demand-planning

name	demand-planning
description	Forecast product or service demand using historical analysis, seasonal patterns, market signals, consensus planning, forecast accuracy metrics, and bias correction. TRIGGER when: user says /demand-planning, "forecast demand", "demand forecast", "predict demand", "sales forecast", "demand plan".
argument-hint	[product/SKU] [time-horizon] [market-or-region]
user-invocable	true

Demand Planning

Build accurate, actionable demand forecasts that align supply with expected customer demand while minimizing forecast error and inventory risk.

---

Input Gathering

Collect the following before starting the forecast cycle:

Input	Description	Required
Historical sales data	Minimum 24 months of shipment or POS data by SKU/region	Yes
Product hierarchy	Category, family, SKU mapping	Yes
Time horizon	Short (0-3 mo), Medium (3-12 mo), Long (12-36 mo)	Yes
Granularity	Weekly, monthly, quarterly	Yes
Promotional calendar	Planned promotions, price changes, launches	No
Market intelligence	Competitor actions, economic indicators, trends	No
Channel breakdown	D2C, wholesale, retail, e-commerce split	No

---

Step-by-Step Process

Step 1 — Data Cleansing and Baseline Preparation

• Remove or flag outliers (stockouts, one-time bulk orders, data entry errors).
• Normalize for calendar effects (trading days, leap years).
• Fill gaps using interpolation or proxy SKU behavior.
• Document every adjustment with reason codes.

Step 2 — Statistical Forecast Generation

Run appropriate models based on demand pattern:

Demand Pattern	Recommended Model
Stable / Smooth	Exponential Smoothing (Holt)
Trending	Holt-Winters, Linear Regression
Seasonal	SARIMA, Holt-Winters Seasonal
Intermittent	Croston, SBA (Syntetos-Boylan)
New Product	Analogous product mapping
Lumpy	Bootstrapping, Willemain

• Generate point forecasts and prediction intervals (80% and 95%).
• Produce forecasts at the lowest meaningful granularity, then aggregate upward.

Step 3 — Demand Sensing and Signal Integration

Layer external signals onto the statistical baseline:

• Leading indicators: Web search trends, social media sentiment, foot traffic.
• Economic signals: GDP growth, consumer confidence, commodity prices.
• Customer intelligence: Pipeline data, customer forecasts, POS sell-through.
• Competitive signals: Product launches, pricing moves, store openings/closures.

Weight each signal by its historical correlation to actual demand.

Step 4 — Consensus Planning and Collaborative Review

Facilitate a structured demand review meeting:

1. Present statistical forecast vs. prior period actuals.
2. Overlay sales team input (bottom-up account-level adjustments).
3. Incorporate marketing input (promotions, campaigns, launches).
4. Reconcile finance top-down targets with bottom-up build.
5. Lock the consensus number for each SKU-location combination.
6. Record assumption log for every manual override.

Step 5 — Forecast Accuracy Measurement

Calculate and track these KPIs each cycle:

Metric	Formula	Target
MAPE	Mean Absolute Percentage Error	< 20% at SKU-month
Bias	(Forecast - Actual) / Actual	Between -5% and +5%
WMAPE	Weighted MAPE by revenue	< 15%
Forecast Value Added (FVA)	Accuracy improvement over naive model	> 0 for every step
Tracking Signal	Cumulative error / MAD	Between -4 and +4

Step 6 — Bias Correction and Continuous Improvement

• Identify persistent over- or under-forecasting by product, region, or planner.
• Apply bias correction factors to the statistical baseline.
• A/B test new models or signals against the incumbent.
• Archive each forecast vintage for retrospective analysis.
• Update model parameters quarterly; retrain models annually.

---

Output Format

## Demand Forecast — [Product/Region] — [Period]

### Executive Summary
- Forecast horizon: [start] to [end]
- Total forecasted demand: [units/revenue]
- Forecast confidence: [high/medium/low]
- Key assumptions: [list]

### Forecast by Period
| Period | Statistical Forecast | Adjusted Forecast | YoY Change | Confidence Interval |
|--------|----------------------|-------------------|------------|---------------------|
| ...    | ...                  | ...               | ...        | ...                 |

### Accuracy Scorecard (Prior Cycle)
| Metric | Value | Target | Status |
|--------|-------|--------|--------|
| MAPE   | ...   | < 20%  | ...    |
| Bias   | ...   | +/- 5% | ...    |

### Assumptions and Risks
- [Numbered list of key assumptions]
- [Identified upside/downside risks with probability]

### Recommendations
- [Actionable next steps]

---

Quality Checklist

• Historical data covers at least 24 months and has been cleansed
• Outliers documented and handled with reason codes
• Appropriate model selected for each demand pattern
• Prediction intervals generated alongside point forecasts
• External signals evaluated and weighted
• Consensus review completed with cross-functional input
• Every manual override has a documented rationale
• Accuracy metrics calculated against prior forecast
• Bias correction applied where persistent drift detected
• Forecast locked and communicated to supply planning

---

Edge Cases

• New product launch with zero history: Use analogous product mapping; identify 2-3 similar past launches and scale by expected distribution and marketing spend.
• Demand during supply-constrained periods: Back-calculate unconstrained demand using lost-sales estimates; do not train models on constrained actuals.
• Promotional cannibalization: Model baseline and promotional lift separately; account for pull-forward and pantry-loading effects.
• Market discontinuities (pandemic, regulation change): Segment data into pre- and post-event; weight recent data more heavily or restart baseline.
• Highly intermittent / spare-parts demand: Switch to Croston or bootstrapping; report demand probability and expected order size separately.
• Currency or price-driven volume shifts: Decompose demand into price-driven and organic components; forecast each independently.