Skip to main content

1. Demand Planning Excellence

Teaching rewrite of demand-planning best practices, organised around a five-step framework. Original wording and examples; numbers are illustrative.

Introduction​

A supply chain makes thousands of decisions every day: what to buy, how much to produce, where to ship, how much to keep in stock. Almost all of them depend on one unknown — how much customers will want in the future. The better that estimate, the better the decisions: fewer shortages, less obsolete stock, more sales, lower cost.

A demand planner is like the lookout on a ship with a spyglass. The lookout does not steer; the job is to report what is on the horizon — accurately and without bias — so the crew on deck can decide what to do.

This lesson covers:

  • What demand planning excellence means: efficacy and efficiency
  • Why accuracy alone is not the goal
  • The five-step framework: Objective → Data → Metrics → Baseline model → Review process
  • How the framework changes for short-, mid-, and long-term forecasts

1.1 Efficacy and efficiency​

PillarMeaningQuestion to ask
EfficacyForecasts actually help decision-makers“Does this forecast change a decision for the better?”
EfficiencyThe team spends as little time as possible producing them“Could a model or a rule do this work instead of a person?”

The aim is to work smarter, not harder: automate what can be automated, and point human effort where it adds the most value.

Accuracy is necessary, not sufficient

A forecast can be accurate and still useless — for example, if it is made at the wrong level of detail (country per month when the warehouse needs SKU per week) or if it forecasts the wrong thing (shipments instead of demand). Usefulness is the goal; accuracy is one ingredient.

Worked example: accurate but useless​

A planner forecasts total monthly demand for a product family at 10,000 units and the actual is 10,100 — 99% accurate. But replenishment decisions are made per SKU per warehouse per week. Inside the family, one SKU sold 40% more than its share, another 40% less. The family forecast was excellent; the stock was in the wrong places.

1.2 The five-step framework​

flowchart LR
  A[1. Objective] --> B[2. Data]
  B --> C[3. Metrics]
  C --> D[4. Baseline model]
  D --> E[5. Review process]
StepCore questionKey ideas
1. ObjectiveWhat do you need to forecast, and why?Start from the decision supported; choose aggregation level (product, location, time bucket) and horizon
2. DataWhat data feeds the forecast?Collect unconstrained demand, not constrained sales; gather demand drivers (promotions, prices, launches)
3. MetricsHow do you judge quality?Measure accuracy and bias; compare with benchmarks; weight metrics for broad portfolios
4. Baseline modelHow do you generate a forecast automatically?Time-series, causal, or machine-learning models create a baseline with minimal manual work
5. Review processWho adjusts the baseline, and does it help?Humans add information the model lacks; track it with forecast value added (FVA)

Step 1 — Objective​

Always start with why:

  • Why do we forecast demand?
  • Which decisions does this forecast support?
  • (And therefore) Who will use it?

The decision sets the granularity (SKU × store × week? brand × country × month?) and the horizon (next week? next year?). Different teams may need different forecasts, made with different models and processes.

Step 2 — Data​

The most important data is unconstrained demand — what customers asked for, not what you managed to ship (the next two lessons). Next come demand drivers: promotions, prices, product launches, events, weather.

Step 3 — Metrics​

Track both accuracy (how big are the errors?) and bias (do errors lean consistently one way?). A benchmark — e.g. a simple naive or moving-average forecast — tells you whether your process is actually any good.

Step 4 — Baseline model​

Planners should not type every number by hand. A model generates a baseline for all items, using as many relevant insights as possible. Humans then review only where they can add value.

Step 5 — Review process​

Sales, marketing, planners, and customers may know things the model does not (a lost tender, a competitor’s launch). Their edits should be measured: the forecast value added (FVA) framework compares the accuracy of each step (baseline → planner edit → sales edit → final) to see who improves the forecast and who makes it worse. Without it, forecast reviews turn into influence wars.

Checkpoint: excellence and the framework​

Quiz

Checkpoint — sections 1.1 and 1.2

1 / 4

Which pair defines demand planning excellence here?

Pick an answer to see if you are right.

1.3 Three horizons, three set-ups​

Short termMid term (S&OP)Long term (budget)
DecisionWhat to ship to each store this weekWhat to produce over the next monthsAnnual budget, capacity, launches
GranularitySKU × store × weekSKU (global) × monthBrand/segment × revenue × month/quarter
Update cadenceWeeklyMonthlyYearly (with re-forecasts)
Who reviewsFew planners, only high-value itemsMany stakeholders: sales, marketing, finance, supply, customersManagement, finance, marketing
ModelMachine learning using granular drivers (promotions, prices, shortages, weather)Statistical baseline + reviewed inputs; value-weighted metrics; FVACausal/driver model to build scenarios (price, marketing, new products)
Main riskToo many items to review by handFunctional bias in overridesIntentional and cognitive bias, wishful thinking

Worked example: where to spend review time​

A planner has 5,000 SKU-store combinations to review each week and 10 hours. At 2 minutes each, reviewing everything would take 5,000 × 2 = 10,000 minutes ≈ 167 hours. Instead:

  • Let the model handle the long tail.
  • Review the top 300 items by value or uncertainty: 300 × 2 = 600 minutes = 10 hours.

The team’s time goes where a human is most likely to add value.

Checkpoint: horizons​

Quiz

Checkpoint — section 1.3

1 / 4

A retailer must decide weekly shipments to 400 stores. Which set-up fits?

Pick an answer to see if you are right.

Rank: setting up a forecasting process​

Quiz

Sequence a forecasting project

1 / 2

Rank the steps to set up a new forecasting process.

OrderDrag the handle to rearrange — top should be first.

  1. 1Build a baseline model and compare it with the benchmark.
  2. 2Set granularity and horizon to match those decisions.
  3. 3Collect unconstrained demand history and demand-driver data.
  4. 4Define who may override the baseline and track their FVA.
  5. 5List the decisions the forecast must support and who makes them.
  6. 6Choose accuracy and bias metrics, plus a simple benchmark.

Drag items into the correct order (or use ↑↓), then check.

Check: demand planning excellence​

Quiz

Demand planning excellence

1 / 5

The role of the demand planner is best compared to…

Pick an answer to see if you are right.

Takeaways​

Remember
  • Forecasts exist to support decisions — usefulness beats raw accuracy
  • Excellence = efficacy + efficiency
  • Framework: Objective → Data → Metrics → Baseline model → Review process
  • Start with why and which decision; that sets granularity and horizon
  • Use a model baseline; let humans enrich where they add value, measured with FVA
  • Short, mid, and long-term forecasts need different granularity, models, and reviewers

Further Reading​