Why It Matters

Service level and inventory impact

• Better forecasts reduce stockouts and overstock by aligning safety stock to true uncertainty.

• Probabilistic forecasts calibrate inventory to service targets rather than averages.

Capacity and supply alignment

• Visibility of demand variability stabilizes production schedules and supplier orders, reducing expediting.

Revenue and margin effects

• Fewer lost sales and markdowns through right-time, right-place availability.

• Smarter promo planning lifts incremental margin instead of driving unprofitable volume.

Working capital and cash

• Lower excess and obsolete inventory; faster cash conversion through accurate replenishment.

Speed of response and resilience

• Demand sensing detects short-term shifts; scenario planning prepares responses to shocks.

Typical ROI ranges

• Common outcomes include reduced inventory, improved fill rate, lower expedite costs, and better forecast accuracy within two to three cycles when governed well.

Browse our inventory & forecasting tools to calculate reorder points, inventory days, and stock velocity in minutes.

Demand Planning Process: Step by Step

Define scope, hierarchy, and cadence

• Scope: SKUs, locations, channels, customers, time horizon (e.g., 0–13 weeks, 3–18 months).

• Hierarchy: Product (SKU → category), geography (DC → region), customer/channel.

• Cadence: Weekly for short-term, monthly S&OP for mid-to-long term; lock calendar early.

Data collection and cleansing

• Gather order history, shipments, POS, returns, stockouts, promotions, price changes, lead times.

• Cleanse: Remove outliers, fix calendar effects, label events, align to working days, reconcile units.

Segmentation (ABC/XYZ) and pattern ID

• ABC by revenue or margin; XYZ by variability (coefficient of variation) or intermittency.

• Identify patterns: trend, seasonality, intermittency, lifecycle phase, promo sensitivity.

Model selection and baseline creation

• Choose per-segment models; automate selection where possible.

• Generate baseline forecast with uncertainty (P50, P70, P90) at the right granularity.

Enrichment with market intelligence

• Layer in sales input, marketing plans, distributor feedback, and macro signals.

• Track adjustments as deltas with reasons to audit later.

Consensus, approval, and publication

• Demand review to align on baseline and overrides.

• Publish approved forecast to supply, procurement, and finance with scenario notes and confidence bands.

Forecasting Methods and When to Use Them

Time-series basics (MA, ES, ARIMA)

• Moving Average (MA): Stable demand without trend/seasonality; quick and robust.

• Exponential Smoothing (ES/Holt-Winters): Trend/seasonality handling with minimal data.

• ARIMA/SARIMA: When autocorrelation structure matters and sufficient history exists.

Use when:

• You have at least 18–24 months of clean history for seasonal items.

• Data is stable or follows identifiable trend/seasonal patterns.

Causal/ML models and feature selection

• Multiple regression, gradient boosting, random forests, and neural networks for promo, price, events, weather, and macro drivers.

• Start with parsimonious features: price index, promo flag/depth, cannibalization variables, holiday/wave indicators, weather (temp, precipitation), macro indices.

Use when:

• Demand is promotion- or price-driven, or sensitive to exogenous factors.

• You can maintain a reliable feature pipeline and avoid leakage.

Intermittent demand (Croston, SBA)

• Croston: Separates demand size and interval; good for spare parts with zero-heavy history.

• SBA (Syntetos–Boylan Approximation): Bias-corrected version of Croston.

• Bootstrapped or negative binomial models for count-based sporadic demand.

Use when:

• Many periods are zero with occasional spikes.

• Traditional ES/ARIMA over-forecasts.

New products (analogs, diffusion, cannibalization)

• Use analogs from similar products/features/channels, adjusted by price and marketing.

• Diffusion models (e.g., Bass) for adoption curves in consumer tech.

• Model cannibalization by allocating category demand and accounting for shelf space and promo overlaps.

Use when:

• Limited or no history; leverage pre-launch indicators and leading signals.

Short-term demand sensing vs planning horizon

• Demand sensing: Near-term updates (0–8 weeks) using POS, e-commerce clicks, weather, and inventory positions; higher frequency, lower latency.

• Medium/long-term planning: Use time-series/causal models, aggregated and reconciled monthly.

Model monitoring and re-tuning

• Track accuracy, bias, and drift by segment and horizon.

