Demand Forecasting: Planning Hiring, Capacity and Inventory Without Confusing Sales Targets with Reality

A distributor wants to sell 1,560 units over the next six weeks. Its evidence supports expected demand of 1,200, but purchasing must commit before customers reveal the answer. Ordering for the target risks surplus stock; ordering for the average risks missed sales. The difficult question is not which number sounds more confident. It is how much the business should put at risk now, given what it knows and what it can still change.

A service company faces the same decision without a warehouse. Its probability-weighted pipeline exactly matches the team’s available hours, yet several projects starting together could overwhelm delivery. Hiring for every possibility could be equally damaging. Both companies need to translate uncertain demand into commitments that have real costs, deadlines and consequences.

Demand forecasting becomes valuable when it changes a decision. This article follows that connection from evidence to purchasing, staffing, cash and customer promises. Two worked examples show why the appropriate commitment can differ from the forecast, while a practical decision brief explains how to make the discipline part of running a business. The aim is neither perfect prediction nor cautious growth: it is to pursue opportunity without making obligations that the evidence cannot support.

1. Demand Forecasting: Separating Targets, Forecasts, Plans and Commitments

A target expresses what management wants to achieve. A forecast estimates what is likely under stated assumptions. A plan describes the actions intended to improve the outcome. Hyndman and Athanasopoulos make this distinction in Forecasting: Principles and Practice: goals, forecasts and planning should work together, but they serve different purposes. Confusing them makes it harder to see whether ambition has an operational foundation.

A fourth element makes the distinction commercially useful: the commitment. A purchase order, employment contract, production booking or delivery promise converts a plan into an obligation. Management can reasonably commit above or below the central forecast, depending on the cost of shortage, the consequences of surplus and the flexibility available. That is a business decision, not a reason to alter the forecast until it matches the desired expenditure.

Suppose the target is 30% growth while the current forecast suggests 12%. The difference should become a work programme: which customers, offers or channels could close the gap, what must happen first, and what evidence would show progress? Raising the forecast to 30% removes the warning without creating demand. Treating 12% as an unavoidable ceiling would make the opposite mistake by ignoring the company’s ability to change its prospects.

Maintain a visible distinction between the current outlook and the additional demand expected from planned initiatives. A funded campaign with relevant evidence deserves different treatment from an idea awaiting approval. Record its expected timing, assumptions and incremental contribution without counting activity already included in the baseline. If an initiative slips, management can then revise the forecast and commitments without losing sight of the original growth objective.

Also distinguish customer demand from the sales the business can fulfil. If customers would buy 1,400 units but the company can supply only 1,100, report both figures and the gap. Quietly reducing the demand forecast to 1,100 conceals the commercial opportunity; promising all 1,400 without securing supply conceals the operating risk. The purpose of the planning discussion is to resolve that gap, not make it disappear on paper.

2. Forecast the decision you actually have to make

Choose the unit, the date and the level of detail

Begin with a specific question: how much of this product must be ordered today, how many qualified delivery hours will be needed next month, or whether sustained demand justifies another permanent role. Identify the resource consumed, the person making the decision and the last useful date for acting. A forecast that becomes accurate after that deadline cannot improve the commitment already made.

Revenue alone is rarely enough to answer these questions. Two contracts with the same value can consume very different quantities of labour, materials or equipment time. Forecast the relevant operating units, then translate them into financial outcomes using the expected price and product mix. A revenue increase driven by higher prices, for example, does not automatically justify an equivalent increase in inventory or staffing.

Preserve four separate dates where they differ: order placement, delivery or service, revenue recognition and cash collection. The same transaction may appear in different periods across sales, operations and finance. Trace it through those periods rather than asking one monthly sales number to perform every job; otherwise a strong order book can conceal a near-term delivery or funding problem.

