Demand forecasting is the estimation of how many units of a product will sell in a future period. In an inventory business it exists to answer one operational question: how much do I need to have, and when do I need to have committed the cash to get it. A forecast that does not connect to a purchase order date is a report, not a plan.
The term gets used loosely for two different things. Forecasting total company demand is a planning exercise. Forecasting one SKU on one channel is an inventory exercise. They use similar math and behave very differently.
Forecast horizon. How far ahead the estimate reaches. For inventory purposes the horizon that matters is lead time plus review period, not a calendar quarter.
Lead time demand. Expected units sold between placing a purchase order and having the goods available to sell. This is the number a reorder point is built from, and it is a distribution rather than a single value.
Review period. How often you actually look and decide. A seller who places orders monthly has to cover lead time plus 30 days, not lead time alone.
Service level. The probability of not stocking out during lead time, chosen deliberately. Ninety-five percent and 99 percent sound similar and imply very different safety stock, because the tail of the demand distribution is where the cost sits.
Safety stock. Units held to absorb variance in demand and in lead time. It is a function of variability, not of average demand, which is why fast-moving stable SKUs often need less than slow erratic ones.
MAPE. Mean absolute percentage error. The common accuracy measure. It behaves badly on intermittent demand, because dividing by a near-zero actual produces enormous percentages, and most long-tail ecommerce SKUs have intermittent demand.
Bias. The average signed error. Consistently forecasting high is bias. Forecasting wildly in both directions is error. A model can have near-zero bias and terrible error, or low error and persistent bias, and the fixes are different.
Intermittent demand. Sales series with many zero days. Standard time series methods handle these poorly, which is why the long tail of a catalog is the hardest part to forecast and usually the part nobody forecasts at all.
Cannibalization. Demand shifting between your own listings, most often when a bundle or multipack competes with the single unit. Forecasting each SKU independently misses this entirely.
Naive baseline. The simplest possible forecast, usually the last period repeated. Any method that cannot beat it on your data has not earned its price.
Inventory is a cash position that happens to be shaped like boxes. The U.S. Census Bureau's Monthly Retail Trade and Food Services data, retrieved from FRED, put the retail inventories to sales ratio for retail trade excluding motor vehicle and parts dealers at 1.08 in May 2026, seasonally adjusted, down from 1.12 in January 2026. That ratio means U.S. retailers as a group were carrying roughly a month of sales in stock.
For a seller holding a month of sales in inventory, a forecast error of 20 percent on the wrong SKUs is a working capital problem before it is a service problem. That is the honest framing. Forecasting exists to reduce the amount of cash sitting in the wrong boxes.
A company-level demand series is smooth because it is the sum of hundreds of independent series. Individual errors offset. That is a statistical property, not a management achievement.
A single SKU on a single channel has none of that smoothing. It responds to search rank, advertising spend, a competitor going out of stock, a review, a deal event, a listing suppression. Its variance is high, its zero days are frequent, and the same SKU on a second channel has a different pattern entirely.
The consequence: aggregate accuracy figures quoted by any forecasting vendor tell you almost nothing about how the tool will perform on the decision you actually make. Ask for accuracy at the item and location level, which is where the M5 research found that the advantage of complex methods shrinks. That finding, and what it means for a reorder decision, is unpacked in how AI changes inventory planning.
One SKU, sold on Amazon and Shopify. Landed cost $7.85.
Trailing 12 weeks of combined demand, in units per week: 940, 1,110, 860, 1,290, 1,020, 780, 1,450, 990, 1,130, 870, 1,210, 1,060.
Mean is 1,059 units per week. Standard deviation is 189.
Lead time demand. Supplier lead time is 42 days, or 6 weeks. Expected demand over lead time is 6,354 units. Standard deviation over lead time, assuming independence, is 189 times the square root of 6, or 463 units.
Safety stock at 95 percent service. The z value is 1.65, so safety stock is 764 units, or $5,997 of committed cash.
Safety stock at 99 percent service. The z value is 2.33, so safety stock is 1,079 units, or $8,470.
Four percentage points of service level costs $2,473 in permanently committed cash for this one SKU. Across 300 SKUs with similar profiles the decision is worth hundreds of thousands of dollars, and it is a decision, not a calculation.
Reorder point. Lead time demand plus safety stock: 7,118 units at 95 percent service. When available stock, meaning on hand plus inbound, falls to that number, the order goes out.
Now change one thing. Suppose lead time is not reliably 42 days but has run from 38 to 71. Recomputing with lead time variance included pushes the reorder point past 9,000 units. The forecast did not change. The answer moved by 27 percent.
A forecast is only as good as the two numbers it sits on: how many units you actually have, and what each one cost.
Unit availability has to be tracked by location, including at marketplace fulfillment centers, with transfers treated as in transit until received so units are neither double counted nor invisible. The inventory layer inside ConnectBooks works this way, with FIFO valuation and a restock report driven by sales history and velocity with lead times and inbound stock included. Seasonality is not currently part of that calculation, which is worth knowing if your demand curve is holiday-shaped.
Unit cost has to include freight, duty, and inbound handling. A forecast that recommends buying more of a SKU whose real margin is eight points lower than the books show is an expensive kind of accurate. Profit reporting at SKU and channel level is what makes that check possible, and it depends on settlement data reconciled at transaction level rather than posted as monthly summaries.
Sell-through rate. Units sold divided by units received, over a period. A retrospective measure that constrains what a forecast should look like.
Forecast value added. The improvement a forecasting process delivers over the naive baseline. Often negative in practice, which is why measuring it is worth the hour.
Demand sensing. Short-horizon adjustment using recent signals such as traffic or add-to-cart rates. Useful for the next two weeks, not for a container order.
Consumption forecast versus shipment forecast. What customers buy versus what you move into a channel. Marketplace sellers often conflate these, and inbound shipments then look like demand.
Forecast to size a purchase order, not to produce a report. Pick a service level on purpose and know what it costs in cash. Measure your method against the naive baseline before believing it. And check that the units and costs underneath the forecast reconcile to something real, because every forecasting technique in this glossary assumes the current position is correct.
Running an e-commerce business comes with plenty of challenges, but ConnectBooks is here to make your life easier. With real-time insights, seamless integrations, and detailed tracking of your profitability and inventory, you can stay ahead of the game. Whether you’re selling on Amazon, Shopify, Walmart, TikTok or eBay, ConnectBooks helps you manage your finances with 100% accuracy and confidence, so you can focus on growing your business.
Ready to level up? Start making smarter, data-driven decisions every step of the way. Try ConnectBooks Free Today or Schedule a Demo