• Trigger re-tuning on threshold breaches, structural breaks, or new data availability.

Metrics and Governance

Accuracy metrics (MAPE, WAPE, sMAPE)

• MAPE: Mean(|Actual − Forecast| / Actual). Sensitive to low volumes.

• WAPE: Sum(|A − F|) / Sum(A). Stable across hierarchies; good portfolio metric.

• sMAPE: Symmetric version for mixed scales; useful when zeros exist.

Example:

• Periods A: 100, 120, 80; F: 90, 130, 100

• Absolute errors: 10, 10, 20; WAPE = (10+10+20)/(100+120+80) = 40/300 = 13.3%

Bias and over/under-forecasting

• Bias = Mean(F − A). Percent bias = Sum(F − A)/Sum(A).

• Example: Sum(F − A) = −15 on Sum(A) = 300 → Bias = −5%; systematic under-forecasting.

Linking forecast error to safety stock

• Safety stock = z × σLTD, where σLTD includes forecast error and lead-time variability.

• For periodic review with demand uncertainty, approximate σLTD = sqrt(L × σFE²), where σFE is forecast error std dev and L is lead time in periods.

Example:

• Weekly item, lead time L = 4 weeks, σFE = 30 units/week, service target 95% (z ≈ 1.65)

• Safety stock ≈ 1.65 × sqrt(4 × 30²) = 1.65 × 60 = 99 units

Hierarchy, aggregation, and reconciliation

• Forecast at the level where signal-to-noise is highest, then reconcile across SKU, channel, and geography.

• Use bottom-up for granular promos; top-down for strategic long-range; middle-out for blended approaches.

• Ensure additivity: the sum of children equals parent after reconciliation.

RACI, roles, and meeting structure

• Roles:

  • Demand planner: Owns baseline, metrics, and exceptions.

  • Sales/marketing: Provide event and promo inputs, validate assumptions.

  • Supply planning: Translates demand to capacity/material plans.

  • Finance: Aligns with revenue/margin targets and risks.

  • Executive sponsor: Resolves trade-offs and sets policy.

• Cadence:

  • Weekly demand sensing huddle (30–45 min).

  • Monthly demand review (60–90 min).

  • Monthly supply review (60–90 min).

  • Executive S&OP (60 min).

• RACI:

  • Baseline forecast: R = Demand planner, A = Demand planning lead, C = Sales, I = Finance.

  • Consensus and publication: R = Demand planner, A = S&OP lead, C = Sales/Supply/Finance, I = Execs.

  • KPI governance: R = Analytics, A = S&OP lead, C = Finance, I = All.

Exception management and alerts

• Define tolerances by segment (e.g., WAPE > 25% for A items).

• Auto-generate exceptions for large deltas, low service risk, or capacity breaches.

• Track reason codes for overrides to enable learning.

Data and Tooling

Internal data and master data quality

• Ensure SKU, UOM, customer, and location master data integrity.

• Capture stockouts and substitutions to avoid underestimating true demand.

• Maintain event calendars with promo depth, price, and media.

External data (weather, macro, events)

• Weather: Temperature, precipitation, extreme events by location.

• Macro: Disposable income, housing starts, industrial production by segment.

• Events: Holidays, sporting events, school calendars, local festivals.

Feature store and reproducibility

• Centralize engineered features with versioning and lineage.

• Log model parameters, code, and data snapshots for reproducibility.

System architecture and integrations

• Data lake/warehouse as source of truth.

• Forecasting engine with APIs to ERP, APS, CRM, and POS/e-commerce.

• Workflow layer for approvals, audit trails, and exception routing.

Tool selection criteria

• Supports probabilistic forecasts and scenario planning.

• Native hierarchy management and reconciliations.

• Strong feature engineering, automation, and MLOps capabilities.

• Transparent metrics, bias tracking, and override governance.

• Ease of integration and security/compliance readiness.

Scalability and governance considerations

• Start with a pilot scope, then scale to full portfolio.

• Apply data access controls and model governance (approvals, drift monitoring).

• Establish clear ownership for data pipelines and model stewardship.

Advanced Planning Under Uncertainty

Probabilistic forecasting and percentiles

• Replace single-point forecasts with distributions (e.g., P50, P70, P90).

