKU has no history, so borrow one. Pick an analog: a similar product in the same category, price band and channel mix, and use its first months as the starting curve. Buy a conservative first batch, watch sell-through daily for the first two weeks, and re-forecast weekly. Speed of reorder matters more than precision of the first guess.

SECTION 08Measuring forecast accuracy

In brief: Track WAPE, the total absolute error divided by total actual sales, because it weights big sellers properly, and bias, which shows whether you consistently over- or under-forecast. MAPE is common but distorted by slow movers. There is no universal target: beat a simple "same as last period" forecast, then keep improving.

Three measures cover almost everything:

MAPE = average of ( |actual − forecast| ÷ actual ), across SKUs or periods
WAPE = sum of |actual − forecast| ÷ sum of actual
Bias = sum of (forecast − actual) ÷ sum of actual

• MAPE (mean absolute percentage error) is easy to explain but explodes on slow movers: miss a 2-unit SKU by 2 and it counts as a 100% error, the same as missing a 2,000-unit SKU by 2,000.
• WAPE (weighted absolute percentage error) weights each error by volume, so it reflects what the misses cost the business. It is the better headline measure for eCommerce catalogs with long tails.
• Bias shows direction. A forecast that is always 10% high quietly builds dead stock; always 10% low, and you live with stockouts. Aim for bias near zero.

Our demand forecasting guide goes deeper on accuracy measures and forecast reviews. How accurate does a forecast need to be? Accurate enough to beat the naive benchmark and to keep stockouts and excess within what your margins can absorb. Focus accuracy work on the SKUs that carry the revenue. Good inputs move the number most: Wholsum Foods reached 99.99% demand forecast accuracy within six months on EasyEcom, as its customer story describes.

SECTION 09Turning forecasts into reorders

In brief: A forecast becomes action through the reorder point, which is expected demand during the supplier's lead time plus safety stock, and the order quantity, which covers demand until the next order arrives minus stock on hand and on order. Recalculate both whenever the forecast changes, and allocate each receipt by warehouse.

Reorder point = (forecast daily demand × lead time in days) + safety stock
Order quantity = forecast demand over the cover period + safety stock − stock on hand − stock on order

Worked example. A SKU is forecast at 40 units a day, the supplier takes 10 days, and you hold 120 units of safety stock. The reorder point is 40 × 10 + 120 = 520. When stock falls to 520, order. If you order every 30 days, the quantity is 40 × 30 + 120 = 1,320, minus whatever is on hand and already on order.

Two habits keep this honest. First, recalculate reorder points when the forecast moves, especially in the run-up to a sale; a reorder point set in a quiet month will be far too low in a festive one. Second, decide allocation at receipt, not after: route the inbound stock to the warehouses whose forecast needs it, instead of unloading everything at one location and transferring later.

Done consistently, this is how forecasting pays for itself. Across brands in EasyEcom's Festive Report 2025, dead stock fell from 23.2% to 11.7% even as order volumes doubled.

SECTION 10Where EasyEcom fits

In brief: Forecasting is only as good as the stock and sales data behind it. EasyEcom keeps one real-time stock position across every channel and warehouse, records sales by SKU, channel and location, flags low stock and expiry, and feeds forecasts into planning, so replenishment happens ahead of demand rather than after a stockout.

The hardest part of forecasting for most brands is not the math; it is getting clean, connected data out of several channels and warehouses. EasyEcom's order management system and warehouse management run on one inventory record, so sales history, stockouts, stock on hand and stock in transit come from the same source. At Wholsum Foods, the forecast sets the plan, and low-stock and expiry alerts flag what needs attention. RedTape replenishes ahead of demand. For quick commerce, dark-store-level stock and 24 to 48 hour replenishment cycles keep fill rates where the platforms expect them.

If your next sale season is already on the calendar, book a demo and we will walk through your forecast-to-reorder flow with your own data.