Match the horizon to the time needed to respond

Suppose replenishment takes four weeks and purchasing reviews orders every two weeks. An order placed at the next review arrives six weeks from today. The present stock decision therefore needs to consider demand until that later replenishment becomes available, while checking existing stock and dated receipts along the way. A six-week total does not solve a shortage in week two if the incoming goods arrive in week four.

Hiring requires the same reasoning over recruitment, notice periods, onboarding and the period for which the role will remain an obligation. Equipment and premises require a longer view again. Use enough detail to make the decision, but avoid forecasting every small item with equal intensity. A scarce specialist or low-cost component that can stop an entire delivery deserves closer attention than its share of revenue would suggest.

3. Build evidence that reflects demand, not just completed sales

Understand what the historical record leaves out

Sales history records what the business managed to sell under the conditions that existed. It may not reveal requests rejected because stock, appointments or capacity were unavailable. Research by Christopher Conlon and Julie Holland Mortimer demonstrates that ignoring product availability can bias demand estimates and obscure the effect of stockouts. A model trained on constrained sales can therefore mistake a supply problem for weak customer demand.

Flag stockouts, rejected work, cancellations, substitutions and deferred requests. Keep observed transactions separate from estimates of unmet demand, because a waiting list does not guarantee a later purchase. Record requested dates as well as fulfilled dates, and avoid counting a customer twice when they switch products. This creates an evidence trail that can explain both the sale and the opportunity the business could not serve.

Protect that record from a second problem: double counting. When an opportunity becomes a signed order, move it from uncertain pipeline into the appropriate backlog or delivery schedule. Do not leave it contributing to both. Likewise, a historical run-rate forecast already contains some normal flow of new business; adding the entire pipeline on top without reconciling the two can manufacture growth that does not exist.

Separate real signals from convenient assumptions

For B2B work, distinguish three questions: will the customer buy, when will delivery begin, and how much work will fall inside the planning window? An opportunity’s probability of eventually closing does not answer all three. Estimate conversion and timing from comparable opportunities, checking customer type, deal stage, size and age where the sample allows. A probability attached to a CRM stage should not be mistaken for evidence of calibration.

Treat enquiries, quotations and customer activity as candidate signals rather than automatic demand. Test whether they improve forecasts early enough to influence the decision. A signal that precedes invoices by a week may still be too late for a six-week purchasing commitment. When modelling external drivers, distinguish values genuinely known in advance from values that would themselves need forecasting; using later information in a historical test creates an unfair advantage.

Retain the context around unusual periods: promotions, price changes, customer shutdowns, delayed orders and product replacements. For Gulf businesses, test the effects of Ramadan, Eid, local working calendars and relevant events against the actual customer base rather than applying one regional adjustment. Calendar variables and temporary or permanent interventions are established forecasting tools, but the size and direction of their effects require evidence.

A sudden change also deserves investigation before the model is retrained. Missing records, lost availability, a customer departure and genuine market growth can produce similar-looking movements for different reasons. Annotate what is known, keep uncertainty visible and use a scenario when the evidence cannot yet distinguish a temporary disruption from a lasting change. More historical data is not helpful if the business quietly assumes that every past period still describes the present.

4. Use the simplest method that survives a fair test

Match the approach to the way customers buy

For frequent, relatively stable demand, begin with a recent average or last-period benchmark. For a seasonal series, compare with the equivalent previous season. These simple methods establish what a more complex approach must improve upon. Hyndman and Athanasopoulos use such benchmarks throughout their forecasting guidance because sophistication by itself does not demonstrate predictive value.

When the level changes, exponential smoothing can place greater weight on recent observations, with extensions for trend and seasonality. Where demand has identifiable drivers, test a model that incorporates them. Choose between approaches using forecasts made for unseen periods at the relevant horizon, rather than selecting whichever curve most closely traces the historical data.

Recurring services benefit from separating existing customers, renewals, expected departures, expansions and new customers. Project businesses need opportunity-level timing and workloads. Products ordered in occasional large batches need attention to both the intervals between orders and their size; a smooth weekly average can hide the event that creates the real requirement. Specialist intermittent-demand methods exist, but their assumptions and uncertainty estimates need checking.

