Selling 45 units when you could deliver only 45 does not prove demand was 45. 200 ad impressions do not prove there will be 200 purchases, either.

An operational forecast supports concrete decisions: how many ingredients to buy, which shifts to open, how many orders to accept, and which promises to communicate. It does not claim to predict the future with absolute precision. IBM describes demand forecasting as using historical information and relevant signals to anticipate needs and support inventory and capacity decisions [R01]. Sales and operations planning combines that estimate with supply constraints [R02]. This article focuses on the operating horizon, not on sizing the total market or proving a business opportunity.

1. Separate the market, interest, orders, and unmet demand

Not every signal measures the same phenomenon:

Signal What we know What we still do not know
Potential market People or organizations who could need the offer How many will buy in a given period
Visit or impression Someone was exposed to the message or channel Whether there is intent and ability to purchase
Qualified inquiry Someone expressed interest compatible with the offer Whether price, timing, and terms will be accepted
Confirmed order A defined commercial commitment exists Whether it will be fulfilled and paid as expected
Fulfilled sale A specified transaction or delivery was completed How much demand was excluded by capacity limits
Inquiry declined for lack of slots Someone requested something we could not accept Whether that request would have converted into a purchase

The operational demand we seek to estimate is expressed in units and time—for example, lunches requested for tomorrow. Observed sales may be capacity-constrained: when inventory runs out or reservations close, units sold no longer represent every potential request. Logging lost orders helps, but it does not turn every inquiry into a guaranteed purchase.

If you still need to estimate how many buyers may exist before operations begin, use the market size and demand estimation guide; that estimate does not replace an order forecast.

2. Define a unit, time period, and decision

Before calculating, write one sentence: “I need to forecast confirmable lunches per 30-minute time slot over the next three days, in the area we actually cover, to plan preparation and delivery.” Do not mix customers, orders, and units: one person can request two lunches, and one order may contain several items.

Use a decision-based horizon: hourly for slots, daily for production, weekly for work schedules and purchasing, monthly for supplier reviews. The lead time needed depends on how long supplies take to arrive or a resource takes to adjust; a daily forecast does not solve a purchasing decision requiring months of notice. IBM recommends choosing the forecast horizon according to the decision it will inform [R03].

Segment only where a meaningful operational distinction exists: delivery zone, meal period, product type, channel with different fulfillment times, or customers under recurring contracts. Overly narrow segments with only two observations each can create the appearance of rigor without useful evidence.

3. Collect data that explain what happened and why

Begin with a record containing the date, slot, units requested, accepted, fulfilled, canceled, declined for lack of capacity, and the reason. Add the current price or promotion, holidays, unusual circumstances, inventory shortages, relevant weather, and coverage changes.

Systems that record only completed sales leave blind spots. A restaurant may sell out at 12:15 and appear stable because it still finishes each day with 45 lunches. An app may show fewer registrations because it was down. A B2B service may receive fifteen requests, close three contracts, and need months of work to deliver them. The underlying units and time horizons differ.

Data quality before sophistication: identify duplicates, missing observations, cancellations, and classification changes. Do not add inquiries and confirmed orders when the latter are a subset of the former. Preserve the source and who made a correction.

4. Forecast with little history without manufacturing certainty

When operations have just begun, do not pretend you have a robust statistical series. Combine verifiable advance orders, documented ranges from experienced operators, interviews, or limited operational tests, clearly separating observed evidence from judgment. A customer saying they may need 10 units per month does not guarantee orders; a signed contract may create obligations unlike an informal conversation.

Construct at least an operational range: “between 25 and 40 lunches, with more uncertainty on Friday.” If the decision is reversible, avoid committing perishables or fixed capacity for the high estimate without enough supporting signals. If the decision requires advance planning, document what information will arrive and what would change the plan.

A simple method may be preferable to a sophisticated model: the naïve forecast uses the previous comparable period; a moving average uses the most recent n periods. IBM describes these as common starting points, with limitations when the underlying pattern shifts [R04].

5. Calculate a moving average and recognize its limits

Suppose we have three comparable operating days with 45, 45, and 42 fulfilled lunches. A three-period moving average is:

Fulfillment forecast = (45 + 45 + 42) / 3 = 44 lunches per day.

This is a reproducible estimate of observed fulfillment, not unconstrained demand, if orders were declined on two of those days. Moving averages smooth fluctuations but can react slowly to a promotion or schedule change. Compare like periods: Mondays with Mondays, peak hour with peak hour, or separate segments when there are enough observations.

Keep the input values and the calculation cutoff date. When adding a new period, remove the oldest one. If you switch from a three-day to a seven-day average, record why; do not retrospectively choose the method that best explains data you have already seen.

6. Incorporate time-slot variation, seasonality, and announced changes

A daily average hides the fact that 30 of 45 orders may be requested for the same window. Record demand by slot and compare it to that slot's capacity. Fridays, holidays, paydays, vacations, relevant weather, or promotions may affect demand, but two things changing together does not establish causation.

Before adjusting the number, identify what changed, when it changed, which segments were affected, the plausible impact, and the supporting evidence. For an upcoming campaign, show a hypothetical scenario of additional inquiries rather than a promise of orders. When the price changes, distinguish those periods from earlier ones until comparable observations accumulate.

