Forecasting Playbook & Planning Template

Welcome — purpose and how to use this playbook

This playbook helps teams produce forecasts that are evaluated, understood, and actually used in planning and operations. It focuses on methods you can apply repeatedly: frame the problem, choose appropriate horizons, pick and test models, generate actionable scenarios, reconcile forecasts across hierarchies, settle differences with domain owners, and embed forecasts into regular planning workflows.

Who this is for

Team leads, planners, analysts, product owners, finance partners, supply chain and operations managers who need forecasts to inform staffing, inventory, purchasing, budgeting, or capacity decisions.

How to use the playbook

1. Start with problem framing and the decision you must support.
2. Pick horizons and models suited to that decision.
3. Establish evaluation rules and backtest before trusting a model.
4. Use reconciliation + a structured consensus process to create an operational forecast.
5. Embed results into monthly reviews that create actions and accountability.

Problem framing

Good forecasts begin with a clear decision. Ask:

• What decision will this forecast drive? (e.g., staffing next quarter, monthly purchase orders, cash planning)
• What is the tolerable error for that decision? Which errors are more costly (over- or under-forecast)?
• Which data sources are reliable and which are noisy or delayed?
• Who must accept the forecast and act on it?

Document answers as part of every forecast run so assumptions remain visible.

Horizon selection — map horizon to decision

Choose forecast horizons based on the lead times and decision cadence:

• Nowcasting / days to 2 weeks: tactical operational adjustments.
• Monthly (1–3 months): replenishment, short-term staffing, near-term cash.
• Quarterly (3–12 months): budgeting, hiring, seasonal planning.
• Annual / multi-year: strategic capacity, capital investment, product portfolio.

Use overlapping horizons as needed (e.g., an ensemble of short-term and medium-term models) but keep record of which horizon informs which decision.

Model selection heuristics

Pick models to match data quality, horizon, and business dynamics. Simple models are often more robust than complex ones.

• Baseline / Naïve: Last period, moving average — good baseline and often hard to beat for short horizons.
• Seasonal decomposition: Trend + seasonality + residuals — useful when seasonal patterns are stable.
• Exponential smoothing / ARIMA / ETS: Well-suited for time series with autocorrelation and stable structure.
• Regression / causal models: Use when reliable leading indicators or drivers (promotions, price, macro variables) are available.
• Machine learning: Consider when many predictors exist and you have sufficient history, but prefer explainable methods for operational adoption.
• Hybrid / Ensemble: Combine models (weighted average) to reduce single-model risk.

Heuristics: prefer simpler models for short horizons, only add complexity where backtests show consistent value, and validate that features are available in production.

Error metrics and backtesting patterns

Choose metrics that reflect your decision cost and the statistical properties you care about.

• Absolute error: MAE — easy to interpret, robust to outliers.
• Percentage error: MAPE — useful but problematic near zero; prefer sMAPE or MAPE with guardrails.
• Squared error: RMSE — penalizes large errors, useful when big misses are costly.
• Probabilistic / calibration: CRPS, prediction interval coverage — essential when you produce probabilistic forecasts.
• Bias: Mean signed error — shows systematic over- or under-forecasting.

Backtesting patterns:

1. Holdout and rolling-origin evaluation: simulate how the model would have performed in production by re-training and forecasting across many rolling windows.
2. Avoid data leakage: ensure predictors are available at forecast time and do not use future information.
3. Compare to simple baselines: if a model can't beat the naïve baseline, don't deploy it.
4. Report multiple metrics and confidence intervals; capture both accuracy and calibration.

Scenario generation

Scenarios turn uncertainty into decision-ready alternatives. Use three layered scenarios as a minimum:

• Base case: most likely, used for standard planning.
• Upside: plausible better outcome and its implications for capacity and inventory.
• Downside / stress: plausible adverse outcome and mitigation steps.

Steps to generate scenarios:

1. Identify key drivers (price, promotion, supply constraints, macro indicators).
2. For each driver, specify a plausible range and probability (qualitative or quantitative).
3. Construct scenario narratives and translate into model inputs or multiplicative adjustments.
4. Estimate operational impacts (inventory, staff hours, cash flow) for each scenario and list trigger points for escalation.

Forecast reconciliation across hierarchies

Reconciling forecasts makes the numbers coherent across levels (SKU → category → site → region → total). Common methods:

• Bottom-up: forecast at lowest level and aggregate. Good when local signals are strong.
• Top-down: forecast at aggregate level and allocate down using proportions. Good when aggregate signal is more stable.
• Middle-out: forecast at an intermediate level and reconcile up and down.
• Optimal reconciliation: statistical methods (e.g., MinT) that adjust forecasts to be coherent while minimizing error.

Practical rule: start with the method that matches your data strengths, and make reconciliation rules explicit and auditable.

Consensus process with domain owners

A forecast that isn’t owned won’t be used. Use a structured process to build consensus while preserving accountability.

Roles

• Forecast owner: maintains model runs, metrics, and recommended forecast.
• Domain owner / product owner: validates assumptions, provides judgement inputs (promotions, product launches).
• Planner / operations lead: assesses feasibility against capacity and inventory.
• Sponsor / decision owner: accepts the forecast and signs off actions or escalations.

Consensus meeting agenda (recommended)

1. Review last period actuals vs forecast and highlight root causes for misses.
2. Present model results, metrics, and scenario impacts.
3. Domain owners raise known events (promotions, supply issues); update assumptions.
4. Agree on reconciled forecast for each decision horizon and record ownership and actions.
5. Record key risks, triggers, and next review date.

Embedding forecasts into planning workflows

Make forecasts actionable by linking them to concrete processes:

• Define which forecast horizon drives which operational action (PO creation, hiring, capital approvals).
• Create explicit handoffs and SLA for updating forecasts (who updates, when, and how).
• Include forecast metrics (bias, coverage) in monthly reviews and performance dashboards.
• Use triggers tied to forecast variance to start contingency plans (e.g., safety stock release).
• Audit decisions: store reconciliations and assumptions so you can trace decisions back to their forecast basis.

Template: Monthly Forecast Review

Use this template in your monthly meeting. Capture the decisions and who will act.

Monthly review checklist

• Confirm actuals are final for prior period.
• Review model performance (baseline vs current) and record metrics.
• Agree adjustments based on domain inputs and reconcile hierarchies.
• Translate forecast into operational actions and assign owners.
• Record assumptions and triggers for scenario escalation.

Common pitfalls (mal-hungers) to avoid

• Relying on a single point forecast without communicating uncertainty.
• Using models that cannot be reproduced in production or depend on unavailable features.
• Allowing consensus meetings to defer responsibility — establish clear sign-off.
• Ignoring backtesting and deploying unvalidated models.

Small experiments to improve confidence

Try one experiment per month:

• Run an A/B backtest comparing your production model to a simple baseline for the last 12 months.
• Add a short-term demand sensing input (POS, web traffic) and measure improvement for 0–14 day horizon.
• Create a probabilistic forecast for one product and track interval coverage over three months.

References and further reading

• Practical Time Series Forecasting — Hyndman & Athanasopoulos (selected chapters)
• Forecast Value Add (FVA) methodology — supply chain planning communities
• Articles on forecast reconciliation and probabilistic forecasting

Next steps — operationalizing this playbook

Begin by running one documented monthly review using the template above. Capture the forecast owner, metrics, assumptions, and actions. Iterate: shorter cycles increase learning and reduce surprise.