Demand Planning
Build a usable demand plan from history, patterns, business signals and review. Learn baseline forecasting, error and bias checks, promotions, new products and consensus decisions.
Each block is planned for 60 minutes: about 35 minutes to learn, 15 minutes for the worked case and 10 minutes for notes or review. Take breaks between blocks. You can complete the lesson over several days.
The idea in simple words
Demand planning asks: what might customers request, where, and when? The answer is an estimate with uncertainty, not a promise. This lesson shows how to build a sensible starting forecast, add evidence carefully, measure what was useful, and explain what remains uncertain. We use a fictional reusable-bottle business throughout.
Start here: Start with the business question before selecting a model. A forecast is useful only when its product, location, time bucket and decision horizon match the choice a planner must make.
1. Define the decision and forecast grain
Write down the decision, the owner, the horizon and the unit before looking at a chart. “How many units should the north warehouse replenish next week?” is usable; “forecast the business” is not. Choose a grain such as SKU–location–week when that is how stock is ordered. A finer grain gives detail but creates more noisy series; a coarser grain can hide local shortages. Keep the time bucket consistent with the lead time and decision cycle.
Worked example
Northstar plans one SKU for one regional warehouse, one week at a time, across a 12-week horizon. A daily forecast would add noise because replenishment is reviewed weekly; a company-wide monthly total would hide which warehouse needs stock.
Try it · 10 minutes
Write one forecast question for a product you know. Name its SKU, location, time bucket, horizon, decision owner and the action the answer will support.
Watch for: Do not select a method until the grain and business decision are clear.
2. Prepare history and understand what was observed
Sales history is not always the same as unconstrained customer demand. A stockout can make recorded sales low even when customers wanted more. Returns, late postings, duplicate transactions, unit changes and one-off bulk orders can also distort a series. Keep the original data, flag corrections, and record the reason. Mark zero sales as a real zero only when the item was available and the period was open. Separate lost sales, cancelled orders and deferred orders where the process can identify them.
Worked example
The bottle sold 30 units in a week but was out of stock for three days. The 30 is observed sales, not necessarily the full demand. If inventory records show 12 requests could not be filled, a simple demand estimate for that period might be 42; the team should label that estimate and its source.
Try it · 10 minutes
Take ten periods of sample data. Mark one possible stockout, one unusual transaction and one missing value. Decide which should be corrected, flagged or left untouched—and document why.
Watch for: Never silently replace awkward history. Keep a traceable distinction between source facts and estimated adjustments.
3. Read the pattern before choosing a method
A series can have a level, trend, seasonality, intermittent demand or a mixture. Level is its typical height; trend is a sustained direction; seasonality repeats on a calendar; intermittent demand contains many quiet periods and occasional orders. Plot the series and compare it with a relevant calendar before fitting a model. Check whether a pattern repeats across more than one cycle. A single spike is not automatically a trend or seasonality. Compare at the same aggregation at which the decision is made.
Worked example
Monthly demand for the bottle is usually around 40, rises each spring, and grew slowly from 36 to 42 over the last year. The spring increase repeats and may be seasonal; the gradual rise may be trend. An isolated 95-unit order from one event should be explained before a model treats it as normal demand.
Try it · 10 minutes
Sketch a 12-period series with a trend and a repeating seasonal peak. Label which evidence supports each pattern and what extra data would increase your confidence.
Watch for: A chart helps the planner ask questions; it does not prove the cause of a pattern.
4. Build and compare simple baselines
Start with a transparent baseline. A naïve forecast repeats the last actual. A moving average uses the mean of recent periods; a weighted average gives selected periods more influence. A seasonal-naïve method repeats the comparable prior season. Exponential smoothing updates a level by giving more weight to recent observations, while trend and seasonal methods add components. Compare methods with a time-ordered holdout: fit on earlier periods, predict later periods, and never let future data leak into the past. Keep the baseline as a benchmark even when a more complex model is available.