For a new product or market, combine carefully chosen analogies, customer evidence and small-scale tests. Make differences from the comparison case explicit, and distinguish expressions of interest from purchasing behaviour. Forecasting research treats new-product estimates as particularly dependent on judgment and warns that stated intentions do not reliably translate into purchases. Start with transparent assumptions and update them as behavioural evidence arrives.

Express uncertainty without inventing precision

A point estimate gives one summary of demand; a prediction interval describes a range of outcomes under the model’s assumptions. An 80% interval is not a guarantee, and a label such as “conservative” does not create a probability. Where the history cannot support a credible distribution, clearly described scenarios are more honest than attaching exact percentages to intuition.

Keep the views of the business consistent. Product forecasts should reconcile with category and company totals, and the workload plan should reflect the same customer assumptions. Forecast reconciliation addresses these aggregation relationships, but agreement in the total does not prove accuracy at the level where stock or specialist hours are allocated. That distinction becomes important in the examples that follow.

5. Inventory: why the right stock level can differ from the forecast

Start with the cost of being wrong in each direction

The following distributor example is hypothetical. All prices, probabilities and costs are teaching assumptions, not market benchmarks. The company must prepare stock for one six-week campaign, with all supply available before it begins and no replenishment during it. Management targets 1,560 units, while the model assumes demand is approximately normally distributed around 1,200 units, with a standard deviation of 180 units.

Under that assumption, the central 80% of demand outcomes falls between approximately 969 and 1,431 units. The range shows uncertainty, not an instruction to order its upper endpoint. The normal distribution is used to make this example transparent; it should not be imposed on intermittent, highly skewed or otherwise unsuitable demand simply because its calculations are convenient.

Each unit sells for AED 200 and costs AED 120 landed. An unsold unit has a net recovery value of AED 100 after the campaign. A shortage therefore loses AED 80 of contribution, whereas a leftover unit loses AED 20 against its purchase cost. Assume no other incremental costs, discounts, penalties or constraints. This is the classical single-period inventory problem, commonly called the newsvendor problem, which balances shortage and surplus costs.

Consider one additional unit. It earns AED 80 if demand reaches it and loses AED 20 if it remains unsold. Under these assumptions, stocking that unit is worthwhile while its chance of selling exceeds 20%. The corresponding balance is reached at the 80th percentile of demand: 80 ÷ (80 + 20) = 80%. The forecast remains 1,200; the economics justify a higher stock position.

Translate the stocking position into an order

The 80th-percentile demand level is approximately 1,351.5 units. With purchases in packs of ten, 1,350 units is the cost-minimising feasible level under this model. It gives approximately an 80% probability of meeting the campaign’s entire demand. That is not the same as fulfilling 80% of units or customer orders; the probability of any shortage and the proportion of demand served are different measures.

The company already has 600 usable units, with 400 confirmed to arrive before the campaign. Another 100 are reserved for older orders outside this forecast. That leaves 900 units available for the campaign. Reaching 1,350 therefore means buying 450 more units, committing AED 54,000. Existing inventory, new purchasing and the demand forecast now each have a distinct role in the decision.

Planning basis	Total campaign supply	Additional purchase	New cash commitment
Forecast mean	1,200 units	300 units	AED 36,000
Cost-balanced stock level	1,350 units	450 units	AED 54,000
Sales target	1,560 units	660 units	AED 79,200

Buying to the target commits AED 25,200 more than the cost-balanced alternative. That is a difference in cash committed, not an automatic profit increase. Buying only to the mean commits less but accepts more shortage exposure. The decision requires both a funding view and an economic view; neither can be inferred simply from which order is smaller.

Figure 1. Calculated expectations under the hypothetical model, not observed losses or guaranteed savings. Purchase outlays are shown separately in the table.

