Demand forecasting for a China-based ecommerce operation should estimate the units customers will request during the replenishment horizon, not simply extend the warehouse’s recent shipments. Shipments can fall because stock ran out, a listing was paused or a product failed inspection. Treating those interruptions as weak demand creates the next shortage.
Begin with a small set of SKU-level forecasts that can be checked against actual orders. A complicated model is difficult to justify until the team has a reliable demand history and a meaningful baseline.
Build a demand history that preserves interruptions
Keep order creation, cancellation, shipment and return dates separately. Count the demand event chosen for the forecast consistently. If the objective is to reserve stock at checkout, a shipment-date series will place demand too late.
Mark periods of stockout, restricted selling, unusual promotions and channel outages. Keep the observed orders rather than silently replacing them with estimates. If lost demand is estimated, store that estimate in a separate field with its method and assumptions.
Product changes need similar treatment. A revised garment size chart, a new bundle or a different pack count may change the meaning of one unit. Joining their histories under the same SKU can make a demand jump look like a model failure when the product itself changed.
Match the forecast horizon to the decision
A picking roster needs a short-horizon order and line forecast. A replenishment purchase needs expected unit demand over production, receiving and release time. A shipping-mode decision may need the stock deficit under several possible arrival windows.
Write down the decision before selecting the model. Weekly units per SKU may be appropriate for replenishment, while an hourly order-arrival profile is more useful for a flash sale. Do not force both decisions into one monthly sales total.
Separate the underlying forecast from a management adjustment. A promotion uplift or launch assumption should have an owner, an explanation and an expiry date so that its effect can be evaluated later.
Compare a simple baseline with the proposed method
Useful baselines include the recent average and the corresponding period in a previous seasonal cycle, where sufficient history exists. New products without comparable history need explicitly judgmental scenarios rather than a confidence interval that looks statistically established.
Evaluate forecasts using information that would have been available at the time. The forecasting textbook Forecasting: Principles and Practice distinguishes errors on future observations from residuals on data already used to fit a model. This difference matters when a vendor presents an impressive historical fit.
For an illustrative three-week test, actual demand is 90, 120 and 110 units. Forecast A predicts 100 each week, giving absolute errors of 10, 20 and 10; its mean absolute error is 13.3 units. Forecast B predicts 95, 115 and 105, giving a mean absolute error of 5 units. That small test favors B for those weeks. It does not prove B will remain better through a promotion or product change.
Check both bias and costly misses
A model can average out to the correct total while repeatedly underforecasting popular SKUs. Review signed error for bias, absolute error for typical miss size and the largest shortage-producing errors.
Avoid relying only on percentage error for slow-moving products. A week with zero demand makes some percentage measures undefined. Use a suitable unit-based or scaled measure and inspect the actual series.
Inventory outcomes also depend on lead-time variability, order constraints and inventory policy. A forecast with lower average error may still produce worse service if its largest misses occur on critical replenishment dates.
Translate the forecast into a controlled stock decision
Suppose an illustrative four-week forecast is 400 units. The operation has 230 saleable units, 60 units committed to existing orders and 100 confirmed incoming units. The replenishment calculation must use a clearly defined available-stock figure and avoid subtracting commitments twice. Shopify’s inventory states provide one example of why these quantities should remain distinct.
Document the expected demand, usable inventory, confirmed receipts and chosen uncertainty allowance in the purchase recommendation. Then review what changed when actual demand arrives. A forecast becomes operationally valuable when the buyer can explain the resulting order quantity, including the assumptions that could make it wrong.
Sources and verification links
Operational rules can change. Check the linked primary sources and confirm shipment-specific requirements before booking.