Worked example
Recent weekly sales were 40, 44 and 46. A three-week moving average is (40+44+46)/3 = 43. If the same week last year was 51 and demand clearly peaks each spring, a seasonal baseline of 51 may be more relevant. The better choice depends on the pattern and out-of-sample performance, not on which formula sounds advanced.
Try it · 10 minutes
Calculate a three-period moving average and a naïve forecast using the last value. Compare both to a held-out actual of 48. Record absolute errors: 5 and 2 units.
Watch for: Use the same forecast horizon and evaluation periods when comparing methods.
5. Add events, price and collaborative signals carefully
A baseline describes recurring demand under ordinary conditions. A documented promotion, price change, launch, holiday or customer event may justify a separate adjustment. Estimate uplift from comparable events where possible and note the assumptions. Check for cannibalization: a promoted variant may take demand from another variant instead of adding fully new demand. A sales input is useful when it names the customer, timing, quantity and evidence. Keep the baseline and overrides visible so actual outcomes can teach the next plan.
Worked example
The 43-unit baseline has a confirmed event expected to add 8 incremental bottles, producing a working plan of 51. If the event discount also shifts 5 buyers from another bottle color, the company-wide increase is smaller than 8. Record whether the 8 is an uplift or a gross event forecast.
Try it · 10 minutes
Create a small event record with owner, dates, expected uplift, evidence, affected products and a post-event review date.
Watch for: Do not count the same event twice in both model inputs and a manual override.
6. Measure error, bias and usefulness
Forecast error is actual minus forecast; define the sign convention and keep it consistent. MAE gives average absolute units of miss. RMSE gives extra weight to larger misses. MAPE expresses percentage error but is undefined or unstable when actual demand is zero or very small. WAPE divides total absolute error by total actual volume and is often easier to aggregate, but high-volume items dominate it. Bias reveals a repeated tendency to forecast too high or too low. Measure at the horizon where the decision is made and segment by product, location and lifecycle.
Worked example
Across actuals 50, 60 and 40 with forecasts 55, 55 and 50, absolute errors are 5, 5 and 10; MAE is 20/3 = 6.7 units. WAPE is 20/150 = 13.3%. Forecasts total 160 versus actuals 150, so this sample is 10 units high overall.
Try it · 10 minutes
Compare two forecasts over the same periods using MAE and bias. Explain which one you would prefer for a capacity decision and why the score alone may not decide it.
Watch for: Do not optimize one metric without checking service, inventory, cost and the asymmetry of over- and under-forecasting.
7. Handle new, sparse and changing demand
A new item has little or no own history. Use a documented analogue, product attributes, launch plan and early order signals, then revise quickly as evidence arrives. Intermittent items need methods and service policies that recognize long gaps and occasional demand; a smooth average can create a misleading pattern. When a market shift changes the data-generating process, old history may be less relevant. Keep a change log, use scenarios and shorten review intervals until the pattern stabilizes. Reconcile forecasts across product families carefully so totals do not contradict their detail.
Worked example
A new bottle color has no sales history. An older color suggests 40 weekly units, while the launch plan expects a 20% uplift. The initial 48 is a hypothesis, not measured demand. Compare actual orders after launch and review the analogue and uplift each week.
Try it · 10 minutes
Choose a new item or sparse-demand case. List the analogue, assumptions, early signals and the date when the first estimate will be revisited.
Watch for: A model cannot create evidence that the business has not collected.
8. Run a reviewable demand-planning cycle
A practical cycle has a data cutoff, baseline generation, exception review, event input, consensus discussion, approval and a record of changes. Focus review time on exceptions with value: large error risk, a new event, a capacity implication or a material change. Assign one owner to each override and give it an expiry or review date. Compare forecast value added: did the manual change improve results against the baseline over an appropriate sample? Share a range or scenario when uncertainty matters more than a single point.
Worked example
Northstar presents 43 baseline units, an 8-unit event assumption, and a 51-unit working plan. The sales owner confirms the event; the planner checks comparable events; operations sees the resulting volume; after the event, the team records actuals and whether the adjustment helped.
Try it · 10 minutes
Draft a one-page review: baseline, material changes, evidence, uncertainty, owner, downstream impact and next review date.