Expected mismatch cost is approximately AED 7,181 at 1,200 units, AED 5,039 at 1,350 units and AED 7,353 at 1,560 units. These figures combine lost contribution from shortages and loss on surplus stock. They are not additional supplier invoices or predictions of one campaign’s actual result. They explain why the middle choice has the best expected outcome under the stated conditions.

Challenge the assumption that could change the answer

Suppose the net recovery value falls from AED 100 to AED 40, while customer demand remains unchanged. Surplus now costs AED 80 per unit, equal to the shortage cost. The cost-balanced position moves to the 50th percentile: 1,200 units. Management should buy less even though its demand forecast has not fallen. This is why an accurate forecast cannot substitute for a sound decision model.

Other conditions can change the answer too: cash limits, minimum orders, shelf life, promised service levels or the consequences of disappointing a strategic customer. Add those explicitly rather than forcing them into the demand estimate. In an ongoing replenishment system, also account for dated receipts, lead-time uncertainty and repeated ordering; the single-campaign calculation is not a universal safety-stock formula.

6. Staffing: why an average workload can hide a delivery crisis

Calculate the work that could actually arrive

Consider a hypothetical service business with eight specialists, each providing 30 client-delivery hours per week after allowing for other responsibilities. Over six weeks, capacity is 1,440 hours. Existing work consumes 840, leaving 600 hours for new projects. For this calculation, assume staff are interchangeable and work can be spread evenly across the period; these simplifications will need relaxing in a real staffing plan.

Three opportunities could use that capacity. Their probabilities refer to the stated workload actually falling inside the six-week window, not just to eventual contract wins. They are illustrative assumptions, not recommended CRM percentages. The weighted calculation is useful for estimating the mean workload, but the company will deliver whole project outcomes rather than the fractions used in that calculation.

Opportunity	Hours within the window	Probability in this example	Expected hours
Project A	480	50%	240
Project B	360	50%	180
Project C	240	75%	180
Total	Up to 1,080	Not additive	600

Expected demand exactly matches available capacity. Nevertheless, assuming independent project outcomes, demand exceeds 600 hours in three cases: A and C without B, A and B without C, or all three together. Their probabilities sum to 43.75%. The company has an apparently balanced average and a substantial chance of overload at the same time.

Figure 2. Enumerated outcomes under the stated independence assumption. Hatched bars exceed available capacity; the worst case requires 480 additional hours.

One extra specialist would add only 180 hours across six weeks, even if available immediately. That would not cover the largest shortfall. More importantly, a temporary concentration of projects does not by itself justify an ongoing salary. Compare permanent recruitment against the expected duration of the workload, while investigating staged starts, qualified partners or other ways to meet the immediate requirement.

Put a value on flexibility before purchasing it

The three overload cases require an additional 120, 240 and 480 hours respectively. Weighting them by their probabilities gives 127.5 expected overflow hours. This is the average amount of shortage across possible outcomes, not a recommended staffing quantity. Like the 600-hour mean, it describes the distribution; it does not describe a schedule that will necessarily occur.

Now suppose a qualified partner guarantees up to 480 hours of suitable overflow capacity for a reservation fee of AED 3,000, plus AED 120 per hour actually used. Assume overflow can be assigned without delay or indivisibility problems, and each otherwise unserved hour would forfeit AED 300 of contribution after all non-partner variable costs. These additional assumptions allow the company to compare two complete responses.

Without the partner, expected foregone contribution is 127.5 × AED 300 = AED 38,250. With the reservation, expected extra cost is AED 3,000 + 127.5 × AED 120 = AED 18,300. The arrangement improves expected contribution by AED 19,950 under this model. The reservation is justified by the value of the flexibility it provides, not by a claim that all three projects will arrive.

The condition is revealing: each outsourced overflow hour preserves AED 180 after the partner’s hourly fee. The AED 3,000 reservation therefore needs more than about 16.7 expected overflow hours to pay for itself under these assumptions. The forecast can change substantially before the reservation stops being worthwhile. That threshold is more useful for the decision than demanding a single perfectly accurate workload figure.