• Use percentiles for decisions: plan inventory to P90 for high service, P50 for capacity planning.

Service-level targets and inventory policies

• Map service classes to z-values:

  • Class A (97–99%): critical SKUs or channels.

  • Class B (95%): standard items.

  • Class C (90%): low-margin or long-tail.

• Calibrate safety stock using demand variability and lead-time risk.

Scenario planning (price, promo, supply limits)

• Create scenarios for promo depth (10% vs 20%), price elasticity, and media support.

• Overlay supply constraints (supplier caps, labor) and measure revenue/margin trade-offs.

Promotion and price elasticity modeling

• Estimate uplift = base × elasticity × price change + media effects.

• Control for cannibalization within category; compare gross and net lift.

• Validate with hold-out tests or matched-market designs.

Explainability and stakeholder trust

• Use interpretable features and SHAP-like importance summaries.

• Publish reason codes and confidence bands with each forecast.

• Track override performance to learn which inputs add value.

Risk and resilience playbooks

• Define triggers (e.g., POS drop >15% week-over-week) and predefined actions.

• Build alternate sourcing and capacity flex options linked to demand signals.

• Maintain scenario libraries for macro shocks and rapid recovery.

Implementation Playbook

30-60-90 day roadmap

• Days 0–30:

  • Define scope, governance, and cadence.

  • Build data pipelines for history, events, and master data.

  • Segment SKUs (ABC/XYZ) and select initial models.

  • Establish KPI baselines (WAPE, bias, service, inventory).

• Days 31–60:

  • Deploy baseline models by segment (time-series, Croston/SBA, causal where needed).

  • Stand up demand review and exception workflows.

  • Pilot probabilistic outputs for A items; link to safety stock.

• Days 61–90:

  • Expand to promotions and demand sensing.

  • Reconcile forecasts across hierarchies; publish to supply and finance.

  • Run first end-to-end scenario through S&OP; document ROI and lessons.

Quick wins and KPI baselines

• Clean master data and label promotions for top SKUs.

• Eliminate obvious over-forecasting with guardrails and bias checks.

• Focus on A-XYZ items first; protect service where it matters most.

Change management and training

• Role-based training for planners, sales, and finance.

• Standardize reason codes and override etiquette.

• Celebrate accuracy improvements and service gains to build trust.

Templates and checklists

• Diagnostic checklist:

  • Do we have at least 24 months of clean history per seasonal SKU?

  • Are promos and price changes labeled with depth and dates?

  • Is WAPE reported weekly by segment and horizon?

  • Do we track bias and override impact with reason codes?

  • Are safety stocks computed from forecast error and service targets?

• Glossary (concise):

  • P50/P90: Demand percentiles; 50th and 90th confidence points.

  • WAPE: Weighted absolute percentage error, portfolio accuracy metric.

  • Bias: Systematic over/under-forecasting tendency.

  • Demand sensing: Short-horizon, high-frequency forecasting.

  • Reconciliation: Making forecasts add up across hierarchies.

Data remediation plan

• Prioritize top 20% SKUs/channels by revenue for data fixes.

• Automate outlier detection and event labeling.

• Backfill missing calendars, prices, and stockout flags.

Common pitfalls to avoid

• One-size-fits-all modeling across segments.

• Ignoring uncertainty and planning to single points.

• Uncontrolled overrides without reason codes.

• Measuring accuracy only at aggregate level.

• Not tying forecasts to inventory and service policies.

Industry Nuances

CPG and retail seasonality

• Strong seasonality and promos; use hierarchy reconciliation and price/promo features.

• Demand sensing with POS and inventory position for short-term accuracy.

Discrete manufacturing and options

• Configurable products; forecast at option or module level and reconcile to BOM.

• Long lead times magnify forecast error impact; scenario planning is critical.

Aftermarket/spares intermittent demand

• Use Croston/SBA or count models; aggregate across compatible parts where possible.

• Service-oriented safety stock with higher percentiles.

Pharma and regulatory constraints

• Long, regulated lead times; sample-based demand and tender cycles.

• Portfolio-level planning with strict shelf-life and serialization data.

E-commerce and short horizons

• High-frequency signals (clicks, carts); rapid demand sensing and automation.

• Price elasticity and promo responsiveness vary by cohort and channel.

