Inventory Optimization

Problem

Determine optimal stock levels, reorder points, and order quantities across a SKU-location hierarchy to minimise total cost: holding cost + stockout cost + ordering cost + waste (perishables). Inventory optimisation translates demand forecasts into replenishment decisions, accounting for lead times, demand uncertainty, supplier constraints, and service level objectives.

This is the decision layer downstream of demand forecasting — it translates a probabilistic forecast into an order quantity using inventory policy theory (EOQ, newsvendor, safety stock calculations).

Users / Stakeholders

RoleDecision
Supply chain plannerApprove/adjust system-generated order recommendations
Category buyerNegotiate MOQ (minimum order quantities) and lead times
Warehouse managerCapacity planning; slotting optimisation
FinanceWorking capital target; inventory provision
Store managerMaintain in-store availability; manage waste

Domain Context

  • Lead time uncertainty: Supplier lead times vary. Safety stock must buffer demand variation AND lead time variation.
  • Perishability: Fresh food, pharmaceuticals have shelf-life constraints. Waste cost can equal or exceed stockout cost. FEFO (First Expired, First Out) rotation is operational constraint.
  • MOQ constraints: Suppliers have minimum order quantities that may exceed optimal reorder quantities. Integer programming required.
  • Multi-echelon: Warehouse → DC → store. Stock allocation across echelons affects service levels. Multi-echelon optimisation is complex but high-value.
  • Seasonal demand: Safety stock must increase ahead of seasonal peaks. Dynamic safety stock calculation required.
  • Data freshness: Inventory positions must be current. Point-of-sale data and warehouse management system data latency directly impacts decision quality.

Inputs and Outputs

Inputs:

Demand forecast: point forecast + prediction intervals (P10/P50/P90) per SKU per location
Current inventory: on_hand, on_order, in_transit, allocated
Supplier data: lead_time_mean, lead_time_std, MOQ, case_pack_size, cost
Policy parameters: service_level_target (e.g., 97.5%), holding_cost_rate, ordering_cost
Product attributes: shelf_life, units_per_case, weight, volume, supplier_id

Output:

reorder_point:    Inventory level that triggers a new order
order_quantity:   Recommended order quantity (EOQ-based or min-cost)
safety_stock:     Buffer stock recommendation vs current safety stock
expected_service_level: Predicted fill rate under recommended policy
days_of_cover:    Expected inventory coverage in days
action:           ORDER / HOLD / CANCEL_PENDING / EMERGENCY_ORDER

Decision or Workflow Role

Daily batch run (after demand forecast refresh)
  ↓
Safety stock calculation: σ_demand × σ_lead_time × z(service_level)
  ↓
Reorder point: demand_during_lead_time + safety_stock
  ↓
For SKUs below reorder point:
  → Calculate EOQ: √(2 × annual_demand × ordering_cost / holding_cost)
  → Apply MOQ constraints
  → Generate purchase order recommendation
  ↓
Planner review: approve / adjust / reject (exception-based)
  ↓
Orders placed via ERP → supplier fulfils → inventory received
  ↓
Actual demand + actual replenishment → forecast accuracy feedback

Modeling / System Options

ComponentApproachNotes
Demand forecastLightGBM / ARIMA (see Demand Forecasting)Feeds into inventory model
Safety stockAnalytical (cycle service level formula)z × σ_dL formula
Order quantityEOQ + integer constraintsClassic operations research
Lead time modelEmpirical distribution from supplier historyNon-parametric sufficient
Multi-echelonStochastic dynamic programming / simulationHigh value; high complexity

Note: Inventory optimisation is primarily an operations research problem once the demand forecast is available. ML is in the demand forecast; classical OR is in the replenishment decision.

Deployment Constraints

  • Integration: Must connect to ERP (SAP, Oracle) for inventory positions and order placement. Data latency from ERP affects decision quality.
  • Planner override: System generates recommendations; planners override for supplier relationship reasons, promotion knowledge, etc. Override rate should be monitored.
  • Seasonal pre-build: Safety stock must increase ahead of seasonal peaks. Requires forward-looking demand signal.

Risks and Failure Modes

RiskDescriptionMitigation
Forecast error propagationPoor forecast → wrong reorder point → stockout/overstockForecast MASE monitoring; robust safety stock formula
Lead time surpriseSupplier fails to deliver → stockoutEmergency order trigger; lead time distribution update
MOQ trapMOQ forces overorder of slow-moving SKUMOQ negotiation; consignment stock agreements
Seasonal under-stockSafety stock not increased for peakPre-peak safety stock review; seasonal uplift logic

Success Metrics

MetricTargetNotes
In-stock rate (fill rate)> 97.5%Service level KPI
Inventory turnoverImprovement vs prior yearCapital efficiency
Waste / expired stock< 2% of sales (perishables)Waste KPI
Working capital-5–15% reductionFinance KPI
Planner intervention rate< 20% of ordersAutomation quality

References

  • Silver, E. et al. (2017). Inventory and Production Management in Supply Chains. CRC Press.
  • Zipkin, P. (2000). Foundations of Inventory Management. McGraw-Hill.

Modeling

Reference Implementations

Adjacent Applications