Prescriptive Optimization Recipe — Inventory Example

Why this playbook matters

Inventory decisions sit at the intersection of uncertainty, cost, and service. This recipe helps teams move from static rules or spreadsheets to safe, auditable prescriptive systems that produce implementable recommendations—without ignoring real operational constraints, stakeholder incentives, or data limits.

When to use this recipe

• To reduce cost while maintaining or improving service levels
• When demand and lead-times show measurable variability
• When recommendations must be auditable and operationally feasible
• When you plan to automate replenishment suggestions into ERP/WMS or operator dashboards

Assumptions & scope

This playbook focuses on single-item or SKU-level replenishment policy design and validation. It assumes you can access historical demand, lead-time, and cost data. It covers both heuristic approaches (reorder point, (s,Q), periodic review) and prescriptive approaches (stochastic optimization, simulation-based recommenders).

Quick data checklist

• Historical demand series at the operational cadence (daily/weekly)
• Lead time distribution (purchase/production, plus variability)
• Unit costs: purchase, holding, shortage/backorder, ordering/setup
• Service-level definitions used by the business (cycle service level, fill rate, etc.)
• Business constraints: batch sizes, minimum orders, storage capacity, shelf-life, supplier lot constraints
• Operational rules: review frequency, approval process, exception handling

Stepwise recipe

1) Define the objective

Clarify whether the primary objective is minimizing total cost, achieving a target service level, maximizing fill rate, or a hybrid (cost + penalty for stockouts). Make the objective measurable and agreed by stakeholders.

• Cost objective example: minimize expected annual holding + ordering + backlog costs
• Service objective example: achieve ≥ 95% cycle service level while minimizing holding costs

2) Data needed and preparation

Prepare cleaned time series for demand and lead times. Where data are sparse, use pooling, hierarchical modeling, or borrow information across similar SKUs.

• Impute missing periods consistently; preserve seasonality and promotions where possible
• Segment SKUs by demand profile (fast/slow, intermittent) and choose methods accordingly

3) Model selection: heuristics vs prescriptive

Choose a model class that matches the hunger and data quality.

• Simple heuristics: safety-stock + reorder point (works well with stable demand)
• Periodic review policies: (R, S) for periodic replenishment cycles
• Stochastic optimization: explicit objective and constraints, good when you can model uncertainty and want optimal trade-offs
• Simulation-based recommenders: useful to evaluate policies under realistic stochastic scenarios

Practical formula (heuristic) — safety stock & reorder point

When demand per day is approximately normal, a common safety-stock formula is:

SS = z * sigma_d * sqrt(L)

and reorder point R = mean_d * L + SS, where:

• mean_d = average demand per day
• sigma_d = standard deviation of daily demand
• L = average lead time in days
• z = normal quantile for desired service level (e.g., z≈1.645 for 95% cycle service level)

Worked numeric example

Suppose mean daily demand = 50 units, sigma_d = 20 units, L = 5 days, target cycle service level = 95% (z≈1.645).

SS = 1.645 × 20 × sqrt(5) ≈ 1.645 × 44.72 ≈ 73.6 → round to 74 units

Reorder point R = 50 × 5 + 74 = 324 units

Interpretation: when on-hand + on-order falls to 324, create a replenishment action that respects lot sizes and other constraints.

4) Explicit constraints and business rules

Encode hard and soft constraints up-front. Examples:

• Hard: storage capacity, expiration/shelf-life, minimum order quantity
• Soft: preferred suppliers, cost tiers, consolidation windows
• Operational: who must approve changes above X%, how frequently models rerun

5) Validation using historical simulations/backtests

Don’t trust theoretical optima without testing. Use backtesting and Monte Carlo simulation to estimate how the policy performs under realistic patterns:

1. Replay historical demand with your policy and record realized service, stockouts, and costs.
2. Run Monte Carlo simulations using fitted demand and lead-time distributions to estimate variability in outcomes.
3. Report confidence intervals for key metrics (service level, average inventory, stockout frequency).

6) Productionization checklist for recommendation delivery

Make recommendations actionable and safe to deploy:

• Deliver recommendation formats that operators can consume (ERP change requests, CSV, API)
• Include rationale and key drivers for each recommendation (e.g., expected demand shift, lead-time increase)
• Implement human-in-the-loop approvals for first deployments; define rollback conditions
• Respect batching and supplier constraints when converting continuous recommendations into discrete orders
• Version models and log inputs + outputs for auditability

7) Monitoring, guardrails and alerts

Continuously monitor operational KPIs and set automated guardrails:

• KPIs: fill rate, cycle service level, average inventory days, stockout events, forecast error (MAPE), recommendation adoption rate
• Drift detection: monitor changes in demand variance or lead-time distributions and trigger model retraining
• Anomaly alerts: spikes in purchase price, supplier lead-time jumps, or rapid inventory builds
• Safety limits: never increase recommended order sizes above a business-specified cap without human review

8) Example runs and interpretation

When you run the model, produce a short explanation with each SKU recommendation:

• Recommendation: reorder point or quantity, with confidence band
• Key drivers: PoP demand change, supplier lead-time change, forecast uncertainty
• Actionability: required order size after batching; who needs to approve
• Interpretation checklist: if the recommended safety stock is much larger than current policy, explain trade-offs in carrying cost vs stockouts

Common pitfalls and how to avoid them

• Missing constraints: Recs that ignore lot sizes or storage caps are useless—encode constraints early.
• Overfitting: Complex models trained on past promotions may recommend excessive safety stock—segment and exclude promotion windows where appropriate.
• Opaque recommendations: Provide explanations and simple fallback rules so operators can understand and trust suggestions.
• No monitoring: Without KPIs and drift detection, models silently degrade.

Quick rollout plan

1. Prototype on a small SKU cohort (high-volume, non-perishable).
2. Run parallel test: model recommendations vs current policy for an A/B period.
3. Evaluate business KPIs and operator feedback, iterate on constraints and communication.
4. Gradually expand and automate delivery with human approval gates and monitoring dashboards.

Playbook checklist (one-page)

• Objective defined and agreed
• Data readiness verified
• Model class selected with justification
• Constraints and business rules encoded
• Validation plan (backtest + simulation) documented
• Recommendation delivery format defined
• Approval workflow and rollback plan in place
• Monitoring KPIs and alerts configured

Next steps & experimentation ideas

Try adding a simulation-driven tuner that searches policy parameters (s, Q) across the trade-off space of expected cost vs service level and returns Pareto-optimal candidate policies for the business to choose among. Consider lightweight explainability: include the top 3 drivers (forecast change, lead-time change, recent stockout) with each recommendation.