Before acting, test the limits. Projects may share a funding source and therefore move together, the partner may lack a required skill, or work may be concentrated in one week. Independence is not required to add expected hours, but it is required for the particular combination probabilities used here. Recalculate the distribution and costs when those assumptions change, and preserve service quality and contractual feasibility as constraints rather than treating all hours as interchangeable.

7. Connect the decision to cash and the rest of the business

A commercially attractive plan still needs funding at the right time. Convert the operational plan into a dated cash schedule covering deposits, supplier balances, payroll, expected collections and other obligations. Apply the actual payment terms and plausible delays. Expected revenue is not cash available today, and expected cost is not necessarily the largest amount the business may have to pay.

The distributor’s additional order requires AED 54,000 if paid upfront, before allowing for obligations attached to the 400 units already inbound. Those receipts belong in the supply calculation, but any unpaid supplier balance also belongs in the cash forecast. The company should not count incoming stock as economically free because its purchase order was placed in an earlier planning cycle.

The service reservation illustrates the same distinction. Its expected cost is AED 18,300, but the maximum payment in this model is AED 60,600 if all 480 partner hours are needed. Management must check whether customer deposits and collection timing can fund that exposure. An arrangement can improve expected profitability and still be infeasible without working capital or revised commercial terms.

Where funding or capacity is constrained, make the trade-off explicit. Compare the contribution preserved by competing commitments, service obligations, strategic relationships and the consequences of failure. Do not allocate every scarce resource in proportion to revenue or choose the cheapest individual response in isolation. The decision must remain workable for the company as a whole, including outcomes that are less favourable than the average.

8. Decide what to commit, what to reserve and what to reconsider

Turn scenarios into actions with deadlines

Separate a prediction interval from a scenario. An interval describes uncertainty under a model; a scenario changes the conditions, such as a supplier arriving late or a major customer postponing its start. A downside scenario does not become a 10th-percentile forecast merely because it is labelled “low”. Its purpose is to expose consequences and prepare responses, including cases too poorly understood to assign reliable probabilities.

For each material scenario, identify what remains committed, what can change and the latest date for changing it. An early hiring discussion, a provisional production slot and a signed employment contract have different reversibility. Match the strength of evidence to the cost of reversing the decision. Preparing an option may be sensible well before the evidence supports exercising it.

A practical trigger includes an observation, an action, an owner and a deadline. For the service company, newly confirmed start dates may activate partner hours; a deteriorating six-month workload outlook may pause permanent recruitment. For the distributor, a changed supplier arrival date may require an alternative shipment. “Review next month” is insufficient if the useful response window closes next week.

Buy better information only when it can improve the decision

Imagine a customer will confirm on Friday while a normal-price supplier window closes on Thursday. Waiting provides information but may remove a valuable option. Investigate whether a deposit, reservation or staged order preserves the ability to respond. The commercial question is not whether more information would be interesting, but whether it could change an affordable action before the deadline.

The distributor example also provides a useful bound. With the 900 available units already fixed and every other assumption unchanged, even perfect knowledge of campaign demand could eliminate no more than the approximately AED 5,039 of expected mismatch cost associated with the selected policy. The real value would be lower because some excess existing stock cannot be undone, and ordinary forecasts are not perfect. Paying more purely to avoid those same campaign mismatch costs would not be justified within this model.

This does not make better forecasting unimportant. Information can affect several campaigns, pricing, service and other decisions beyond this narrow example. It does mean that its value should be connected to the choices it improves. Sometimes a more responsive supplier or flexible contract changes the economics more than another small improvement in predictive accuracy.

9. Measure forecast quality and decision quality separately

Evaluate the forecast that existed before the commitment

Save dated forecast snapshots. An outlook updated two days before delivery should not be used to defend an order placed six weeks earlier. Test methods by moving through history in time order, using only information available at each forecast date and scoring the horizon that matters. This rolling-origin approach is designed to prevent future information from leaking into the evaluation.

