Last month's sales are not next month's demand: building a usable wholesale forecast

A customer buys three months' stock in one order. Another cannot buy anything because the line is unavailable. Both events appear in the sales history, but neither tells the planner what an ordinary month should look like. This is where a wholesale forecast can go wrong long before anyone chooses a forecasting model.

The business needs a repeatable process that distinguishes recorded sales, expected demand and the quantity it should purchase. Those numbers are related, but they answer different questions. A sales target is another number again: what the commercial team hopes to achieve is not automatically the most likely outcome.

Start with the decision the forecast must support

Define the item, location, customer group and time period needed for replenishment. A monthly forecast for an entire category may help budgeting but be too broad to decide which size or variant to order next Tuesday. Equally, producing daily forecasts for every slow-moving line may add noise without improving a purchasing decision.

The planning horizon must cover the period during which the business cannot respond easily. If supplier lead time is six weeks and orders are reviewed weekly, the buyer needs a view beyond the next few days. The horizon should reflect the review cycle and the actual replenishment process, including receiving and release time where relevant.

Agree ownership before the first review. The planner maintains the baseline, sales provides specific customer intelligence, purchasing confirms supply constraints, and the authorised commercial manager approves material changes under the business's policy. One person can hold several roles in a small firm, but the reasons for each change should still be visible.

Keep the original sales record intact

Do not rewrite transaction history to make it look normal. Retain the sales record and add a separate planning view that identifies promotions, product launches, stockouts, unusual bulk orders and changes in distribution.

A zero-sales week during a stockout is not proof of zero demand. At the same time, it is not permission to invent the missing sales. Use available evidence such as unfulfilled orders, customer enquiries and comparable in-stock periods. Record an estimate as an estimate, with its basis and uncertainty.

Returns require care too. A large return posted this week may relate to sales made months earlier. Netting it blindly against current demand can obscure both the original selling pattern and the current returns problem. Keep the commercial event traceable while choosing a consistent forecasting treatment.

Separate the baseline from identifiable events

Hyndman and Athanasopoulos, in Forecasting: Principles and Practice, discuss the value and limitations of judgmental adjustments to statistical forecasts. Their guidance supports using relevant information that the model lacks, while documenting adjustments and avoiding changes made merely to express confidence or ownership.

In practical terms, retain the unadjusted forecast alongside the final version. An account manager's statement that a retailer has confirmed a promotion is information the model may not contain. “We need a strong quarter” is a target, not evidence of additional demand.

Suppose a line normally sells 100 units a week. A four-week promotion is expected to add 40 units a week for a defined customer group. The planning record can show a baseline of 400 units plus a promotional assumption of 160. These are illustrative quantities, not a recommended uplift.

If the promotion is cancelled, the planner can remove the 160-unit event without rebuilding the baseline. If the uplift is already included in the model, adding it manually would count it twice. The review must establish what information is already present before an override is approved.

Translate demand into an order only after checking stock

A forecast of 560 units does not mean the purchase order should contain 560 units. The buyer also needs usable stock, confirmed inbound quantities, outstanding commitments, the replenishment policy and supplier constraints.

For a simplified example, assume the relevant planning period needs 560 units, usable uncommitted stock is 220, and a reliable inbound order will provide 180 in time. The uncovered requirement is 160 units before any additional buffer or order-rounding rule. If the supplier sells in cases of 24, the operational order may need to be rounded under the agreed policy.

Committed customer orders must be treated consistently. If they are already included in the forecast, adding them again inflates demand. If the business uses a process in which firm orders consume the forecast, explain and configure that process explicitly. The planner and buyer should be able to reconcile the final requirement.

Measure the effect of adjustments

Keep a record of the forecast as it stood when the buying decision was made. Replacing it with a later, better-informed version makes accuracy look better without improving the original decision.

Compare baseline and adjusted forecasts over the same horizon. A change can be sensible and still turn out wrong, so review a series of decisions rather than punishing one miss. Look for persistent bias: are particular accounts, planners or event types repeatedly adding more demand than materialises?

Use measures that suit the data. Percentage errors become awkward when actual demand is zero or very small. An absolute unit error, an aggregate weighted measure and a signed bias measure can answer different questions. The point is to make the review understandable, not to select the metric that produces the most flattering score.

Give uncertainty an operational consequence

A forecast range is useful only if it changes a decision. For a volatile new line, the response might be a smaller opening order, an earlier review or an agreed supplier call-off arrangement. For a stable line with a long replenishment time, the business may choose a different buffer.

Record the commercial trade-off and who approved it. Sales may prefer availability, finance may prefer less stock, and purchasing may face a minimum order. The plan should resolve those tensions before the order is placed.

A workable forecast does not remove uncertainty. It makes the assumptions visible, links them to a buying decision and leaves enough evidence to improve the next one. That is more valuable than a precise-looking spreadsheet whose largest numbers came from undocumented optimism.

Reference: Hyndman and Athanasopoulos, Forecasting: Principles and Practice, judgmental forecasts. The wholesale workflow and numerical examples are original illustrations.