Demand forecasting

A forecast is not a prediction contest; it is an input to a decision. This path fixes the decision first, establishes a reproducible baseline, and governs human judgement so that overrides can be measured rather than argued about.

Start a decision

6 steps · 6 cited sources · free during beta

The question this path answers

How should we forecast demand so that plans, staffing and stock follow reality?

The path, step by step

Name the decision the forecast serves

Write down the decision, its horizon, the level of detail required and the cost of being high versus low. A weekly staffing decision and a quarterly purchasing decision need different forecasts.

Why it worksForecast accuracy only means something relative to a use. Competition evidence shows no single model family wins everywhere, so the benchmark has to be the decision, not a leaderboard.

EvidenceMakridakis et al. (2020) · Silver et al. (2016)

CheckpointA one-page forecast brief: decision, horizon, granularity, error cost and current baseline.

Common failureOptimising statistical accuracy on a series that no planner actually uses.
Classify the demand pattern

Split the portfolio into smooth, seasonal, intermittent and lumpy demand, and mark promotions, new products and one-off orders. Forecast each class with a method suited to it.

Why it worksIntermittent demand behaves differently from fast-moving demand; treating every SKU with the same model produces systematic error. Judgement can help, but only when its direction and size are recorded.

EvidenceSyntetos et al. (2005) · Fildes & Goodwin (2007) · Fildes et al. (2009)

CheckpointShare of volume and revenue in each demand class, with the method assigned to each class.

Common failureApplying one model, or one override habit, to the whole portfolio.
Establish a reproducible baseline

Build a seasonal-naive benchmark and one exponential-smoothing or ARIMA challenger. Backtest both on a holdout of at least two full seasonal cycles before touching the live plan.

Why it worksForecasting competitions show that simple, well-tuned statistical benchmarks are hard to beat and that ensembles are useful, but only when evaluated out of sample. A model that only fits history has learned nothing about the next period.

EvidenceMakridakis et al. (2020) · Makridakis et al. (2022)

CheckpointOut-of-sample error and bias for each model, at the horizon the decision actually uses.

Common failureChoosing a model by training fit, or by how sophisticated it sounds.
Reconcile levels and publish uncertainty

Make SKU, family and total forecasts add up deliberately, and publish a range with the point forecast. Tell planners which range they should plan against.

Why it worksHierarchical forecasts are often incoherent because each level is produced separately. Reconciliation and prediction intervals make the forecast usable for risk decisions rather than false precision.

EvidenceMakridakis et al. (2022) · Silver et al. (2016)

CheckpointTotals reconcile by construction, and every planning decision has a stated range.

Common failurePublishing a single number with no sense of its error.
Govern overrides with exception rules

Allow human edits only for named exceptions, with a reason code, direction and owner. Compare the edited forecast with the untouched baseline every month.

Why it worksLarge-scale evidence from supply-chain planning found that some judgemental adjustments help and many small or upward adjustments hurt. Measurement is what separates expertise from optimism.

EvidenceFildes et al. (2009) · Fildes & Goodwin (2007)

CheckpointOverride rate, reason mix, and forecast value added for edited versus unedited forecasts.

Common failureLetting the loudest planner edit anything, then blaming the model when the plan misses.
Run a monthly forecast review that retires complexity

Review error, bias and value added by model and override. Delete models, hierarchy levels and reports that do not improve a decision.

Why it worksCompetition evidence rewards disciplined benchmarking, not accumulated complexity. A smaller forecasting process that is measured is easier to trust and maintain.

EvidenceMakridakis et al. (2020) · Syntetos et al. (2005)

CheckpointA short value-added table and a decision log showing what was removed, kept or retested.

Common failureAdding a new model every quarter while nobody can name the one the business plans on.

Use this when

Plans, stock or rotas keep missing the actual level of demand.
Forecast accuracy is reported, but nobody can say which decisions improved.
Planners override the system constantly, or have stopped using it.

Measures to run

Forecast bias: Mean signed error over a rolling period; shows systematic over- or under-planning.
Error versus baseline: MAE or MASE relative to a seasonal-naive benchmark at the decision's horizon.
Forecast value added: Whether each model, hierarchy level or override improves on the simpler alternative.
Override rate: Share of forecasts edited by a human, with reason codes and the error of edited versus untouched forecasts.

Pitfalls

Treating forecast accuracy as the goal instead of a means to a better decision.
Using human judgement without recording it, so its value can never be evaluated.
Reporting one aggregate accuracy number that hides systematic bias in the items that matter most.

The evidence base

Each step above cites hand-selected research. These are the studies behind the path.

Makridakis et al. (2020) — The M4 Competition: 100,000 time series and 61 forecasting methods
Silver et al. (2016) — Inventory and Production Management in Supply Chains
Syntetos et al. (2005) — On the categorization of demand patterns
Fildes & Goodwin (2007) — Against Your Better Judgment? How Organizations Can Improve Their Use of Management Judgment in Forecasting
Fildes et al. (2009) — Effective forecasting and judgmental adjustments: An empirical evaluation and strategies for improvement in supply-chain planning
Makridakis et al. (2022) — M5 accuracy competition: Results, findings, and conclusions

Guides for this path

Forecast bias, accuracy and value added: what to measure and what to do about overrides

Other playbooks

Capacity, flow and scheduling — Where is work waiting, what is the real constraint, and how should we release and staff work?
Quality assurance and process control — How do we define quality, stabilise the process and hold the gain?
Inventory and replenishment — How much stock should we hold, where, and what should trigger a replenishment?
Supply chain and supplier resilience — Where is the network exposed, and which capability or buffer actually reduces the exposure?
Service operations and customer experience — Where does the customer experience break, and what should we change in the service system?
Maintenance and asset reliability — Which assets matter, how do they fail, and which maintenance tasks are actually justified?
Continuous improvement and change adoption — How do we make improvement routine, and how do we get a change to stick?
Measurement and operating cadence — Which measures should we run the operation on, and how should we review them?
Incident response and on-call — How should we handle incidents and on-call so the service recovers quickly without burning out the team?

The path, step by step

Name the decision the forecast serves

Classify the demand pattern

Establish a reproducible baseline

Reconcile levels and publish uncertainty

Govern overrides with exception rules

Run a monthly forecast review that retires complexity

The evidence base

More from the library

Guides for this path

Other playbooks