Measure the size of errors and their direction separately. Average absolute error answers how far the forecast was from the outcome in the relevant units. Average signed error reveals persistent over- or under-forecasting, but positive and negative errors can cancel. Percentage measures also need care where demand is zero or very small. A single accuracy percentage cannot carry all these meanings.

The model’s objective matters as well. A method evaluated for mean demand and a policy stocking at a higher percentile are doing different jobs. Do not label the extra stock in the distributor example “forecast bias”: it is a deliberate response to asymmetric costs. Evaluate the demand estimate against demand and the stocking policy against its costs and service outcomes, without pretending that they should always produce the same number.

Inspect the mix, not just the total

Suppose a company forecasts 600 units each of two non-substitutable products. Actual demand is 900 for A and 300 for B. The total is exactly right at 1,200 units, yet the allocation leaves 300 units too few against one product and 300 too many against the other. Aggregate agreement offers little protection when the wrong goods, locations or skills receive the resources.

Figure 3. A hypothetical mix error. Total absolute error is 600 units despite zero error in the company total.

Here, total absolute error divided by total actual demand is 600 ÷ 1,200 = 50%, a descriptive measure commonly called weighted absolute percentage error, or WAPE. It is useful for illustrating this mismatch, not a universal accuracy score. Its denominator can distort comparisons across changing demand levels, and it is undefined if total actual demand is zero; forecasting researcher Rob Hyndman explains these limitations in detail.

Assess uncertainty too. Across enough comparable cases, an 80% prediction interval should contain approximately 80% of outcomes if well calibrated. Inspect its width and the severity of misses rather than celebrating a range so broad that almost anything fits. Distributional scoring methods explicitly consider these issues; one successful or unsuccessful campaign cannot establish calibration.

Review why the outcome differed

Classify a miss before deciding what to repair. Did customer demand differ, did delivery timing move, did supply fail, did productivity change, or did a planned commercial action never happen? A perfectly forecast order can still be delivered late because a required skill was unavailable. That is a delivery-planning failure, not evidence that the demand model needs replacing.

Likewise, evaluate policies over repeated comparable decisions, not just the latest outcome. A cost-balanced order may leave stock after one weak campaign and still have been the sensible choice when made. Conversely, an unjustified purchase can be rescued by an unusually strong result. Preserve the information and rationale available at the decision date so that the business learns from both good and bad outcomes without rewriting history.

10. Make human judgment and AI earn their place

Preserve the baseline when people adjust it

Commercial teams can add information missing from a model: a confirmed customer closure, an agreed product replacement or an unusually large order. Require each material adjustment to record the evidence, the affected period and an accountable owner. Keep the original and adjusted versions. Structured judgment adds a testable contribution; an undocumented increase made to reassure management does not.

In research published in the 2025 volume of the International Journal of Forecasting, Robert Fildes, Paul Goodwin and Shari De Baets analysed approximately 147,000 forecasts from six studies. Adjustments improved accuracy and bias for only just over half of stock-keeping units, with variation between datasets; upward adjustments were more likely to worsen performance. This is a reason to evaluate intervention, not to exclude useful commercial knowledge.

Compare the simple benchmark, selected model and final managerial adjustment at the same horizon on the same observations. This is the practical purpose of forecast-value-added analysis: discovering which stages improve the result. A meeting that repeatedly worsens forecasts should be redesigned, while a valuable source of customer information should become a documented input rather than remain a personal advantage.

Separate estimating demand from permission to act

Apply the same standard to AI-enabled forecasting. Ask whether it improves performance on unseen periods, represents uncertainty usefully and produces better decisions after real costs and constraints are considered. Require traceable inputs and reproducible outputs. A persuasive explanation of a forecast is not evidence that its probabilities are reliable, nor does a marginal accuracy improvement establish a return on the software investment.