B2B project-driven demand

• Hybrid of forecast and order book; use opportunity pipeline and stage probabilities.

• Scenario plans for project timing shifts and supply allocation.

Worked Examples and Mini Case

MAPE and bias calculation

• Actuals: 500, 600, 400; Forecast: 550, 540, 420

• Absolute percentage errors: 10%, 10%, 5%

• MAPE = (10% + 10% + 5%) / 3 = 8.33%

• Bias: Sum(F − A) = (50 − 60 + 20) = 10; Percent bias = 10 / 1500 = 0.67% (slight over-forecast)

Safety stock from forecast error

• Monthly item, lead time L = 2 months, σFE = 200 units/month.

• Target service 98% (z ≈ 2.05).

• Safety stock ≈ 2.05 × sqrt(2 × 200²) = 2.05 × 283 = 580 units (rounded).

Promo uplift estimation

• Base weekly demand = 1,000 units.

• Price drop = 10%; elasticity = −1.5; media uplift factor = 1.2.

• Uplift = base × (|elasticity| × price drop) × media = 1000 × (1.5 × 0.10) × 1.2 = 180 units.

• Promo forecast = 1,180 units; adjust for cannibalization if needed.

Scenario comparison (P50 vs P90)

• P50 monthly forecast = 10,000; P90 = 11,500.

• Margin per unit = $5; holding cost per unit-month = $0.20.

• Inventory set to P90 adds 1,500 units, holding cost = $300/month.

• If stockout penalty or lost margin exceeds $300 equivalent, P90 is justified; otherwise consider P70.

SKU-Location segmentation example

• ABC by revenue: Top 20% SKUs = 80% revenue.

• XYZ by variability (CV):

  • A–X: Stable, high-value → time-series, tight tolerances, P90 stocking.

  • A–Z: Volatile, high-value → causal/ML or probabilistic focus, frequent review.

  • C–Z: Volatile, low-value → aggregate planning, higher thresholds, make-to-order where feasible.

ROI estimation checklist

• Baseline metrics captured? (WAPE, bias, service, inventory turns, expedite cost)

• Target improvements set by segment?

• Inventory reduction from safety stock optimization quantified?

• Service gains translated to revenue protection?

• Opex impacts (planning time saved, expedite reduction) estimated?

• Tooling and data costs included with payback period modeled?

FAQs

What’s the difference between demand forecasting and demand planning?

Demand forecasting generates statistical predictions of future demand. Demand planning uses those forecasts, plus market and business inputs, to set policies and make decisions on inventory, capacity, supply, and financial commitments within S&OP.

How do I choose the right forecasting model for my SKU portfolio?

Segment first. Use ES/ARIMA for stable, seasonal items with history; Croston/SBA for intermittent demand; causal/ML where price, promo, or weather drive demand. Automate selection with a decision tree and monitor performance by segment and horizon.

What is a good MAPE and how should I measure accuracy and bias?

It depends on volume and volatility. Track WAPE and bias by segment and horizon. Use WAPE at portfolio level, MAPE for like-for-like comparisons, and always pair with bias to detect systematic over/under-forecasting.

How can I forecast new products or items with little history?

Use analogs adjusted for price and marketing, diffusion models for adoption, and early signals from pre-orders or digital engagement. Account for cannibalization within the category and update frequently post-launch.

How do promotions and price changes affect the forecast?

Model uplift using price elasticity, promo depth, timing, and media support. Control for cannibalization and forward-buying. Validate against hold-out periods and refine over time.

What is demand sensing and how is it different from forecasting?

Demand sensing updates near-term demand using high-frequency signals like POS and web traffic. It complements mid/long-term forecasting, improving responsiveness within the next few weeks.

How do I link forecast uncertainty to safety stock and service levels?

Use probabilistic forecasts to estimate demand variability, then compute safety stock with z-values aligned to service targets. Update σFE and lead-time assumptions regularly and adjust by segment.

How often should I update the forecast and run consensus meetings?

Update near-term forecasts weekly for sensing and monthly for mid/long-term planning. Hold a monthly demand review, supply review, and executive S&OP, with a short weekly sensing huddle for fast-moving items.

Discover why teams move from spreadsheets to Verve AI for end-to-end inventory forecasting and purchase order management.