For scheduled services and B2B offerings, account for bookings, renewals, and existing contracts. Do not automatically treat an open sales opportunity the same as a signed purchase order; use explicit ranges or probabilities when useful without implying certainty.

7. Worked example: detect demand hidden by slot limits, stockouts, and declines

Hypothetical case. Over five days we record fulfilled lunches and additional inquiries from distinct customers who wanted to buy at the advertised price and time, but were declined solely because there were no slots:

Day Fulfilled lunches Qualified requests without a slot Requests recorded as a minimum signal of interest
Monday 32 0 32
Tuesday 36 0 36
Wednesday 45 8 53
Thursday 45 5 50
Friday 42 0 42

The eight and five additional requests were verified as distinct and are not duplicated in fulfilled orders. Average fulfillment over the last three days is (45 + 45 + 42)/3 = 44. The average recorded interest signal over those days is (53 + 50 + 42)/3 = 48.33. The difference of 4.33 does not prove how many more items would have sold: some people might not have completed payment. But ignoring those requests can also understate opportunities.

If declined inquiries were recorded incompletely, uncertainty increases. Never automatically call 53 the full demand: other people may not even have attempted an order after the site showed “sold out.” Your estimate must state what was observed, not claim knowledge of demand that cannot be measured directly.

8. Create scenarios that lead to different decisions

For the next comparable period, we might propose—only as planning assumptions—low demand of 42, a base case of 48, and high demand of 56 lunches. These are not statistical confidence intervals or validated forecasts: the base case rounds an historical interest signal, while the extremes express decision-relevant uncertainty.

Once the scenarios are defined, compare them with effective capacity and time-slot demand before expanding reservations or shifts.

Hypothetical scenario Estimated requests Operating implication if slots remain capped at 45
Low 42 May be manageable if slots, supplies, and delivery fit
Base 48 Up to 3 slots short in that period if every request competes for the same capacity
High 56 Up to 11 slots short; review schedules, campaign limits, and alternatives

Check materials, people, time windows, costs, and receipts separately. The high case does not authorize hiring permanent staff. Do not add low, base, and high cases together: they are alternatives, not three simultaneous order streams.

9. Measure error and bias without penalizing transparency

Once a period ends, compare the forecast recorded before the result was known to the later observation. Define whether the target is requested, accepted, or fulfilled units; switching targets week to week invalidates comparisons. Absolute error is |actual − forecast|; a signed error can be defined as actual − forecast, so a positive value means underforecasting.

In a hypothetical later period, we forecast 48 requests and record 52 comparable requests. Absolute error is 4, and signed error is +4. A sequence of positive errors across periods might reveal a systematic downward bias; one result does not establish that. Also record capacity restrictions, since judging fulfilled sales is different from judging a forecast of requested units.

Avoid percent error formulas when the actual denominator is zero. With small datasets, a difference of two orders may produce a large percentage without justifying a complete change of direction. Compare the economic consequences of each kind of error: wasting perishables and turning away customers do not necessarily cost the same.

10. Create a review routine and action thresholds

Before each operating day, record forecast, confirmed orders, capacity by slot, possible shortages, planned action, and owner. Afterward, record sales, unmet requests, errors, and relevant conditions. Once a week, review assumptions and decide whether to purchase differently, move schedules, reserve capacity, or retain existing rules.

When the forecast points to replenishment needs, translate it into controlled purchasing and inventory, accounting for supply lead times.

Do not expand capacity merely because the limit was exceeded on two days. Ask whether the additional demand is repeatable, at what price, with what contribution, in which window, and under what constraints. When information is insufficient, a limited trial—such as opening a second slot for a few days—may produce evidence without a large investment. Formal hypothesis testing and continuous improvement are covered elsewhere in the Atlas; the goal here is to make explainable operating decisions.

Template: operational demand forecast sheet

Field What to record Example or note
Unit and horizon Product/service, place, period, and slot Lunches, covered zone, tomorrow 12:00–13:00
Source and date Sales, reservations, declined inquiries, calendar Order system and slot record
Observed quantities Requested, accepted, fulfilled, declined, and why Do not mix categories
Method and assumptions Moving average, last period, or documented judgment Three comparable periods
Low / base / high Three separate values and their meanings 42 / 48 / 56 requests
Critical constraint Slots, supplies, staffing, cash, or transportation 45 slots per period
Review Prior forecast, actual, error, decision, and owner Revisit before the next cycle

The sheet does not need fifteen decimal places. It needs to identify what was counted, what was assumed, what the business could serve, and what changes when new evidence arrives.

Common mistakes and limits

Treating impressions as purchases; adding duplicated orders and inquiries; using sales capped by stockouts as total demand; ignoring hourly distribution; changing estimates without preserving the original forecast; turning scenarios into certainties; extrapolating three days into a year; and buying permanent resources after an isolated spike. This illustrative method does not replace safety, perishability, staffing, or industry-specific regulatory analysis.

What you should have by the end

A forecast by unit and period, observations that separate requests from fulfillment, a reproducible calculation, three scenarios with explicit limits, an error comparison, and a decision about purchasing, slots, or shifts. The next step is not producing as much as possible, but comparing uncertain demand with the capacity the business can actually deliver.

References