Keep three activities distinct: estimating demand, recommending a response and authorising a commitment. Start with reviewed recommendations, defined spending limits and explicit exception handling. Consider automation only where the decision rules, monitoring and rollback arrangements are appropriate to the exposure. The business should always be able to identify what changed, why an action followed and who remains accountable.

11. Give the business one decision brief, not another dashboard

A forecasting process becomes useful when people can see what it changes. The following brief brings the essential questions together without requiring every reader to understand the statistical machinery. Use it for one consequential decision, attach the supporting calculations and retain a dated version. It is a management record rather than a demand for every department to agree on one optimistic number.

Decision field	What the brief should establish
Decision and deadline	What must be committed, by whom, and the last useful date to act.
Target and demand outlook	The ambition, the evidence-based estimate and the assumptions separating them.
Uncertainty	A supported probability range or clearly labelled scenarios, including shared risks.
Existing position	Stock, dated receipts, backlog, usable capacity and obligations already accepted.
Response and trade-offs	What to commit, reserve or defer; shortage, surplus and service consequences.
Cash exposure	Expected payments, peak funding requirement and the timing of collections.
Trigger and owner	The evidence that would change the action, the response deadline and accountable person.
Review record	The forecast snapshot, actual outcome, decision costs and explanation of the difference.

Set a working rhythm around the speed of the decisions. Review material exceptions frequently enough to respond, and periodically reconcile demand, capacity and cash across the business. Sales should explain commercial evidence, operations should establish delivery feasibility, finance should test funding and exposure, and the decision owner should approve the trade-off. Accountability is clearer when estimating demand is not treated as negotiating a departmental budget.

Protect honest forecasts from the incentive to provide reassurance. A lower outlook should prompt investigation and commercial action, not automatic punishment for pessimism. Equally, avoid allowing people to lower expectations merely to make later performance look better. Evaluate the quality of the evidence and the execution of the response separately, keeping the growth target visible throughout.

Target’s June 2022 inventory announcement shows why category-level decisions matter. The company planned markdowns, excess-stock removal and order cancellations, alongside rapid revisions to forecasts and plans by category. It expected strength in food, household essentials and beauty while taking a more conservative view of home products. The lesson is not that one model could have prevented every problem, but that changing evidence required changes to specific commitments.

12. Start with one decision that matters

In the first week, choose a product family, project team or recurring service where better planning could change an important commitment. Define the units, horizon, decision deadline and owner. Assemble the relevant history, current obligations and availability constraints. Sketch the demand pattern and record the assumptions that could materially change the decision; begin with a usable record rather than waiting for every system to be perfect.

In the second week, compare a simple baseline with a suitable alternative using historical decision dates. Build a range where the evidence supports it, or explicit scenarios where it does not. During the third week, translate those outcomes into workload, inventory and cash, then compare feasible responses. Include an option that changes flexibility, not just options that increase or decrease the total commitment.

In the fourth week, hold the first decision review, preserve the forecast and record the trigger for revisiting it. This establishes a working process, not proof of forecasting excellence. Improvement requires repeated outcomes at the relevant horizon. Expand the system when the first area produces useful decisions, and keep reducing effort spent on information that never changes an action.

Forecast for the decision, not for the presentation

The distributor did not need to choose between believing in its target and trusting its forecast. It needed to price the consequences of shortage and surplus, then determine the stock commitment it could fund. The service company did not need to hire for every possible project. It needed to recognise the workload distribution, value flexible capacity and secure that option before the response window closed.

Both examples reveal the same discipline: estimate what customers may require, understand what the company has already promised, compare feasible responses and preserve the ability to change course. Forecasts inform that process; they do not replace management judgment. The standard is not whether a number looks precise, but whether the resulting decision is defensible under the evidence available at the time.

Ambition sets the direction. Evidence shapes the commitment. A business can pursue substantial growth while remaining explicit about what must become true, how much it is prepared to risk and what it will do if circumstances change. That is the point at which demand forecasting stops being a reporting exercise and becomes part of how a company grows.