Demand Forecasting & Replenishment¶

SKU-store level demand prediction, automated replenishment, and seasonal planning for retail operations.

Priority: P0 — Immediate ROI
Time to Value: 4-6 weeks
Category: Planning & Supply Chain

Business Problem¶

Retail demand is driven by a complex mix of seasonality, promotions, trends, weather, local events, and competitive actions. Inaccurate forecasting at the SKU-store level results in:

Stockouts — empty shelves that drive customers to competitors; estimated 4-8% of revenue lost to out-of-stocks
Overstock — excess inventory leading to markdowns, spoilage (grocery/fresh), and tied-up working capital
Bullwhip amplification — small demand signals at the store level get amplified through the supply chain, causing wild swings in warehouse and supplier orders
Promotion chaos — promotional demand lifts are guessed rather than modeled, leading to either product shortages or excess post-promotion inventory
Long-tail blindness — high-SKU-count retailers (10K-100K SKUs) cannot manually forecast slow-moving or new items

Capabilities¶

SKU-Store-Day Demand Forecasting¶

Granular demand prediction at the SKU × Store × Day level for the next 4-12 weeks, accounting for base demand, trend, seasonality, promotions, price changes, and external factors.

Promotion Demand Modeling¶

Predict the volume lift from planned promotions by type (BOGO, % off, bundle) based on historical promotional response curves per category and store cluster.

Automated Replenishment¶

Convert demand forecasts into store-level replenishment orders considering lead times, minimum order quantities, shelf capacity, and delivery schedules — pushed to ERP for execution.

New Product Forecasting¶

Predict demand for new SKUs with no sales history by leveraging similar-item analogues, category trends, pre-launch digital signals (search interest, social buzz), and test-store performance.

Weather & Event Overlay¶

Automatically adjust forecasts for weather-sensitive categories (beverages, ice cream, outdoor gear) and local event impacts (concerts, sports, festivals, school holidays).

Data Sources & Ontology Mapping¶

flowchart LR
    subgraph Data Plane
        POS["POS / Commerce"]
        ERP["ERP / Supply Chain"]
        DIGITAL["Digital Analytics"]
        SOCIAL["Social & Reviews"]
    end

    subgraph Ontology Entities
        SALES["Sales History"]
        INVENTORY["Inventory Positions"]
        PROMO["Promotion Calendar"]
        TRENDS["Market Trends"]
        PRODUCT["Product Master"]
    end

    subgraph AI Workflow
        FE["Feature Engineering"]
        MODEL["Forecast Models"]
        REPLEN["Replenishment Engine"]
    end

    POS --> SALES
    ERP --> INVENTORY
    ERP --> PRODUCT
    ERP --> PROMO
    DIGITAL --> TRENDS
    SOCIAL --> TRENDS

    SALES --> FE
    INVENTORY --> FE
    PROMO --> FE
    TRENDS --> FE
    PRODUCT --> FE

    FE --> MODEL
    MODEL --> REPLEN

Ontology Entity	Source System	Key Fields
Sales History	POS / E-commerce	SKU, Store, Date, Units Sold, Revenue, Transaction Count, Basket Context
Inventory Positions	ERP Warehouse/Store	SKU, Location, On Hand, In Transit, On Order, Days of Supply
Promotion Calendar	ERP / Marketing	Promo ID, SKU/Category, Mechanic, Start/End Date, Store Scope, Discount %
Market Trends	Digital Analytics + Social	Search Volume, Social Mentions, Review Sentiment, Trending Topics
Product Master	ERP Item Master	SKU, Category, Subcategory, Brand, Pack Size, Shelf Life, Launch Date

AI Workflow¶

Sales Decomposition — Decompose SKU-store sales history into base demand, trend, weekly/yearly seasonality, promotion lift, and residual components
Feature Engineering — Enrich with calendar features (day-of-week, holidays, pay cycles), weather forecasts, local events, competitive signals, and digital trend indicators
Model Training — Ensemble of LightGBM (cross-learning across SKUs) and DeepAR (time-series deep learning) models; SKU-store level with hierarchical reconciliation (store → region → national)
Promotion Overlay — Apply learned promotion response curves per promo mechanic × category × store cluster to adjust base forecast during promotional windows
New Item Handling — For SKUs with <13 weeks history, use analogue-based forecasting (similar category, price point, brand) enriched with pre-launch digital signals
Replenishment Conversion — Convert daily demand forecast into replenishment orders: forecast → safety stock → reorder point → order quantity, respecting lead times, MOQs, and shelf capacity
Output — Forecast dashboard for planning team; replenishment orders pushed to ERP; promotional demand estimates for merchandising; new product forecasts for buying team

Dashboard & Alerts¶

Key Metrics¶

KPI	Description	Target
Forecast Accuracy (WAPE)	Weighted Absolute Percentage Error at SKU-store-week	< 30%
On-Shelf Availability	% of SKU-store combinations with stock available for sale	> 97%
Overstock Rate	% of SKU-stores with >6 weeks supply on hand	< 8%
Promotion Forecast Accuracy	Predicted vs. actual promotional lift	± 15%
Waste / Markdown Rate	% of inventory sold at markdown or wasted (perishables)	< 4%
Auto-Replenishment Adoption	% of SKU-stores on automated replenishment	> 80%

Alert Rules¶

Alert	Trigger	Severity	Action
Stockout imminent	SKU-store projected to reach zero stock within 2 days	Critical	Emergency replenishment; transfer from nearby store if possible
Demand spike	Forecast exceeds 3σ above baseline for a SKU cluster	High	Investigate driver (viral trend, competitor OOS); expedite supply
Promotion under-supply	Promotional demand forecast exceeds available inventory by >20%	High	Increase allocation; notify merchandising team
Forecast degradation	WAPE exceeds 45% for a category for 3 consecutive weeks	Medium	Retrain model; investigate data quality or assortment change
New product underperformance	New SKU sells <30% of forecast in first 4 weeks	Info	Reassess analogue selection; review placement and pricing

ROI Model¶

Metric	Before	After	Impact
On-shelf availability	93%	97.5%	4.5 pt improvement → $6.8M recovered lost sales
Overstock inventory	14% of SKU-stores	7%	50% reduction → $3.2M working capital freed
Markdown/waste	6.2% of inventory	3.8% of inventory	39% reduction → $3.6M margin recovery
Planning effort	12 planners manual forecasting	4 planners exception-based	67% effort reduction
Promotion waste	18% of promo stock unsold post-promotion	8%	56% reduction

Estimated Annual ROI

$10M - $18M annually from recovered sales, reduced waste, working capital optimization, and planning efficiency — across a mid-size retailer with $250M annual revenue and 15K+ SKUs.

Implementation Notes¶

Requires minimum 104 weeks (2 years) of SKU-store-day sales history from POS for baseline model training
Promotion calendar must be available 4+ weeks in advance with mechanic details (% off, BOGO, bundle) for promotional demand modeling
ERP inventory feeds should be daily or near-real-time for automated replenishment to function accurately
Weather API integration (5-day forecast) adds 2-4% accuracy improvement for weather-sensitive categories
Hierarchical forecast reconciliation ensures store-level forecasts sum consistently to regional and national plans

← Back to Catalogue | Previous: Customer 360 | Next: Personalization →