Watch for: A consensus number is useful only when its assumptions and decision owner are visible.
Further reading and scope
These sources provide background for the planning concepts. The explanations, diagrams and Northstar examples on this page are written for this site. Official software manuals describe the named product and configuration; they do not define a universal process. No source text or figures are reproduced here.
- Hyndman, R. J., and Athanasopoulos, G. (2021). Forecasting: Principles and Practice, 3rd ed., OTexts. Online edition updated 28 October 2026. Open textbook; not peer-reviewed research.
The reference covers time-series patterns, benchmark methods, train/test evaluation, accuracy measures and judgmental forecasting. The lesson’s Northstar quantities and wording are original examples. - Forecast accuracy evaluation and benchmark methods, Forecasting: Principles and Practice, Chapter 5.
Useful for comparing a forecast against held-out periods and understanding common error measures.
Quick knowledge check
Choose an answer before opening the explanation. These questions review the key ideas from all eight blocks.
1. What must be defined before choosing a forecast method?
- The color of the dashboard
- The decision, grain, horizon and time bucket
- The most complex available model
- The annual budget
Reveal answer and explanation
Answer: A. The color of the dashboard A forecast must match the decision and the level at which it will be used.
2. Why may recorded sales understate demand?
- Sales are always lower than orders
- A stockout can prevent some demand from becoming a sale
- Returns increase demand
- Forecasts remove sales
Reveal answer and explanation
Answer: B. A stockout can prevent some demand from becoming a sale When product is unavailable, observed sales can be capped by available stock.
3. What does seasonality mean?
- A one-time large order
- A pattern that repeats at a similar calendar time
- A data correction
- A supplier lead time
Reveal answer and explanation
Answer: B. A pattern that repeats at a similar calendar time Seasonality is recurring timing, such as an annual holiday peak.
4. A three-period moving average uses which observations?
- The next three actuals
- Three recent periods selected by the method
- Only the largest period
- All periods in the future
Reveal answer and explanation
Answer: B. Three recent periods selected by the method It averages recent historical periods to form a baseline.
5. Why keep baseline and event uplift separate?
- To hide the total
- To prevent double-counting and learn whether the adjustment helped
- To make the plan larger
- Because events cannot affect demand
Reveal answer and explanation
Answer: B. To prevent double-counting and learn whether the adjustment helped Separate components can be reviewed and evaluated after the event.
6. When is MAPE difficult to use?
- When demand is always positive and large
- When actual demand is zero or close to zero
- When units are bottles
- When forecasts are stored weekly
Reveal answer and explanation
Answer: B. When actual demand is zero or close to zero Division by zero is undefined and small actuals create unstable percentages.
7. What does a persistent forecast bias describe?
- Random variation around actuals
- A repeated tendency to be high or low
- A product identifier
- A supplier capacity limit
Reveal answer and explanation
Answer: B. A repeated tendency to be high or low Bias is directional over- or under-forecasting across periods.
8. How should a new product forecast be treated?
- As certain because an analogue exists
- As a documented hypothesis to revise as evidence arrives
- As zero forever
- As a copy of a competitor’s private data
Reveal answer and explanation
Answer: B. As a documented hypothesis to revise as evidence arrives An analogue supports an initial estimate but cannot replace the new item’s own evidence.
9. What is a useful reason to review a forecast exception?
- It is the first row in a spreadsheet
- It has material risk, new evidence or downstream impact
- It uses a common SKU
- It has no owner
Reveal answer and explanation
Answer: B. It has material risk, new evidence or downstream impact Exception review should focus effort on changes that may matter to decisions.
10. What does forecast value added test?
- Whether a manual change improves results versus a baseline
- Whether the chart uses color
- Whether inventory is positive
- Whether demand is seasonal
Reveal answer and explanation
Answer: A. Whether a manual change improves results versus a baseline It evaluates whether interventions improve forecast performance compared with a chosen benchmark.
Educational content only. All quantities, products, scenarios and costs are illustrative. Real methods depend on data quality, local policy, system design, contracts and operating constraints.